vol2 no 2

(IJCNS) International Journal of Computer and Network Security, Vol. 2, No. 2, February 2010

1

Multiple region Multiple Quality ROI Encoding using Wavelet Transform

P.Rajkumar1 and Madhu Shandilya2

1,2Department of Electronics, Maulana Azad National Institute of Technology, Bhopal, India

[email protected]

Abstract: The Wavelet Transform, which was developed in the last two decades, provides a better time- frequency representation of the signal than any other existing transforms. It also supports region of interest (ROI) encoding. This allows different regions of interests to be encoded at different bit rats with different quality constraints, rather than encoding the entire image with single quality constraints. This feature is highly desirable in medical image processing. By keeping this in mind we propose multiple regions multiple quality (MRMQ) encoding technique which facilitates different ROIs according to the priority of enhancement. The paper utilizes the adaptive biorthogonal wavelet application to keep quality in the highest priority ROI, as they posses excellent reconstruction features. The paper also proposes bit saving by pruning the detail coefficients in the wavelet decompositions and truncating the approximation coefficients outside ROI by using space frequency quantization method. Simulation results obtained shows that, by proposed method the image quality can be kept high up to loss less in the ROI region while limiting the quality outside the ROIs. Keywords: Wavelet Transform, supports region of interest.

1. Introduction to medical imaging Medical images are now almost always gathered and stored in digital format for easy archiving, storage and transmission, and to allow digital processing to improve diagnostic interpretation. Recently, medical diagnostic data produced by hospitals increase exponentially. In an average-sized hospital, many tetra or 1015 bytes of digital data are generated each year, almost all of which have to be kept and archived. Furthermore, for telemedicine or tele browsing applications, transmitting a large amount of digital data through a bandwidth-limited channel becomes a heavy burden [1]. Three-dimensional data sets, such as medical volumetric data generated by computer tomography (CT) or magnetic resonance (MR), typically contain many image slices that require huge amounts of storage. A typical mammogram must be digitized at a resolution of about 4000 x 5000 pixels with 50 µm spot size and 12 bits, resulting in approximately 40Mb of digital data. Such high resolution is required in order to detect isolated clusters of micro calcifications that herald an early stage cancer. The processing or transmission time of such digital images could be quite long. Also, archiving the amount of data generated in any screening mammography program becomes an expensive and difficult challenge. [2] The storage requirement for digital coronary angiogram

video is huge. A typical procedure of 5 minutes, taken at 30 frames per second for 512x512 pixel images results in approximately 2.5GB of raw data. [3]

2. Need for Image compression A digital compression technique can be used to solve both the storage and the transmission problems. An efficient lossy compression scheme to reduce digital data without significant degradation of medical image quality is needed.

Compression methods are important in many medical applications to ensure fast interactivity during browsing through large sets of images (e.g. volumetric data sets, time sequences of images, image databases). In medical imaging, it is not acceptable to lose any information when storing or transmitting an image. There is a broad range of medical image sources, and for most of them discarding small image details might alter a diagnosis and cause severe human and legal consequences. [1]

In addition to high compression efficiency, future moving image coding systems will require many other features. They include fidelity and resolution scalability, region of interest enhancement, random access decoding, and resilience to errors due to channel noise or packet loss, fast encoding/decoding speed, low computational and hardware complexity.

Recent developments in pattern recognition (regions of interest segmentation) and image processing have advanced the state-of-the-art of image processing. During the past few years, we have applied some of these latest image-processing techniques on image enhancement classification and compression.

3. Region of interest (ROI) encoding Support of region of interest (ROI) access is a very interesting feature of image compression, in which an image sequence can be encoded only once and then the decoder can directly extract a subset of the bit stream to reconstruct a chosen ROI of required quality [4] The arbitrarily shaped regions inside an image will be encoded at different quality levels according to their importance or, as per diagnostic relevance. The whole image is transformed and coefficients associated to the ROI are coded at higher precision (up to lossless) than the background. Especially in medical imaging ROI coding help compression method s to focus on those regions that are important for diagnosis purpose.


2

For some applications, only a subsection of the image sequence is selected for analysis or diagnosis. Therefore, it is very important to have region of interest retrievability that can greatly save decoding time and transmission bandwidth. An important part of genetics analysis starts with researchers breaking down the nucleus of a human cell into a jumbled cluster of chromosomes that are then stained with dye so that they can be studied under a microscope. This jumble of stained chromosomes, which carry the genetic code of the host individual, is then photographed, creating a chromosome spread image (see Fig. 1 (a)). This image is subsequently subject to a procedure called chromosome-karyotyping analysis. The result of this procedure is a karyotype image (see Fig. 1 (b)), the standard form used to display chromosomes. In this configuration, the chromosomes are ordered by length from the largest (chromosome 1) to the smallest (chromosome 22 in humans), followed by the sex chromosomes. Karyotype images are used in clinical tests, such as amniocentesis, to determine if all the chromosomes appear normal and are present in the correct number. [5] Unlike some other types of medical imagery, chromosome images (see Fig. 1) have an important common characteristic: the regions of interest (ROIs) to cytogeneticists for evaluation and diagnosis are all well determined and segmented prior to image storage. The remaining background images, which may contain cell nuclei and stain debris, are kept as well in routine cytogenetics lab procedures for specimen reference rather than for diagnostic purposes. Since the chromosome ROIs are much more important than the rest of the image for karyotyping analysis, loss less compression for the former is required while lossy compression for the latter is acceptable. This calls for lossy and loss less region-of-interest (ROI) coding. In contrast, commercial chromosome karyotyping systems fail to utilize the ROI information by compressing entire chromosome spread or karyotype images.

Figure. 1(a) A metaphase cell spread image[5]

Figure. 1(b) A metaphase cell’s karyotype[5]

In the karyotype, all chromosomes in the spread are rotated and copied onto an image with constant background and positioned according to their classes. The label annotation is drawn separately.

4. Wavelet transform Discrete cosine transform and Wavelet transform are more commonly used for compression. The popular JPEG & MPEG uses discrete cosines transform based compression;

While JPEG 2000 uses Wavelet transform based compression. The DCT based compression; the algorithm breaks the image into 8x8 pixel blocks and performs a discrete cosine transform on each block. The result is an 8x8 block of spectral coefficients with most of the information is concentrated in relatively few coefficients. Quantization is performed, which approximates the larger coefficients; smaller coefficients become zero. These quantized coefficients are then reordered in a zig zag manner to group the largest values first, with long strings of zeroes at the end that can be efficiently represented.

While this algorithm is very good for general purposes, it has some draw backs when applied to medical images .It degrades ungracefully at high compression ratios, producing prominent artifacts at block boundaries, and it can not take advantage of patterns larger than the8x8 blocks. Such artifacts could potentially be mistaken as being diagnostically significant. A wavelet approach was also suggested by Li who considered the problem of accessing still, medical image data, remotely over low bandwidth networks. This was accomplished using a region of interest (ROI) based approach that allocated additional bandwidth within a dynamically allocated ROI, whilst at the same time providing an embedded bit stream, which is useful for progressive image encoding.[3] Some of the key features of the wavelet transform which make it such a useful tool are:

• Spatial-frequency localization, • Energy compaction, • Decaying magnitude of wavelet coefficients across

sub-bands. Wavelet based compression schemes generally out

perform JPEG compression in terms of image quality at a given compression ratio, and the improvement can be dramatic at high compression ratios. [9] The Wavelet Transform, which was developed in the last two decades, provides a better time- frequency representation of the signal than any other existing transforms. It supports ROI (region of interest) encoding. Multiple regions multiple qualities ROI encoding facilitates different ROIs according to the priority to be enhanced with different qualities (i.e. at different compression ratios, both by lossless and lossy methods), while limiting the bit size outside ROI by compressing heavily (lossy). This paper proposes the application of adaptive biorthogonal wavelets to keep quality in the highest priority ROI as they posses excellent reconstruction features. 4.1 Classification of wavelets We can classify wavelets into two classes: [7] (a) Orthogonal and (b) biorthogonal.Based on the application, either of them can be used. 4.1.1 Features of orthogonal wavelet filter banks The coefficients of orthogonal filters are real numbers. The filters are of the same length and are not symmetric. The low pass filter, G

0 and the high pass filter, H

0 are related

to each other by


3

H0 (z) = z

-N G

0 (-z

-1) (1)

{Φi,j (t)Ψi,j (t)} and {Φi,j(t) Ψ˜i,j(t)} are the biorthogonal basis functions[8]. a˜j,k and b˜j,k are the scaling and wavelet coefficients respectively; together they form the biorthogonal DWT coefficients of x(t). The DWT starts at some finest scale j = M and stops at some coarsest scale j = N. M - N is the number of levels of decomposition in the biorthogonal DWT of x (t).

4.1.3 Selection of wavelet from the biorthogonal family: Different wavelets will be suited for different types of images. One wavelet basis may produce highest coefficient for a picture, may not necessarily produce highest

coefficients for all pictures for all images while decomposing the image using the wavelet transform.

Keeping this in the mind, Adaptive wavelet basis function is additionally proposed. Amongst the all important biorthogonal wavelet basis functions, a particular wavelet basis function is chosen on the basis of comparison of the highest coefficients produced by all the different wavelet basis functions at the first level of the decomposition process. The proposed MRMQ Encoding uses the Discrete Wavelet Transform, chooses the right Biorthogonal basis function on the basis of the highest coefficient produced in the highest priority ROI amongst of all the ROIs.

5. Multiple Regions Multiple Quality ROI Encoding For certain applications, only specific regions in the volumetric data or video are of interest. For example, in MR imaging of the skull, the physician is mostly concerned about the features inherent in the brain region. In video conferencing, the speaker’s head and shoulders are of main interest and need to be coded at higher quality whereas the background can be either discarded or encoded at a much lower rate. High compression ratios can be achieved by allocating more bit rate for region(s) of interest (ROI) and less bit rate for the remaining regions, i.e. the background.

Region-based image coding schemes using heterogeneous (multiple) quality constraints are especially attractive because they not only can well preserve the diagnostic features in region(s) of interest, but also meet the requirements of less storage and shorter transmission time for medical imaging applications and video transmission.

A main advantage of the proposed technique is that it supports multiple-region multiple-quality (MRMQ) coding. By this method, total bit budget can be allocated among multiple ROIs and background depending on the quality constraints. Experimental results show that this technique offers reasonably well performance for coding multiple ROIs.

To support region of interest (ROI) coding, it is necessary to identify the wavelet transform coefficients associated with the ROI. We have to keep track of the coefficients that are involved in the reconstruction of ROI through each stage of decomposition.

5.1 ROI Coding

We use region mapping which trace the positions of pixels in an ROI in image domain back into transform domain by inverse DWT. The coefficients of greater importance are called ROI coefficients, the rest are called background coefficients.

The coding algorithm, fig.2 needs to keep track of the locations of wavelet coefficients according to the shape of the ROI. To obtain the information about ROI coefficients, a mask image, which specifies the ROI, is decomposed by the wavelet decomposition of the image. In each decomposition stage, each subband of the decomposed mask contains information for specifying the ROI in that subband. By successively decomposing the approximation coefficients

The two filters are alternated flip of each other. The

alternating flip automatically gives double-shift orthogonality between the low pass and high pass filters, i.e., the scalar product of the filters, for a shift by two is zero i.e, ∑G [k] H [k-2l] = 0, where k,lεZ . Perfect reconstruction is possible with alternating flip.

Also, for perfect reconstruction, the synthesis filters are identical to the analysis filters except for a time reversal. Orthogonal filters offer a high number of vanishing moments. This property is useful in many signal and image processing applications. They have regular structure, which leads to easy implementation and scalable architecture. 4.1.2 Features of biorthogonal wavelet filter banks In the case of the biorthogonal wavelet filters, the low pass and the high pass filters do not have the same length. The low pass filter is always symmetric, while the high pass filter could be either symmetric or anti-symmetric. The coefficients of the filters are either real numbers or integers.

For perfect reconstruction, biorthogonal filter bank has all odd length or all even length filters. The two analysis filters can be symmetric with odd length or one symmetric and the other antisymmetric with even length. Also, the two sets of analysis and synthesis filters must be dual. The linear phase biorthogonal filters are the most popular filters for data compression applications

The analysis and synthesis equations for the biorthogonal DWT of any x (t) ε L2(R) are

Analysis

a ̃ j,κ = ∫ x(t)2j/2φ̃(2jt-k)dt (2)

b̃j,κ = ∫ x(t)2j/2ψ ̃(2jt-k)dt (3)

Synthesis M-1

x(t)= 2 N/2Σ a ̃N,k φ(2 Nt-k) + Σ 2j/2

k j=N Σ b ̃j,k ψ(2 jt-k) (4) k


4

(LL subband) for a number of decomposition levels, information about ROI coefficients is obtained.

Figure 2. ROI Encoding using Wavelet Transform

5.2 Outside ROI Space –Frequency Quantization (SFQ) technique exploits both spatial and frequency compaction property of the wavelet transform through the use of two simple quantization modes.

To exploit the spatial compaction property, a symbol is defined, that indicates that a spatial region of high frequency coefficients has zero value. Application of this symbol is referred to as zero-tree quantization. This is done in the first phase called Tree Pruning Algorithm. In the next phase called Predicting the tree, the relation between a spatial region in image and the tree- structured set of coefficients is exploited. Zero tree quantization can be viewed as a mechanism for pointing to the location where high frequency coefficients are clustered. Thus, this quantization mode directly exploits the spatial clustering of high frequency coefficients predicted.

For coefficients that are not set to zero by zero tree quantization, a common uniform scalar quantization, independent of coefficient frequency band is applied. Uniform quantization followed by entropy coding provides nearly optimal coding efficiency.

6. Results

In Fig. 3 an Image is taken showing the abdominal operation. In the above Image MRMQ ROI encoding is applied. The result is shown in fig. 4.Here the no. of ROIs are 2, as shown the area outside the ROI is heavily degraded .The PSNR of this image is 25dB.while the quality is lossless in the ROI. The PSNR vs. bits per pixel of the ROI enhanced Image at various compression points is plotted in fig.5. This shows that although the PSNR difference is less whiles the bits per pixel reduces from 1bpp to 0.5 bpp using Matlab® simulation.

Fig. 6 an image of a newborn baby affected by

sacrococcyged teratoma (before surgery) is taken. In fig. 7 the area outside the ROI is heavily degraded. In fig. 8, Outside the ROI is excluded for further reducing the bit size requirement. The PSNR vs. bits per pixel at various compression points are plotted in fig. 9.

Figure 5. PSNR vs. bits per pixel of the ROI enhanced Image “abdomen”

Note: ROI area is 17230pixels and Image is 448x336 pixels.

Figure 6. Teratoma affected baby

Figure 7. ROI enhanced Teratoma affected baby

(Background degraded)

Figure 8. ROI enhanced Trachoma affected baby (Background excluded)

Figure 3. Original medical Image of abdominal operation (2.32bpp)

Figure 4. ROI Enhanced medical image of fig. 3 Here no. of ROI =2(0.92bpp, PSNR =25.2db)

Bits per pixel

PSNR


5

Figure 9. PSNR vs. bpp of the ROI enhanced Teratoma affected baby

7. Conclusions The simulation results give bits per pixel with corresponding PSNR values for the proposed MRMQ ROI encoding. It has been observed that for still images, the bit size requirement was made less by degrading the image quality out side the ROI, while keeping the quality of image uncompromised in the ROI region.

In the present paper, still image results are simulated with the biorthogonal wavelet transform family for the multiple region multiple quantization (MRMQ) ROI coding, therefore further analysis can be extended in the field of video processing. Also the present analysis is carried on the medical still images, which can be extended for the video processing with graphical and natural images.

References

[1] Shaou-Gang Miaou, Shih-Tse Chen, Shu-Nien Chao , Wavelet-Based Lossy-to-Lossless Medical Image Compression Using Dynamic VQ And SPIHT Coding, Biomedical Engineering applications, Basis &Communications. Pg 235-242 Vol. 15 No. 6 December 2003.

[2] M´onica Penedo, William A. PearmanPablo G. Tahoces, Miguel Souto, Juan J. Vidal, EmbeddedWavelet Region-Based Coding Methods Applied To Digital Mammography IEEE International Conference on Image Processing, vol.3 Barcelona, Spain 2003.

[3] David Gibson, Michael Spann, Sandra I. Woolley, A Wavelet-Based Region of Interest Encoder for the Compression of Angiogram Video Sequences, IEEE Transactions on Information Technology in Biomedicine 8(2): 103-113 (2004)

[4] Kaibin Wang, B.Eng., M.Eng. Thesis: Region-Based Three-Dimensional Wavelet Transform Coding, Carleton University, Ottawa, Canada, May, 2005.

[5] Zixiang Xiong, Qiang Wu and Kenneth R.,Castlemen, Enhancement, Classification And Compression Of Chromosome Images. Workshop on Genomic Signal Processing, 2002 , Cite seer.

[6] Charilaos Christopoulos, Athanassios Skodras ,Touradj Ebrahimi - The JPEG2000 Still Image Coding System:An Overview , IEEE Transactions on Consumer

Electronics, Vol. 46, No. 4, pp. 1103-1127, November 2000.

[7] K.P.Soman and K.I.Ramachandran -Insight Into

Wavelets –From Theory to Practice,PHI Publications. [8] Sonja Grgic, Mislav Grgic, and Branka Zovko-Cihlar,

Performance Analysis of Image Compression Using Wavelets, IEEE Transactions on Industrial Electronics, vol.48, no. 3, June 2001.

[9] Michael W. Marcellina, Margaret A. Lepleyb, Ali Bilgina, Thomas J. Flohrc, Troy T. Chinend, James H. Kasner, An overview of quantization in JPEG 2000 Signal Processing: Image Communication 17 (2002) 73–84.

Authors Profile P.Rajkuma received the M.Tech. Degree in Electronics Engineering from Maulana Azad National Institute of Tech.Bhopal, India in 2006.He is doing his research in the area of medical image processing. Madhu Shandilya received the M.Tech. and Ph.D. degrees in Electronics Engineering from Maulana Azad National Institute of Tech.Bhopal, India in 1994 and 2005, respectively. From last 22 years she is associated with Electronics Engineering from Maulana Azad National Institute of Tech.Bhopal. She has published/presented about 20 papers in various national and international journals /conferences. Her area of specialization is digital image processing.

PSNR (db)

Bits per pixel


6

2- Dimension Automatic Microscope Moving Stage

Ovasit P.1 Adsavakulchai S.1,* Srinonghang W1., and Ueatrongchit, P. 1

1School of Engineering, University of the Thai Chamber of Commerce

126/1 Vibhavadi Rangsit Rd., Bangkok 10400 *Corresponding author, e-mail: [email protected]

Abstract: Currently, microscope stage has controllable by manual. In medical laboratory, the specimen slide examination system processes one at a time for microscopic examination. The main objective of this study is to develop a two-dimension automatic microscope moving stage. This equipment is designed by microcontroller PIC 16F874 using stepping motor as horizontal feed mechanism. There are three function modes, the first one is manual, the second is automatic scan specimen which transfers the specimen slide onto a microscope stage which has controllable X and Y axis positioning to move the specimen slide into the optical viewing field of the microscope and examination over the desired area of the specimen and the last one is to examine specimen slides may be automatically returned to the microscope stage for reexamination. The result of this study can be concluded that the accuracy of this equipment for reexamination the specimen slide is 86.03 % accuracy.

Keywords: Microscope Automatic Moving Stage, microcontroller PIC16F874, stepping motor

1. Introduction The microscope is a conventional laboratory microscope with attached to actuate the stage and control is affected through manual [1] as shown in Figure 1.

Figure 1. Microscope moving stage The basic principal for diagnostic in red blood cell is using microscope manually. In such cases, electronic systems may be used to automatically examine and analyze the optical images of the microscope [2]. Where electronics systems are used for rapid analysis of microscope specimen images it becomes desirable to automatically regularly and rapidly feed the specimens to the microscope optics. After analysis a specimen would be removed to make room for the next specimen and would be collected for either further examination, reference, record keeping or disposal.

An automated slide system is disclosed which organizes microscope slides in cassettes, automatically and positioned each slide under the microscope as provided by the protocol, and after examination returns the slide to its proper cassette [3]. A slot configured for holding slides in spaced parallel configuration using the mechanism for removing and replacing a slide housed. A feed arm containing a longitudinal to draw-out spring wire surrounding an imaginary longitudinal axis having at the first end and a second end, the first and second end being bent orthogonal to one another and to the imaginary longitudinal axis of said draw-out spring wire, said longitudinal draw-out spring wire being positioned in said longitudinal channel in said feed arm such that bent ends protrude from the channel and wherein said longitudinal draw-out spring wire is operatively positioned in said longitudinal channel such that the draw-out spring wire is rotatable therein, allowing for each bent end to change orientation in respect to the feed arm. [1].

The main objective of this study is to develop an automatically returned to the microscope stage for reexamination. These technologies dramatically increased the accuracy of measurement results and contributed greatly to the modernization of testing and medical care (medical testing).

2. Materials and Methods Sample preparation: 2.1 To prepare the blood smear sample and set up the

microscope working area at 1000x with 0.2 mm. dimension [2] as shown in Figure 2. In order to the area of the specimen slide is viewed during examination of the specimen without sliding it.

Figure 2. Microscope working area

2.2 To set up the scope of microscope moving stage

scanning area with 40 x 26 mm. as shown in Figure 3. About the specimen stage this opening in the specimen


7

stage is made as large as possible and exposes the full width of the specimen slide.

Figure 3. Scanning area

2.3 The automatic sequential examination of a group of microscope specimen slides comprising: 2.3.1 Moving stage design: Apparatus comprised a substage directing serves to move stage in a horizontal plane and there is provided further positioning means supporting said translation means and operable for moving said stage with a specimen slide supported therein vertically as shown in Figure 4.

Figure 4. Moving stage design

2.3.2 Digital Electronics Design:

The digital electronics architecture has two main functional blocks, Master Board and Slave Board. ICP (Instrument Control Processor): used a PIC 16F873 processor to perform all instrument control and event processing functions as shown in Figure 5. The ICP will be responsible for the following tasks: processing commands; monitoring source and adjusting the LCD readout mode as required; calculating centroids and transmitting centroid positions [4],[5],[6].

Figure 5. Digital Electronics design

Figure 5 is to illustrate the overall digital electronics using two microcontrollers with serial communication and synchronous Serial Peripheral Interface (SPI).

2.3.3 Microcontroller in Slave Board:

2.3.3.1 Encoder: using sequential logic to control moving stage. The characteristic of encoder 2 signal using microcontroller PIC 16F873 via RA0-RA3 port as shown in Figure 6. The signals moving stage is to set up into 3 statuses 1. No movement 2. Increase the distance and 3. Reduce distance as shown in Figure 7.

Figure 6. The characteristic of encoder 2 signal

Figure 7. Logical control

2.3.3.2 Stepping motor: control moving stage using microcontroller PIC 16F873 via RB0-RB7 port and working together with IC ULN2803 to control stepping motor as shown in Figure 8.


8

Figure 8. Characteristics of stepping motor control

2.3.3.3 Serial Peripheral Interface: using Master Synchronous Serial Port in microcontroller PIC 16F873 using Microcontroller in Slave Board as shown in Figure 9.

Figure 9. Serial Peripheral Interface (SPI)

2.3.4 Microcontroller in Mater Board 2.3.4.1 Input from keyboard using RB1-RB7 in term of matrix 4 x 3 2.3.4.2 Display result using RA0-RA5 for Liquid

Crystal Display 2.3.4.3 Serial Peripheral Interface SCK port 2.3.4.5 Data communication using SPI as shown

in table 1 Table 1 : Data communication using SPI

To design main menu as shown in Figure 10 to control the

microscopic stage.

X 0000 #:ExitY 0000 *:Stop

Microscope Scan.Version 1.00

1:Home 2:Manual3:Scan 4:Config

X 0000 #:ExitY 0000 *:Jump

HomePlease Wait.

1:Posit. #:Exit2:Option Config

3

1 4

Wait

1:Start #:Exit2:Final < Save

X _000 #:ExitY 0000 < Jump

Wait

Start.Please Wait.

X 0000 #:ExitY 0000 *:Scan

*

T: 0sec #:ExitS:>_unit Space

1: 0sec #:Exit2: 0unit Option

Start 0000x0000yFinal 0000x0000y

T:>_sec #:ExitS: 0unit Time.

2 : S

1

1 : T

2

* : X :Y# : 1 or 2

#

2

#

Figure 10. Main menu of control program

3. RESULTS AND DISCUSSION To test the points and the results is shown in table 2.

The microscopic examination of the specimen slide can take place either visually or automatically. Motorized microscope components and accessories enable the investigator to automate live-cell image acquisition and are particularly useful for time-lapse experiments about 20 milliseconds [7],[8]. For this purpose the X and Y positioning systems can be controlled manually or automatically. Thus, the specimen slide carried by the stage may be moved to any desired location relative to the optical axis by actuation of the Y-axis drive 43 and the X-axis drive 44. For automatic examination the drives 43, 44 would be energized under scan or other program control. Finally, the specimen slide reexamination is automatically returned to the microscope stage for reexamination with very high accuracy.

Table 2: Sample testing results

Status 1a 1b 2a 2b

Master

Register SSPBUF

Register SSPBUF

Register SSPSR Register SSPSR

Slave


9

4. Conclusion

After the examination of a particular series of specimen slides has been completed any individual specimen slide that requires re-examination can by either operator signals or by predetermined control signals be fed automatically back into the microscope viewing optics for further examination [9].

Upon completion of the examination of a slide the horizontal positioning Y-axis drive returns the specimen slide on the stage to the position. It can be concluded that the accuracy of this equipment for reexamination the specimen slide is 86.03 % accuracy.

Acknowledgements

This project is supported by University of the Thai Chamber of Commerce grant.

References [1] C.R.David, Microscopy and related methods from

http://www.ruf.rice.edu/~bioslabs/methods/microscopy/microscopy.html

[2] Robert H., et.al, Handbook of Hematologic Pathology, Marcel Dekker, Inc. (2000)

[3] Qin Zhang et al., A Prototype Neural Network Supervised Control System for Bacillus thuringiensis Fermentations, Biotechnology and Bioengineering, vol. 43, pp. 483-489 (1994).

[4] N. Armenise et al., High-speed particle tracking in nuclear emulsion by last generation automatic

microscopes, accepted for publication on Nucl. Instr. Meth. A(2005).

[5] Powell, Power and Perkins, The Study of Elementary Particles by the Photo-graphic Method, Pergamon Press (1959).

[6] W.H. Barkas, Nuclear Research Emulsion, Academic Press Inc. (London) Ltd. (1963).

[7] John F. Reid et al., A Vision-based System for Computer Control and Data Acquisition in fermentation Processes, 38th Food Technology Conference, 1992.

[8] J.M. Shine, Jr., et al., Digital Image Analysis System for Determining Tissue-Blot Immunoassay Results for Ratoon Stunting Disease of Sugarcane, Plant Disease, vol. 77 No. 5, pp. 511-513, 1993.

[9] Jinlian Ren et al., Knowledge-based Supervision and Control of Bioprocess with a Machine Vision-based Sensing System, Journal of Biotechnology 36 (1994) 25-34.

Authors Profile Suwannee Adsavakulchai received the M.S. degrees in Computer Information Systems from Assumption University in 1994 and Doctoral of Technical Science from Asian Institute of Technology in 2000, respectively. She now works as lecturer in the department of Computer Engineering, School of Engineering, University of the Thai Chamber of Commerce.


10

A Large Block Cipher Involving a Key Applied on Both the Sides of the Plain Text

V. U. K. Sastry1, D. S. R. Murthy2, S. Durga Bhavani3

1Dept. of Computer Science & Engg., SNIST,

Hyderabad, India, [email protected]

2Dept. of Information Technology, SNIST,


3School of Information Technology, JNTUH,


Abstract: In this paper, we have developed a block cipher by modifying the Hill cipher. In this, the plain text matrix P is multiplied on both the sides by the key matrix. Here, the size of the key is 512 bits and the size of the plain text is 2048 bits. As the procedure adopted here is an iterative one, and as no direct linear relation between the cipher text C and the plain text P can be obtained, the cipher cannot be broken by any cryptanalytic attack. Keywords: Block Cipher, Modular arithmetic inverse, Plain text, Cipher text, Key.

1. Introduction The study of the block ciphers, which was initiated several centuries back, gained considerable impetus in the last quarter of the last century. Noting that diffusion and confusion play a vital role in a block cipher, Feistel etal, [1] –[2] developed a block cipher, called Feistel cipher. In his analysis, he pointed out that, the strength of the cipher increases when the block size is more, the key size is more, and the number of rounds in the iteration is more. The popular cipher DES [3], developed in 1977, has a 56 bit key and a 64 bit plain text. The variants of the DES are double DES, and triple DES. In double DES, the size of the plain text block is 64 bits and the size of the key is 112 bits. In the triple DES, the key is of the length 168 bits and the plain text block is of the size is 64 bits. At the beginning of the century, noting that 64 bit block size is a drawback in DES, Joan Daemen and Vincent Rijmen, have developed a new block cipher called AES [4], wherein the block size of the plain text is 128 bits and key is of length 128, 192, or 256 bits. In the subsequent development, on modifying Hill cipher, several researchers [5]–[9], have developed various cryptographical algorithms wherein the length of the key and the size of the plain text block are quite significant. In the present paper, our objective is to develop a block cipher wherein the key size and the block size are significantly large. Here, we use Gauss reduction method for obtaining the modular arithmetic inverse of a matrix. In what follows, we present the plan of the paper. In section 2, we have discussed the development of the cipher. In section 3, we have illustrated the cipher by

considering an example. In section 4, we have dealt with the cryptanalysis of the cipher. Finally, in section 5, we have presented the computations and arrived at the conclusions.

2. Development of the cipher Consider a plain text P which can be represented in the form of a square matrix given by P = [Pij], i = 1 to n, j = 1 to n, (2.1) where each Pij is a decimal number which lies between 0 and 255. Let us choose a key k consisting of a set of integers, which lie between 0 and 255. Let us generate a key matrix, denoted as K, given by K = [Kij], i = 1 to n, j = 1 to n, (2.2) where each Kij is also an integer in the interval [0 – 255]. Let C = [Cij], i = 1 to n, j = 1 to n (2.3) be the corresponding cipher text matrix. The process of encryption and the process of decryption applied in this analysis are given in Fig. 1.

Figure 1. Schematic Diagram of the cipher

Here r denotes the number of rounds. In the process of encryption, we have used an iterative procedure which includes the relations P = (K P K) mod 256, (2.4)


11

P = Mix (P), (2.5) and P = P ⊕ K (2.6) The relation (2.4) causes diffusion, while (2.5) and (2.6) lead to confusion. Thus, these three relations enhance the strength of the cipher. Let us consider Mix (P). In this the decimal numbers in P are converted into their binary form. Then we have a matrix of size n x 8n, and this is given by

Here, P111, P112, …, P118 are binary bits corresponding to P11. Similarly, Pij1, Pij2, …, Pij8 are the binary bits representing Pij. The above matrix can be considered as a single string in a row wise manner. As the length of the string is 8n2, it is divided into n2 substrings, wherein the length of each substring is 8 bits. If n2 is divisible by 8, we focus our attention on the first 8 substrings. We place the first bits of these 8 binary substrings, in order, at one place and form a new binary substring. Similarly, we assemble the second 8 bits and form the second binary substring. Following the same procedure, we can get six more binary substrings in the same manner. Continuing in the same way, we exhaust all the binary substrings obtained from the plain text. However, if n2 is not divisible by 8, then we consider the remnant of the string, and divide it into two halves. Then we mix these two halves by placing the first bit of the second half, just after the first bit of the first half, the second bit of the second half, next to the second bit of the first half, etc. Thus we get a new binary substring corresponding to the remaining string. This completes the process of mixing. In order to perform the exclusive or operation in P = P ⊕ K, we write the matrices, both P and K, in their binary form, and carryout the XOR operation between the corresponding binary bits. In the process of decryption, the function IMix represents the reverse process of Mix. In what follows, we present the algorithms for encryption, and decryption. We also provide an algorithm for finding the modular arithmetic inverse of a square matrix. Algorithm for Encryption 1. Read n, P, K, r 2. for i = 1 to r { P = (K P K) mod 256 P = Mix (P) P = P ⊕ K } 3. C = P 4. Write (C)

Algorithm for Decryption 1. Read n, C, K, r 2. K–1 = Inverse (K) 3. for i = 1 to r { C = C ⊕ K C = IMix (C) C = (K–1 C K–1) mod 256 } 4. P = C 5. Write (P) Algorithm for Inverse (K) // The arithmetic inverse (A–1), and the determinant of the

matrix (∆) are obtained by Gauss reduction method. 1. A = K, N = 256 2. A–1 = [Aji] / ∆, i = 1 to n, j = 1 to n

//Aji are the cofactors of aij, where aij are elements of A, and ∆ is the determinant of A

3. for i = 1 to n { if ((i ∆) mod N = 1) d = i; break; } 4. B = [d Aji] mod N // B is the modular arithmetic inverse of A

3. Illustration of the cipher Let us consider the following plain text. No country wants to bring in calamities to its own people. If the people do not have any respect for the country, then the Government has to take appropriate measures and take necessary action to keep the people in order. No country can excuse the erratic behaviour of the people, even though something undue happened to them in the past. Take the appropriate action in the light of this fact. Invite all the people to come into the fold of the Government. Try to persuade them as far as possible. Let us see!! (3.1) Let us focus our attention on the first 256 characters of the above plain text which is given by No country wants to bring in calamities to its own people. If the people do not have any respect for the country, then the Government has to take appropriate measures and take necessary action to keep the people in order. No country can excuse the erratic (3.2) On using EBCDIC code, we get 26 numbers, corresponding to 256 characters. Now on placing 16 numbers in each row, we get the plain text matrix P in the decimal form

Obviously, here the length of the plain text block is 16 x 16 x 8 (2048) bits.


12

Let us choose a key k consisting of 64 numbers. This can be written in the form of a matrix given

by The length of the secret key (which is to be transmitted) is 512 bits. On using this key, we can generate a new key K in the form

where U = QT, in which T denotes the transpose of a matrix, and R and S are obtained from Q and U as follows. On interchanging the 1st row and the 8th row of Q, the 2nd row and the 7th row of Q, etc., we get R. Similarly, we obtain S from U. Thus, we have

whose size is 16 x 16. On using the algorithm for modular arithmetic inverse (See Section 2), we get

On using (3.6) and (3.7), it can be readily shown that

K K–1 mod 256 = K–1K mod 256 = I. (3.8) On applying the encryption algorithm, described in Section 2, we get the cipher text C in the form

On using (3.7) and (3.9), and applying the decryption algorithm presented in section 2, we get the Plain text P. This is the same as (3.3). Let us now find out the avalanche effect. To this end, we focus our attention on the plain text (3.2), and modify the 88th character ‘y’ to ‘z’. Then the plain text changes

only in one binary bit as the EBCDIC code of y is 168 and that of z is 169.

On using the encryption algorithm, we get the cipher text C corresponding to the modified plain text (wherein y is replaced by z) in the form

On comparing (3.9) and (3.10), we find that the two cipher texts

differ in 898 bits, out of 2048 bits, which is quite considerable. However, it may be mentioned here that, the impact of changing 1 bit is not that copious, as the size of

the plain text is very large. Even then it is remarkable. Now let us change the key K given in (3.6) by one

binary bit. To this end, we replace the 60th element 5 by 4. Then on using the original plain text given by (3.3), we get C in the form

On comparing (3.9) and (3.11), we find that the cipher texts

differ in 915 bits, out of 2048 bits. From the above analysis, we find that the avalanche effect is quite pronounced and shows very clearly that the cipher is a strong one.

4. Cryptanalysis In the literature of cryptography, it is well known that the different types of attacks for breaking a cipher are:

(1) Cipher text only attack, (2) Known plain text attack, (3) Chosen plain text attack, (4) Chosen cipher text attack.

In the first attack, the cipher text is known to us together with the algorithm. In this case, we can determine the plain text, only if the key can be found. As the key contains 64 decimal numbers, the key space is of size 2512 ∼ (103)51.2 = 10153.6 which is very large. Hence, the cipher cannot be broken by applying the brute force approach. We know that, the Hill cipher [1] can be broken by the known plain text attack, as there is a direct linear relation between C and P. But in the present modification, as we have all nonlinear relations in the iterative scheme, the C can never be expressed in terms of P, thus P cannot be determined by any means in terms of other quantities. Hence, this cipher cannot be broken by the known plain text attack.


13

As there are three relations, which are typical in nature, in the iterative process for finding C, no special choice of either the plain text or the cipher text or both can be conceived to break the cipher.

5. Conclusions In the present paper, we have developed a large block cipher by modifying the Hill cipher. In the case of the Hill cipher, it is governed by the single, linear relation C = (K P) mod 26, (5.1) while in the present case, the cipher is governed by an iterative scheme, which includes the relations P = (K P K) mod 256, (5.2) P = Mix (P), (5.3) and P = P ⊕ K. (5.4) Further, it is followed by C = P. (5.5) In the case of the Hill cipher, we are able to break the cipher as there is a direct linear relation between C and P. On the other hand, in the case of the present cipher, as we cannot obtain a direct relation between C and P, this cipher cannot be broken by the known plain text attack. By decomposing the entire plain text given by (3.1) into blocks, wherein each block is of size 256 characters, the corresponding cipher text can be obtained in the decimal form. The first block is already presented in (3.9) and the rest of the cipher text is given by

In this analysis, the length of the plain text block is 2048 bits and the length of the key is 512 bits. As the cryptanalysis clearly indicates, this cipher is a strong one and it cannot be broken by any cryptanalytic attack. This analysis can be extended to a block of any size by using the concept of interlacing [5].

References

[1]. Feistel H, “Cryptography and Computer Privacy”, Scientific American, May 1973.

[2]. Feistel H, Notz W, Smith J, “Some Cryptographic Techniques for Machine-to-Machine Data Communications”, Proceedings of the IEEE, Nov. 1975.

[3]. William Stallings, Cryptography and Network Security, Principles and Practice, Third Edition, Pearson, 2003.

[4]. Daemen J, Rijmen V, “Rijdael: The Advanced Encryption Standard”, Dr. Dobb’s Journal, March 2001.

[5]. V. U. K. Sastry, V. Janaki, “On the Modular Arithmetic Inverse in the Cryptology of Hill Cipher”, Proceedings of North American Technology and Business Conference, Sep. 2005, Canada.

[6]. V. U. K. Sastry, S. Udaya Kumar, A. Vinaya Babu, “A Large Block Cipher using Modular Arithmetic Inverse of a Key Matrix and Mixing of the Key Matrix and the Plaintext”, Journal of Computer Science 2 (9), 698 – 703, 2006.

[7]. V. U. K. Sastry, V. Janaki, “A Block Cipher Using Linear Congruences”, Journal of Computer Science 3(7), 556 – 561, 2007.

[8]. V. U. K. Sastry, V. Janaki, “A Modified Hill Cipher with Multiple Keys”, International Journal of Computational Science, Vol. 2, No. 6, 815 – 826, Dec. 2008.

[9]. V. U. K. Sastry, D. S. R. Murthy, S. Durga Bhavani, “A Block Cipher Involving a Key Applied on Both the Sides of the Plain Text”, International Journal of Computer and Network Security (IJCNS), Vol. 1, No.1, pp. 27 – 30, Oct. 2009.

Authors Profile Dr. V. U. K. Sastry is presently working as Professor in the Dept. of Computer Science and Engineering (CSE), Director (SCSI), Dean (R & D), SreeNidhi Institute of Science and Technology (SNIST), Hyderabad, India. He was Formerly Professor in IIT, Kharagpur, India and worked in IIT, Kharagpur during 1963 – 1998. He guided 12 PhDs, and published more than 40 research papers in various international journals. His research interests are Network Security & Cryptography, Image Processing, Data Mining and Genetic Algorithms. Dr. S. Durga Bhavani is presently working as Professor in School of Information Technology (SIT), JNTUH, Hyderabad, India. Her research interest is Image Processing. Mr. D. S. R. Murthy obtained B. E. (Electronics) from Bangalore University in 1982, M. Tech. (CSE) from Osmania University in 1985 and presently pursuing Ph.D. from JNTUH, Hyderabad since 2007. He is presently working as Professor in the Dept. of Information Technology (IT), SNIST since Oct. 2004. He earlier worked as Lecturer in CSE, NIT (formerly REC), Warangal, India during Sep. 1985 – Feb. 1993, as Assistant Professor in CSE, JNTUCE, Anantapur, India during Feb. 1993 – May 1998, as Academic Coordinator, ISM, Icfaian Foundation, Hyderabad, India during May 1998 – May 2001 and as Associate Professor in CSE, SNIST during May 2001 - Sept. 2004. He worked as Head of the Dept. of CSE, JNTUCE, Anantapur during Jan. 1996 – Jan 1998, Dept. of IT, SNIST during Apr. 2005 – May 2006, and Oct. 2007 – Feb. 2009. He is a Fellow of IE(I), Fellow of IETE, Senior Life Member of CSI, Life Member of ISTE, Life Member of SSI, DOEACC Expert member, and Chartered Engineer (IE(I) & IETE). He published a text book on C Programming & Data Structures. His research interests are Image Processing and Image Cryptography.


14

Electrocardiogram Prediction Using Error Convergence-type Neuron Network System

Shunsuke Kobayakawa1 and Hirokazu Yokoi1

1Graduate School of Life Science and Systems Engineering, Kyushu Institute of Technology

2-4 Hibikino, Wakamatsu-ku, Kitakyushu-shi, Fukuoka 808-0196, Japan {kobayakawa-shunsuke@edu., yokoi@}life.kyutech.ac.jp Abstract: The output error of a neuron network cannot converge at zero, even if training for the neuron network is iterated many times. “Error Convergence-type Neuron Network System” has been proposed to improve this problem. The output error of the proposed neuron network system is converged at zero by infinitely increasing the number of neuron networks in the system. A predictor was constructed using the system and applied to electrocardiogram prediction. The purpose of this paper is to prove the validity of the predictor for electrocardiogram. The neuron networks in the system were trained 30,000 cycles. As a result, averages of root mean square errors of the first neuron network was 2.60×10-2, and that of the second neuron network was 1.17×10-5. Prediction without error was attained by this predictor, so that its validity was confirmed.

Keywords: Volterra neuron network, Predictor, Error free, Electrocardiogram.

1. Introduction A neuron network (NN) cannot be learned if there is not a correlation between input signals to each layer and teacher signals of it. Therefore, uncorrelated components between them cannot be learned when correlated and uncorrelated components intermingled between them. NN cannot learn completely correlated components between them when learning capability of NN is low. Elevating learning capability of NN as a means to learn completely these correlated components is thought. However, such NN has not existed. Various researches have been actively done to elevate output accuracy of NN up to now. Improvements on the I/O characteristics of a neuron [1-11,17], the structure of NN [13-23] and NN system [24] are effective as means of the output accuracy elevation. However, output errors of their NNs cannot converge at zero, even if trainings for their NNs are iterated many times. Therefore, usual NNs are used with output error in tolerance. However, it is insufficient when applying NN to predictive coding [10-12] and the flight control for an aircraft [18,19] and a spacecraft etc. which smaller output error is advisable.

Then, NN system with a possibility which this problem can be solved even with NN which learning capability is low if it is used by plural has been proposed. As a result, “Error Convergence Method for Neuron Network System (ECMNNS)” which output error of single output NN system using NNs of multi-step converges and “Error Convergence Neuron Network System (ECNNS)” which is designed using it have been proposed by S. Kobayakawa [25]. The output error is theoretically converged at zero by infinitely increasing the number of NNs in ECNNS. However, it is necessary to devise ECNNS when it is used as plural

outputs, for ECNNS is a single output. The highest output accuracy of NN system cannot be expected for even if ECMNNS is simply applied to BP network (BPN) [26], because the BPN has the mutual interference problem of learning between outputs [18-22]. Then “Error Convergence Parallel-type Neuron Network System (ECPNNS)” which ECNNS was applied to a parallel-type NN which does not have the above-mentioned problem and which outputs accuracies are high to deal with plural outputs has been designed. Furthermore, “Error Convergence-type Recurrent Neuron Network System (ECRNNS)” which ECNNS was applied and “Error Convergence Parallel-type Recurrent Neuron Network System (ECPRNNS)” which ECPNNS was applied also have been designed. It is theoretically shown that output accuracy of NN system is elevated by them.

In general, BPN which learns a time series signal using a teacher signal and input signals is that two teacher signal values or more occasionally correspond to same values of input signals. In such a state, the BPN cannot be used for ECNNS because it cannot be learned. There are means using input-delay NN and Volterra NN (VNN) to eliminate this problem. These means are used for usual researches on compression for nonlinear signal using predictive coding [10,27]. Learning for a nonlinear predictor using NNs is easier by strengthening of causality between signals from past to present and a prediction signal. Therefore, Learning for NN at the first step in ECNNS is comparatively easy. However, learning for NN at high step in ECNNS is difficult because the causality weakens by rising of steps in ECNNS. Then, ECNNS is redesigned for improvement on learning capability. The redesigning is that NN at each step in ECNNS can be used as a predictor to strengthen causality between an input signal and a teacher signal of it. As a result, predictor using ECNNS [28] has been designed. This is called “Error Convergence-type Predictor (ECP)”

The purpose of this paper is to prove the validity of ECP with rounding which is constructed of 2nd-order VNNs (2VNN) of two steps for a normal sinus rhythm electrocardiogram (ECG). This ECP is called “Error Convergence-type 2nd-order Volterra Predictor (EC2VP)”. As a result, prediction without error was attained by this predictor, so that its validity was confirmed.

2. Single Output Error Convergence-type Neuron Network System

2.1 Principle Here, it thinks about a single output NN which error of the output signal to a teacher signal does not converge at zero though it becomes smaller than one before training for the


15

NN under the result of executing until the training converges to all input signals with correlation between the teacher signal.

The training is executed by this NN at the beginning. Next, other training is executed by another NN using an error signal obtained from an output signal and a teacher signal of the NN as a teacher signal. Thus it is thought that an error of an output signal to a teacher signal cannot be converged at zero, even if it is kept training an error signal obtained from training of previous NN one after another by another NN. That is, the error obtained from a difference between the sum of the output signals of all NNs and the teacher signal of the first NN does not converge at zero. This is shown in expressions from (1) to (4).

i i iy z ε= + ( )1, 2, ,i n= L (1)

1i iyε += ( )1, 2, , 1i n= −L (2)

11

n

i ni

y z ε=

= +∑ (3)

lim 0nn

ε←∞

≠ (4)

where y is the teacher signal, z is the output signal, ε is the output error, suffix of each sign is steps of NN.

Moreover, trainings for error signals used as a teacher signal for NN at the second step or more become difficult because it is thought that error signals become small along with the number of steps. Error signals which are amplified are used for teacher signals to NN at the second step or more to improve this problem. As a result, these trainings become easy. An output error of NN after training is reduced when it is restored to signal level before the amplification, if the amplification factor is larger than 1. Moreover, it approaches zero when the amplification factor is very large.

Here, an output error obtained from training under an amplification factor of an error signal which is a teacher signal to NN at the n step as An is assumed to be ε. Output error εn to a teacher signal given to a whole of NN system is shown like expression (5) and (6). Moreover, it is shown like expressions (7) and (8) when steps of NN are infinitely, and the output error converges at zero. Therefore, an error obtained from a difference between a sum of output signals after restoration of all NNs and the teacher signal given to the whole of NN system converges at zero from expressions (3) and (8) when the steps of NN are infinitely, and they become equal. This is shown in expression (9). Thus, this means to improve the above-mentioned problem is effective and necessary to obtain highly accurate output. This means is called ECMNNS, and NN system which applies this is called ECNNS.

nnA

εε = (5)

nε ε< (6)

lim nn

A←∞

= ∞ (7)

TrainingExtracting the error

Gain tuning to the teacher signal

Spreading the errorRestoration to an output signal

Sum of restored output signals

Iteration until the error converges at zero

+-

Teacher signal to ECNNS

Figure 1. Concept of processing for error convergence-type neuron network system

lim 0n n nAε

ε←∞

= = (8)

1lim

n

in iy z

←∞ =

= ∑ (9)

2.2 Single Output Neuron Network System Figure 1 shows a concept of processing for ECNNS. Figure 2 shows that ECNNS is designed based on this concept. Symbol NN in this figure also contains the state of a neuron. This ECNNS can be built in freely selected NN and the learning rule of the processible type for the I/O relation which internal each NN should be achieved. Here, the general I/O characteristics of ECNNS are discussed without touching concerning the I/O characteristic of NN concretely applied to ECNNS.

ECNNS equips with amplifiers to tune amplitudes of input signals and a teacher signal to NN at each step, to execute the appropriate training. Furthermore, ECNNS also equips with amplifiers for restoration which amplification factor is a reciprocal of an amplification factor for amplification of the teacher signal to each NN on its output part, to restore its output signal level. The I/O characteristics of ECNNS are shown in expressions from (10) to (16). Moreover, the relation of their teacher signals is shown in expressions from (17) to (20). Here, conditions of the amplification factors used by expressions (16), (18) and (19) are expressions (21) and (22).

( )

in1 2, , ,x nx x x= L (10)

( )inin 1 2, , ,A i i i ina a a= L ( )1, 2, ,i n= L (11)

Aij ij jx a x= ( )in1, 2, , ; 1,2, ,i n j n= =L L (12)

( )in1 2, , ,i Ai Ai Ainx x x=x L ( )1, 2, ,i n= L (13)

( )xAi i iz f= ( )1, 2, ,i n= L (14)


16

-

-+

+

-+

-+

-

-

-1

1

n

ii

z−

=∑

A2

1/A2

1/A3

A3

1/An

An

1/A1-+

A1

+

+

+

Ain2

Ain3

Ainn NNn

x Ain11x

2x

3x

nx

1ε

2ε

3ε

nε

1Az1Ay

2Az

3Az

Anz

1z

2z 3z

2Ay

3Ay

nzAny

y

z

+

+

+

+

NN1

NN2

NN3

2y

3y

ny

1y

Figure 2. Error convergence-type neuron network system

1

n

ii

z z=

= ∑ (15)

Ai

ii

zz

A= ( )1, 2, ,i n= L (16)

1y y= (17)

1 1 1Ay A y= (18)

1

1

i

Ai i jj

y A y z−

=

= −

∑ ( )2,3, ,i n= L (19)

i Ai Aiy zε = − ( )1, 2, ,i n= L (20)

0iA ≠ ( )1, 2, ,i n= L (21)

lim nn

A←∞

= ∞ (22)

where x is the input signal vector, x is the input signal, xAi is the input signal after amplification at the ith step, xi is the input signal vector after amplification at the ith step, Aini is the input signal amplification factor vector at the ith step, ai is the input signal amplification factor at the ith step, suffixes of x, xAi and a i are input signal numbers, zAi is the output signal of NN at the ith step, fi is a nin variables function which shows the I/O relation of NN at the ith step, y is the teacher signal to ECNNS, z is the output signal of ECNNS, yAi is the teacher signal to NN at the ith step after amplification, zi is the output signal to NN at the ith step after restoration, Ai is the teacher signal amplification factor for NN at the ith step, εi is the output error of NN at the ith step.

x ECNNS1

ECNNS2

y1z1y2z2

ECNNS3y3z3

ECNNSmymzm

Figure 3. Error convergence parallel-type neuron network system

3. Plural Outputs Error Convergence-type Neuron Network system

3.1 Principle ECNNSs are applied to the parallel-type NN when applying ECNNS to a plural outputs NN system, for ECNNS is a single output. NN system of this type is called ECPNNS. Moreover, it can be said a general type of ECNNS. Figure 3 shows ECPNNS. This is constructed of parallel units which apply ECNNS of the same number as the outputs. Therefore, it is thought that high output accuracy is obtained, because the mutual interference problem of learning between outputs of BPN is not caused either, for training for each output is executed independently by ECNNS.

3.2 Plural Outputs Neuron Network System I/O characteristics of ECPNNS are shown in expressions from (23) to (28). Moreover, relation of teacher signals is shown in expressions from (29) to (32). Here, conditions of amplification factors used by expressions (28), (30) and (31) are expressions (33) and (34).

( )inin 1 2, , ,A ij ij ij ijna a a= L

( )1, 2, , ; 1, 2, , ii m j n= =L L (23)

Aijk ijk kx a x= in

1, 2, , ; 1,2, ,1, 2, ,

ii m j nk n

= = =

L L

L

(24)

( )in1 2, , ,ij Aij Aij Aijnx x x=x L

( )1, 2, , ; 1, 2, , ii m j n= =L L (25)

( )xAij ij ijz f= ( )1,2, , ; 1, 2, , ii m j n= =L L (26)

1

in

i ijj

z z=

= ∑ ( )1, 2, ,i m= L (27)

Aij

ijij

zz

A= ( )1,2, , ; 1, 2, , ii m j n= =L L (28)

1i iy y= ( )1, 2, ,i m= L (29)

1 1 1Ai i iy A y= ( )1, 2, ,i m= L (30)


17

1

1

j

Aij ij i ikk

y A y z−

=

= −

∑ ( )1,2, , ; 2,3, , ii m j n= =L L

(31) ij Aij Aijy zε = − ( )1,2, , ; 1, 2, , ii m j n= =L L (32)

0ijA ≠ ( )1, 2, , ; 1,2, , ii m j n= =L L (33)

lim inn

A←∞

= ∞ ( )1, 2, ,i m= L (34)

where the first figure of suffix of each sign is the parallel unit number, the second figure of it is the step number in ECNNS, the third figure of it is the input signal number. However, xk is the kth input signal to ECNNS.

4. Applications 4.1 Simulator

Here are two application network systems using ECNNS and ECPNNS. For example, control signals and state signals are necessary for input when achieving a simulator for nonlinear plant. Furthermore, the state signals output from the simulator every moment decided depending on the input signals are recurrently needed for the state input signals. When achieving the simulator with high accuracy, it is advisable that errors are not included in output of the simulator. Therefore, it is thought that the simulator which works by the specified significant figure obtains a highly accurate output by the plant model using ECNNS and a rounding to eliminate the output error. Concretely, the network system which the output signal of ECNNS recurs to the input through a rounding shown in Figure 4 is thought about the simulator in case of a state signal. This is called ECRNNS. Furthermore, the network system which output signal vector of ECPNNS recurs to the input through roundings as shown in Figure 5 is thought about the simulator in case of plural state signals. This is called ECPRNNS. In these figures, y is the teacher signal vector, and z is the output signal vector.

4.2 Nonlinear Predictor A nonlinear predictor used for predictive coding and its principle must be improved to obtain high accuracy. Here is a means for improvement on learning capability of NN at each step in ECNNS for a nonlinear predictor used for predictive coding. This is called ECP. Learning for a nonlinear predictor using NNs is easier by strengthening of causality between signals from past to present and a prediction signal. Therefore, learning for NN at the first step in ECNNS is comparatively easy. However, learning for NN at the high step is difficult because it is guessed that the causality weakens by rising of the steps. Then, NN at each step in ECNNS is used as a predictor to strengthen causality between input signals and a teacher signal of it.

Moreover, learning capability of the NN can be elevated by increasing the number of input signals [29]. Figure 6 is redesigning of ECNNS in Figure 2 to realize it. Furthermore, expression (13) is changed to expressions from (35) to (38). Expression (37) shows initial conditions for NNs from the second step. I/O relations of NNs in ECP are

shown in expression (39). An output signal of ECP is shown in expression (40). A teacher signal to ECP is shown in expression (41).

( ) ( ) ( ) ( )( )in1 11 12 1, , ,A A A nx x xτ τ τ τ=x L (35)

( ) ( ) ( ) ( ) ( )( )inin 1 2, , , ,i i ij Ai Ai AinAe x x x xτ τ τ τ τ=x L

( )in2,3, , ; 1,2, ,i n j n= =L L (36)

( ) 0ijx τ =

( )in0; 2,3, , ; 1, 2, ,i n j nτ ≤ = =L L (37)

( ) ( ) ( )1

1

1

i

ij j kk

x x zτ τ τ−

−

=

= − ∑

in

02,3, ,1, 2, ,

i nj n

τ > = =

L

L

(38)

( ) ( )( )Ai i iz fτ τ= x ( )1, 2, ,i n= L (39)

( ) ( )1ˆ jz xτ τ += ( )in1, 2, ,j n= L (40)

( ) ( )1

jy xτ τ += ( )in1, 2, ,j n= L (41) where xij is the input error signal at the ith step to input signal xj to ECP, Aeini is an amplification factor of the input error signal at the ith step, Ai

is an amplification factor of the teacher signal at the ith step, fi is the nin variables function when i is 1 or the nin + 1 variables function when i is 2 or more to show I/O relation of NN at the ith step, x̂ is the prediction.

5. Computer simulations

5.1 2nd-order Volterra Neuron Network Figure 7 shows 2nd-order Volterra neuron (2VN) in discrete-time. I/O characteristics of 2VN are shown in expressions from (42) to (44).

( ) ( ) ( )

1

n

i ii

u w xτ τ τ

=

= ∑ (42)

( ) ( ) ( )

10

( ) ( ) ( ) ( )2

0

( )

( , )

Qp

p

Q Qp q

p q p

s p u

p q u u h

τ τ τ

τ τ τ τ

σ

σ

−

=

− −

= =

=

+ −

∑

∑∑

(43)

( ) ( )( ) 1( ) tan ( )z f s A sτ ττ −= = (44)

where u is the input weighted sum, xi is the ith input signal, wi is the ith connection weight, s is the input sum, D is the delay, Q is the prediction order, σ1 is the prediction coefficient of the 1st-order term corresponding to the signal obtained from between from an input of the 1st delay to an output of the Qth delay, σ2 is the prediction coefficient of the 2nd-order term corresponding to the product of all combinations of two sig-


18

yz

xECNNS

Rounding

Figure 4. Error convergence-type recurrent neuron network system

yz

x

Roundings

ECPNNS

Figure 5. Error convergence parallel-type recurrent neuron network system

( )3τε

NN1

NN2

-

-+

+

-+

- +

-

-

-( )1

1

ni

iz τ−

=∑

A2

1/A2

1/A3

A3

1/An

An

1/A1

-+

A1

+

+

+

NNn

( )τx

( )1τε

( )nτε

( )1Az τ

( )1Ay τ

( )2Az τ

( )3Az τ

( )Anz τ

( )1zτ

( )2z τ

( )3z τ

( )2Ay τ

( )3Ay τ

( )nz τ

( )Any τ

( )y τ

( )z τ

+

+

++

( )2τε

+-

+ -

+-

Ainn

( )1τx

( )2τx

( )nτx

D( )11zτ −

D( )12z τ −

D( )13z τ −

( )jx τ

Ain2

Ain3

( )3τx

NN3

( )2 jx τ

( )3 jx τ

( )njx τ

Ain1

Aein2

Aein3

Aeinn

( )2y τ

( )3y τ

( )ny τ

( )1y τ

Figure 6. Predictor using error convergence-type neuron network system

nals included in combinations of the same signal obtained from between from an input of the 1st delay to an output of the Qth delay, h is the threshold, z is the output signal, f is the output function, A is the output coefficient. wi , h , σ1 a n d σ2 are changed by training.

Figure 8 shows a three-layer 2VNN of one input one output which is constructed of 2VNs. This 2VNN is used for ECP.

5.2 Method


19

+

+

+

M

M

M

M

M

M

++−

+

( )h τ

( )s τ( )z τ

f（･）

( )1 (0) τσ

( )1 (1) τσ

( )1(2) τσ

( )1( )Q τσ

( )2 (0, 0) τσ

( )2 (0,1) τσ

( )2 (0, 2) τσ

( )2(0, )Q τσ

( )2 (1,1) τσ

( )2 (1, 2) τσ

( )2 (1, )Q τσ

( )2 (2, 2) τσ

( )2 (2, )Q τσ

( )2 ( , )Q Q τσ

DDD( )Qu τ−( )1u τ − ( )2u τ −( )u τ

( )1wτ

( )2wτ

( )nwτ

( )1x τ

( )2x τ

( )nx τ

Figure 7. 2nd-order Volterra neuron

Input Output( )iτx ( )

Aiz τ

2VN

2VN

Figure 8. 2nd-order Volterra neuron network

In computer simulations, EC2VP constructed of 2VNNs of two steps shown in Figure 9 is trained using combinations of an input signal x(τ) and a teacher signal y(τ) = x(τ+1) in the time series pattern of one dimension in space direction. The teacher signal is shown in Figure 10. This is a normal sinus rhythm ECG signal of MIT-BIH No.16786. This ECG signal is that the sampling frequency is 128 Hz, the significant figure is four-digit, the quantization step size is 0.005 mV, and the number of data for the input signal without an initial input and the teacher signal is 640, respectively. Here, training for 2VNN at each step in EC2VP is completed sequentially from the first step.

At the beginning, 2VNN at the first step (2VNN1) is trained using combinations of an input signal x(τ) and a teacher signal y1

(τ)=x(τ+1) in the time series pattern of one dimension in space direction. Here, computer simulations for the training are executed as A1=a11=1/3.275 according to the following procedure from 1) to 4).

1) A pair of the input signal or signals and the teacher signal is given once after 1,580 initial data are inputted into the 2VNN at the training. This process is defined as one training cycle.

2) Table 1 shows conditions for computer simulations to valuate prediction accuracy for the 2VNN. Initial values of prediction coefficients of the 2VNN are decided by exponential smoothing, and the other initial values are

Rounding2VNN1

2VNN2

-

-+

A2

1/A2

1/A1

-+

A1

( )1

τε

( )1Az τ

( )1Ay τ

( )2Az τ

( )1zτ

( )2z τ

( )2Ay τ

( )y τ

( )z τ++

( )2τε

+-

( )11Ax τ

( )2x τ

D( )11z τ −

( )x τ

a21

( )21x τ

a11( )2y τ

( )1y τ

( )2Ax τ

( )21Ax τ

Aein2

Figure 9. Error convergence-type 2nd-order Volterra Predictor

-1.0-0.50.00.51.01.52.02.53.0

Time [s]

Vol

tage

[mV

]

543210

0

Figure 10. Teacher signal used for training for error convergence-type 2nd-order Volterra predictor

decided by pseudo-random numbers at the training process a time. Gradient descent method is used for learning rule for the 2VNN. The number of middle layer elements and filter length of the 2VNN has been decided by studying experiences of ECG prediction using 2VNN.

3) The trainings for searches are executed as a parameter to set a condition for the computer simulations which is the learning reinforcement coefficient.

4) Averages of root mean square errors (RMSEs) obtained from the searches of three times are compared.

2VNN achieved the minimum average of RMSEs by the searches is NN at the first step in EC2VP. A part of training signals to 2VNN at the second step (2VNN2) in EC2VP is an error signal obtained from a difference between an output signal of 2VNN1 and a teacher signal to EC2VP.

Next, 2VNN2 is trained using combinations of input signals x21

(τ) and x(τ) in the time series pattern of two dimensions in space direction and a teacher signal y2

(τ)=x21(τ+1) in the time series pattern of one dimension in

space direction. A signal to which gain tuning which adjusts the maximum absolute value of error signal to 1 is performed are used for xA2

(τ) and yA2(τ). Here, computer

simulations for the training are executed as Aein2=A2 and a21=1/3.275 according to the above-mentioned procedure from 1) to 4).


20

Table 1: Conditions for computer simulations to train 2nd-order Volterra neuron network at each step

10-5～110 times

30,0003

-0.3～0.3-0.3～0.3

Prediction coefficents 0.7×0.3p

Training cycles

Learningreinforcement

coefficients

Range

Processing times

Connection weightsThresholds

Gradient-basedmethod

Initialconditions

Filter length

Interval

σ2

σ1

Momentum 0Output coefficents 1

0.7×0.3p×0.7×0.3q

Learning roule Learning roule for Volterra neuron network

Number of middle layer elements

Steps The 1st step The 2nd step

469

1064

Time [s]

Vol

tage

[mV

] ]

543210

00.2

1.0

-0.2

×10-1

0.40.60.8

-0.4-0.6-0.8-1.0-1.2

Figure 11. Teacher signal for 2nd-order Volterra neuron network at the second step

Training cycles

2VNN1

10-6

10-9

10-3

1

0 10,000 20,000 30,000

Eval

uatio

n fun

ctio

n val

ue

2VNN2

Figure 13. Relation between training cycles and average evaluation function values at the minimum average of root mean square errors concerning 2nd-order Volterra neuron network at each step

Learning reinforcement coefficent

RM

SE

Minimum (1.17×10-5)

10-1 110-210-310-410-5

10-1

101

10-2

10-3

10-4

10-5

N2VNN1

Minimum (2.60×10-2)

N2VNN2

10-6

Figure 12. Averages of root mean square errors and their standard deviations to learning reinforcement coefficient in gradient descent method term

-1.0-0.50.00.51.01.52.02.53.0

Time [s]

Vol

tage

[mV

]

系列1 系列2

543210

2VNN1 2VNN2

0

Figure 14. Output signal of 2nd-order Volterra neuron network at each step


21

Then, 2VNN achieved the minimum average of RMSEs by the searches is NN at the second step in EC2VP. An output of the 2VNN2 is restored at a level of the teacher signal to EC2V. The restored output signal is added to the output of 2VNN1 at the same time as it. Finally, the fourth decimal place of the output after the addition is rounded off. The obtained output is an output of EC2VP as a result. Its prediction accuracy is evaluated.

5.3 Results Results of computer simulations are shown in figures from 11 to 14. A result of making training signal for 2VNN2 is shown in Figure 11. Averages of RMSEs and their standard deviations obtained by the computer simulations for 2VNN at each step are shown in Figure 12. Fig. 13 shows a relation between training cycles and average of evaluation function values of 2VNN at each step which is recorded before beginning to train and at the first time and every 100 times of the training cycle when the minimum average of RMSEs is obtained by searches of three times. This figure shows that prediction errors when training for 2VNN1 is saturated can be decreased more by using 2VNN2. Moreover, gradient of average evaluation function value of 2VNN2 at 30,000 cycles is surmisable like being able to train more. Figure 14 shows output signal of 2VNN at each step. An output signal of EC2VP which is obtained as a result of adding output signals of 2VNN1 and 2VNN2, and rounding it. This signal is error free at all, and is equal to the teacher signal.

Thus, complete learning accomplishment capability of EC2VP could be demonstrated. An excellent learning capability and validity of EC2VP to normal sinus rhythm ECG can be confirmed.

6. Discussion

6.1 Accuracy Up for Output and Speed Up for Training

It is thought that a highly accurate output is obtained by ECNNS at a few training cycles because its error is converged by plural NNs. Here, error signal of NN at previous step which NN at final step uses to train is a very small signal if steps of NN are set infinitely. That is, sum of output signals of all NNs becomes a very highly accurate output signal because output error of NN at final step becomes equal in nil compared with a teacher signal to a whole of ECNNS. The training cycle as the whole of ECNNS can be theoretically one time when thinking only one datum is trained under this condition. For example, when the probabilistic descent method is used for learning rule for NNs in ECNNS, it is expected that a steady output signal is obtained at any training cycles because the highly accurate training can be executed every one datum as shown in the above-mentioned. Moreover, a component ratio of output signal of NN at each step to the output signal of ECNNS to grow large at more former step as the training cycle increases is thought. It is thought that learning capability of ECNNS is remarkable as steps of NN increases and the output accuracy and the training speed elevate.

6.2 Improvement on Learning Capability Capability of discrete-time neuron network (DTNN) is improved if sampling frequency of signal processed with DTNN is upped. However, there is a limit in such means to improve the capability of DTNN because the limit of the sampling frequency is caused by restriction to operating frequency fordigital circuit, memory capacity and data processing time of it.

Moreover, the lowest sampling frequency to process efficiently data is decided by the sampling theorem. However, training for DTNN must be executed in a state which the data are increased by upping the sampling frequency if excellent training for DTNN is not obtained by the data. As a result, data processing for DTNN becomes no efficiency. Here, it is thought that learning capability of DTNN can be improved without upping the sampling frequency by ECNNS.

6.3 Instability Instability of ECP is how to process components of random signals which are included in inputs signal and teacher signals of ECP. There is a means which the random signals are processed as noises. This means eliminates these noises by filtering an input signal and a teacher signal to ECP as the preprocessing. There are Fourier transform, Bayesian model and trend model etc. [30] as the means which can be used. However, there is also necessary information like signals of nature, vital and economy etc. in the random signals. This means must be performed with attention so as not to miss the information.

6.4 Validity It thinks about training for 2VNN at the second step in EC2VP. ECP is an improvement on its learning capability by using NN which makes the causal relation between input signal and teacher signal available as an internal predictor in ECP to enable trainings for NNs from the second step where the trainings becomes difficult. Furthermore, two input signals are used for strengthening unique concerning the output signal to the input signal. Actual training also is shown excellent results, and the effect concerning the above-mentioned to make easily to train is seen. It is thought that this is because a relation between the input signals and the teacher signal is a correlation.

These computer simulations are insufficient as a demonstration for a highly accurate predictor because generalization capability to unlearning signals is not confirmed, though it demonstrated the complete learning accomplishment capability of EC2VP. Therefore, demonstrations for the generalization capability using ECGs of arrhythmia etc. are necessary.

7. Conclusions In this study, it was shown to obtain a highly accurate output by improving the learning capability using single output NNs which an error of an output signal to a teacher signal does not converge at zero though it becomes smaller than one before training under a result of executing until the training converges to all input signals with correlation between the teacher signal and which the inputs are common in style connected with multi-steps. This is a method which the error to the teacher signal of a whole of


22

NN system can be converged at a small value by amplifying an error signal obtained from NN at each step and training it as a teacher signal of NN at the next step one after another. Moreover, an error to a teacher signal of NN at the first step converges at zero by infinitely setting steps of NN. This method is called ECMNNS. It explained ECNNS applied this method and means to use by ECPNNS which ECNNS is applied to the parallel-type NN for NN of plural outputs. Furthermore, it also explained ECRNNS and ECPRNNS which can be expected to use as the simulator for a nonlinear plant as applications using ECNNS and ECPNNS.

Moreover, ECP improved learning difficulty when there is little causality between input signals and a teacher signal to NN at each step in ECNNS by structuring NN at each step in ECNNS as a predictor and strengthening causality between the input signals and the teacher signal was designed. Using 2VNN for NN at each step in ECP was proposed.

Finally, computer simulations to train EC2VP constructed of 2VNNs of two steps were executed using a normal sinus rhythm ECG signal, and prediction accuracy of EC2VP was evaluated. As a result, learning capability obtained an output without error, that is, ECP having complete learning accomplishment capability was demonstrated. To be a validity which can use EC2VP as a highly accurate predictor by this demonstration was confirmed. It can be said that ECNNS will have an enough capability to be the leading system of means to construct NN in the future because of excelling theoretically concerning accuracy up for output, speed up for training and improvement on learning capability of DTNN. On the other hand, an enough demonstration as a highly accurate predictor has been not able to perform because generalization capability to unlearning signals is not confirmed. The future work is demonstrating the generalization capability of EC2VP using ECG of arrhythmia etc.

Acknowledgments We wish to express our gratitude to members in our laboratory who cooperate always in the academic activity.

References

[1] C. L. Giles and T. Maxwell, “Learning, Invariance and Generalization in High Order Neural Networks,” Applied Optics, Vol.26, No.23, pp. 4972-4978, 1987.

[2] K. J. Lang and G. E. Hinton, “A Time-Delay Neural Network Architecture for Speech Recognition,” Carnegie Mellon University Computer Science Technical Report, CMU-CS-88-152, pp. 1-37, 1988.

[3] T. Possio and F. Girosi, “Networks for Approximation and Learning,” Proc. of the IEEE, Vol.78, No.9, pp. 1481-1497, 1990.

[4] S. Iwamoto, T. Yosh ida and H. Yokoi, “Basic Investigation Associated with Neural Control of Biped Walking Robot,” Technical Report of IEICE, MBE93

-106, pp. 23-30, 1994. [5] Y. Fujisue, E. Inohira and H. Yokoi, “Robotic Control

by Volterra Network,” Technical Report of IEICE, NC2003-78, pp. 39-43, 2003.

[6] S. Uota and H. Yokoi, “A Realization of Motion Diversity of the Robotic Hand by the Hierarchical Motion Schema,” Technical Report of IEICE, NC2003-75, pp. 25-28, 2003.

[7] J. Miyoshi and H. Yokoi, “An Improvement of a Neural Network for Learning a Slip Angle of a Four-Wheel Steering Car,” Technical Report of IEICE, NC2004

-107, pp. 87-90, 2004. [8] S. Shigemura, T. Nishimura and H. Yokoi, “A

Method of Removing Blink Artifacts from EEG Signals Using Neural Networks with Volterra Filters,” Technical Report of IEICE, MBE2004-87, pp. 57-60, 2005.

[9] S. Suematsu and H. Yokoi, “A Motion Generating System for Multi-Fingered Myoelectric Hand,” International Congress Series 1291, pp. 257-260, 2006.

[10] S. Kobayakawa, T. Fujii and H. Yokoi, “Evaluation of Nonlinear Prediction Capabilities of Neuron Networks for Electrocardiogram,” Proc. of the 20th Annual Meeting of Biomedical Fuzzy Systems Association, pp. 9-12, 2007.

[11] S. Kobayakawa, T. Fujii and H. Yokoi, “Evaluation of Prediction Capabilities of Neuron Networks Used for Electrocardiogram,” Proc. of the 5th International Symp. on Management Engineering, Kitakyushu, Japan, pp. 156-161, 2008.

[12] S. Kobayakawa, T. Fujii and H. Yokoi, “Nonlinear Prediction for ECG by 2nd-order Volterra Neuron Network,” Journal of Biomedical Fuzzy Systems Association, Vol.11, No.2, pp. 101-111, 2009.

[13] A. G. Ivakhnenko, “The Group Method of Data Handling-A Rival of the Method of Stochastic Approximation,” Soviet Automatic Control, Vol.13 c/c of Avtomatika, 1, 3, pp. 43-55. 1968.

[14] A. D. Back and A. C. Tsoi, “FIR and IIR Synapses, A New Neural Network Architecture for Time Series Modeling,” Neural Computations, Vol.3, pp. 375-385, 1991.

[15] M. Hoshino, T. Kitamura, T. Masuda, M. Suzuki and J. Chao, “On Multilayer RBF Networks and a Novel Pyramid Network,” Proc. of the Society Conf. of IEICE, Nagoya, Japan, p. 28, 2000.

[16] N. Kinoshita and K. Nakamura, “Two-D Spreading Associative Neural Network Recognizes the Shape and Position of An Object Presented in the Two-D Space,” Technical Report of IEICE, NC97-166, Vol.97, No.623-624, pp. 209-216, 1998.

[17] S. Kobayakawa and H. Yokoi, “The Volterra Filter Built-in Neural Network for the Aircraft Pitch Attitude Control,” The Lecture Proc. of the 2005 Fiscal Year Electricity Relation Institute Kyushu Branch Association Convention, Fukuoka, Japan, p. 429, 2005.

[18] S. Kobayakawa and H. Yokoi, “Application to Prediction Problem of Parallelized Neuron Networks in the Aircraft,” Technical Report of IEICE, SANE2006-119 -133, Vol.106, No.471, pp. 43- 45, 2007.

[19] S. Kobayakawa and H. Yokoi, “Evaluation for Prediction Capability of Parallelized Neuron Networks,” Proc. of the 8th SOFT Kyushu Chapter Annual Conf., Kitakyushu, Japan, pp. 3-6, 2006.

[20] S. Kobayakawa and H. Yokoi, “Evaluation of the Learning Capability of a Parallel-type Neuron Network,” Proc. of the First International Symp. on


23

Information and Computer Elements 2007, Kitakyushu, Japan, pp. 43-47, 2007.

[21] S. Kobayakawa and H. Yokoi, “Experimental Study for Dominance to Accuracy of Prediction Output of Parallel-type Neuron Network,” Technical Report of IEICE, NC2008-1-10, Vol.108, No.54, pp. 29-34, 2008.

[22] S. Kobayakawa and H. Yokoi, “Evaluation for Prediction Accuracies of Parallel-type Neuron Network,” International MultiConf. of Engineers and Computer Scientists 2009 Proc., Hong Kong, China, Vol.I, pp. 156-161, 2009.

[23] H. Yokoi and T. Kimoto, “Multilayered Neural Networks with Intermediate Elements,” Journal of Biomedical Fuzzy Systems Association, Vol.1, No.1, pp. 87-97, 1999.

[24] H. Sori and T. Yasuno, “Several-Hours-Ahead Wind Speed Prediction System Using Hierarchical Neural Network,” Journal of Signal Processing, Vol.12, No.6, pp. 507-514, 2008.

[25] S. Kobayakawa and H. Yokoi, “Proposal of Error Convergence-type Neuron Network System,” Presented Proc. to 2008 International Symp. on Intelligent Informatics, Kumamoto, Japan, pp. 1-10, 2008.

[26] D.E. Rumelhart, G.E. Hinton and R.J. Williams, “Learning Representations by Back-propagating Errors,” Nature, Vol.323, No.6088, pp.533-536, 1986.

[27] A. Chatterjee, A. Nait-Ali and P. Siarry, “An Input-delay Neural-Network-Based Approach for Piecewise ECG Signal Compression,” IEEE Transactions on Biomedical Engineering, Vol.52, No.5, pp. 945-947, 2005.

[28] S. Kobayakawa and H. Yokoi, “Proposal of Predictive Coding Using Error Convergence-type Neuron Network System,” The Proc. of the ISCA 22nd International Conf. on Computers and Their Applications in Industry and Engineering, San Francisco, USA, pp. 169-174, 2009.

[29] S. Kobayakawa and H. Yokoi, “Evaluation of Learning Capabilities of BP Networks to Number of Input Signals,” Technical Report of IEICE, SANE2007-102-124, Vol.107, No.442, pp. 83-86, 2008.

[30] M. Onodera, Y. Isu, U. Nagashima, H. Yoshida, H. Hosoya and Y. Nagakawa, “Noise Filtering Using FFT, Bayesian Model and Trend Model for Time Series Data,” The Journal of Chemical Software, Vol.5, No.3, pp. 113-127, 1999.

Authors Profile

Shunsuke KOBAYAKAWA received the B.Eng. degree in 1986, accomplished credits for the master's course in 1989 in electrical engineering from Okayama University, Okayama, Japan, completed auditor in faculty of engineering in 1995 and received the M.Sc. degree in biological functions and engineering in 2003 from Kyusyu Institute of Technology (KIT), Kitakyushu, Japan, respectively. He has

been working as a part-time lecturer at Research Course of Telecommunication System, Subaru Professional College in 2006, a research assistant since 2006 and a teaching assistant in 2008 at Graduate School, KIT. He also obtained Associate Professional Engineer in electrical and electronics engineering, First-Class Technical Radio Operator for On-The-Ground Services, Class I Information Technology Engineer and Aerospace Products Inspector etc. as national qualifications in Japan. He is a student of doctoral program, Department of Biological Functions and Engineering, Graduate School of Life Science and Systems Engineering, KIT and a director of KOBAYAKAWA Design Office at present. His present research interests include control for aerospace vehicles using neuron networks. He is a member of Biomedical Fuzzy Systems Association, The Japan Society for Aeronautical and Space Sciences, Information Processing Society of Japan, The Institute of Electronics, Information and Communication Engineers, and The Institute of Electrical and Electronics Engineers.

Hirokazu YOKOI received the B.Eng. degree in 1972, the M.Eng. degree in 1974 in electrical engineering from Nagoya University, Nagoya, Japan, the D.M.Sc. degree in medicine in 1985 and the D.Eng. degree in electronics in 1989 from The University of Tokyo, Tokyo, Japan, respectively. He works as a professor of

Department of Biological Functions and Engineering, Graduate School of Life Science and Systems Engineering, Kyusyu Institute of Technology, Kitakyushu, Japan at present. His present research interests include development of ultra large-scale neurochips and their applications to biped walking robots, intelligent assistive devices as well as automatic driving, and modeling of human cognitive processes with application to human-centered information equipments. He is a member of Biomedical Fuzzy Systems Association, Japan Society for Fuzzy Theory and Intelligent Informatics, The Institute of Electronics, Information and Communication Engineers, and Human Interface Society.


24

Segmentation Method Based On Circular Edge Magnitude for Classification of Textures

1 Dr. A. Nagaraja Rao , 2K. Lavanya, 3G. Aruna Kumari, 4M.N. Hima Bindu

1 Professor, Dept. of IT, Lakireddy BaliReddy College of Engg.

Mylavaram, Krishna Dt., A.P., India. [email protected]

2 Asst. Prof., Dept. of IT, Lakireddy BaliReddy College of Engg.


3 Associate Prof., Dept. of CSE, Vidya Jyothi Institute of Tech.,

Moinabad, RR Dist, A.P., India. [email protected]

4 Asst. Prof. Dept. of IT, Lakireddy BaliReddy College of Engg.


Abstract: Texture segmentation is one of the most important techniques for image analysis, understanding and interpretation. The task of texture segmentation is to partition the image into a number of regions such that each region has the same textural properties. The present paper proposes a new segmentation method which is an alternative to non-maximal suppression based on edge magnitude. The relative edge magnitudes are calculated by considering the spatial information of 5X5 mask with 3X3 circular neighborhood masks. The central pixel under consideration in 5X5 mask is whether an edge pixel or not, can be identified by manipulating all the edge magnitudes from circular neighboring masks. The proposed method is applied on various Brodatz textures and the experimental results shows effective segmentation which is useful in classification of textures.

Keywords: Texture, Segmentation, Edge, Magnitude, Non-Maximal Suppression, Neighborhood, Circular Mask. 1. Introduction Texture is defined as a pattern that is repeated and is represented on the surface of an object. Segmentation is a fundamental low-level operation on images and to separate textures into a single texture type, first we need to preserve spatial information for each texture. A homogeneous region refers to a group of connected pixels in the image that share a common feature. This feature could be brightness, color, texture, motion etc. For instance, the manual grey level thresholding which does not provide the spatial information for each texture [12] could generate inappropriate segmentation result. Depending on the number of images is known in advance, the techniques for image segmentation can be classified as supervised [1-3] or unsupervised [4-6]. Since in most of real applications the number of images is generally unknown, the unsupervised approach is

considered more useful. There are three main classes of image texture segmentation that belong to the above two techniques. They are cluster-based, edge-based and region-based methods [7]. The edge-based segmentation exploits spatial information by detecting the edges in an image, which correspond to discontinuities in the homogeneity criterion for segments. Edges in a given depth map are defined by the points where changes in the local surface properties exceed a given threshold. The local surface properties mostly used are surface normals, gradients, principal curvatures, or higher order derivatives. Edge detection techniques used on texture image could result in noisy and discontinuous edges and therefore segmentation process becomes more complicated [13]. As edge detection methods look for abrupt changes, they are very sensitive to the noise in the range data. Moreover, as only the measurements near the edges are used to make major decisions, the available information is not optimally utilized. In many situations the edges do not form closed boundary curves and it can be difficult to make correct grouping decisions there by resulting in over or under segmentation. Some of the typical variations on the edge-based segmentation techniques are reported by many researchers [8, 10, 11]. The organization of this paper as follows: Section 2 gives the related work and the proposed method and algorithm described in section 3. Experimental results of texture segmentation using the circular edge magnitude are shown in the section 4 and section 5 gives the conclusions.


25

2. Related Work In the edge image, no zero value edge pixels are present, but small edge values correspond to non-significant grey level changes resulting from quantization noise, small lighting irregularities. Simple thresholding of an edge image can be applied to remove these small values based on an image of edge magnitudes [14]. A problem with simple detector is the thickening that is evident where there should only be a simple boundary. This can be partially rectified if edges carry directional information by performing some form of non-maximal suppression to suppress multiple responses in the neighborhood of single boundaries [15]. In the proposed method, the maximum edge magnitude in 8-connectivity is calculated leading to the edge direction. The pixel under consideration in 3X3 window is displayed only if its grey level value is greater than the maximum of the two adjacent pixels of the edge direction. 3. Proposed Segmentation Method Edge based segmentation is one of the common approaches to segment the image and it remains very important over period of time. All edge based segmentation methods depend on edges identified in an image by any edge detection techniques. Edge based segmentation techniques differ in strategies and the necessary priori information is incorporated into those methods. Since the edge detection is very difficult in texture images when compared to image segmentation, the present study proposes the new technique which is based on relative edge magnitude in 5X5 neighborhood using circular rotating masks and the procedure and algorithm is described as follows. A practical method that has chosen in the present approach is to determine whether a central pixel is an edge pixel or not depending on its neighboring rotating masks inside the running window, in horizontal, vertical, right diagonal and left diagonal directions. Note that, since the proposed method checks the relative Difference of Edge Magnitude (DEM) values of rotating masks using 5X5 window as shown in Fig.1 instead of 3X3 window as applied in other methods. This will eliminate some false edge pixels on seemingly smooth regions. If any one of the maximum or average or minimum of the adjacent masks of maximum DEM value of rotating masks is less than the central pixel grey level value, the pixel is identified as being not located for segmentation. The stated method is explained in the following algorithm.

(a) (b) (c) (d) Figure 1. (a),(b),(c) and (d) represents the rotating 3x3 masks that contain central pixel of 5x5 window in Horizontal(H), Vertical(V),Right Diagonal(RD) and Left Diagonal (LD) Directions. Algorithm 1. BEGIN. Step 1: Read the 5x5 window from original image. Step 2: Select circular 3X3 masks that contain the central pixel of a 5X5 window of the image as shown in Fig.1. Step 3: The sum of the grey levels of the circular masks are calculated. Step 4: The difference between the two Right Diagonal, Left Diagonal, and Vertical and Horizontal direction’s circular rotating masks are computed Step 5: To find edge, Maximum of the above differences is obtained and Central Pixel of 5X5 window is replaced by 1 iff any one of the adjacent differentiating masks maximum/mean/ minimum grey level is greater than or equal to Central Pixel of the current 5x5 mask, otherwise the Central Pixel is replaced by 0. Step 6: Repeat step 1 to step 5 by 5X5 running window throughout the image. Step 7: Display the resultant segmented image. END. 4. Experimental Results The Brodat’z Texture Images, Plastic Bubbles, Straw, Pig Skin and Raffia, are taken to apply the proposed algorithm and these Textures are shown in Fig.2. The segmented texture images are shown, when all suppression parameters like maximum, mean and minimum, in Fig.4, 5 and 6 respectively. In Fig. 3, the segmented texture images of non-maximal suppression on 3X3 neighborhood are shown. For any texture classification, feature extraction is most important step. The proposed algorithm segments the textures by extracting the linear, horizontal and vertical edge features.


26

The edge boundaries are extracted by non-maximal suppression in 5X5 neighborhood using the proposed algorithm and the segmented results are shown in Fig.4. Similarly the parameters like non-average and non-minimum are suppressed in 5X5 neighborhood for

segmentation and the results are shown in Fig.5 and Fig.6 respectively. Only the strong edge features are extracted when non-average method is applied and over segmentation, i.e. decomposing into more inner regions, is achieved when non-minimal suppression is applied.

(a) (b)

(c) (d)

Figure 2. Original Brodatz Textures (a) Plastic Bubbles (b) Straw (c) Pig Skin (d) Raffia

(a) (b)

(c) (d)

Figure 3. Segmentation results when non-maximum is suppressed in 8-connectivity. (a) Plastic Bubbles (b) Straw (c) Pig Skin (d) Raffia.


27

(a) (b)

( c) (d)

Figure 4. Segmented Images when Non-Max is considered using proposed method. (a) Plastic Bubbles (b) Straw (c) Pig

Skin (d) Raffia.

(a) (b)

(c) (d) Figure 5. Segmentation results when non-average is suppressed in 8-connectivity. (a) Plastic Bubbles (b) Straw (c) Pig

Skin (d) Raffia


28

(a) (b)

(c) (d)

Figure 6. Segmented Images when non-minimum is suppressed (a) Plastic Bubbles (b) Straw (c) Pig Skin (d) Raffia.

5. Conclusions Detecting edges and then eliminating irrelevant ones and connecting (grouping) the others are the key to a successful edge-based segmentation. The segmentation scheme exploited spatial information by detecting edges in an image that correspond to discontinuities in the homogeneity criterion for segment. As the mask size increases in texture segmentation by non-maximum and average suppression, noise reduces and the borders are more exact. The proposed segmentation algorithm proposed, is simple to implement and shows better results than the existing non-maximal suppression algorithm.

From the Fig.4 and 5, the segmented results clearly show the local boundaries in all textures. By the technique of rotating mask, local boundaries in all directions are made clearly visible. The segmentation with continuous boundaries are more visible in Fig.4 when compared to Fig.3, Fig.5 and Fig.6. The segmentation results of textures

indicate the domination of linear, circular and horizontal patterns. By the above topological patterns, one can classify textures easily. With the proposed segmentation method the classification of the textures based on topological structures becomes more effective.

6. Acknowledgements We would like to express our gratitude to the management of LBR College of engineering for providing facilities in the college. We are thankful to the Director, Dr. L.S.S. Reddy, who has given motivation and constant encouragement to complete this paper. Also we would like to thank Dr. R. Chandrasekaran and Prof. C. Nagaraju for their invaluable suggestions to improve the quality of this paper. References [1] D.Dunn, W. E. Higgins and J. Wakeley, Texture

segmentation using 2-D Gabor elementary functions,


29

IEEE Trans. Pattern Analysis Mach. Intell. 16(2), 130-149 , 1994.

[2] Y. M. Zhu and R. Goutte, Analysis and comparision of space/spatial – frequency and multi scale methods for texture segmentation, Opt. Eng. 34(1), 269-282(1995).

[3] M. Unser, Texture classification and segmentation using wavelet frames, IEEE Trans. Image Process. 4(11), 1549-1560, 1995.

[4] J. Mao and A. K. Jain, Texture classification and segmentation using multiresolution simultaneous autogressive models, Pattern Recognition 25(2), 173-188,1992.

[5] Y. Hu and T. J. Dennis, Textured image segmentation by context enhance clustering, IEEE Proc. Vis. Image Signal Process. 141(6), 413-421, 1994.

[6] J. L. Chen and A. Kundu, Unsupervised texture segmentation using multichannel decomposition and hidden Markov models, IEEE Trans. Image Process. 4(5), 603-619,1995.

[7] R.C. Gonzalez, and R.E.Wood,.. Digital Image Processing., Wesley Publishing Company, pp. 458-461,2002.

[8] O.R.P.Bellon, A.I. Direne, and L. Silva, Edge detection to guide range image segmentation by clustering techniques. In: International Conference on Image Processing (ICIP ’99), Kobe, Japan, pp. 725–729, 1999.

[9] X. Jiang, A. Hoover, G. Jean-Baptiste, D. Goldgof, K. Boywer, and H. Bunke, A methodology for evaluating edge detection techniques for range images. In: Proc. Asian Conf. Computer Vision, pp. 415–419, 1995.

[10] A.D. Sappa, and M. Devy, Fast range image segmentation by an edge detection strategy. In: Third International Conference on 3-D Digital Imaging and Modeling, pp. 292–299, 2001.

[11] M.A. Wani and H.R. Arabnia, Parallel edge-region-based segmentation algorithm targeted at reconfigurable multiring network. Journal of Supercomputing 25(1), pp. 43–62, 2003.

[12] C.E. Honeycutt and R. Plotnick , Image analysis techniques and gray-level co-occurrence matrices (GLCM) for calculating bioturbation indices and characterizing biogenic sedimentary structures, Computers & Geosciences 34, pp. 1461-1472, 2008.

[13] S. Zheng, J. Liu and J.W. Tian, A new efficient SVM-based edge detection method, Pattern Recognition Letters 25, pp. 1143-1154, 2004.

[14] A. Kundu, Mitra, A new algorithm for image edge extraction using a statistical classifier approach. IEEE Transactions on Pattern Analysis and machine Intelligence, 9(4): 569-577, 1987.

[15] Milan Sonka, Vaclav Hlavac and Roger Boyle, Image Processing,Analysis and machine vision, Second Edition, Vikas publishing House,pp. 135-137,2001.

Authors Profile

Agastyaraju NagarajaRao received the M.Sc. (Computer Science) Degree from S.V. University in 1999. He received his Ph.D. degree in Computer Science from University of Mysore in 2009. He is having 10 years of teaching experience at graduate and

postgraduate level. Present he has been working with LBR college of Engineering, Mylavarm, as a professor in Department of IT. He has published more than 6 international journal papers, 14 national and international conferences. His research interests include Image Processing, Pattern Recognition, Data Mining and Texture Segmentation. He is a life member for CSI and ISCA.

K. Lavanya completed her B.Tech. (CSE) From JNT University. Now she is pursuing her M.Tech. (CSE) from JNT University, Hyderabad. Currently she is working as Assistant Professor in department of IT, LBR College of Engineering, Mylavaram. She has overall 5 years of teaching experience at

graduate level. Presently she is doing her project in image processing.

G.Aruna kumari received B.E. degree in Electronics and Communication Engineering from Andhra University in 1995, and also in 1999 she obtained M.Tech degree in Computer Science and Technology from the same university. She has 10 years of teaching

experience at graduate and postgraduate level. She is pursuing Ph. D. (CS) from Jawaharlal Nehru Technological University, Hyderabad, in the field of image processing. Her research interests include of image processing and pattern recognition.

M.N.Hima Bindu completed her B.Tech (CSIT) from JNT University,Hyderabad. She is Pursuing M.Tech. (SE) From JNT University, Kakinada. She has been working as Assistant Professor in department of IT, LBR College of Engineering, Mylavaram. She has 4 years of teaching experience at graduate level.


30

Design of a 3.3GHz Band Application LC VCO for Low Phase noise and Low power Using Current

Mirror

Namrata Prasad1, R. S. Gamad2 1Electronics & Instrumentation Engineering Department, SGSITS, 23, Park Road, Indore, M.P., India – 452003

[email protected]

2 Electronics & Instrumentation Engineering Department, SGSITS, 23, Park Road, Indore, M.P., India – 452003

[email protected]

Abstract: In this paper a novel methodology is used to implement an LC tank Voltage Controlled Oscillator (VCO) with low power consumption, low phase noise and low tuning range. Here a relationship between, the phase noise, power consumption and the bias current is fully analyzed to improve the performance of VCO. The present design architecture of VCO has employed two current mirrors one at the top and other at the bottom end of cross coupled VCO, to balance the impedance and gives exact replica of current in both the arm of current mirror circuit. The phase noise measured is -148.59dBc/Hz at 3.3GHz carrier frequency with offset voltage of 1V. It will consume power of 5.68mW and the bandwidth is 4.31GHz, It also has the tuning range of 1.82%. From the simulation results we have seen that the power consumption, phase noise has reduced with current mirror. Finally we have compared the results with the earlier published work and got improvement in the present results as given in table 1.

Keywords: LC-VCO, current mirror, cadence, low power, low phase noise.

1. Introduction Voltage Controlled Oscillators (VCOs’) are used in wireless application. It is one of the most significant block for the Radio Frequency (RF) communication system because it define all the performance parameter of the system with the rapid development of RF communication application in the field of cellular telephony, codeless phone, wireless data networks, two way paging etc. the performance parameter Low phase noise, Low power consumption and wide tuning range are the basic requirement which are interrelated and for the better system performance a tradeoff has to be achieved among these crucial requirements [1].There are mainly two confurigation of cross coupled VCO. Here, complementary cross coupled VCO is used to achieve Low phase noise, in RF communication phase noise is reduced by the drawback of high power because it degrade the system integrity by reducing the signal integrity of the output of a transceiver. Therefore by proper analyzing the circuit with well controlled current flow a low phase noise and a Low power consumption is obtained [2]. The paper is described in following section 2. Briefly described the analysis and design of cross coupled VCO, section 3 describe the proposed VCO with the addition of current mirror at the top and bottom end of the VCO architecture. The result is presented in section 4. And finally the conclusion is presented in section 5. 2. Design and analysis of cross coupled VCO The core component of the VCO consists of cross coupled PMOS and NMOS that is (M1, M2) and (M0, M3) from VCO core to

generate a negative resistance to cancel the loss in the LC tank. It consists of a tail current mirror to provide bias current in a design. A typical resonance frequency of cross coupled VCO is given by [5]:

12oscF

LCπ=

(1) Where, L is the inductor in Henry of LC tank and C is the capacitance. The passive element i.e. the on-chip spiral inductor L and the two capacitor forms the frequency tuning network and Vcon is the controlled voltage. In case of a single current mirror the impedance is unbalanced resulting in different currents in the mirror arms which increase the power consumption and also has a negative effect on the phase noise performance of the circuit. Figure 1 shows schematic view of the earlier VCO design without using current mirror [5]. To overcome this problem we have used current mirror in the proposed VCO design.

Figure 1. Schematic view of the earlier VCO design

3. Proposed VCO Design


31

In order to reduce the power consumption, phase noise, tuning

range because they are the key parameter for the performance of a transreceiver system. In the proposed VCO design the top and bottom current controlled architecture is employed i.e. current mirror which balance the impedance in both the arm of circuit and hence the current become the exact replica of the bias current. In this design a current controlled mechanism is used to reduce even harmonics in the drain current which has a direct impact on the phase noise component which results reduction in phase noise it also achieves the negative resistance from the active devices by drawing minimum amount of current from the supply and reducing the power consumed by the circuit. Here, novels current controlled architecture is used to shift the wave form and control shape of the output waveform by adjusting the transistor sizes for the current mirror. The new proposed design schematic is presented in fig. 2

.

Figure 2. Schematic view of the present VCO design To reduce phase noise and power consumption, we have used

current mirror architecture in our proposed design in cross coupled form. Phase noise is most important parameter in VCO design therefore; phase noise performance is optimized by using lessons’ formula [4].

2

202

2( ) 10log [1 ( ) ](1 )2 2

c vco

m m av

f f KTRKfKTL FMf Q f P f m

= + + +

(2)

Where, L (FM) phase noise in dBc/Hz, Fm is the frequency offset from the carrier in Hz, f0 is central frequency in Hz, fc is flicker noise corner frequency in Hz, Q is the loaded quality factor of the tuned circuit, F is noise factor, K is Boltzmann's constant in J/K, T is temperature in K, Pav is average power at oscillator output, R is the equivalent noise resistance of the varactor and Kvco is oscillator voltage gain in Hz/V. From equation (2), Kvco dominates the phase noise performance in the modified Lesson's formula, thus phase noise performance can be improved by reduction Kvco. Maximum d. c. power dissipation= (Vsupply ) x (Ibias ) (3) Tuning range can be determined as follows [5]:

0 max 0 min

0

% 100W WTunningrange xW−

= (4)

Where, W0max is the maximum frequency of operation, W0min is the minimum frequency of operation and W0 is the frequency of operation. 4. Simulation result and discussion This work is carried under the environment of cadence software and schematic editor is used for design entry. In this design we have used specter RF simulator for Simulation, by using TSMC 0.18µm technology. The design is simulated with different architecture i.e. without current mirror, with tail current, and with current mirror. The applied voltage is 2V at the center frequency of 3.3GHz, with the Bandwidth of 4.31GHz. We have compared our simulation results with earlier work done and got improvement in this reported results and are shown in the table 1. Simulated output voltage responses of the present design are presented in fig. 3 and 4. Phase noise is given in fig. 5.

Figure 3. Simulation result of the output voltage


32

Figure 4. Simulation AC response of the output Voltage

Figure 5. Results of the Phase noise

Table 1: Comparison of present results with earlier Work done

Parameters Ref.

[6] With tail current mirror only [5]

This design

Without current mirror

With current mirror

Operating Voltage

2V

2V

2V

2V

Technology (CMOS)

0.35µm (TSMC)

0.18µm (UMC)

0.18µm (UMC)

0.18µm (UMC)

Power consumption

(mW)

18

15.76

8.69

5.68

Operating Frequency

6GHz

3.3GHz

3.3GHz

3.3GHz

Tuning Range

17%

3.207%

29.8%

1.82%

Phase Noise (dBc/Hz)

-94

-151

(mdB/Hz)

69.63

-148.59

Bandwidth

(GHz)

-

-

1.611

4.31

5. Conclusion

In this paper, we have presented a novel, low phase noise, low power Cross coupled VCO using 0.18µm UMC technology. The proposed VCO will consume 5.68mW of power at 2V supply and achieves the phase noise of -148.59dBc/Hz at 3.3GHz and the tuning range is 1.82%. The result shows that the power consumption is reduced by 10% as compared to the design with tail current. Here we have used the technique for balanced the impedance in the circuit for better result. Present VCO design will become more useful where the low phase noise and low power consumption are the main requirements. Finally present results are compared with the earlier reported work and got improvement in this reported results as given in table 1. Acknowledgment This work has been carried out in SMDP VLSI laboratory of the Electronics and Instrumentation Engineering department of Shri G. S. Institute of Technology and Science, Indore, India. This SMDP VLSI project is funded by Ministry of Information and Communication Technology, Ref. Technology (CMOS) Government of India. Authors are thankful to the Ministry for the facilities provided under this project.


33

References

[1] P. Dudulwar, K. shah, H. Le, J. Sing, “Design and Analysis of Low Power Low Phase Noise VCO”, International Conference Mixed Design 2006.

[2] Lin Jia, Jian-Guo Ma, Kiat Seng Yeo, Manh Anh Do, “A Novel Methodology For The Design Of LC Tank VCO With Low Phase Noise”, IEEE International Symposium On circuit systems, Vol.1, 23-26 May 2005, pp. 376-379.

[3] Maria del Mar Hershenson, Ali Hajimiri and Sunderarajan S. Mohan, Stephen P. Boyd, Thomas H. Lee, “Design and Optimization of LC oscillator”, 1999 IEEE/ACM International Conference on Digital Object Identifier,7-11 Nov. 1999 pp. 65-69.

[4] R. M. Weng and J. Y. Lin, “A 2.4GHz Low Phase noise Voltage Controlled Oscillator”, Department of Electrical Engineering, National Dong Hwa University, Taiwan, R.O.C. PIERS Proceedings, Beijing, China, March 23-27, 2009.

[5] Namrata Prasad, R. S. Gamad and C. B. kushwah, “Design of a 2.2-4.0 GHz Low Phase Noise and Low Power LC VCO”, International Journal of Computer and Network Security Vol.1, N0.3, 2009, pp. 15-18.

[6] B. Razavi, “A study of phase noise in CMOS oscillator” IEEE Journal of Solid-State circuits Vol. 31. no.3, Mar.1996, pp. 331-343.

Authors Profile

Namrata Prasad received the B. E. Degree in Electronics and communication Engineering. From S.A.T.I. Vidisha in 2008 and pursuing M.Tech degree in Microelectronics and VLSI Design from S.G.S.I.T.S. Indore, India in 2008-2010. Recently she is working with a project on VCO design and analysis.

R. S. Gamad received the B. E. in Electronics & Communication Engineering from V. University, India in 1995 and M.E. degrees in Digital Techniques & Instrumentation Engineering with honors from Rajiv Gandhi Technical University Bhopal, India in 2003. He has been working in teaching and research professions since 1996. He is now working as Asst. Prof. in

Department of Electronics & Instru. Engineering of S. G. S. I. T. S. Indore, India. His interested field of research is Dynamic testing of A/D Converter, Design of an A/D converter and communication.


34

New protocol to increase the reliability in WSNs

Ali HosseinAlipour1, Habib Moti Ghader 2, Mojtaba GorbanAlizadeh3 1 Islamic Azad University- Tabriz Branch, Iran

[email protected] 2 Islamic Azad University- Tabriz Branch – Young Research Club, Tabriz, Iran

[email protected] 3 Sama organization(offilated with Islamic Azad University)-khoy Branch, Iran

[email protected] Abstract: During the recent years we observe the use of wireless sensor networks in applications like phenomenon management, Battlefield Recognition, guardianship of borders and safe observing. Considering that all these distributions were randomly and data-based protocols work properly in these methods. In this article we have proposed new algorithms with automatic learning. As the results show in new algorithm networks lifetime increases and reliability increases as well.

Keywords: Data-Centric protocol, Reliability, Network

lifetime, Wireless Sensor Networks

1. Introduction Recent improvements in complex integrations (IC) have made a new generation of micro called Sensors. That from economical aspect they are commodious and also they are used in two military groups and non military [2, 3]. To consider having group of limitations such as battery life time, calculating and memory significantly they have been predicted as non recyclable and also they live until their powers fade away. So power is something rare for systems like sensor. During a special mission the correct consumption for sensors lifetime should be managed knowingly [1]. Considering the use of sensor networks, fault tolerance is something important for them. This importance especially in uses like military and nuclear laboratory is seen significantly. Whereas I such environments information are really vital, first it's necessary to deliver the information to destination. And the second is to deliver the correct information. Because deciding on the basis of incorrect information in worse than not to decide. the main focus of most researches in this field has been tend to fault tolerances which during them nodes are completely corrupted and less efforts have been done for incompatible errors. Data incompatibility errors happen due to changes in binary contents when it is processing. In this article we mix three protocols TinyLAP and FDDA. In a manner that we increased the lifetime of the network and reliability by TinyLAP and other automatic learner called FDDA. And also we corrected the incompatibility errors. In the next part done works are showed and in the third part proposed protocol has been explained and ultimately in parts 4 and 5 simulation and conclusion are mentioned.

2. Related works

2.1 TinyLAP protocol

In this part TinyLAP protocol [7] which is based on our proposed protocol is going to be partly explained. In TinyLAP protocol we allocate a Learning Automata for each node. And they use this for routing the appropriate path with their own circumstance. TinyLAP protocol includes two levels, "Distribution" and "Routing and learning”. Distribution level starts by the node which has a data to be sent. The node makes a stack called FLOOD and sends it to its neighbor. This stack includes data attributes that is very low-sized .and Neighbors by receiving this stack send to their neighbors again. When the Base Station (Sink) receives the FLOOD stack, makes another stack called FEEDBACK and distributes it through the network. And when nodes receive this stack add the new path to their table and distribute it. This level finishes by approaching the FEEDBACK stack for all the nodes in the network. At the end of this level any node has diverse paths to the central station. Each potential node for choosing the path evaluates according to 1-2 relation. In this relation

ih is the number

of steps to Sink for the ith and ngh is the number of paths in the routing table. Each node is armed to a learning automata its actions is equal with amount of paths exist in that node to the base stationand possibility of choosing each action is equal with probability of choosing the opposite path by that action in the routing table. In fact, Learning Automata actions have one to one opposition with the paths in the routing table.

∑=

−= ngh

jj

i

h

hiP

1

1)(

)2 -1(

Routing and learning starts when the source node receives FEEDBACK stack. Source node chooses the path that has the maximum probability and sends the stacks with the chosen path. Middle nodes also do so and continue until delivering the stack to central station. Each node after sending the stack waits for the respond from the receiver node. If they received positive answer the path receives a reward. TinyLAP also uses Warn stack. Warn stack sends when the energy of the nod is lower than 70% (for the first time the primary energy is assumed) each I node if receives


35

Warn stack from j node and if j node is exist in routing table of I node, penalizes the same action in the route of j node.

2.2 FDDA protocol

the main focus of most researches in this field has been tend to fault tolerances which during them nodes are completely corrupted and less efforts have been done for incompatible errors. Data incompatibility errors happen due to changes in binary contents when it is processing. This error is called soft error [5]. This error happens when the content of pktd

stack is received by the n node similar to content of pktd

stack which is not sent by the n node. The incompatibility error can happen in rode in the manner of temporarily or permanently. Internal problems in hardware components such as processor or memory units can appear incompatibility errors. In contrast with corruptions of energy resource, a node that has been involved with incompatibility errors still does its services correctly. But in doing some other services they encounter errors. From few efforts that have been done in the field of incompatibility fault tolerance is the protocol offered by Ssu et al and his colleagues [4]. The working procedure of this protocol that we call it FDDA is to first make two distinct routes between the source node and central node. And then the data in two copies and by the two routes is being sent. In the central node datum are compared with each other. And if they are similar they can be admitted. Else a new route (without any shared point with the rest routes) between source node and central node is being made and then data in three copies and from two previous route and new route is being sent. To recognize the correct data stack central node uses the quorum in vote. in FDDA protocol to recognize the corrupted node, the central node sends a message for all the routes of corrupted node(knowing that each node maintains returning route) that with this message the error amount of corrupted route increases and when a node gets errors more than threshold automatically becomes a corrupted node and has to be inactive. Central node sends a correcting message for the correct routes. And the scope of errors of this route becomes zero.

Figure 1. data sending flow in FDDA protocol

2.3 Learning Automata

A Learning automata is a conceptual mode that chooses randomly one of its actions and applies that to its environement.environement evaluates the chosen action by the automata and by a make up signal sends them to learning automata. Learning automata update its own internal situation by the chosen action and make signals. and then chooses the next action. Figure 2 shows the relation between learning automata and environment. [6]

Figure 2. Learning automata connection with environment

[6]. The environment shown by where },...,,{ 21 rαααα = is a set of inputs },...,,{ 21 rββββ = is a set of outputs and

},...,,{ 21 rcccc = is penalty probabilities. Environment can be shown by three },,{ cE βα= and set of inputs },...,,{ 21 rαααα = is a set of outputs },...,,{ 21 rββββ = and

},...,,{ 21 rcccc = set of penalty probabilities. Whenever β set has two members, environment is type P. in such an environment 11 =β is the penalty and 02 =β is considered as reward. In an environment type Q, β set has infinite members. And in an environment type S, β set has infinite members as well. ci Is the penalty probability of iα .learner automats are divided into two fixed and changing groups.

3. Proposed method TinyLAP has some problems: 1) in contrary to what designers of this protocol claim, what they call Learning automata is not a Learning automata. Instead of taking samples they use probabilities graph to choose the most probable. 2) This protocol uses Warn stack to adjust the energy consumption. That this job can be done with the responding stack. 3) In TinyLAP protocol according to the 1-2 relation, for the nodes using more than two routes to base stationthe sum of probabilities becomes more than one. (Relation 1 -3) in 1 -3, relation

ih the number of steps to base stationfor the ith route and ngh is the number of routes to the routing table. Proposed protocol is mentioned in following part that solves these problems. 4) the reliability is not offered in this part.

1)1()(

1

1

1

1

1−=−=−=

∑

∑∑

∑∑

=

=

=

=

=

nghh

hngh

h

hiP ngh

jj

ngh

jjngh

ingh

jj

ingh

i

)3 -1(

FDDA Protocol has its own problems such as sending datum from three random routes. And it can not be optimum. And the second problem is that in FDDA after comparing the first if they were not equal. Send from another three routes. Is main difficult this protocol no optimal sending path and choice randomize. Purpose of this protocol is to make balance in energy consumption in network nodes. This work happens by the learner automats. In fact, proposed protocols are combination of three methods. This method first finds all the routes like TinyLAP or EAR. Then like the FDDA method transmits the data with two routes finally these become compared in the Sinks. If they were equal, admits one of them as ultimate output.

)(nβ

Random Environment

Learning Automata

)(nα


36

Figure 3. graph of proposed algorithm

In proposed protocol each node has learning automata. And they act in this manner. Each node in each time by using the automata chooses one or two routes. If the chosen route is an appropriate route is gives reward. lese it should be punished. Purpose of this rewarding or punishing is to increase or decrease the choosing rate. Proposed protocol includes to levels “making routing tables” and “routing and learning". In level of making tables, the routing table of each node is being made. In level of routing and learning the source node starts to send the data from two or three diverse routes. The middle node does the same job. The data tables update according to their response stack. In this protocol each node has Learning automata with its routes to Sink (Sink). The route that is used to balance the energy. In following parts we will explain them in details. Making Routing Tables This level starts by Sink. This node makes FLOOD stack. And sends this stack to all it neighbors. This stack contains three fields. Number of nodes sender, count of steps to Sink, and energy stage of sender node. Before distribution the FLOOD stack by the Sink, changes the sender nodes number to its own number, changes the number steps to zero. Other nodes while receiving FLOOD stacks start to make routing tables that are like this: If a node received just one FLOOD, adds information of received stack to its own routing table and makes the routing table one. If a node received more than one FLOOD stack, puts all the information in its own routing table.(each record for one route). Choosing probability of each from this route is on the basis of relation 2-3.

∑∑==

−+=≤∀ m

ii

im

i i

ii

lenergyleve

lenergyleveh

numhop

numhophPmii

11

)1(1

1

)3 -2(

In this relation, it is number of received stack. numhopi is the number of steps for ith stack. energyleveli is the amount of I stack and m is the number of received stacks. h is the amount of constant that is a number between 1 and 0. This parameter is the displayer of effective ratio of steps in front of energy of chosen node. The more parameter reaches to one, the more steps become effective. And as it reaches to Zero, energies effect will be more to all the probable routes. FLOOD stacks receiver, quantifies fields of the FLOOD stack and distributes them in the network. Quantifying is to introducing its own number as sender nodes number and puts its energy stage in that field. To give amount for

the steps, adds one unit to the routing unit that has the least step number. Whenever considering to a routing table, one learning automata that its actions is equal by the number of its routing table. In fact, there is a one to one relation between actions of automata and routes of routing table. Probability of choosing each action is equal with probability of the same route in the table. Whenever learning automata chooses a practical node, the same route with that action goes to be chosen to be sent to the Sink. If chosen action is appropriate, the probability of selecting that increases. If not, selecting probability according to algorithm decreases. At the end of this level each node has one Learning automata and routing table to guide the data to Sink Routing and learning level In this level, each node that had a data to be sending sends data according to its routing table. By the help of learning automata chooses two different routes to send data stack. Data stack in addition to that data, includes primary source fields, final destination and sender node. The source node quantifies the data stacks fields and then sends the stacks on the chosen routes. Each node that sends a data stack, if it is not that destination, guides that stack to the Sink by the help of routing table and learning automata. This middle node prior to send the stack makes the receiver sources number same as its own number. Also this node makes another stack called ACK sends is to the node that is going to receive data stack. ACK stack includes sender's field and energy stage. ACK stacks sender node, changes receivers' numbers to senders numbers. To reduce the transmitted messages between two nodes and economizing in energy consumption, P parameter is being introduced. Each node by choosing each route sends P number of data on that route. Also, each node while receiving P numbers of data sends just one ACK stack to the sender node. This job avoids lots of ACK stacks and reduces energy consumption. Every node, by receiving an ACK stack rewards the route that ACK stack has navigated. If energy of on node that has sent the stack: Be lower than 50% from the average energy of first nodes in the route, choosing action gives penalty according to 3β that β is based on 3 -5 relation. More than 50% and less than 80% average of energy in contrast with first nodes; this action gives penalty by the relation of 3-3.

β)3.0

5.01(

−+= avgenergy

lenergyleve

V

i

p )3 -3(

In relation 3-3, avenergy is the energy of primary nodes and energyleveli is the energy stage of ACKs sender. And β is being selected by the 5-3 relation. 0/3 is the result of difference between 50% and 80%. The result of relation is between β and 2β. More than 80% and less than 100% from the average of primary nodes and route steps means that this action should be rewarded by a parameter. A is evaluated by relation 5-3. If it is more than the energy of other primary nodes, this action will be rewarded by parameter a.

Sink

Source


37

numhoplenergylevenumhopnumhoplenergyleve ii

max*)(max*

11 +−+

+=γ

γψλα (4-3)

numhopavgenergynumhoplenergyleveavgenergy ii

max*)(

22 +

+−+=

γ

γψλβ (5-3)

In relation 5-3 and 4-3, energylevel and numhopi are energy level and ACK stack senders steps. Energylevel is the starting energy of node, maxnumhop the greatest number of steps from the receiver node, and avgenergy is the average of primary nodes energy. λ1 and λ2 are displayers of the least acceptable amount for rewarding and punishing parameters. Considering the difference in scale between steps (numhop) and remaining energy (energylevel) γ parameter is being selected in a way that balances the different scale. ψ1 and ψ2 are being selected in a way that they don’t let a and β parameters rise more than a specific boundary. And finally, to make sure that received data is correct in Sink they are being compared. If both of the data were the same, one of them is going to be selected as main received data. And throw away the other. If the data is not correct the Sink sends a FEEDBACK from both routes. As soon as receiving FEEDBACK message, the node sends the data by the third route. And when it receives them again, asks for majorities vote to accept them. To receive a third incorrect data happens rarely. In this way this algorithms continues until receiving the correct data. But we just send the data from 3 possible routes.

4. Simulation Results We stimulated a wireless sensor network in a 100*100 space and with an equal distribution of 100 sensors randomly by using MATLAB software. In this simulation the central node at the end of area with the co ordinations has been put. The primary energy of sensors is 0.5 j .duration of simulation is 1000 cycle and consumption energy is equal with table 1.

Table 2: used Radio characteristics in our simulations Energy Dissipated Operation

Eelec=50nJ/bit Transmitter/Receiver Electronics EDA=5nJ/bit/signal Data Aggregation

Єƒs=10pJ/bit/m2 Transmit Amplifier if dmaxtoBS ≤ d0

єmp=0.0013pJ/bit/m4 Transmit Amplifier if dmaxtoBS ≥ d0

0

200000

400000

600000

800000

1000000

0 100 200 300 400No Nodes

No

pack

dt

Propose protocol tinyLAP FDDA

Figure 4. Count of received data stacks by several protocol

The results show that TinyLAP some times sends by three routes but has the same lifetime in network. Also there is fault tolerance in new method. Comparing with EAR acts

better and the lifetime of network increases. As you see in the Figure 4, the numbers of sent stacks have been compared with the corrupted nodes. And also fault tolerance has been added to the method. In comparing with FDDA both methods get majority vote. But in this new method to send uses the least energy using routes. So in contrasting with FDDA the lifetime of network increases significantly. That Figure 5 shows this.

0

0.5

1

0 20 40 80 100 120 140 160 180 200Failure node

scal

e re

ceiv

e di

rect

to

sum

all d

ata

FDDA Proposed protocol

Figure 5: Comparison of delivered stacks in contrast with corrupted nodes in proposed protocol and FDDA

5. Conclusion and future works In this article one protocol that uses aware automats to find the appropriate routes to send the data stacks to balance the energy consumption and correct sending of datum among nodes proposed. The results of simulation showed that proposed protocol from balancing aspect among nodes and lifetime of network and reliability have the good performance than TinyLAP, EAR and FDDA. New protocol contrary to EAR protocols does not need local information.

Reference [1]Gaurav Gupta, Mohamed Younis "Fault-Tolerant

Clustering of Wireless Sensor Networks"2003 IEEE [2]Yongxuan Lai, Hong Chen "Energy-Efficient Fault-

Tolerant Mechanism for Clustered Wireless Sensor Networks" 2007 IEEE.This work is supported by the National Natural Science Foundation of China under Grant.

[3]Ameer Ahmed Abbasi,Mohamed Younis,Saudi Arabia"A survey on clustering algorithms for wireless sensor networks" Computer Communications30(2007)2826-2841 WWW.ScienceDirect.com

[4]Chessa S. and Santi P., “Comparison-based system-level fault diagnosis in ad hoc networks”, in: Proceedings of 20th IEEE Symposium on Reliable Distributed Systems, pp. 257–266, 2001.

[5]Ssu K. F., Chou C. H., Jiau H. C. and Hu W. T., “Detection and diagnosis of data inconsistency failures in wireless sensor networks”, in: Proceedings of the Computer Networks, Vol 50, Issue 9, Pages 1247-1260, 20 June 2006.

[6] D. Chen and P. K. Varshney, "QoS support in wireless sensor networks: a survey", in Proc. of International Conference on Wireless Networks (ICWN '04), pp. 227-233, Las Vegas, Nev., USA, June 2004.

[7] M. Ankit, M. Arpit, T. J Deepak, R. Venkateswarlu and D.Janakiram. “TinyLAP: A Scalable learning automata-based energy aware routing protocol for sensor networks”.Communicated to IEEE Wireless and Communications and Networking Conference to be held in Las Vegas, NV USA. 2006.


38

New method to decrease probability of failure nodes in WSNs

Ali HosseinAlipour 1, Davood KeyKhosravi 2, Abbas Mirzaei Somarin 3

1Islamic Azad University- Tabriz Branch, Iran [email protected]

2Islamic Azad University-Osku Branch Iran [email protected]

3 Islamic Azad University-Tabriz Branch, Iran [email protected]

Abstract: Clustering in wireless sensor networks is one of the crucial methods for increasing of network lifetime. There are many algorithms for clustering. One of the important cluster based algorithm in wireless sensor networks is LEACH algorithm. In this paper we proposed a new clustering method for increasing of network lifetime. We distribute several sensors with a high-energy for managing the cluster head and to decrease their responsibilities in network. The performance of the proposed algorithm via computer simulation was evaluated and compared with other clustering algorithms. The simulation results show the high performance of the proposed clustering algorithm. Keywords: Network Clustering, Nodes failure, Energy-Aware

Communication, Wireless Sensor Networks

1. Introduction Recent improvements in integrated circuits (IC) have fostered the emergence of a new generation of tiny, called Sensors. That from economical aspect they are commodious and also they are used in non military (for instance environmental managing: temperature, pressure, tremor, etc) To consider having group of limitations such as battery life time, calculating and memory significantly they have been predicted as non recyclable and also they live until their powers fade away. So power is something rare for systems like sensor. During a special mission the correct consumption for sensors lifetime should be managed knowingly. The power of sensor can not support more than far connection. Therefore to transmit they need the architecture of multi sectional. A useful way to decrease the system lifetime is to divide them to diverse clusters [2]. Parts of a cluster-based network sensor are base stations and sensors. In this method sensors relay the data flow by head clusters. The central station always stays far from where sensors are expanded. In this manner saving the consumption energy and awareness of that to communicate with central station has various methods. Two methods of routing in articles have been proposed [5, 6]. These methods because of their route detecting and finding optimum steps in relation with central station have head load. In addition, they will have extra load on nodes that are located around

central station, so most of the traffic will be from them. To avoid these overheads and unbalanced consumption of energy some high-energy nodes called “Gateways” are deployed in the network [2]. These sensors are used as head clusters due to decrease the failure probability of head clusters. And this increases the lifetime of the network. But since this method takes a lot of expenditure so in this article we just use these sensors as manager for a number of head clusters. In this manner each one becomes gatewayamong each head cluster. This method decreases both networks lifetime and failure probability. In the second part, the architecture of two networks and the relevant tasks will be explained. In the third part, the proposed protocol has been explained and in the fourth part the results of simulating and tests evaluation can be seen. The last part involves conclusion of the article and discussing about pattern of future researches.

2. Related works System architecture for clustered sensor networks has been shown in figure 1. There just two sorts of nodes, cluster joint sensors and head cluster with tolerance of energy shortcoming. Joint sensors and homogeneous head clusters with a same identity have been assumed as similar. All the connections are wireless. The connection of joint nodes with the main station is possible only with head cluster. For sending information schedule we use TDMA (Time-Division Multiple Access) protocol. During starting the process a unique ID, primary energy and TDMA scheduling are attributed for all the sensors and gateways. We suppose that the entire node are aware from others place by the GPS. In the beginning all of the connective bridges are assumed in connection area. as the energy consumption of GPS is high , it is On at the beginning of the clustering and on the other states it is in the sleep mode. Connection scheduling among connective bridges first appears with head cluster when it establishes. The central station always stays far from where the sensors are expanded. in this order, maintaining the consumption energy and being aware of that in relation with central station have different methods: such as LEACH (Low-


39

energy Adaptive Clustering Hierarchy) [1] and SEP (Stable Election Protocol) [7] and also two other routing method have been explained in articles [5, 6].these methods due to detecting the path and finding the optimum steps in relation with command node have head load. In addition having extra load on nodes, which is located around central station, most of the traffics will be because of them. To avoid this head loads and unstable energy consumption some of nodes have been expanded through the networks by a high-energy that called gateway [2].these sensors act as gateway among clusters and central stations. And mage the entire network in cluster. Each sensor with a high-energy belongs just to one cluster. And the connection with the central station just takes place through cluster Gateway. In this method failure probability decreases and networks lifetime increases.

Figure 1. network style with clustering

3. Proposed method These methods because of their route detecting and finding optimum steps in relation with central station have head load. In addition, they will have extra load on nodes that are located around central station, so most of the traffic will be from them.

Figure 2. Multi-gateway clustered sensor network

To avoid this extra load and unstable consumption of energy

some of the nodes have been expanded with a high-energy called Gateway [2] sensors are used as head clusters due to decrease the failure probability of head clusters. And this increases the lifetime of the network but since this method takes a lot of expenditure so in this article we just use these sensors as manager for a number of head clusters. To do so, we expand some of nodes according to lifetime, space and number of exist sensors in network. While clustering we don’t need this work node. We can cluster the network with the algorithms like SEP, LEACH and TEEN (Threshold-sensitive Energy-Efficient sensor Network Protocol).afterward the clustering is done, each head cluster sends a signal to these sensors. And with these signals the sensors specify which cluster is appropriate to manage. And with the hypothesis of network they choose some of the cluster to in order to manage them. And each closer is being managed just by one of these sensors. After establishing the network the role of these sensors as gateways between head clusters and central stations, by the hypothesis network chooses some of clusters to manage. And each cluster is being controlled by just one of the sensors. After establishing the network, the sensors have the role of a gateway between central stations and head clusters. To be attentive that head clusters to transmit to central stations and data assembling and calculating in protocol consume a great deal of energy. All the responsibility of head cluster is given over to joint cluster sensors or Gateway. Then after receiving data from its joint nodes without any calculating delivers them to gateway. And its gateway that transmits them to base station after doing necessary works and calculations. This method can be used in two ways. One that we spread high-energy sensors beside other sensors. And another practical way is to put them between root station and head clusters. In both aspects both network lifetime increases and extra load eliminates from head clusters and also failure probability decreases. That other cluster heads don’t have connection with Sink station. And this connection is accomplished via Gateway and these nodes with high-energy contain the rule of Gateway. And these Gateways to lifetime termination managing same cluster heads. But similar to LEACH algorithm in any time period the cluster head is changing. When the cluster node is changing, the cluster head tell to gateway via a signal. This protocol is resumed to end of lifetime.

4. Simulation Results We stimulated a wireless sensor network in a 100*100 space and with an equal distribution of 100 sensors randomly by using MATLAB software. In this simulation the central node at the end of area with the co ordinations has been put. And we spread 4 sensors with high power in network. The primary energy of typical sensors is 0.5 J and sensors with high-energy are 1.0 J. we adjust the execution of the simulation for 1000 cycle and also consumption energy is evaluated based on table number 1.


40

Table 1: used Radio characteristics in our simulations Energy Dissipated Operation

Eelec=50nJ/bit Transmitter/Receiver Electronics

EDA=5nJ/bit/signal Data Aggregation Єƒs=10pJ/bit/m2 Transmit Amplifier

if dmaxtoBS ≤ d0 єmp=0.0013pJ/bit/m

4 Transmit Amplifier

if dmaxtoBS ≥ d0 The results of simulation show that new method in comparison with LEACH and SEP acts better and also increases the networks lifetime significantly. We test this protocol and LEACH and SEP with different sensors (50,100,200,300,400,500) and as seen in figure 3 the results show that the new method is better than exist methods. And the lifetime of the network is more than the same lifetime in LEACH and SEP. both LEACH and SEP die with 100 sensors when they see the first sensor and live for another 200 time. While in the proposed protocol after observing the first died sensor that itself observes later than LEACH and then lives for another 300 times.

Figure 3. Comparing proposed algorithm with others

5. Conclusion and future works The node of Gateway with a high-energy through the sensors is used as a central manager is just a step away from the central station. Ultimately after simulating we found out that proposed protocol plays an indispensable role in increasing network lifetime and could have been increased the lifetime in comparison with SEP and LEACH. In this article it is supposed that sensor nodes and gateways are fixed and motionless. On the other program we will research the mobile gateways.

Reference [1] Kazem Sohraby, Daniel Minoli, Taieb Znati "Wireless

Sensor Networks Technology, Protocols, and Applications" Published by John Wiley & Sons, Inc., Hoboken, New Jersey. Published simultaneously in Canada. 2007.

[2] Gaurav Gupta, Mohamed Younis "Fault-Tolerant Clustering of Wireless Sensor Networks" 2003 IEEE

[3] Yongxuan Lai, Hong Chen "Energy-Efficient Fault-Tolerant Mechanism for Clustered Wireless Sensor Networks". 2007 IEEE. This work is supported by the National Natural Science Foundation of China under Grant.

[4] Ameer Ahmed Abbasi, Mohamed Younis, Saudi Arabia "A survey on clustering algorithms for wireless sensor networks" Computer Communications 30(2007)2826-2841 WWW.ScienceDirect.com

[5]S. Singh, M. Woo and C. S. Raghavendra, "Power-Aware Routing in Mobile Ad Hoc Networks", Proc. of ACM MOBICOM'98, Dallas, Texas, October 1998

[6] D. Estrin, R. Govindan, J. Heidemann, and S. Kumar. "Scalable coordination in sensor networks" Proc. of ACM/IEEE MobiCom 1999, Seattle, Washington, August 1999.

[7] Georgios Smaragdakis Ibrahim Matta Azer Bestavros” SEP: A Stable Election Protocol for clustered heterogeneous wireless sensor networks” Technical Report BUCS-TR-2004

[8] Piraeus Tillapart, Sanguan Thammarojsakul, Thanachai Thumthawatworn, Pratit Santiprabhob”An Approach to Hybrid Clustering and Routing in Wireless Sensor Networks” 2005 IEEE.

Author Profile

Ali HosseinAlipour received the B.S. degrees in Computer Engineering from Islamic Azad University- Shabestar Branch in 1999 & 2002 and M.S. degrees in Computer architecture engineering from Islamic Azad University- Tabriz Branch in

2007 & 2010, respectively. My interests about researches include WSN and schedule into multiple processes. .

800

900

1000

1100

1200

1300

1400

10 100 200 300 400 500

?

Net

wor

k Li

fetim

e

NEW Protocol SEP LEACHNo.Nod


41

A New Clustering Algorithm for Increasing of Lifetime in sensor Networks

Abbas Mirzaei Somarin1, Habib Motee Ghader2, Amir Masoud Rahmani 3 and Ali Ghafari4

1 Islamic Azad University-Ardabil Branch,Ardabil, Iran

[email protected]

2 Islamic Azad University- Tabriz Branch- Young Research Club, Tabriz, Iran [email protected]

3 Islamic Azad University- Oloom Tahgigat Branch-, Tehran, Iran

[email protected]

4 Islamic Azad University- Tabriz Branch , Tabriz, Iran [email protected]

Abstract: One of the crucial problems in sensor networks is limitation of energy, which affects in network lifetime. There are many algorithms for increasing network lifetime. One of the methods in this case, is network clustering. A good clustering increases the lifetime of network. In this paper a new clustering algorithm based on Learning Automata for increasing network lifetime is proposed. The proposed algorithm based on LEACH Protocol. Performance of proposed algorithm is compared with previous methods by implementing simulation. The obtained results from proposed algorithm is better than others. Results are indication of high efficiency of suggestion algorithm.

Keywords: Network Clustering, Algorithm.

1. Introduction Wireless networks are networks, which are, consist of many little nodes by low facility. These nodes, which are called a sensor, can feel special feature such as wetness, temperature, and pressure in their environment and send this for their neighbors. In other words two major facilities of these sensors are the feeling of special parameter around environment is ability of connection. In some operations these nodes might be join to each other by connective cables, but in most cases a wireless network is completely wireless. In this network nodes are generally fixed or fixed limited motions. Despite of fixed networks and other wireless networks, which the quality of service is completely, clear in sensor networks there is no fixed recounting. Some of these recounting are network proper covering, amount of active nodes in per time, accuracy of received information in Sink (central node) and the time of transferring information to Sink. Some of these recounting such as proper covering and amount of active nodes in per time are depended to operation and others like accuracy of received information and time of transferring information Sink intended the feature of network. One of the sensor networks important features is probability of devastation in some nodes specially because of losing energy. For this reason there are so many nodes in wireless networks. In this case if some of these nodes has been destroyed others can replace them so some of these nodes should be active but others should be inactive in order not consume their energy. Therefore, quality of service

can be recounting according to active nodes because in this case networks lifetime will be increased. One of the effective ways in increasing of networks lifetime is clustering. In sensor network clustering, sensor network divided into some branches and one of the nodes have been chosen as a top branch. Duty of top branch is receiving information from other branches and sends them to the Sink. In dynamic clustering network can be cluster just once and nodes of branches never transfer to others but network nodes can be the members of other branches. In this paper by using of learning automata we will change the LEACH algorithm that increases network lifetime. In the rest of this paper we will talk about learning automata in section 2. Then in section 3 LEACH algorithm will be explained. In section 4 suggested algorithm will be introduced and in section 5 the rest of dramatization had been shown.

2. Learning Automata Learning automata is an abstract model that chooses randomly an operation from a set of finite operations and then applies it on the environment to the selected operation environment is evaluated by learning automata and then informs the evaluated result by the help of a reinforcement signal to the learning automata. The learning automaton updates its interior situation by utilizing selected operation and reinforcement signal and selects the next operation afterwards. Interconnection between learning automata and the environment is shown in Fig.1 [1].

Figure 1. Learning automata connection with

environment [2]. The environment shown by where },...,,{ 21 rαααα = is a set of inputs },...,,{ 21 rββββ = is a set of outputs and

},...,,{ 21 rcccc = is penalty probabilities. When β is a set of binary, so the environment is a P type. In this kind of environment 11 =β is considered as penalty and 02 =β as

)(nα )(nβ

Random Environment

Learning Automata


42

reward. In Q type environment, β includes numerous finite members and in S type β has a number of infinite members. ci is a penalty probability of operation. Learning automata is divided into two groups as stable and unstable structure. Learning automata can be utilized in many different ways such as: routing in communication network, face recognize, programming processes in a computer network. queue theory, access control in asynchronous transmit network, partitioning objects and finding an optimal structure for nerve systems [2, 7-11].

3. LEACH Protocol One of the first and famous presented hierarchical protocols to the sensor networks is LEACH protocol [3]. In this protocol, the time span of the nodes activity divides into some periods. At the beginning of each period some of the nodes be selected haphazardly as a cluster head (CH). To do this, each node produces a haphazard number between 0, 1. Where as this number from the amount of T (n) that results in by the use of formula (1), be less, the stated node be presented as cluster head. In formula 1, p is the correlation of the number of clusters to the total nodes of the networks, r is the number of period and G the number nodes that in the last period of 1/p are not selected as cluster head.

∈×−=

otherwise

Gnif

prP

P

nT

,0

,)1mod(1)(

Relation 1. Random number generation between zero and one

After determining the nodes of cluster head, the other nodes on the basis of the power of the received signal from each cluster, decide to be accepted as a member of which cluster. The cluster head node divides its responsibility glen in to some time slots (Figure 2). This time slots on the basis of TDMA mechanism are cooperated between the members of cluster. In each time slot, the cluster head connect with on of the members of the cluster and receives the information packs of that member. The cluster head in every some slots sends its received information from its members to the Sink. In order to the distribution of the load on different nodes after finishing of a period, to start a new period, the cluster head by the declared mechanism above are exchange.

Figure 2. Period and Time Slice in LEACH protocol

4. Improved Protocol (LALEACH) As mentioned in the previous section LEACH algorithm do the act of clustering and is selected a cluster head to each

cluster. And the responsibility of each cluster head is gatherings the information’s from the other nodes of the same cluster and sending of to the Sink. In LEACH protocol the cluster head never be changed. And because this reason the node that is considered as cluster head, consumes more energy. And this itself causes the declining of the life long of the network. In order to overcome this problem we can exchange dynamically the cluster head node between the members of cluster nodes. In suggested algorithm, we change the LEACH node in a way that the node of cluster head in each period exchange dynamically between the clusters nodes. The way of selecting the cluster head node between the nodes of a cluster forms by the learning automata. In suggested algorithm there is considered a learning automata for each cluster the number of the actions of automata equals with the number of clusters nodes. Each action by automata corresponds with a node from the cluster. In figure 3 a cluster and its corresponded automata are showed.

Example of Sensor network

and cluster

Equivalence Learning

automata for cluster (1)

Figure 3. The assigning the cluster nodes to correspond learning automata

The Way of choosing the clustering After clustering in LEACH protocol, turns to the selection of cluster head to each clusters. As it mentioned, for each of the clusters there is considered a learning automata that the number of its actions equals with the number the nodes of the some cluster. The learning automata between its actions choose an action that its number is more than the other. After choosing a node as a cluster head, in order to transfer the information of its clusters nodes by the TDMA mechanism connects with them and again the learning automata choose another node as the cluster. (It chooses the node as a cluster that the corresponding action that has more amount than the other actions). The amounts of automata in each time are exchange by the formula 2.

i

i

in

nn d

eP =

Relation 2. action value Relation

N11

N6

N2 N1

N9

N10

N3

N4 N5

N8 N7

LA

a1

a3 a4

a2 a5

N3 N4

N8

N7 N5


43

In this formula: n : The number of cluster nodes.

in : The thi node.

ine : The remain energy of in node.

ind : The average of distance of in node with the other

nodes of the same cluster.

inp : The amount of thi action of automata.

In figure 4, period and the way of time slot in suggested protocol is showed. By comparing the figure 2 and 4 the way of exchanging the node of cluster head for a cluster is showed (clustering in each period is done and the exchange of the cluster head node is done inside the clusters).

Figure 4. Period and Time slice in LELEACH protocol.

Figure 5 shows a sensor network. At first this network by the LEACH algorithm is divided into three clusters. By the next period that a new algorithm will be occurs, the suggested algorithm will exchange the cluster head dynamically.

Figure 5. an Example of clustered sensor network At first after clustering by the use of LEACH algorithm, the node of cluster head is appear. After connecting the cluster head node with the all of its clusters node and

transferring their outputs to the Sink, turns to the exchange of cluster head node. The way of exchanging the clusters nodes of number (1) as the cluster head is occurs like this: First stage: in this stage by the information of the number (1) cluster the exists in figure of table 6, the number of actions of the cluster (1) automata is appeared.

AUTOMATA ACTIONS VALUES

Average Distance Remain Energy

Node No

55.043.53

1==np 34.5

3862

1 =++

=d Jen 31

= 1n

83.065

2==np 6

31242

2 =++

=d Jen 52

= 2n

13.067.71

3==np 67.7

33128

3 =++

=d Jen 13

= 3n

46.03.4

24

==np

3.43

3464 =

++=d Jen 2

4= 4n

Figure 6. The figure (5) network cluster number (1) nodes information.

After determining the amount of the automata’s actions related to the cluster (1) as table of figure 6, now it is selected a node as a cluster head that its amount of action is more than the others, so it is selected as the cluster head node. After selecting the node of number (2) as a cluster head, it declares its being as cluster head to the all nodes exist in its cluster. The other cluster head contents of the table of figure 6m which are being up-to-date in selecting of cluster head are being selected. After connecting the node ( 2n ) by all nodes of its cluster and transferring its information to the cluster head node, it makes up-to-date its remained energy in table (6). After getting up-to-date the information of the table, the node, which its amount of action is more than the others, is selected as cluster head. This process continues until the beginning of next stage. 5. The Result of Simulations In this section the result of suggested algorithms simulation is comparing and assessing with the result of previous algorithm. The results of suggested algorithm is compared and assessed in respect of lifetime by the algorithm of LEACH, HEED, Extended HEED clustering. In the tested simulation of the sensor environment is considered

150150× and the radio range of the sensor environment is considered 30 meters. The first energy of the nodes is considered 2J. Trials to some of the different sensor nodes is done by N={100, 150, 200, 250, 300, 350, 400, 450, 500} the shown results is the result of min of results gained from 10 performing different simulations. The results from different algorithms are compared with each other respects of lifetime that is one of the main standards of the quality of service in sensor networks. The results of assessing are showed in Figure (7) it is clear that the lifetime of network in the suggested algorithm is more than the other ways.

N11

N6

N2

N1

N9

N10

N3

N4

N5

N8

N12

Cluster 1

Cluster 2

Cluster 3

2m

12m

6m 4m

3m

8m

CH

CH

CH

N19 N18

N16

N17

N13

N23 N20

N15

N22 N21

N14 N7

N25

N24

Cluster 4 Cluster 5

Cluster 6

CH

CH CH


44

Figure 7. Comparison of suggested algorithm with others.

6. Conclusion In this paper a new algorithm based on LEACH algorithm is suggested. The goal of suggested algorithm is the increasing the lifetime of the sensor network. In LEACH algorithm the nodes of cluster head are selecting with out considering their remained energy and because the nodes of cluster head consume more energy than the other node, so LEACH algorithm in this respect that chooses haphazardly the nodes of cluster head causes decreasing the lifetime of the network. In suggested algorithm instead of selecting haphazardly the node of cluster head, it is selecting between the nodes their remained energy is more than the other nodes and also near to its neighbor nodes. So, by this way nodes with high energy are selecting as cluster head and increase the lifetime of network. The effectiveness of suggested algorithm compared with cleared clustering LEACH, HEED, Extended HEED algorithms has better results. Reference [1] D. Chen and P. K. Varshney, "QoS support in wireless

sensor networks: a survey", in Proc. of International Conference on Wireless Networks (ICWN '04), pp. 227-233, Las Vegas, Nev., USA, June 2004.

[2] H.MotieGhader, S..Parsa, M.Hossein Nejad, “Application of Learning Automata for DAG Scheduling on Homogeneous Networks", 17th Iranian Conference on Electrical Engineering, Iran University of Science and Technology, ICEE2009.

[3] S. M. Abolhasani, M. Meybodi, “Usage of Learning Automata for Routing, Fault Tolerance and Topology Control in Wireless Sensor Network”, MSC Thesis, March 2008.

[4] W. R. Heinzelman, A. Chandrakasan, and H. Balakrishnan, “An Application-Specific Protocol Architecture for Wireless MicrosensorNetworks,” IEEE Transactions on Wireless Communications, vol.1 ,no.4 , pp.660 –670 , October2002 .

[5] O. Younis and S. Fahmy, "Distributed Clustering in Ad-hoc Sensor Networks: A Hybrid, Energy-Efficient Approach", In Proc. of IEEE INFOCOM, March 2004 .

[6] M. Esnaashari, M. R. Meybodi1, “A novel clustering algorithm for wireless sensor networks using Irregular Cellular Learning Automata”, IEEE, 27-28 Aug. 2008.

[7] Meybodi, M. R. and Beigy, H., New Class of Learning Automata Based Scheme for Adaptation of Backpropagation Algorithm Parameters, Proc. Of EUFIT-98, Sep. 7-10, Achen, Germany, pp. 339-344, 1998.

[8] Oommen, B. J. and Ma, D. C. Y., Deterministic Learning Automata Solution to the Keyboard Optimization Problem, IEEE Trans. On Computers, Vol. 37, No. 1, pp. 2-3, 1988.

[9] Beigy, H. and Meybodi, M. R., Optimization of Topology of neural Networks Using Learning Automata, Proc. Of 3th Annual Int. Computer Society of Iran Computer Conf. CSICC-98, Tehran, Iran, pp. 417-428, 1999.

[10] Beigy, H. and Meybodi, M. R."Optimization of Topology of Neural Networks Using Learning Automata, Proc. Of 3th Annual Int. Computer Society of Iran Computer Conf. CSICC-98, Tehran, Iran, pp. 417-428, 1999.

[11] Hashim, A.A., Amir, S.and Mars, p. Application of Learning Automata to Data Compression, In Adaptive and Learning Systems, K. S. Narendra (Ed), New York: Plenum Press, pp. 229-234, 1986.


45

Securing Digital Information using Quasigroups

Saibal K. Pal1 and Shivam Kapoor2

1Scientific Analysis Group, DRDO, Metcalfe House, Civil Lines, Delhi – 110 054, India

[email protected]

2Department of Computer Science, University of Delhi, Delhi – 110 007 India

[email protected]

Abstract: The non-associative property of quasigroups has been recently found to be useful in many information security applications. In particular, quasigroups have been used for ensuring confidentiality and integrity of the data transmitted over insecure public channels. Quasigroups operations are computationally simple and can be efficiently used for protection of voluminous media like images, audio, video and different forms of multimedia. They are also suitable for securing data transmitted from and to mobile and miniature resource-constrained devices with limited capability. We first describe schemes for construction of large sized quasigroups that can be used to ensure confidentiality without complicating the operations or increasing the processing of data by introducing additional rounds. These quasigroups are generated using isotopies, product of smaller quasigroups, affine and non-affine mappings, theta-mappings, keyed permutations, T-functions etc. Using these concepts, design of fast encryption schemes for different media is presented. Ensuring data integrity is also of vital importance for many present day applications. Schemes for generation of highly non-associative quasigroups are described and their use for secure hashing of data is explained. Computer implementation of these schemes demonstrates their simplicity and power for efficiently securing digital information.

Keywords: quasigroup, isotopy, non-associativity, encryption,

hashing, digital media.

1. Introduction 1.1 Latin Square A Latin square [1], [2] of order n is an n x n square matrix whose entries consist of n symbols such that each symbol appears exactly once in each row and each column. Examples of a 4 x 4 and 5 x 5 Latin square are given below

Order 4:

cbadbadcadcbdcba

This Latin square generated by using elements of the set {a, b, c, d} is called a reduced Latin square as the elements in the first row & the first column are in monotonically increasing order.

Order 5:

3145223514152434213554321

This specific Latin square is not a reduced Latin square as the rows are not arranged in order.

Number of Latin Squares: The total number of Latin squares N [3] of order n are computed using the formula:

N(n, n) = n! (n-1)! L(n, n) (1)

Table 1. The number of reduced Latin squares n L(n , n)

1 1 2 1 3 1 4 4 5 56 6 9408 7 16 942 080 8 535 281 401 856 9 377 597 570 964 258 816

Here, L(n, n) is the number of reduced Latin squares of order n. For large values of n, the number of reduced Latin squares is difficult to compute & hence the total number of Latin squares of high order is unknown. 1.2 Quasigroup (QG)

1.2.1 Definition: A quasigroup (Q,*) [4], [5] is a set Q of elements along with a binary operation ‘*’ having the following properties: (a) For all a, b є Q, a * b є Q (Q is closed under *) (b) For all a, b є Q, there exist unique x, y є Q so that

a * x = b and y * a = b i.e. ( (Q, *) has unique solubility of equations).

Because of the unique solubility of equations, each element will appear exactly once in each row and exactly once in each column of the multiplication table of (Q,*). That is, each row and column is a permutation of the elements of Q. If |Q| = n, then the interior of the Cayley table for (Q, *) forms an n x n Latin square.


46

Examples of Quasigroups:

• The set of Integers with the subtraction operation ( - ) form a quasigroup.

• The non-zero Rationals and non-zero Reals with division operation ( ÷ ) form a Quasigroup.

• Let Q = Z6 = {0, 1, 2, 3, 4, 5} and let x * y = (x + y) mod 6. Then ( Q , *) is addition modulo 6 on Z6 and the Cayley table for ( Q ,* ) is given by

432105532105442105433105432205432115432100543210*

Figure 1. Cayley table for (Q ,*)

(Q,*) is a quasigroup because its interior is a 6 x 6 Latin square. This quasigroup is also a group. 1.2.2 Some Properties of QGs • Quasigroups have cancellation properties that is if ab =

ac, then b = c. This follows from the uniqueness of left division of ab or ac by a. Similarly, if ba = ca then b = c.

• Left and right multiplication: By definition of a quasigroup Q, the left and right multiplication operators are defined by

L( x )y = xy R( x )y = yx (2) These are the bijections from Q to itself. The inverse

maps are given in terms of left and right division by L( x ) ˉ¹ y = x\y R( x ) ˉ¹ y = y/x (3) • A quasigroup is not required to be associative,

commutative, or to have an identity element.

1.2.3 Use of QGs in Coding & Cryptography Apart from their applications in many diverse areas, quasigroups have been recently used as cryptographic primitives in the design of encryption schemes, for pseudo-random number generation, for construction of hash functions and for design of error-detecting codes. As the basic operations on quasigroups are very simple, it is possible to construct fast and efficient schemes for different applications. 1.2.4 Basic Encryption and Decryption Using QGs Suppose that we are given a quasigroup with the operation * as follows:

312030213220311130203210*

Figure 2. Quasigroup used for Encryption

The following convention is normally used: Leader - Any of the symbols of Quasigroup can act as a leader. Here we have four choices: 0, 1, 2, 3 for the leader. The leader decides which row would contribute to the encryption process. Key - The leader and the quasigroup itself (with the unique arrangements of elements) may be used as keys during the encryption process. The basic encryption scheme [6], [7] using a leader is as follows. 1.2.5 Encryption Scheme: Let the plain message be represented by M = (x1, x2,…..,xn). We choose a leader L and pass on this secret information as a key to the receiver. Assuming that the quasigroup generation mechanism is available to the receiver, encryption is performed as follows

EL(x1, x2,…….,xn) = (y1, y2,……..,yn) where y1 = L * x1 and yi = yi-1 * xi (for all i > 1) (4) For example if M = (3 0 2 1 2 3 3 1), then x1 = 3, x2 = 0, x3 = 2 and so on. Then, with the Leader as L = 0, the quasigroup given in Figure 2, and the encryption formula given above, we have y1 = L * x1 = 0 * 3 = 1

y2 = y1 * x2 = 1 * 0 = 0

y3 = y2 * x3 = 1 * 2 = 0 and so on.

Thus, the encrypted message EL(M) = (1 0 0 0 3 3 3 2). Since there are various options for choosing a leader, the encryption scheme can be made stronger by using different leaders for encryption of different blocks of the message. The Quasigroup can also be changed periodically by permuting its rows and columns to increase the complexity of encryption scheme. 1.2.6 Decryption Scheme: For each quasigroup operation ‘*’ we can associate a new quasigroup operation ‘o’ defined by: x o y = z iff x * z = y (5) The “dual” operation ‘o’ is used for decryption and the “inverse” quasigroup is generated using the above equation:

312030213213021203103210o

Figure 3. Quasigroup used for Decryption

Decryption is carried out with the following formula:

DL(y1, y2,…….,yn) = (x1, x2,……..,xn)

where x1 = L o y1

and xi = yi-1 o yi (for all i > 1) (6)


47

By performing calculations using (6) and the previous quasigroup, the encrypted message is converted back to the original plain message.

DL(EL(M)) = (3 0 2 1 2 3 3 1) This scheme is computationally simple and is based on basic lookup operations on quasigroups. To make this scheme more powerful and practical, construction of large quasigroups is required and is reported in Section 2. 1.3 Non-Associativity of Quasigroups

In addition to the requirement of constructing large and unstructured quasigroups for cryptographic purposes, it is also important to generate quasigroups with high degree of non-associativity. These find applications in the design of secure keyed hashing schemes or message authentication codes [8]. To measure the degree of non-associativity of quasigroups, Meyer [9] borrowed the concept based on measuring the non-commutativity of groups. For a group (G,o), the multiplication group is the subgroup generated by all the left and right multiplications of G. As the left and right multiplication by an element would be the same for a commutative group, the multiplication group would be smaller. In contrast, for a non-commutative group, the multiplication group would be larger. Therefore, the size of the multiplication group of a group indicates the degree of non-commutativity of the group. A similar approach can be used to measure the degree of non-associativity of a quasigroup (Q,o). Here also, a larger multiplication group indicates that the quasigroup has a higher degree of non-associativity. 2. Generation of QGs

Construction of large quasigroups [10], [11] from smaller ones is an important problem for many applications. In cryptography, larger the size of the Latin square, larger is the number of choices that can be made by the communicating party and higher levels of security can be provided. Moreover, generation of Latin squares based on key permutations [12] help to regenerate the Latin square used by the sender at the receiving end with minimal information exchange. We present below different schemes for generation of huge Latin squares.

2.1 Simple Product of Quasigroups The notion of simple product of quasigroups is important for cryptographic applications since, if two quasigroups are given, it permits to construct a quasigroup of the order equal to the product of the orders of these quasigroups. Let Q1 = < R , • >, Q2 = < S , ~ > be two arbitrary quasigroups. Let the elements of two quasigroups be given by

R = {0, 1, 2, . . . , n1 − 1}, S = {0, 1, 2, . . . , n2 − 1}

Then the simple product Q of these quasigroups is the algebraic system Q = Q1 × Q2 = < T , * > (7) where T = {0, 1, . . . , (n1n2) − 1} To define the operation * in the set T, let us first assume that: Tcp = R × S = {t0, t1, . . . , tn1n2-1} (8)

which is the Cartesian Product (CP) of the sets R and S. Further, let ti = (ri, si), tk = (rk, sk), where i, k ∈ T; ri, rk ∈ R; si, sk ∈ S. Noting that the quasigroup with elements belongs to the set Tcp, Qcp = < Tcp , * > is the simple product of quasigroups, we can define the operation * as follows:

ti * tk = (( ri • rk ), ( si ~ sk )) (9) But we want to represent the simple product of quasigroups as a quasigroup. To this end, we must convert the set Tcp into the set T. There are many ways of doing such a conversion. The mapping h : tx →X, where tx ∈ Tcp, X ∈ T defined by the function h(tx) = h(rx, sx) = n2rx + sx (10) gives one of the simplest solutions. 2.2 Generation using Linear Mapping Using this scheme, we generate two linear functions f and g and then elements associated with each function is stored in one dimensional array of size equal to size of the permutation. Now, in order to generate the ( i , j )th element of the huge Latin square, the ith element of the first function is added to the jth element of the second function and modulus operation w.r.t the size of the Latin square to be generated is applied to the addition of ith element of 1st function and jth element of 2nd function. In this way we get a huge Latin square with elements in the range 0 to the size of Latin square. Let Q = Zn = {0,1…, n-1} and let the group operation be addition modulo n. Then we can create a Quasigroup (Q , ±) from (Zn , +) by defining f(x) = px + a (11) g(x) = qx + b where p and q are relatively prime w.r.t. the order of the quasigroup and a, b are positive integers less than the size of quasigroup, and then further defining h( x , y ) = ( f( x ) + g( y ) ) % n (12) For example: Let Q = Z8 = {0, 1, 2, 3, 4, 5, 6, 7} and let the group operation be addition modulo 8. Then we can create a quasigroup (Q , ±) from (Z8 , +) by defining f(x) = 5x + 6 and g(x) = 3x + 7, where a = 4 & b = 1 and p = 3 & q = 5. Since we are creating (Q , ±) by linear mapping, we will define h(x, y) = (f(x) + g(y)) % 8. Then (Q , ±) is as shown below

52741630705274163630527416563052741416305274341630527274163052127416305076543210±

Figure 4. Quasigroup using Linear Mapping We find that (Q , ±) is not associative, because h(h(0, 2), 4) = h(3, 4) = 0 but h(0, h(2, 4)) = h(0,3) = 6. It is not commutative, because h(0,1) = 0 but h(1,0) = 2. There is no identity element and hence no inverses.


48

Limitations of the above method: The problem with this scheme is that (Q , ±) is still quite structured. The elements in each row always appear in the same order. Similarly, the elements in each column are also arranged in the same order. In addition, each row is a copy of the previous row, shifted one space to the right. Each column is a copy of the previous column, shifted one space down. As we see in above example: 1st and 2nd row are (5 2 7 4 1 6 3 0) and (0 5 2 7 4 1 6 3) respectively. Hence by determining any of the rows, one can guess the entire Latin square. 2.3 Generation using Keyed Permutation In the previous scheme, one row could be easily derived from the other row. We remove this problem by using non-linear mapping. In this scheme two random permutations [12] f and g are generated and then each permutation is stored in one dimensional array of size equal to size of permutation. Now, in order to generate the (i, j)th element of the huge Latin square, the ith element of the first permutation is added to the jth element of the second permutation and modulus operation with respect to the size of the Latin square to be generated is applied to the addition of ith element of 1st permutation and jth element of 2nd permutation. In this way we get a huge Latin square with elements in the range 0 to the size of Latin square. Let Q = Zn = {0, 1…., n-1} and let the group operation be addition modulo n. Then we can create a quasigroup (Q , ±) from (Zn , +) by supposing

f(x) = any random permutation {6, 2, 8…} g(x) = any random permutation {5, 3, 9…}

then defining h( x , y ) = ( f( x ) + g( y ) ) % n (13) In the example below, to generate a Latin square of order 8 first we will generate two permutation f and g and then apply h( x , y ) = ( f( x ) + g( y ) ) % 8 for x and y ranging from 0 to 7

X 0 1 2 3 4 5 6 7 f(x) 2 5 0 4 6 7 1 3 g(x) 7 4 3 2 1 0 6 5

Then the quasigroup (Q , ±) created from f and g by using the above equation is

01345672767123450645701236534670125412456703356012347223567014170234561076543210±

Figure 5. Quasigroup using Keyed Permutation We find that (Q, ±) is not associative, because h(h(0, 2), 4) = h(5, 4) = 0 but h(0, h(2, 4)) = h(0, 1) = 6. It is not commutative, because h(0, 1) = 6 but h(1, 0) = 4. There is no identity element and hence no inverses. Unlike the previous case, all the rows are independent of each others.

2.4 Generation using T-Functions T-functions, proposed by Klimov and Shamir [13], [14] are a relatively new class of invertible mappings using a

combination of arithmetical operations like addition, subtraction, multiplication and negation together with Boolean operations like OR, XOR and NOT. This helps to design cryptographic schemes resistant to many present day attacks. Example of such a mapping is )2)(mod( 2 nVCxxx +→ (14) where C is a constant and V represents the logical operation OR. This turns out to be a permutation of single cycle of length 2n given certain conditions. In general, a T-function is a mapping in which the ith bit of the output depends on 0, 1, … , ith input bits. Using a small number of such primitive operations over n-bit (32, 64 etc.) words, efficient cryptographic building blocks can be designed. It is interesting to note that composition of two T-functions will also be a T-function. Because we will be using T-functions to create a quasigroup, it is convenient to choose k = l, so that f takes as input and produces as output k many binary strings of length n. In order to use a T-function f to define a quasigroup operation, we will see that f needs to be a permutation. Therefore, we need to know which T-functions are invertible. Example: Let 3

232

32 *: zzzv → be given by

v(x , y) = x2y2 + 3(x V y) (15)

here V represents the Boolean OR operation and addition and multiplication are defined for mod 23 = mod 8. Let c = < 1 , 0 , 1 > ∈ Z2

3. We define

x o y = c + (x + y) + 2v(x , y) = 5 + x + y + 2x2y2 + 6(x V y) (16)

Based on this, the quasigroup is created where the binary representatives of the quasigroup elements are used for ease of notation.

o 0 1 2 3 4 5 6 7 0 5 4 3 2 1 0 7 6 1 4 7 2 5 0 3 6 1 2 3 2 5 4 7 6 1 0 3 2 5 4 7 6 1 0 3 4 1 0 7 6 5 4 3 2 5 0 3 6 1 4 7 2 5 6 7 6 1 0 3 2 5 4 7 6 1 0 3 2 5 4 7

Figure 6. The quasigroup created using a T-function v(x, y) One of the disadvantages of using a T-function in creating a quasigroup can be seen in the resulting structure present in the quasigroup. From the given example we observe that the entries in each row and column alternate between even and odd numbers in the quasigroup. That is, if x o y is even, then x o (y + 1) and (x + 1) o y will be odd and vice versa. It is evident that this property holds in any quasigroup created from a T-function. We notice that both x and y are either even or odd. Clearly, if x is even, then x + 1 is odd and vice versa. Since x o y = c + x + y + 2v(x, y) and x o (y + 1) = c + x + y + 1 + 2v(x , y + 1) (17)


49

Because 2v(. , .) is always even, the parity of xo(y+1) will be different than the parity of xoy.

Other useful schemes for generation of quasigroups are based on isotopies and complete mappings. In addition, if we choose affine isotopies or affine complete mapping, the Quasigroup created has special properties. 2.5 Theta Mapping

A quasigroup can also be created from a group using a special kind of mapping called a complete mapping [9].

Definition: Let (G, +) be a group and let i : G à G denote the identity map on G. Ө : G à G is a complete mapping if Ө is a bijection and i - Ө is a bijection where (i - Ө)(x) = x - Ө(x). Example: Let (G, +) = (Z9, + ) where Z9 = (0 , 1, 2, 3, 4, 5, 6, 7, 8) and addition is performed modulo 9. Then Ө(x) = 5x + 4 is a complete mapping as seen in Figure 7 because both Ө and i - Ө are bijections x 0 1 2 3 4 5 6 7 8 Ө(x) 4 0 5 1 6 2 7 3 8 i - Ө(x) 5 1 6 2 7 3 8 4 0

Figure 7. A complete mapping on Z9 Creating Quasigroups using Complete Map: Sade [15] suggested creating a quasigroup (Q, o) from an admissible group (Q, +) and a complete mapping Ө by defining x o y = Ө(x - y) + y , for x , y ∈ Q (18)

Example Let (Q, +) = (Z2

3 , o ) and let Ө (< X2 , X1 , X0 >) is given as < X1 XOR 1, X0 , X2 > if X2 = 0

< X1 XOR 1 , X0 XOR 1, X2 > if X2 = 1 (19) Then Ө and (i XOR Ө) are permutations as shown in the following figure

><><><><><><><><><><><><><><><><><><><><><><><><

011100111101110011000101101110111001100010110010000010111011100001001000

))(

,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,

x(iXORθθ(x)x

Figure 8. Ө and i XOR Ө on Z23

If the elements of (Z23 , XOR) are represented as integers

corresponding to binary strings for ease of notation , then Ө and i - Ө are shown in Figure 9 and the quasigroup created by Ө(x XOR y) XOR y is shown in Figure 10.

65031274))((13572064)(76543210

xXORix

x

θθ

Figure 9. Values of Ө and i XOR Ө on integer corresponding to elements of Z2

3

56302147774120356621746503530564721403657412321475630247213056165031274076543210o

Figure 10. Quasigroup created using the complete map Ө 3. Generation of Non-associative QGs In Section 1.3, we explained the need for generation of highly non-associative quasigroups for cryptographic purposes and measurement of the non-associativity of a given quasigroup. We present schemes with examples for generation of such structures.

Let g : Q à Q be a transposition on a cyclic group (Q , +). Then (Q , o) given by x o y = x + g(y) is a highly non-associative quasigroup. Here Mult(Q) = Sym(Q), where Mult(Q) is the multiplication group of (Q , +) and Sym(Q) represent the symmetric group of (Q, +). Both f and g appear in Mult (Q) as (Q , o) is generated from isotopy from f = id and g. g is a permutation of order 2 as it is a transposition. In addition, if Lx , Rx denote the left and right multiplication by x ε Q and Lx and Rx are in Mult (Q) for all x € Q. Since for g(a)=a, we see that La ∈ Mult (Q) for a ∈ Q. This now implies that La is a cycle of length n = |Q|. g and La together generate the entire symmetric group of size n. Therefore, Sym(Q) = Mult(Q) and this implies that (Q , o) is highly non-associative. For example, let (Q , +) = (Z8 , +) and let g : Q → Q be a transposition such that g interchanges the elements 0 and 1. Then (Q , o) created by x o y = x + g(y) is a highly non-associative quasigroup. Here, we have g(0) = 1 , g(1) = 0 ,and g(a) = a for a >1 xoy = (x + g(y)) mod 8 (20) So a structured highly non-associative quasigroup can be created as

65432170754321067643210756532107645421076534310765423207654312176543201076543210o

Figure 11. A Structured Non-associative Quasigroup

This quasigroup is highly non-commutative and non-associative. We can see from the above example that under the operation ‘o’, For x = 0 and y = 1 x o y = 0 and y o x = 2 For x = 1 and y = 2 x o y = 3 and y o x = 2


50

4. Application of QG to Encryption

Quasigroups are suitable for designing efficient as well as secure encryption schemes due to simplicity of the basic operation. Moreover, large and unstructured quasigroups can be generated using one of the schemes explained in the previous sections. Use of such quasigroups helps to improve the security of the scheme. A basic encryption and decryption scheme based on table lookup has been described in Sections 1.2.4 to 1.2.6. In this scheme the leader L acts as a key to the encryption algorithm and has to be changed frequently. In addition to the leader, changing the quasigroup also helps to improve the security of the scheme. This can be practically made possible [11] without having to transmit much of information as the key to the receiver. The observations below is of (a) Original Tank Image (b) Encrypted Image using same quasigroup but 256 different leaders (one for encrypting each row) and (c) Encrypted image using different leaders and different Quasigroups. To do away with the effects of high levels of redundancy found in visual data, a new scheme has been proposed [12] using lookup and shift operations so that subsequent chunks of similar data in the plain image are mapped into different blocks using the same encryption function. Other efficient schemes using a combination of quasigroup operations together with basic cryptographic primitives are also under development.

Figure 12. Original Tank Image

Figure 13. Encrypted Image using same QG but 256

Leaders

Figure 14. Encrypted Image using different Leaders and

Quasigroups 5. Application of QGs to Hashing

5.1 Definitions A function H( ) is said to be One Way Hash Function (OWHF) if it maps an arbitrary length message M to a fixed length hash value H(M) and satisfies the following properties [17]: • No secret information is required for operation of H( )

and description of H( ) is publicly known. • It is easy to compute H(M) if M is known. • If H(M) is given in the range of H( ), it is hard to find a

message M for given H(M), and given M and H(M), it is hard to find a message M’(≠ M) such that H(M’) = H(M).

A function H( ) is said to be Collision Free Hash Function (CFHF) if it maps an arbitrary length message M to a fixed length hash value and satisfies the following property in addition to the above three properties • It is hard to find two distinct messages M and M’ that

hash to the same result (H(M) = H(M’)). 5.2 Construction of Hash Function Based on QG For a quasigroup (Q, .), a hash function HQ( ) can be defined as )....).)..((...()...( 2121 nnQ qqqaqqqH = (21) where ‘a’ is a fixed element of Q. The original multiplication table of a quasigroup may be modified using homotopy that permutes the rows and columns resulting in a different multiplication table. Qvuvuvu ∈∀= ,))().((. βαγ (22) The above three permutations are generated and used for construction of new quasigroups for hashing. It is possible to calculate the results without storing the table, therefore, large quasigroups may be used for design of hash functions suitable for cryptographic applications. Other schemes based on generation of random quasigroups with multiple lookups have been successfully used for hashing. New quasigroups are constructed using non-linear transformations involving logical OR, XOR and circular shift operations. Computationally efficient hashing schemes based on these simple operations have been designed that also ensure collision & pre-image resistance.


51

6. Design & Implementation Our design and implementation work includes the following:

(1) Generating Orthogonal Latin Square for a given Latin Square.

(2) Encryption and Decryption using Key-based Random Permutation.

(3) Generation of Isotopes of a Latin Square. (4) Construction of a huge Latin Square using the Simple

product of two Latin Squares, Affine Mapping, etc. (5) Encryption/Decryption using Latin Square and Look-

up Table method. (6) Generation of a non-associative Quasigroup with help

of Cayley’s table and T-functions. (7) Improving the basic encryption scheme to handle

redundancy in multimedia. These schemes have been implemented under the Microsoft VC++ and Matlab programming environment on a high-end stand-alone personal computer.

7. Conclusions Suitability of quasigroup based structures for encryption and hashing were established in this paper with examples, implementations and favorable observations and results. Construction of large and unstructured quasigroups is useful for design of encryption schemes with large key space. Similarly, highly non-associative quasigroups find application in the design of efficient hash functions. Modifications in the basic operations and combination with other cryptographic primitives in order to further improve the security would be taken up as our future work in this direction. References [1] R. Bose, B. Manvel, Introduction to Combinatorial

Theory, John Wiley & Sons, 1984. [2] C.F. Laywine, G.L. Mullen, Discrete Mathematics

using Latin Square, Wiley Interscience, 1998. [3] R. Alter, “How Many Latin Squares Are There?”,

American Mathematical Monthly, 82, pp. 632-634, 1975.

[4] S. Markovski, D. Gligoroski, V. Bakeva, “Quasigroup String Processing: Part 1”, Proc. of Maced. Acad. of Sci. and Arts for Math. and Tech. Sci., XX 1-2, pp. 13–28, 1999.

[5] S. Markovski, V. Kusakatov, “Quasigroup String Processing: Part 2”, Proc. of Maced. Acad. of Sci. and Arts for Math. and Tech. Sci., XXI, 1-2, pp. 15–32, 2000.

[6] S. Markovski, D. Gligoroski, S. Andova, “Using Quasigroups for One-one Secure Encoding”, Proc. VIII Conf. Logic and Computer Science, LIRA ’97, Novi Sad, pp. 157–162, 1997.

[7] S. Markovski, D. Gligoroski, B. Stojˇcevska, “Secure Two-way On-line Communication by using Quasigroup Enciphering with Almost Public key”, Novi Sad Journal of Mathematics, 30, No 2, 2000.

[8] B. Schneier, Applied Cryptography (Second Edition), John Wiley & Sons, 1996.

[9] K.A. Meyer, “A New Message Authentication Code Based on the Non-associativity of Quasigroups”,

Doctoral Dissertation, Iowa State University Ames, Iowa, 2006, [Online] Available: http://orion.math.iastate.edu/dept/thesisarchive/PHD/ KMeyerPhDSp06.pdf [Accessed: Feb. 24, 2008].

[10] C.Z. Koscielny, “Generating Quasigroups for Cryptographic Applications”, International Journal of Applied Mathematics & Computer Science, Vol. 12, No. 4, pp. 559-569, 2002.

[11] S.K. Pal, S. Kapoor, A. Arora, R. Chaudhary, J. Khurana, “Design of Strong Cryptography Schemes based on Latin Squares”, Proceedings of the Pre-ICM International Convention on Mathematical Sciences, New Delhi, 2008.

[12]S.M. Hussain, N.M. Ajlouni, “Key Based Random Permutation”, Journal of Computer Science, Vol. 2, No. 5, pp. 419-421, 2006.

[13] A. Klimov, A. Shamir, “A New Class of Invertible Mappings”, CHES, LNCS-2523, 2002.

[14] A. Klimov, A. Shamir, “Cryptographic Applications of T-functions”, Selected Areas in Cryptography, SAC-2003, LNCS-3006, Springer Verlag, pp. 248-261, 2003.

[15] A. Sade, “Quasigroupes Automorphes par le Groupe Cyclique”, Canadian Journal of Mathematics, 9, pp. 321-335, 1957.

[16] S.K. Pal, Sumitra, “Development of Efficient Algorithms for Quasigroup Generation and Encryption”, Proceedings of the 2009 IEEE International Advance Computing Conference, pp. 2529-2534, 2009.

[17] V. Snasel, A. Abraham, J. Dvorsky, P. Kromer, J. Platos, “Hash Function Based on Large Quasigroups”, International Conference on Computational Science (ICSS 2009), Lousiana, USA, LNCS 5544, pp. 521-529, 2009.

[18] S. Markovski, D. Gligoroski, V. Bakeva, “Quasigroup and Hash Functions”, Discrete Mathematics & Applications, Proceedings of the 6th ICDMA, Bansko, pp. 43-50, 2001.

[19] S. K. Pal, D. Bhardwaj, R. Kumar, V. Bhatia, “A New Cryptographic Hash Function based on Latin Squares and Non-Linear Transformation”, Proceedings of the 2009 IEEE International Advance Computing Conference, pp. 2529-2534, 2009.

Authors Profile Saibal K. Pal received the M.S. degree in Computer Science from University of Allahabad in 1990 and PhD from University of Delhi in the area of Information Security. He is presently with DRDO, Delhi. His areas of interest include Cryptography, Information Hiding, Signal Processing and Soft Computing. Shivam Kapoor is presently pursuing his Masters degree in Computer Applications (MCA) from the Department of Computer Science, University of Delhi. His areas of interest include Discrete Mathematics & Combinatorics, Cryptography and Design of Algorithms.


52

Column Vectorizing Algorithms for Support Vector Machines

Chen ZhiYuan1, Dino Isa2 and Peter Blanchfield3

1University of Nottingham, School of Computer Science,

Jalan Broga 43500 Semenyih Selangor Malaysia [email protected]

2School of Electronic Engineering, University of Nottingham,

Jalan Broga 43500 Semenyih Selangor Malaysia [email protected]

3School of Computer Science, University of Nottingham,

Nottingham, NG8 1BB, UK [email protected]

Abstract: In this paper we present the vectorization method for support vector machines in a hybrid Data Mining and Case-Based Reasoning system which incorporates a vector model to help transfer textual information to numerical vector in order to make the real world information more adapted to the data mining engine. The main issue of implementing this approach is two algorithms; the discrete vectorization algorithm and continuous vectorization algorithm. The basic idea of the vectorization algorithm is to derive X value from the original column value and where the vector value is unavailable; the algorithm builds a vector table based on the X value by using appropriate functions. Subsequently, the vector model is classified using a support vector machine and retrieved from the case based reasoning cycle using a self organizing map.

Keywords: Vectorization, Support Vector Machine, Data Mining, Artificial Intelligence, Case-Based Reasoning.

1. Introduction The problem faced by traditional database technology developer today is lack of intelligence support, while artificial intelligence techniques [1] were limited in their capacity to supply and maintain large amount of factual data. This paper provides a method to solve this problem. From a database point of view, there was an urgent need to address the problems caused by the limited intelligent capabilities of database systems, in particular relational database systems. Such limitations implied the impossibility of developing, in a pure database context, certain facilities for reasoning, problem solving, and question answering. From an artificial intelligence point of view, it was necessary to transcend the era of the operating on numerical signals to achieve the real information management system able to deal with large amounts of textual data. Our approach was explicitly designed to support efficient vectorization techniques by providing multiple number resources with minimum inter-dependencies and irregular constraints, yet under strict artificial intelligence considerations. It features a table in a relational database through two types of vectorizing functions, supporting to the

construction of the support vector machine. The rest of this paper is organized as follows: Section 2 presents objectives and related techniques. Section 3 describes in detail the architecture of the hybrid system. Section 4 provides the procedure of vectorization. Section 5 explains the conducted experiments. The conclusion is discussed in section 6.

2. Objectives and Foundation Our research group works on the designing of flexible and adaptable user oriented hybrid systems which aims to combine database technology and artificial intelligence techniques. The preprocessing procedure related to data vectorization step of a classification process, going from low level data mining processes [2] to high level artificial intelligence techniques. Many domain specific system such as user modeling systems [3] or artificial intelligence hybrid systems have been described in literature [4] [5] [6]. Even when the applied strategies are designed as generic as possible, the illustration given for the system are limited to the text document and do not develop any vectorizing algorithm to quantitate the input raw textual data set into numeric data set. Actually, to the best of our knowledge, no such complete and generic vectorization process exists because of the necessity to have an excellent know-how in the implementation of a hybrid intelligent system. Many existed systems have been developed on the basis of using artificial intelligence techniques to provide semantic support to a database system, or database techniques to aid an artificial intelligence system to deal with large amounts of information. The key factors they concerned reside in the exploitation of the equivalence between database field and the knowledge representation system of artificial intelligence. In our hybrid system, vector is the unique representation of data considering the system consistency. On the other hand, for both data mining process and case-based reasoning cycle


53

[7], vectorization and consistency are crucial. The role of vectorization is to convert text table which stored in SQL server, into numerical vector form. Traditional vectorization method concentrates on image object into a raster vector or raw line fragments. While we focus on these table column features and describe how they can be vectorized by applied automatically approach using two kinds of vectorization functions. In order to describe the foundation of the vectorization, the framework of our hybrid system is simply described in the following section.

3. Hybrid System Architecture Overview The concepts of this project are as follows: • To develop a hybrid data mining and case-based

reasoning user modeling system • To combine data mining technology and artificial

intelligence pattern classifiers as a means to construct a Knowledge Base and to link this to the case-based reasoning cycle in order to provide domain specific user relevant information to the user in a timely manner.

• To use the self organizing map [8] in the CBR cycle in order to retrieve the most relevant information for the user from the knowledge base.

Based on these concepts the architecture has been designed which is illustrated in Figure 1. The hybrid system contains five main components: • Individual models, comparable to the blackboard

containing the user information from the real world. • Domain database integrated the preselected domain

information [9]. • A data mining engine which classified both user class

and domain information vectors. • A knowledge base, containing the representation of

classified user information and combined with interested domain knowledge.

• A problem-solving life-cycle called case-based reasoning cycle, assisting in retrieve reuse revise and retain the knowledge base.

Data Mining CBR

Human expert

User ModelData mining engineSVM

Vectorization

User interface

User Model

RetrievedCase

KnowledgeBase

ConfirmedSolution

ProposedSolution

User IDQuery

SOM

RETRIEVE

RETAIN

REVISE

REUSE

Domain Database

Individual Model

Data Mining CBR

Human expert

User ModelData mining engineSVM

Vectorization

User interface

User Model

RetrievedCase

KnowledgeBase

ConfirmedSolution

ProposedSolution

User IDQuery

SOM

RETRIEVE

RETAIN

REVISE

REUSE

Domain Database

Individual Model

Figure 1. The architecture of the system

4. Vectorization As can be seen from the hybrid system architecture, in order to classify individual models and domain information into user model the support vector machine are applied. Individual models are user information which took table format and stored in the SQL server. Domain information in the database is also sorts of tables which stored the preselected user-preferred knowledge. The support vector machine [10] [11] is one of AI techniques which serve as classifier in the system. The main idea of a support vector machine is to construct a hyper plane as the decision surfaces in such a way that the margin of separation between positive and negative features is maximized. The vectorization step is the data preprocessing for the support vector machine which provides the numeric feature vector.

3.1 Feature Type For vectorization task to be as accurate as possible we predefined two type table columns or we called feature type; discrete columns (feature) and continuous columns (feature). Discrete feature contains discrete values, in that the data represents a finite, counted number of categories. The values in a discrete attribute column do not imply ordered data, even if the values are numeric; the distinct character is values are clearly separated. Telephone area code is a good example of discrete data that is numeric. Continuous feature contains values that represent a continuous set of numeric and measurement data, and it is possible for the data to contain an infinite number of fractional values. An income column is an example of a continuous column. The numeric value is not the vital factor to determine the feature type, but if the value is a word then it must be a discrete feature.

3.2 Vectorization algorithm From the technology point of view, vectorization is an approach modeling relationships between the data set and the vectorizing variable. We provide a more flexible approach by allowing some of the features (columns) to be independent and some of the features to be interdependent. Constructing two parallel algorithms to avoid time consuming and save a large amount of effort. The schema of the algorithm is specified in Figure 2 which derives the numeric vector by implementing different functions. The schema is not exhaustive and can evolve with new data, according to user need. Furthermore, once the type of the column has been determined, adding a new record is quite straightforward. These functions are also well suited to dealing with incomplete data. Instances with missing attributes can be handled by summing or integrating the values of other attribute. We represent each column as a data point in a dimensional space, where Z is the total number of attributes (columns). The algorithm computes the vectorizing value (or representation value) between each feature which was denoted by abscissa axis and the vector denoted by y-axis, and all the feature values determine its own vectorizing values. Once the vectorizing value list is obtained, the vector model will be classified based on the implementation of


54

support vector machine so that the core of the hybrid system the knowledge base will be constructed completely.

Figure 2. The schema of the vectorization algorithm

The detailed vectorization algorithms are described in the Table 1 and Table 2 according to discrete columns and continuous columns.

Table 1. The discrete column vectorization algorithm

1: Let V be the representation of Vectors, D be the whole set of the vector model and d be the set of discrete columns.

2: FOR each data point Z DO 3: Select dZ , the discrete features of all data point,

4: Compute dV = ( dxV , dyV ), the corresponding value

between Z and every vector, ( dxV , dyV ) D.

5: ddx nV = , dyV = dnn ×1 ; ]1,0[∈dyV .

6: END FOR

Table 2. The continuous column vectorization algorithm

1: Let V be the representation of Vectors, D be the set of vector model and c be the set of continuous columns.

2: FOR each data point Z DO 3: Select cZ , the continuous features of all data point,

4: Compute cV = ( ′cxV , ′

cyV ), the corresponding value

between Z and every vector, ( ′cxV , ′

cyV ) D.

5: ′cxV = cxcxcx MaxVAvgVV ）（ − ,

xx

xx

cy eeeeV −

−

+−=′ , ].1,1[ +−∈′cyV

6: END FOR

The key computation of these two algorithms is the vectorization value formula given in step 5 of the both table. Formula 1:

ddx nV =

dyV = dnn ×1

Formula 2: ′

cxV =cx

cxcx MaxVAvgVV )( −

xx

xx

cy eeeeV −

−

+−=′ , ].1,1[ +−∈′cyV

In Formula 1, n is the weight parameter associated with the discrete columns which is the sum of value type. dyV is a

combination of the unit value ( n/1 ) multiply the sequence

of the current value type ( dn ). This is a regression-like expression [12]. Regression is used to make predictions for numerical targets. By far the most widely used approach for numerical prediction is regression, a statistical methodology that was developed by Sir Frances Galeton [13]. Generally speaking Regression analysis methods include Linear Regression, Nonlinear Regression. Linear Regression is widely used, owing largely to its simplicity. By applying transformations to the variables, we can convert the nonlinear model (text table column information) into a linear one according to the requirement of the support vector machine. In order to get the negative X value and at the same time keep the same distance among original X value, in Formula 2 we minus average value to all x value and then get the proportion compare with the maximum original X value, after that get the new X value and by means of Hyperbolic Tangent function [14] to map these new value into (-1, +1) scale. In order to explain these algorithms clearly, we show the experiment procedure in the following section.

5. Experiments The vectorization algorithm was tested on the census-income data set extracted from the 1994 and 1995 current population surveys conducted by the U.S. Census Bureau. The data contains 41 demographic and employment related variables. In order to explain how to apply our approach clearly, we choose 8 discrete columns and 8 continuous columns which can be found in table 3 to explain the implementation in details. In Table 4 we list the n value of the discrete columns. For example the worker class n value, because there are 9 kinds of worker class, so n is equal to 9. Parts of the experiment results implemented the proposed algorithms which contain 27 records are shown in figure 3. The input for the algorithm was given 8 discrete features and 8 features and asked to give the vectorized value as output. The discrete attributes were decomposed into n equidistances, which yielded corresponded vector value scaling to the range of (0, 1). For the continuous attribute, firstly the raw attribute value was transferred into the whole x-axis, so that the new x value contain the negative value


55

and by using Hyperbolic Tangent function the vector value was calculated. The Hyperbolic Tangent function make sure the vector value to be projected into the (-1, 1) scale which is required by support vector machine.

Figure 3. Part of the experiment results T The recommend vector value range is (0, 1) or (-1, 1) for support vector machine [15]. One reason for this is to avoid vector value in great numeric ranges dominates those in smaller numeric ranges. Another reason is to avoid the numerical difficulties during the calculation. Because kernel values usually depends on the inner products of feature vectors. For example the linear kernel and the polynomial kernel, large vector values may cause numerical problems [16]. Another reason why we proposed two kinds of algorithm to vectorize discrete columns and continuous columns is to preserve the character of the column for the sake of the later analysis.

Table 3. Parameters for experiments

Discrete columns Continuous columns class of worker* age* education* wage per hour marital stat capital gains sex capital losses reason for unemployment dividends from stocks family members under 18 person for employer live in this house 1 year ago weeks worked in year veterans benefits instance weight*

Table 4. The discrete column n value

Discrete columns n Value class of worker 9 education 17 marital stat 7 sex 2 reason for unemployment 6 family members under 18 5 live in this house 1 year ago 3 Veterans benefits 3

All the experiment results was created on PC computer, CPU Intel(R) Core(TM) Duo CPU T2250 @ 1.73GHz 4.6 2.3, 2GB RAM DDR2 667 MHz, with WinXP. Program was compiled with NetBeans 6.0.

6. Conclusions The proposed hybrid Data Mining and Case-Based Reasoning User modeling system is a multi purpose platform and is characterized by three major processes. The vectorization processing unit communicate through the raw data set the SQL table and the output is the numeric vector, such an approach avoid the data inconsistency usually met in classifying documents chain when implement artificial intelligence tools. In this paper we built vectorization model by applying two algorithms: The discrete vectorization algorithm and continuous vectorization algorithm. The advantage of using discrete algorithm is that each record in the whole table was assigned a vector value in an easily expression calculation. While for the continuous column we choose a relatively complicated formula that is the Hyperbolic Tangent function to achieve the vector value. In designing the algorithm, the key consideration is to bring up easy scientific numerical transformation. Therefore, the formulas in the algorithm are quite basic but the impressive part is it also provides a reasonable balance between a satisfactory result and reasonable processing time. Secondly due to the modular structure of the algorithm it can be adapted easily for application. The results of the algorithm in the experiments labeled clean and the vector points generated by our algorithm have a standard coverage (0, 1) and (-1, 1) which is useful in fulfilling the classification task by means of support vector machine for the hybrid system.

References [1] S. J. Russell, P. Norvig, Artificial Intelligence A

Modern Approach, Prentice-Hall International Inc, 1995.

[2] U. Fayyad, G. Paitetsky-Shapiro, P. Smith, “knowledge discovery and data mining: Towards a unifying framework”, proceedings of the International Conference on Knowledge Discovery and Data Mining, 1996, pp. 82-22.

[3] J. Vassileva, "A practical architecture for user modeling in a hypermedia-based information system", Proceedings of Fourth International Conference on User Modeling, Hyannis, MA, August 1994, pp 15-19.

[4] I.V. Chepegin, L. Aroyo, P. D. Bra, “Ontology-driven User Modeling for Modular User Adaptive Systems”, LWA, 2004, pp.17-19.

[5] I. Watson, Applying Case-Based Reasoning: Techniques for Enterprise Systems, Morgan Kaufmann Publishers, Inc., San Francisco, CA, 1997.

[6] K. Sycara, “CADET: A cased-based synthesis tool for engineering design”, International Journal for Expert System, 4(2), 1992, pp.157-188.


56

[7] A. Aamodt, E. Plaza, “Case-based reasoning: foundational issues, Methodological variations, and system approaches”, AI communications, 7(1), 1994, pp. 39-59.

[8] Kohonen, “self-organizing map using contiguity-constrained clustering”, Pattern Recognition Letters, 1995, pp. 399–408.

[9] B. Hjorland, H. Albrechtsen, “Toward A New Horizon in Information Science: Domain Analysis”, Journal of the American Society for Information Science, 1995, 46(6), 400-425.

[10] E. Osuna, “Support Vector Machines: Training and Applications”, Ph.D thesis, Operations Research Center, MIT, 1998.

[11] V.N. Vapink, Statistical Learning Theory, New York:Wiley.

[12] D.V. Lindley, "Regression and correlation analysis," New Palgrave: A Dictionary of Economics, v. 4, 1987, pp. 120-23.

[13] F. Galeton,“Typical laws of heredity", Nature 15,1877, pp. 492-495, 512-514, 532-533.

[14] M.A. Abdou, A.A. Soliman, “Modified extended tanh-function method and its application on nonlinear physical equations”, Physics Letters A, Volume 353, Issue 6, 15 May 2006, pp. 487-492

[15] E. Osuna, R. Freund, F. Girosi, “Improved training algorithm for support vector machine”, IEEE Neural Networks in Signal Processing 97,1997.

[16] C. Cortes, V. Vapnik, “Support-vector network”, Machine Learning , 1995, pp. 273–297.

Authors Profile

Chen ZhiYuan received the B.A. in Economics from University of HeiLongJiang in 2001 (China). During 2006-2010, she stayed in University of Nottingham, Malaysia Campus to do PhD research in imitate human experts (especially in manufacturing and medical field) to perceive the environment and to make

decisions which maximize the chance of success. From 2007 to 2009, she stayed in Supercapacitor Research Laboratory (SRL), which is supported by Ministry of Science Technology and Inovation of Malaysia to study knowledge management system for manufacturing enviroment.

Dino Isa is a Professor in the Department of Electrical Electronics Engineering, University of Nottingham Malaysian Campus. He obtained a BSEE (Hons) from the University of Tennessee, USA in 1986 and a PhD from the University of Nottingham, University Park Nottingham,

UK in 1991.nnnThe main aim of his research is to formulate strategies which lead to the successful implementations of “Intelligent Systems” in various domains.

Peter Blanchfield is a senior tutor in the School of Computer Science, University of Nottingham. From September 2005 to July 2009 he was Director of the IT Institute in the School, before which he was the Director of Computer Science and IT Division at the Malaysia Campus of the

University of Nottingham. In that role he was involved in setting up the activities of the School there along with the activities of what has become the Engineering Faculty on that campus.


57

Impact and Performance Analysis of Mobility Models on Stressful Mobile WiMax Environments

N Vetrivelan1 Dr. A V Reddy2

1Research Scholar, NIT, Trichy, India,caz0307@ nitt.edu

Assistant Professor, Department of Computer Applications Periyar Maniammai University, Thanjavur, Tamilnadu, India

[email protected] 2Professor, Department of Computer Applications

National Institute of Technology, Trichy, India [email protected]

Abstract: In this study, we have used the design and implementation of a new medium access control protocol with WiMax, based on the IEEE 802.16. We have compared and analysed three different mobility models and the impact of mobility on mobile WiMax environments. We have chosen Gauss-Markov, Manhattan Grid and Random WayPoint Mobility models with DSR routing protocol in WiMax environments. The parameter metrics Packet Delivery Fraction, Routing load, Throughput and Latency have been taken into account. Our ns-2 simulation result shows that the functioning of mobility models will greatly influence the performance of WiMax environments. The result reveals that the throughput, latency and routing load are high in Gauss Markov Mobility model. It also shows that packet delivery fraction is high in Manhattan Grid model. Compared to Manhattan and Gauss Markov models the random waypoint models is stable. Keywords: Mobility, DSR, WiMax, MAC and Simulation

1. Introduction

WiMax [3] is the short form of the Worldwide Interoperability for Microwave Access. Typically, fixed WiMax networks have a higher-gain directional antenna installed near the client which results in greatly increased range and throughput. Mobile WiMax networks are usually made of indoor customer premises equipments (CPE) such as desktop modems, compared to directional antennas but they are more portable. The mobility model [1] is designed to describe the movement pattern of mobile users, and how their location, velocity and acceleration change over time. Since mobility patterns plays a significant role in determining the protocol performance, it is desirable for mobility models to emulate the movement pattern of targeted real life applications in a reasonable way. We have provided a categorization for various mobility models onto several classes based on their specific mobility characteristics. For some mobility models, the movement of the WiMax node is likely to be affected by its movement history. The authors are aware that this performance comparison of mobility scenarios has not attempted in WiMax Environments or IEEE 802.16 module. That is why, the performance of mobility scenarios using DSR

wireless Routing protocol in WiMax module has been chosen and compared.

In Section 2, the related works have been discussed. Brief description of Mobile node and WiMax Module has been presented in Section 3. The Mobility models described in 4. Protocol description has been given in Section 5. The evaluation methodologies have been given in Section 6. In Section 7, the Simulation Parameters and Parameter Values of WiMax V2.03 have been described. Results and Discussion presented in Section 8. The conclusion has been presented in Section 9.

2. Related Work

Tracy Camp, Jeff Boleng [1] surveyed the mobility models that are used in the simulations of Ad hoc networks. Authors described several mobility models that represent mobile nodes whose movements are independent of each other (i.e, entity mobility models) and several mobility models that represent mobile nodes whose movements are dependent on each other ( i.e. group mobility models.) This paper presents a number of mobility models in order to offer researchers more informed choices when they are deciding upon a mobility model to use in their performance evaluations. Illustrated how the performance results of an ad hoc network protocol drastically change as a result of changing the mobility model simulated.

Per Johansson, Tony Larsson compared three routing protocols for wireless mobile ad hoc network. They have done simulation on a scenario where nodes move randomly. Furthermore, three realistic scenarios were introduced to test the protocols in more specialized contexts. In most simulations the reactive protocols (AODV,DSR) performed significantly better than DSDV. At moderate traffic load DSR performed better than AODV for all tested mobility values, while AODV performed better than DSR at higher traffic loads.

Leonardo Betancur [5] performed experiments for WiMax Channel-Phy Model in ns-2. Their work has been on a novel proposal for PHY layer and propagation model that allowed faster and more detailed the link level execution. They described the development process for model and the


58

implementation details in ns-2 for PMP topologies. They described the development of physical layer model based on

the IEEE 802.16 standard also known as WiMax using ns-2. Their work presented a statistical equivalent model for ns-2, which consider the channel effects in WiMax networks for upper layer simulations. Through simulation they reached the conclusion that their model reduced computational effort in link level simulations, without using complex bit-bit simulations also it is necessary to analyze the performance of WiMax networks and found more realistic coverage areas.

Jenhui Chen, Chih-Chieh Wang, [4] presented detailed design and implementation of WiMax module based on the IEEE 802.16 broadband wireless access networks (BWANs) or WiMax module with the point-to-multipoint (PMP) mode for the ns-2. They implemented modules comprised fundamental functions of the service-specific convergence sub layer (CS), the MAC Common part sub layer (CPS), and the PHY layer. A simple call admission control (CAC) mechanism and the scheduler have also been included in this module.

Figure 1. Conversion of Mobile node in to WiMax node

Network Interface

Propagation Model

Queue

Link Layer ARP

Channel

Application Routing Agent

PKT Sent

MAC decides OK ?

Not ok ok

Traffic Generating Agent

Link Layer

Queue

PROPOSED WIMAX

Network Interface Radio Propagation Model

Channel


59

3. Conversion of Mobile Node in to WiMax Node Each mobile node has made use of a routing agent for the purpose of calculating routes to other nodes in the ad-hoc network. Packets were sent from the application and were received by the routing agent. The agent decided a path that the packet must travel in order to reach its destination and stamped it with this information. It then sent the packet down to the link layer. The link layer level used an Address Resolution Protocol to decide the hardware addresses of neighboring nodes and map IP addresses to their correct interfaces. When this information was known, the packet was sent down to the interface queue and awaited a signal from the Multiple Access Control (MAC) protocol. When the MAC layer decided it was ok to send it on to the channel. The propagation model used the transmit and received stamps to determine the power with which the interface would receive the packet. A mobile node has been converted to WiMax node as discussed subsequently. The 802.16 based WiMax module [3][4] consisted of Mac 802_16, Common and Queue have been in accordance with the specifications of the IEEE 802.16-2004 standard and based on the ns-2 version 2.29. An Object oriented programming language C++ were developed for classes. The relationship between WiMax module and ns-2 modules was represented in the stack of the ns-2 is as shown in Fig.1. It consists of the type of objects for the traffic generating agent (TGA), the link layer (LL), the interface queue (IFQ), the designed MAC layer (WiMax module), and the PHY layer (Channel). The implemented module comprised fundamental function of the service-specific convergence sublayer (CS), the MAC Common part sublayer (CPS) and the PHY layer. A simple call admission control mechanism and the scheduler have also been included in this module

4. Mobility Models

4.1 Gauss Markov Model This model [1] was designed to adapt to different levels of randomness. Initially each node is assigned a current speed and direction. At fixed intervals of time, movement occurs by updating the speed and direction of each node. Specifically, the value of speed and direction at the nth instance is calculated based upon the value of speed and direction at the (n-1)th instance. The main advantages of this model are that it eliminates the sudden stops, sharp turns present in Random way point mobility model and is close to being realistic. 4.2 Manhattan Mobility Model The Manhattan model [1] is used to emulate the movement pattern of mobile nodes on streets defined by maps. The Maps are used in this model too. The map is composed of a number of horizontal and vertical streets. Each street has

two lanes for each direction (North and South direction for vertical streets, East and West for horizontal streets). The WiMax node is allowed to move along the grid of horizontal and vertical streets on the map. At an intersection of a horizontal and vertical street, the WiMax node can turn left, right or go straight. This choice is probabilistic: the probability of moving on the same street is 0.5, the probability of turning left is 0.25 and probability of turning right is 0.25. The velocity of a mobile node at a time slot is dependent on its velocity at the previous time slot. Also, a node’s velocity is restricted by the velocity of the node preceding it on the same lane of the street. 4.3 Random Way Point The Random Waypoint Mobility model includes pause times between changes in direction and/or speed. An WiMax [3] node begins by staying in one location for a certain period of time (i.e pause time) Once this time expires, the WiMax node chooses a random destination in the simulation area and speed that is uniformly distributed between minimum and maximum speed. The node then travels toward the newly chosen destination at the selected speed. Upon arrival, the node pauses for a specified time period before starting the process again. This is a memory-less mobility pattern because it retains no knowledge concerning its past locations and speed values. The current speed and direction of a node is independent of its past speed and direction. This characteristic can generate unrealistic movements such as sudden stops and sharp turns. 5. Protocol description 5.1 Dynamic Source Routing The key distinguishing feature of DSR [10] is the use of source routing. That is, the sender knows the complete hop-by-hop route to the destination. These routes are stored in a route cache. The data packets carry the source route in the packet header. When a node in the ad hoc network attempts to send a data packet to a destination for which it does not already know the route, it uses a route discovery process to dynamically determine such a route. Route discovery works by flooding the network with route request (RREQ) packets. Each node receiving an RREQ rebroadcasts it, unless it is the destination or it has a route to the destination in its route cache. Such a node replies to the RREQ with a route reply (RREP) packet that is routed back to the original source. RREQ and RREP packets are also source routed. The RREQ builds up the path traversed across the network. The RREP routes itself back to the source by traversing this path backward. The route carried back by the RREP packet is cached at the source for future use. If any link on a source route is broken, the source node is notified using a route error (RERR) packet. The source removes any route using this link form its cache. A new route discovery process must be initiated by the source if this route is still needed. DSR makes very aggressive use of source routing and route caching.


60

6. Evaluation Methodology To evaluate the three mobility models in WiMax, we used four performance metrics to compare and analyse the realistic movements

6.1 Packet Delivery Fraction

The ratio of number of data packets successfully delivered to the destination, generated by CBR Sources.

PDF=(Received Packets/Sent Packets)*100

6.2 Routing Overhead

It is an important metric for measuring scalability of a protocol. The number of routing packet transmitted per data packet delivered at destination. Each hop wise transmission of a routing packet is counted as one transmission.

Routing load = Packets sent/Received packet

6.3 Throughput

These are the number of packets sent from the source and the number of packets received at the destination

6.4 Latency

The time, it takes for a packet to cross a network connection from sender to receiver

7. Simulation parameters

Table 1: Simulation Parameter Values Routing Protocol DSR

MAC Layer IEEE 802.16 Number of Mobile

WiMax Nodes 100,200,300,400

Mobility Model Gauss Markov

Manhattan Grid Random Waypoint

Transmission Bandwidth of each

line 20

Simulation time 300Secs Traffic Type Constant Bit Rate

Antenna Type Omni Antenna

Table 2 : Important Parameter Values of WiMax V2.03

Bandwidth 20MHz Service flow Scheduling Type 11.13.11

SS_MAC_Address 48 Transaction_ID 16

Downlink frequency 32 Uplink channel_ID 8

7.1 Simulation Model

In this section, the network simulation was implemented using the NS-2 simulation tool. The Network Simulator NS-2 was a discrete event simulator. For simulation Scenario and network topology creation it used OTCL (Object Tool Command Language). To create new objects, protocols and routing algorithm or to modify them in NS-2, C++ source code used. The WiMax module consisted of the type of objects for the traffic generating agent (TGA), the link layer (LL), the interface queue (IFQ), the designed MAC layer (WiMax module), and the PHY layer. The simulations were conducted on Due Core processor at speed 3.0 GHz, 1 GB RAM running Gygwin Environment.

Simulation Mobility Model

Table 3 : Simulation Parameter of Mobility Models

Gauss Markov Manhattan Grid

Random waypoint

x=1000.0 x=1000.0 x=1000.0 y=1000.0 y=1000.0 y=1000.0 Duration =300.0

Duration =300.0

Duration =300.0

Update Frequency=2.

5

Update Dist=5.0 Dim=3

Maxspeed=4.0

TurnProb =0.5

Minspeed =1.0

AngleStdDev 0.39269909

SpeedChangeProb=0.2

Maxspeed =4.0

SpeedStdDev =0.5

MinSpeed =1.0

Maxpause =60.0

MeanSpeed =1.0

SpeedStd Dev=0.2

8. Results and Discussion

We have used ns-2 with WiMax to compare the performance of the mobility models. Four sets of results have been presented vide 100,200,300 and 400 WiMax nodes. In all stressful situations with communication channel 2,4,6,8 and 10 have been chosen. The routes of packets are accomplished with the DSR.


61

Figure 2. Mobility Models – PDF (%) with varied communication channel in 100 WiMax Environments



Figure 5. Mobility Models - PDF (%) with varied communication channel in 400 WiMax Environments

As shown in the Figures 2,3,4,5, with respect to the PDF, of all 100,200,300 and 400 WiMax nodes situation Manhattan Grid Model out performs where as in the Random waypoint there is decreasing trend. In the Manhattan Grid Model the nodes have been chosen at a random destination either in the horizontal or vertical directions. Due to the more restricted movement of the nodes in the network which leads to slightly lesser number of broken links and subsequently lower the chance of getting a stray route error message. At the same time it is found out that Gauss Markov model shows steadiness in performance.

Figure 6. Mobility Models - Routing Load (packets) with

varied Communication channel in 100 WiMax Environments

Figure 7. Mobility Models - Routing Load (packets) with varied communication channel in 200 WiMax Environments

100 Nodes

0

10

20

30

40

50

2 4 6 8 10Communication Channel

PDF

GAUSS-MARKOVMANHATTAN GRIDRANDOM WAY POINT

200 Nodes

0

5

10

15

20


PDF

GAUSS-MARKOV MANHATTAN GRIDRANDOM WAY POINT

300 Nodes

0

5

10

15

20

2 4 6 8 10Communication channel

PDF


0

10

20

30

2 4 6 8 10

Routing Load

Communication Channel

100 Nodes


0

10

20

30

2 4 6 8 10

PDF

Communication Channel

400 Nodes


200 Nodes

05

101520253035


Rou

ting

Load



62

Figure 8. Mobility Models - Routing Load (packets) with varied communication channel in 300 WiMax

Environments

Figure 9. Mobility Models - Routing Load (packets) with varied communication channel in 400 WiMax

Environment As shown in the Figures 6,7,8,9, in all the four WiMax situations, as far as the routing load is concerned the Manhattan model outperforms compared to other two models. It is more likely for the route error messages to be dropped by some intermediate nodes before they actually reach the intended source node. So, the source node, which is unaware of the loss of route error messages, attributes any genuine timeouts due to broken links. This mobility model is choosing random destination either in the horizontal or vertical directions and also more restricted movement of the nodes.

Figure 10. Mobility Models – Throughput (packet) with varied Communication channel in 100 WiMax

Environments

Figure 11. Mobility Models - Throughput (packets) with varied communication channel in 200 WiMax

Environments

Figure 12. Mobility Models – Throughput (packets) with varied communication channel in 300 WiMax

Environments

Figure 13. Mobility Models – Throughput (packets) with varied Communication channel in 400 WiMax

Environments

As shown in the Figures 10,11,12,13, in all 100,200,300 and 400 WiMax nodes Gauss Markov model gives high throughput whereas the Random waypoint and Manhattan models perform with slight decreasing trend compared to Gauss Markov model. This is because the Gauss Markov Mobility Model can eliminate the sudden stops and sharp turns encountered. This Gauss Markov is a more realistic mobility model when compared with the Random Waypoint model. Because of this the chances of getting unrelated route error messages is comparatively less in the case of Gauss Markov model.

300 Nodes

0

10

20

30

40

50

2 4 6 8 10Comminucation Channel

Rou

ting

Load


400 Nodes

0

20

40

60

80


Rou

ting

Load


100 Nodes

0

100

200

300

400

500

600

700


Thro

ughp

ut


200 Nodes

0

100

200

300

400

500

600

700


Thro

ughp

ut


300 Nodes

0100200300400500600700


Thro

ughp

ut


400 Nodes

0100200300400500600700


Thro

ughp

ut



63

Figure 14. Mobility Models-Latency (time) with varied communication channel in 100 WiMax Environments

Figure 15. Mobility Models- Latency (time) with varied communication channel in 200 WiMax Environments



As shown in the Figures 14,15,16,17, in all 100,200,300 and 400 nodes Gauss Markov model gives high latency whereas the Random waypoint and Manhattan models perform with slight decreasing trend compared to Gauss Markov model. This Gauss Markov is a more realistic mobility model when compared with the Random Waypoint model. Because of this, the chances of getting unrelated route error messages is lower in the case of Gauss Markov model.

9. Conclusion In this paper, the three mobility models have been taken up and their performance have been compared with different stressful WiMax environments with ns-2. The performance metric of PDF, Routing load, Throughput and Latency for DSR protocols under the different simulation environment with varying communication channel has been computed. As a result, through simulation, it reveals that the throughput, latency and routing load are high in Gauss Markov Mobility model. This is due to a more realistic mobility model. Because of this the chances of getting unrelated route error messages is lower in the case of Gauss Markov Model when compared Random Waypoint mobility Model. It also shows that packet delivery fraction is high in Manhattan Grid model. This is due to the more restricted movement of the nodes in the network, which leads to slightly lesser number of broken links and subsequently lowering the chances of getting a stray route error message. Compared to Manhattan and Gauss Markov models the Random waypoint model is stable. References

[1] T.Camp, J.Boleng, V.Davies, “ A Survey of Mobility

Models for Ad Hoc Network Research”, in Wireless Communication and (WCMC): Special issue on Mobile Ad Hoc Networking: Research, Trends and Appications, Vol.2, no 5, pp. 483-502, 2002.

[2] C. Bettsltter, G. Resta, P.Santi, “ The Node Distribution of the Random Waypoint Mobility Model for Wireless Ad Hoc Networks”, IEEE Transactions on Mobile Computing, July - September 2003, pp 257-269

[3] Chen, J, Wang, C., Tsai, F., Chang, C., Liu, S.,Guo,J., Lien, W., Sum, J., Hung, C., “The Design and Implementation of WiMax Module for ns-2 Simulator”, ACM Valuetools 2006,Pisa,Italy,ACM Press, New York (2006)

[4] Jenhui Chen, Chih-Chieh Wang,“The Design and Implementation of WiMax Module for ns-2 Simulator”,WNS2-06.

[5] Cicconetti,C., Erta, A., Lenzini,L.,Mingozzi, E. “Performance Evaluation of the Mesh Election Procedure of IEEE 802.16/WiMax”, MSWiM’07,October 22-26,2007,Chania

100 Nodes

0

0.005

0.01

0.015

0.02


Late

ncy


200 Nodes

00.005

0.010.015

0.02


Late

ncy


300 Nodes

00.005

0.010.015

0.020.025


Late

ncy


400 Nodes

0

0.02

0.04

0.06


GAUSS-M ARKOVM ANHATTAN GRIDRANDOM WAY P OINT


64

[6] Leonardo Betancur, Roberto C. Hincaple, Roberto Bustamante, “WiMax Channel-Phy Model in Network Simulator 2”, Oct 10, 2006.

[7] Shiang Ming Huang, “NCTUns simulation tool for WiMax modeling”, WICON 2007, Oct 22-24, 2007.

[8] Bohnert, T.M.,Jakubiak, J.,Katz, M.,Koucheryavy, Y.,Monteiro,E., Borcoci,E, “On Evaluating a WiMax Access Network for Isolated Reserch and Data Networks Using NS-2”, Springer-Verlag Berlin Heidelberg 2007.

[9] Cao,M., Raghunathan,V and Kumar, P.R A tractable algorithm for fair and efficient uplink scheduling of multi-hop Wimax mesh networks. Proc. WiMesh 2006, Reston(VA),USA, Sep.25,2006,pp 101-108.

[10] Elizabeth M.Royer and C.K Toh “A Review of Current Routing Protocols for Ad Hoc Mobile Wireless Networks,” IEEE Personeal Communications, April 1999,pp 46-55.

[11] S.Azad,A Rahman and F.Anwar,”A Performance Comparison of Proactive and Reactive Routing Protocols of Mobile Ad-hoc NET work(MANET),” Journal of Engineering and Applied Sciences 2(5),2007, pp 891-896.

[12] Perkins C.E and Royer.E.M, “Ad Hoc On-demand Distance Vector Routing” In Proceedings of the 2nd IEEE Workshop on Mobile Computing Systems and Applications,New Orleans,LA,February 1999,pp.46-55

[13] Network Simulator 2 (NS-2), http://www.isi.edu/nsnam/ns/

[14] NDSL WiMax Module for ns2 Simulator, http://nds1.csie.cgu.edu.tw/wimax_ns2.php

[15] The WiMax forum.available at http://www.wimaxforum.org/home/.

[16] The WiMax module for ns-2.available at http://ndsl.csie.cgu.edu.tw/wimax_ns2.php.

[17] BonMotion Scenario Generation. www.informatik.unibonn.de/IV /BonnMotion/

Authors Profile

N Vetrivelan Research Scholar, Department of Computer Applications, National Institute of Technolgoy, Trichy. India. Working as an Assistant Professor in Periyar Maniammai University, Thanjavur, India With 15 Years of Engineering Collegiate experience. Presented five International Papers and published two papers in the refereed

journals. Presented paper in IAENG International Conference at Hong Kong. Established WiMax Broad Band connection with Six rural villages in and around Thanjavur district under PURA.

Dr. A V Reddy received Ph.D in II.Sc Bangalore, India. Working as a Professor, Department of Computer Applications, National Institute of Technology, Trichy, Tamilnadu, India with 25 Years of Academic experience. Published Six International journal Papers in refereed journal and also ten International papers presented.


65

Partial Aggregation for Multidimensional Online Analytical Processing Structures

Naeem Akhtar Khan1 and Abdul Aziz2

1Faculty of Information Technology,

University of Central Punjab, Lahore, Pakistan [email protected]

2Faculty of Information Technology,

University of Central Punjab, Lahore, Pakistan [email protected]

Abstract: Partial pre-computation for OLAP (On-Line-Analytic Processing) databases has become an important research area in recent years. Partial pre-aggregation is implemented to speed up the response time of queries that are posed for the array-like decision support interface, subject to the different constraint that all pre-computed aggregates must fit into storage of a pre-determined calculated size. The target query workload contains all base and aggregate cells that are stored in a multidimensional structure (i.e. cube). These queries are in fact range queries pre-defined by users for the support of decision makers. The query workload of an OLAP scenario is the set of queries expected by the users. Most of the published research only deals with the optimization for the workload of views in the context of ROLAP (Relational OLAP). Many researchers have criticized partial-computation schemes, optimized for views that lack of support to ad-hoc querying. The other main aspect is that a view may be too large for pre-computation that calculate very small answers. In this paper, we study the problems of partial pre-computation for point queries, which are best for MOLAP (Multidimensional OLAP) environment. We introduce multidimensional approach for efficiency of the cover-based query processing and the effectiveness of PC Cubes. Keywords: OLAP, MOLAP, Data-Cubes, Data warehouse.

1. Introduction A data warehouse (DW) is centralized repository of summarized data with the main purpose of exploring the relationship between independent, dimensions, static variables and dependent, dynamic, variables facts or measures. There is a trend within the data warehousing community towards the separation of the requirements for preparation and storage necessary for analyzing the accumulated data and the requirements for the exploration of the data with the necessary tools and functionality required [1]. In terms of the storage necessities, a convergent tendency is towards a multi-dimensional hypercube model [2]. On the other hand in terms of analysis and the tools required for On-Line Analytic Processing (OLAP), there is a trend towards standardizing this as well; e.g., the efficient OLAP Council’s Multi-Dimensional Application Programmers Interface (MD-API). Although the trends are for separating the storage from the analysis, the actual physical implementation of DW/OLAP systems reconnects them. This is an evident from the parade of acronyms used today, e.g., MOLAP, ROLAP, DOLAP,

HOLAP, etc., where all physical implementation determines the advantages and disadvantages of storage access an analysis capabilities and also determines any possible extensions in future to the model. In the models quoted above, the two most common in practice are the Multidimensional On-line Analytic Processing (MOLAP) model and the Relational On-line Analytic Processing (ROLAP) model. The main advantage of ROLAP, which depends on relational database (RDB) technology, is that the database technology is well standardized (e.g., SQL2) and is readily available too. This permits for the implementation of a physical system, based on readily available technology and open standards. As this technology is well studied and researched, there are mechanisms which allow for transactions and authorization schemes, thus allowing for multi-user systems with the ability to update the data as required. The main disadvantage of this technology is that the query language as it exists (SQL) is not so sufficiently powerful or flexible enough to support true OLAP features [1]. Furthermore, there is an impedance difficulty, that the results returned, tables, always required to be converted to another form before further programming abilities can be performed. The main advantage of MOLAP, which depends on generally proprietary multi-dimensional (MDD) database technology, is based on the disadvantages of ROLAP and is the major reason for its creation. MOLAP queries are quite powerful and flexible in terms of OLAP processing. The physical model further closely matches the multidimensional model, and the impedance issue is remedied within a vendor’s side. However, there are disadvantages of the MOLAP physical model: 1st) There is no real standard for MOLAP; 2nd) there are no off-the-shelf MDD databases per se; 3rd) there are scalability problems; and 4th) there are problems with authorizations and transactions. As the physical implementation ultimately determines the abilities of the system, it would be advised to find a technology that combines and maximizes the advantages of both ROLAP and MOLAP while at the same time minimizing the dis-advantages. Online Analytical Processing (OLAP) has become a basic component of modern decision support systems. As claimed in [3] introduced the data-cube, a relational operator as well as model used for computing summary views of data that can, in turn, significantly improve the response time of core OLAP operations such as roll-up, drill down, and slice and


66

dice, drill down approach is reflected in Figure 1. Typically developed on top of relational data warehouses these summary views are formed by aggregating values across different attribute combinations. For a d-dimensional input set R, there are 2d probable group bys. A data-cube as well as a lattice which is often used for representing the inherent relationships between group-bys [5], there are two standard data-cube representations: ROLAP (set of relational tables) and MOLAP (multi-dimensional OLAP). The main differences between ROLAP and MOLAP architectures are shown in Table 1 [4].

Figure 1. Hierarchies in data

The array-based structure, MOLAP (Multi-dimensional OLAP), has the advantage that native arrays provide an immediate form of indexing for queries of cube. Research has shown, however, that MOLAP has scalability problems [6]. For example, high-dimension data-cubes represent tremendously sparse spaces that are not easily adapted to the MOLAP model. Hybrid indexing schemes are normally used, significantly diminishing the power of the model.

Table 1: ROLAP Vs MOLAP

Feature ROLAP MOLAP

Usage Variable performance Good performance

Relational engine Multidimensional engine Storage and Access Tables/tuples Proprietary arrays

SQL access language

Lack of a standard language

Third party tools Sparse data compression Database Size Easy updating Difficult updating

Large space for indexes 2% index space

Gigabyte-Terabyte Gigabyte

Moreover, since MOLAP requires to be integrated with standard relational databases, middleware of some form must be employed for handling the conversion between relational and array-based data representations. The efficiency of relational model, ROLAP (Relational OLAP), does not suffer by such restrictions. In standard relational tables its summary records are stored directly without any need for data conversion. ROLAP table based data representation does not pose scalability problems. Yet, many current commercial well-known systems use the MOLAP

approaches. The main issue, as outlined in [6] is the indexing problem for the fastest execution of OLAP queries. The main problem for ROLAP is that it does not offer an immediate and fast index for OLAP queries. Many well-known vendors have chosen the sacrifice scalability for performance. Query performance issue is discussed for ROLAP and proposed a novel, distributed multi-dimensional ROLAP efficient indexing scheme [7]. They showed that the ROLAP advantage for high scalability can be maintained, while at the same time providing a rapid index for OLAP queries. They proposed a distributed indexing efficient scheme which is a combination of packed R-trees with distributed disk striping and Hilbert curve based data ordering. Their method requires very meager communication volume between processors and works in very low bandwidth connectivity multi-processor environments such as Beowulf type processor clusters or workstation farms. There is no requirement of a shared disk and scales well with respect to the number of processors used, and for further improving the scalability of ROLAP with respect to the size and dimension of the data set (which is already better than MOLAP’s scalability), they extend their indexing scheme to the partial cube case. The large number of group-bys, 2d, is a major problem in practice for any data-cube scheme. They considered the case where they do not wish to build (materialize) all group-bys, but only a subset. For example, a user definitely wants to only materialize those group-bys that are frequently used, thereby saving disk space and time for the cube construction. The problem was to find a best way to answer effectively those less frequent OLAP queries which required group-bys that had not yet been materialized. Solving this problem they presented an indexing scheme, based on “surrogate group-bys”, which answers such queries effectively. Their experiments showed that their distributed query engine is almost as efficient on “virtual” group-bys as it is on ones that actually exist. In summary, they claimed that their method provides a framework for distributed high performance indexing for ROLAP cubes with the following properties [7]. In practical, it shows lower communication volume, fully adapted to external memory. There is no requirement of shared disk, maintainable, incrementally; it is efficient for spatial searches in various dimensions, scalable with respect to data sizes, dimensions, and number of processors. They implemented their distributed multi-dimensional ROLAP indexing scheme in STL and MPI, C++ and tested it on a 17 node Beowulf cluster (a frontend and 16 compute nodes). While easily extendable for sharing everything multi-processors, their algorithms performed well on these low-cost commodity-based systems. Their experiments showed that for RCUBE index construction and updating, close to optimal speed has been achieved. A RCUBE index having fully materialized data cube of ≈640 million rows (17 Giga-bytes) on a 16 processor cluster can be generated in just within 1 minute. Their method for distributed query resolution also exhibited good speedup achieving, for example, a speedup of 13.28 on just 16 processors. For distributed query resolution in partial data-cubes, their experiments showed that searches against absent (i.e. non-


67

materialized) group-bys can typically be easily resolved at only a small additional cost. Their results demonstrated that it is possible to build a ROLAP data-cube that is scalable and tightly integrated with the standard relational database approach and, at the same time, provide an efficient index to OLAP queries.

2. DW and OLAP Technologies The prominent definition of Data Warehouse is a "subject-oriented, integrated, nonvolatile and time-variant collection of data in support of management's decisions" [8]. A data warehouse is a well organized single site repository of information collected from different sources. In data warehouse, information is organized around major subjects and is modeled so as fast access to summarize data. "OLAP" as discussed in [9],[10] refers to analysis functionalities generally used for exploring the data. Data warehouse has become a most important topic in the commercial world as well as in the researcher’s community. Data warehouse technology has been mostly used in business world, in finance or retail areas for example. The main concern is to take benefits from the massive amount of data that relies in operational databases. According to [11], the data-modeling paradigm for a data warehouse must fulfill the requirements that are absolutely different from the data models in OLTP environments, the main comparison between OLTP and OLAP environment is reflected in Table 2.

Table 2: OLTP Vs OLAP

Feature OLTP OLAP Amount of data retrieved per transaction

Small Large

Level of data Detailed Aggregated

Views Pre-defined User-defined

Age of data Current (60-90 days)

Historical 5-10 years and also current

Typical write operation

Update, insert, delete

Bulk insert, almost no deletion

Tables Flat tables Multi-Dimensional tables

Number of users

Large Low-Med

Data availability

High (24 hrs, 7 days)

Low-Med

Database size Med (GB- TB)

High (TB – PB)

Query Optimizing

Requires experience

Already “optimized”

The data model of the data warehouse must be simple for the decision maker to understand and for write queries, and must get maximum efficiency from queries. Data warehouse models are called hyper-cubes or multidimensional models and have been prescribed by [12]. The Models are designed for representing measurable indicators or facts and the

various dimensions that characterize the facts. For example, in a area of retail, typical indicators are price and amount of a purchase, dimensions being location, product, customer and time. A dimension is mostly organized in hierarchy, for example the location dimension can be aggregated in city, division, province, country. The "star schema" molds the data as a simple cube, where hierarchical relationship in a dimension is not explicit but is rather encapsulated in attributes; the model of star schema with dimensions is reflected in Figure 2.

Figure 2. Star Schema

The dimension tables are normalized by “snowflake schema”, and make it possible for explicitly representing the hierarchies by separately identifying dimension in its different granularities. At last, when multiple fact tables are required, the "fact constellation" or "galaxy schema" model allows the design of collection of stars. OLAP architectures adopt a multi-tier architecture as reflected in Figure 3 where first tier is a warehouse server, implemented by a relational DBMS. Data of interest must be extracted from OLTP systems (operational legacy databases), extracted, cleaned and transformed by ETL (Extraction, Transformation, Loading) tools before going to load in the warehouse.

Figure 3. How does an OLAP piece fit together

This process aims to consolidate heterogeneous schema (structure heterogeneity, semantic heterogeneity) and for reducing data in order to make it conform to the data warehouse model (by implementing aggregation, dis-cretization functions). Then the data warehouse holds high quality, efficient historical and homogeneous data. The second tier is a data mart, a data mart handles data received from the data warehouse, which is reduced for a selected subject or category. The main focus of data marts is to isolate data of interest for a smaller scope or department, thus permitting the focusing on optimization needs for this data and increase more security control. However this intermediate tier of data mart is optional and not mandatory. The OLAP server is implemented at 3rd level. It optimizes and calculates the hypercube, i.e. the set of fact values for all the relevant tuples


68

of instances for different dimensions (also called embers). In order for optimizing accesses against the data, query results are advance calculated in the form of aggregates. OLAP operators allow materializing different views of the hypercube, allowing interactive queries from decision makers and analysis of the data. Common OLAP operations include drilldown, roll-up, slice and dice, rotate. The fourth tier is an OLAP client, which provides a user interface with various reporting tools, analysis tools and data mining tools for obtaining results. Software solutions are existed for a traditional use.

3. Distributed Index Construction for ROLAP Different methods have been proposed and recommended for building ROLAP data-cubes [13], [14], [15] but there are only very few results available for the indexing of such cubes. For sequential query processing, [16] proposed an indexing model, which is composed of a collection of b-trees. While adequate for low-dimensional data-cubes, b-trees are inappropriate for higher dimensions in that (a) multiple, redundant attribute orderings are required to support arbitrary user queries (b) their performance deteriorates rapidly with increased dimensionality. In [17] proposed the cube-tree, an indexing model which is based on the concept of a packed R-tree [18]. In the dimension of parallel query processing, a typical approach used by current commercial systems like ORACLE 9i RAC for improving throughput by distributing a stream of incoming queries over multiple processors and having each processor answer a subset of queries. But this type of an approach provides no speedup for each individual query. For OLAP queries, which are time consuming, the parallelization of each query is important for the scalability of the entire OLAP system. With respect to the parallelization for general purpose environments of R-tree queries, a number of researchers have presented solutions. As claimed in [19] Koudas, Faloutsos and Kamel presented a Master R-tree structure that employs a centralized index and a collection of distributed data files. Schnitzer and Leutenegger’s Master-Client R-tree [20] improves upon the earlier model by partitioning the central index into a smaller master index as well as a set of associated client indexes. While offering considerable performance advantages in generic indexing environments, neither approach is well-suited for OLAP environment systems. In addition to the sequential problems on the main server node, both utilize partitioning methods that can lead to the localization of searches. Furthermore, neither approach provides the methods for incremental updates. Data reliability is major issue in data warehouses and many solutions have been proposed for its solution. Replication is a widely used mechanism for protecting permanent data loss while replica placement will significantly impact data reliability. Several replica placement policies for different objectives, have been proposed and deployed in real-world systems, like RANDOM in GFS, PTN in RAID, Q-rot in [16], [13], [21] have analyzed how the system parameters, such as object size, system capacity, disk and switch bandwidth, could affect the system’s reliability. However, they only focused on the rough trend about their impact but did not illustrate the correct optimal values of these

parameters. The co-impact of these parameters was not also discussed. Furthermore, for getting the accurate reliability value, some models are so complicated that it is difficult to figure out the best value of each parameter. In [22] worked for designing a reliable large-scale data warehouse or storage system and presented a new object-based-repairing Markov model, which induces many key challenges. One problem was to figure out some basic system parameters, such as the number of nodes, the total number of stored objects and the bandwidth of switch and the node. For designing a reliable system with the optimal system parameters, compared with previous work, their approach makes a significant contribution in two aspects. Firstly, they presented a new object-based Markov model for quantifying the impact of key system parameters on the system reliability in three replica placement strategies. They compared their model with previous complex models; this object-based compact model not only turns to be easier for solving because of its smaller state transition matrix, but also leads to more integrative and practical efficient results. Secondly, they proposed a two-step analyzing process. The first is to find out the comparatively precise optimal value of a system parameter by independently analyzing its impact on the system reliability when other system parameters are fixed. The second is to figure out the best possible combination of these parameters by analyzing their integrated and complex impacts on the system reliability while all of them are tuned. Their analysis results showed that the optimal values do exist and have simple formulas. They presented a new efficient object-based repair model and by analyzing this model, they worked out the individual optimal value of parameters and their optimal combination. The results obtained by them can provide the engineers with direct instructions to design reliable systems.

4. Related Work Here we introduce the lattice view that depicts the relationships between existing views, some algorithms of the pre-computation of views, answering query using the materialized views and view selection.

4.1 The View Lattice Framework The CUBE BY in [23] has resulted in the computation of views which related to SQL queries which has been grouped on all possible combinations of the dimension attributes. Those views are usually denoted by the grouping attributes, e.g. T, L, P, LT, PT, PL, PLT, and All for the example of database with three attributes Time(T), Location(L) and Product(P). As claimed by [24] that a lattice is used to depict the relationship between views. An aggregate view is represented by each node in the lattice. In the lattice view edge existed form node i to node j, view j can be computed from view i to view j which contains one attribute less than view i. In this case, view i is called, parent view of j. in this situation there is a basic view on which every view is dependent. The complete aggregation view “ALL” can be computed from any other view of lattice.


69

Figure 4. Lattice View

4.2 Pre-computation of aggregates In Figure 4 lattice views are shown, for example database which have three dimensions, Time, Location and Product which are reflected as T, L and P respectively. A view is labeled by the dimension’s name, where it is aggregated on view PLT is the basic view while PL is parent view. A lot of aggregation is involved on OLAP queries in data warehouse. Performance can greatly be improved by the pre-computation of aggregates. Many researchers have developed pre-computation algorithms for efficient computing all possible views, which are so called view materialization. In [25] and [26] different efficient view materialization algorithms have been proposed, e.g., Overlap, Pipesoft and PipeHash, which incorporate many efficient optimization techniques, such as using the data sorting in a particular order for computing all views that are prefixes in that order, computing a view from its smallest previously computed parent, and computing the views with most common prefixes in a pipelined technique. In [27] there are some efforts have been made for studying, how the skewed data may affect the pre-computation of aggregates and an approach for dynamically manage the memory usage is recommended. A comparison has been made in [28], difference between the view materialization in MOLAP and ROLAP further an array-based pre-computation efficient algorithm for MOLAP is proposed. In this algorithm the partitions of views is stored in main memory array and further overlaps the computation of different views while using minimum storage memory for each view. In [29] author examined the pre-computation on compressed MOLAP database and some algorithms for the computation of view without de-compression is proposed. In [30] suggested another pre-computation algorithm for MOLAP. One distinct feature in these algorithms was that the aggregation cells are managed in the similar way as the source data. Primary cells and aggregation cells are stored together single data structure, further proposed one multidimensional array, which allows them for quickly accessing. In this algorithm pre-computation considers the points in multidimensional space. The algorithm examines the coordinates of the cells and relies on the relationships among cells for determining how to perform efficient aggregation. A graph-theoretical model is employed for ensuring the correctness of the summation computation.

4.3 View Selection The main issue in full computation of views is storage space; so many researchers study this problem and recommended partial-computation. In [31] efficient approach was proposed for choosing a set of views for materialization under a situation of limited storage space. They introduced a linear cost model; this algorithm assumes that the cost of answering a query is related to the size of view from which the respond of the query can easily be computed. This linear cost model was verified by the experiments. In a greedy view selection algorithm which tries to decide which aggregate view is best for minimizing the query cost. This greedy view selection algorithm first chooses the base view. Materializing one more view can allow some queries which can be easily answered by a smaller metalized view, in this way query cost can be reduced. Their proposed algorithm chooses the view which produces the more reduction in query cost. Research proved that the benefit of the aggregate views selected by this algorithm in no worse than (0.63-f) times the benefit of an optimized selection. The same problem was discussed in [32] and another view selection algorithm PBS (Pick by Size) was proposed. The main difference was that PBS selects the views solely based on the size of the views. In each turn, the view with the smaller size from unselected views is chosen until the total size of the selected views reaches the allocated space limit.

4.4 Query Processing In [33], the traditional query optimization algorithms were generalized to optimize the query in the presence of materialized views. In [34] proposed some techniques for rewriting a given SQL query, such that it uses one or more than one materialized view. They also proposed a semantic approach for determining whether the information existing in a view is sufficient for answering a query. Another query re-writing method was suggested in [35], this technique can utilize the views having different granularities, aggregation granularities and selection regions. Generally, for an OLAP query, their can be many equivalent re-writings using different materialized cubes/views in different ways. Their execution cost is different from one another. An efficient algorithm is also proposed in [35] for determining the set of materialized views used in query re-writing.

5. Recommendations for an Efficient Partial pre-Aggregation

A completely full pre-computation, where all possible aggregates have been pre-computed, can provide the best and most efficient query performance, but in our point of view this approach in not recommended for the following reasons:

• A full pre-computation requires a great storage space for aggregates, so it often exceeds the available space.

• This technique is not based on cost effective use of resources-beneficial, in [32], it is mentioned that the gains from pre-computation outweigh the cost of further disk space after some level of pre-computation.

• Maintenance cost is increased. • It takes long load time.


70

• A full pre-computation is not suitable as very sparse cube (wastage of space) for high cardinality.

However in the situation of fully pre-computation environment, to overcome the storage limitations it is advisable that cube may be partitioned. One logical cube of data should be spread across multiple physical cubes on distinct servers. Divide and Concur approach helps alleviate the scalability limitations of full pre-computation approach. The other approach, which is partial pre-computation of aggregates, can resolve problems of fully aggregation. The main objective of a partial pre-computation technique is to select a certain amount of aggregates for computing before querying time, in this way query answering time can be optimized. But there are two major issues about the optimization process:

• How many partial pre-computed aggregates should be computed?

• What kind of queries is optimized for pre-computation strategy?

The first question depends upon the storage available, we recommends that 40% of all possible aggregates should be computed in advance. Few years back it was recommended by Microsoft, as the vendor of one of the popular OLAP systems that 20% of all possible aggregates should be computed in advance, but now the technology has improved and storage capacity can be achieved by very low cost. The best answer of second question is that these queries should be those that most expected by the decision maker of OLAP application. This type of answer in not so easy because it varies from user to user and application to application, for this purpose and understanding the question systematically, one needs to precisely categorize the chunk of queries as an object for optimization, this type set of queries or expected queries is often called ”query workload”. Pattern of expected queries can be obtained from other same type of case studies and navigational data analysis task. There are many algorithms for partial pre-computation which have already been discussed. For optimized and efficient processing of OLAP queries, most commonly used approach is to store the results of frequently issued queries by decision makers in to summary tables, and further makes use of them for evaluating other queries, this approach is best. In our point of view PBS (Pick by size) algorithm is so fast and best as it will facilitate database administrators for determining the points, where diminishing returns outweigh the cost of the additional storage space. This algorithm also shows how much space should be allocated for pre-computation.

6. Conclusion The partial pre-computation is most popular research area in OLAP environment. Most of the published papers for partial pre-computation are about optimized performance of views as the query workload. Practically users do not care about the processing overhead and the time used in determining the output of the given query, when planning to implement partial pre-computation strategy. For implementation of point queries, the processing overhead is most important fact of consideration. PBS approach is efficient, and selection by PBS is much effective because PC Cube generated by the PBS leads to shorter time required for answering the point

queries. It has most excellent overall performance among variations of member selection algorithms.

References [1] Thomsen, Erik, OLAP Solutions: Building

Multidimensional Information Systems, John Wiley and Sons, 1997.

[2] Agrawal, R., Gupta, A., Sarawagi, S., “Modeling Multidimensional Databases”, Proceedings of the 13th International Conference on Data Engineering, pp. 232-243, 1997.

[3] J. Gray, A. Bosworth, A. Layman, and H. Pirahesh, “Data cube: A relational aggregation operator generalizing group-by, cross-tab, and sub-totals”, Proceeding of the 12th International Conference On Data Engineering, pages 152–159, 1996.

[4] Aejandro A. Vaisman, “Data Warehousing, OLAP, and Materialized Views”, A Survey Technical Report TR015-98, University of Buenos Aires, Computer Science Department, 1998.

[5] V. Harinarayan, A. Rajaraman, and J. Ullman, “Implementing data cubes”, Proceedings of the 1996 ACM SIGMOD Conference, pages 205–216, 1996.

[6] S. Agarwal, R. Agrawal, P. Deshpande, A. Gupta, J. Naughton, R. Ramakrishnan, and S.Sarawagi, “On the computation of multidimensional aggregates”, Proceedings of the 22nd International VLDB Conference, pages 506–521, 1996.

[7] F. Dehne, T. Eavis and A. Rau-Chaplin, “Parallel Multi-Dimensional ROLAP Indexing”, Proceedings of the 3rd IEEE/ACM International Symposium on Cluster Computing and the Grid, (CCGRID’03) 2003.

[8] W.H. Inmon Building the Data Warehouse 3rd Edition, Eds.Wiley and Sons, 1996.

[9] S. Chaudhuri, U. Dayal “An Overview of Data Warehousing and Olap Technolog”, SIGMOD Record 26(1), 1997.

[10] P. Vassiliadis P., T. Sellis, “A Survey of Logical Models for OLAP Databases”, SIGMOD Record Volume 28, Number 1, March, 1999.

[11] R. Kimball, The Data Warehouse Toolkit, J.Wiley and Sons, Inc, 1996.

[12] L. Cabibbo and R. Torlone, “A Logical Approach to Multidimensional Databases”, Proceedings of the 6th International Conference on Extending Database Technology (EDBT'98), Valencia, Spain, 1998.

[13] V. Harinarayan, A. Rajaraman, and J. Ullman, “Implementing data cubes” Proceedings of the 1996 ACM SIGMOD Conference, pages 205–216, 1996.

[14] K. Ross and D. Srivastava, “Fast computation of sparse data cubes”, Proceedings of the 23rd VLDB Conference, pages 116–125, 1997.

[15] S. Sarawagi, R. Agrawal, and A.Gupta, “On computing the data cube”, Technical Report RJ10026, IBM Almaden Research Center, San Jose, California, 1996.

[16] H. Gupta, V. Harinarayan, A. Rajaraman, and J. Ullman, “Index selection for olap”, Proceeding of the 13th International Conference on Data Engineering, pages 208–219,1997.


71

[17] N. Roussopoulos, Y. Kotidis, and M. Roussopolis, “Cubetree: Organization of the bulk incremental up-dates on the data cube”, Proceedings of the 1997 ACM SIG-MOD Conference, pages 89–99, 1997.

[18] N. Roussopolis and D. Leifker, “Direct spatial search on pictorial databases using packed r-trees”, Proceedings of the 1985 ACM SIGMOD Conference, pages 17–31, 1985.

[19] N. Koudas, C. Faloutsos, and I. Kamel, “De-clustering spatial databases on multi- computer architecture”, In Proceedings of Extended Database Technologies, pages 592–614, 1996.

[20] B. Schnitzer and S. Leutenegger, “Master-client r-trees: a new parallel architecture”, 11th International Conf-erence of Scientific and Statistical Database Management, pages 68–77, 1999.

[21] I. Kamel and C. Faloutsos, “On packing r-trees”, Proceedings of the Second International Conference on Information and Knowledge Management, pages 490–499, 1993.

[22] K.Du, Z.Hu, H.Wang,Y.Chen, S.Yang and Z.Yuan, “Reliability Design for Large Scale Data Warehouses”, Journal of Computing, Vol.3, No.10, pp 78-85 October 2008.

[23] J.Gray, A.Bosworth, A.Layman, H.Pirahesh., “Data Cube: A Relational Aggregation Operator Generalizing Group-By,Cross-Tabs, and Sub-Totals”, In Proceedings of International Conference on Data Engineering (ICE’96), New Orleans, February,1996.

[24] V. Harinarayan, A.Rajaraman, and J.D.Ullman, “Implementing Data Cubes Efficiently”, In Proceedings of SIGMOD, pages 205-216,1996.

[25] Sameet Agarwal, Rakesh Agarwal, Prasad M. Deshpandre, Asnish Gupta, Jeffrey F.Naughton, Ragnu Ramakrishnan, Sunita Sarawagi, “On the Computation of Multidimensional Aggregates”, In Proceedings of the 22nd VLDB Conference, Bombay, India, pages 506-521,1996.

[26] P.M.Deshpande, S.Agarwal, J.F.Naughton, and R.Ramakrishnan, “Computation of Multidimensional Aggregates”, Technical Report 1314, University of Wisconsin-Madison, 1996.

[27] Yu,J.X., Hongjun Lu., “Hash in Place with Memory Shifting: Datacube Computation Revisited”, In Proceedings of 15th International Conference on Data Engineering. Page: 254 March, 1999.

[28] Y.Zhao, P.M. Deshpande, and J.F.Naughton., “An Array-based Algorithm for Simulataneous Multidimensional Aggregates”, In Proceedings of ACM SIGMOD, pages 159-170, 1997.

[29] Li,J.,Rotem, D., Srivastava, J., “Aggregation Algorithm for very large compressed Data Warehouses”, In Proceedings of 25th very large Database (VLDB) Conference. Edinburgh, Scotland, 1999.

[30] Woshun Luk., “ADODA: A Desktop Online Data Analyzer”, In 7th International Conference on Database Systems for Advanced Applications (FASFAA,01), Hong Kong, China, April,2001.

[31] V.Harinarayan, A.Rajaraman, and J.D.Ullman, “Impleminting Data Cubes Efficienty”, In Proceedings of SIGMOD, pages 205-206, 1996.

[32] A.Shukla, P.Deshpande, and J.Naughton, “Materialized View Selection for Multidimensional Datasets”, In Proceedings on 24th VLDB Conference, New York, 1998.

[33] S.Chaudhuri,R.Krishnamurthy, S.Potamianos, and K.Shim., “Optimizing Quereis with Materialized Views”, In Proceedings of the 11th IEEE International Conference on Data Engineering, pages 190-200, 1995.

[34] D.Srivastava, S.Dar, H.V.Jagadish, A.Y.Levy., “Answering Queries with Aggregation using views”, In Proceedings of 22nd VLDB Conference, Bombay, India, pages 318-329, 1996.

[35] C.S.Park, M.H.Kim and Y.J.Lee., “Finding an Efficient Rewriting of OLAP Queries Using Materialized Views in Data Warehouses”, Decision Support Systems, vol.32, No.4, pages 379-399,2002.

Authors Profile Naeem Akhtar Khan received the B.S. degree in Computer Science from Allama Iqbal Open University, Islamabad, Pakistan in 2005 and M.S. degree in Computer Science from University of Agriculture, Faisalabad, Pakistan in 2008. He is currently pursuing Ph.D. (Computer Science) degree in University of Central, Punjab, Lahore, Pakistan. His research interests include large-scale data management, data reliability, Data Mining, MOLAP.

Dr. Abdul Aziz did his M.Sc. from University of the Punjab, Pakistan in 1989; M.Phil and Ph.D in Computer Science from University of East Anglia, UK. He secured many honors and awards during his academic career from various institutions. He is currently working as full Professor at the University of Central Punjab, Lahore, Pakistan. He is the founder and Chair of Data Mining Research Group at UCP. Dr. Aziz has delivered lectures in many universities as guest speaker. He has published large number of research papers in different refereed international journals and conferences. His research interests include Knowledge Discovery in Databases (KDD) - Data Mining, Pattern Recognition, Data Warehousing and Machine Learning.

He is member of editorial board for various well known journals and international conferences including IEEE publications. (e-mail: [email protected]).


72

Browser Extensible Secured Hash Password Authentication for Multiple Websites

1T.S.Thangavel and 2Dr. A. Krishnan 1 AP/ Dept. of M.Sc(IT), K.S.Rangasamy College of Technology, Tiruchengode 637215,Tamilnadu,India.

[email protected] 2 Dean, K.S.Rangasamy College of Technology, Tiruchengode 637215,Tamilnadu,India.

[email protected]

Abstract: The techniques such as secured socket layer (SSL) with client-side certificates are well known in the security research community, most commercial web sites rely on a relatively weak form of password authentication, the browser simply sends a user’s plaintext password to a remote web server, often using SSL. Even when used over an encrypted connection, this form of password authentication is vulnerable to attack. In common password attacks, hackers exploit the fact that web users often use the same password at many different sites. This allows hackers to break into a low security site that simply stores username/passwords in the clear and use the retrieved passwords at a high security site. Recently, some collisions have been exposed for a variety of cryptographic hash functions including some of the most widely used today. Many other hash functions using similar constructions can however still be considered secure. Nevertheless, this has drawn attention on the need for new hash function designs. This work developed an improved secure hash function, whose security is directly related to the syndrome decoding problem from the theory of error-correcting codes. The proposal design and develop a user interface, and implementation of a browser extension, password hash, that strengthens web password authentication. Providing customized passwords, can reduce the threat of password attacks with no server changes and little or no change to the user experience. The proposed techniques are designed to transparently provide novice users with the benefits of password practices that are otherwise only feasible for security experts. Experimentation are done with Internet Explorer and Fire fox implementations and report the result of initial user. Keywords: password authentication, secured hash, multi-website password, pseudo random, phishing, cryptographic password 1. Introduction

A random password generator is software program

or hardware device that takes input from a random or pseudo-random number generator and automatically generates a password. Random passwords can be generated manually, using simple sources of randomness such as dice or coins, or they can be generated using a computer. While there are many examples of "random" password generator programs available on the Internet, generating randomness can be tricky and many programs do not generate random characters in a way that ensures strong security. A common recommendation is to use open source security tools where possible, since they allow independent checks on the quality of the methods used. Note that simply generating a password at random does not ensure the password is a strong password, because it is possible, although highly unlikely, to generate an easily guessed or cracked password.

A password generator can be part of a password manager. When a password policy enforces complex rules, it can be easier to use a password generator based on that set of rules than to manually create passwords. In situations where the attacker can obtain an encrypted version of the password, such testing can be performed rapidly enough so that a few million trial passwords can be checked in a matter of seconds. The function rand presents another problem. All pseudo-random number generators have an internal memory or state. The size of that state determines the maximum number of different values it can produce, an n-bit state can produce at most 2n different values. On many systems rand has a 31 or 32 bit state, which is already a significant security limitation.

The main cryptographic hash function design in use today iterates a so called compression function according to Merkle’s [12] and Damgard’s[13] constructions. Classical compression functions are very fast but, in general, cannot be proven secure. However, provable security may be achieved with compression functions designed following public key principles, at the cost of being less efficient. This has been done for instance by Damgard,

where he designed a hash function based on the Knapsack problem. Accordingly, this function has been broken by Granboulan and Joux,[10] using lattice reduction algorithms. The present paper contributes to the hash function family by designing functions based on the syndrome decoding problem, which is immune to lattice reduction based attacks.

Unlike most other public key cryptosystems, the

encryption function of the McEliece cryptosystem is nearly as fast as a symmetric cipher. Using this function with a random matrix instead of the usual parity check matrix of a Goppa code, a provably secure one-way function has been constructed since there is no trapdoor, its security can be readily related to the difficulty of syndrome decoding.

The purpose of this paper is to improve updated

parameters for the hash function. Our paper analyzes asymptotical behavior of their attack. We shall establish that this attack is exponential, such that the design for the hash function is sound.


73

2. Literature Review Computer applications may require random

numbers in many contexts. Random numbers can be used to simulate natural or artificial phenomena in computer simulations, many algorithms that require randomness have been developed that outperform deterministic algorithms for the same problem, and random numbers can be used to generate or verify passwords for cryptography-based computer security systems. The present invention relates to the use of random numbers in such security systems, called as cryptographic applications. Specifically, the present invention pertains to generating a random number in a secure manner for such cryptographic applications. In the context of cryptographic applications[1], there may be an hostile trespasser or agent, who desires to infiltrate the security of cryptographic security system in order to gain access to sensitive, confidential, or valuable information contained therein. For example, banks often encrypt their transactions and accounts.

In order to ensure the utmost security, it is essential

that the security system implements a method for generating a random number that appears completely random. In this manner, a completely random password or cryptographic key presents no opening or prior knowledge that can be exploited by an hostile agent.[2]

Many prior art methods exist for generating

random numbers. These prior art methods typically involve the use of some type of chaotic system. A chaotic system is one with a state that changes over time in a largely unpredictable manner. To use the chaotic system[4] to generate a random number, there is some means of converting the state of the system into a sequence of bits (i.e., a binary number). In the past, chaotic systems were based on various sources, such as the sound of radio static, the output of a noisy diode, output of a Geiger counter, or even the motion of clouds. These chaotic systems can be converted to produce binary numbers by using standard techniques.

For instance, a pseudo-random binary string can be

generated from the digital recording of static noise via a digital microphone. Alternatively, a noisy diode can be sampled at a suitable frequency and converted into a digital signal, or a picture of an area of the sky can be taken and subsequently scanned and digitized. These resulting binary strings that are generated over time are generally random in nature. However, there are several problems associated with simply using a chaotic system as a source of random numbers.[3] First, chaotic systems can be completely or partially predicted over small amounts of time. For example, the position of clouds in some area of the sky at some time can be used to achieve reasonably accurate predictions of the position of clouds in the same area a short time into the future.

Furthermore, the behavior of chaotic systems [6] can be far from completely random. For instance, a digitized picture of a cloud formation will not look like a picture of random

information, but instead, will look like a cloud formation. Moreover, chaotic systems may be biased by outside sources which may be predictable. As an example, a radio signal can be affected by a strong external signal, or the behavior of a noisy diode can be changed by the surrounding temperature. All of the above problems arise because the behavior of a chaotic system may not be completely random. More specifically, an adversary observing or wishing to affect the random number source can take advantage of certain localities that may be inherent in chaotic systems. These localities can occur either in space or time.

Finally, a number of existing applications including Mozilla Firefox provide convenient password management by storing the user’s web passwords on disk, encrypted under some master password. When the user tries to log in to a site, the application asks for the master password and then releases the user’s password for that site. Thus, the user need only remember the master password. The main drawback compared to PwdHash is that the user can only use the web on the machine that stores his passwords. On the plus side, password management systems do provide stronger protection against dictionary attacks when the user chooses a unique, high entropy password for each site. However, many users may fail to do this. 3. Methodology

Random password generators normally output a string of symbols of specified length. These can be individual characters from some character set, syllables designed to form pronounceable passwords, or words from some word list to form a passphrase. The program can be customized to ensure the resulting password complies with the local password policy, say by always producing a mix of letters, numbers and special characters. The strength of a random password can be calculated by computing the information entropy of the random process that produced it. If each symbol in the password is produced independently, the entropy is just given by the formula

where N is the number of possible symbols and L is the number of symbols in the password. The function log2 is the base-2 logarithm. H is measured in bits. An eight character password of single case letters and digits would have 41 bits of entropy (8 x 5.17). Thus a password generated using a 32-bit generator has a maximum entropy of 32 bits, regardless of the number of characters the password contains.

3.1 Secure Hashing

The proposed methodology of the secure hash password system contains one-way hash functions that can process a message to produce a condensed representation called a message digest. This algorithm enables the determination of a message’s integrity, any change to the message will, with a very high probability, results in a different message digest. This property is useful in the generation and verification of digital signatures and


74

message authentication codes, and in the generation of random numbers.

The algorithm is described in two stages, preprocessing and hash computation. Preprocessing involves padding a message, parsing the padded message into m-bit blocks, and setting initialization values to be used in the hash computation. The hash computation generates a message schedule from the padded message and uses that schedule, along with functions, constants, and word operations to iteratively generate a series of hash values. The final hash value generated by the hash computation is used to determine the message digest.

The design principle of hash functions is iterating a compression function (here denoted F), which takes as input s bits and returns r bits (with s > r). The resulting function is then chained to operate on strings of arbitrary length(Fig 1). The validity of such a design has been established and its security is proven not worse than the security of the compression function.

Fig 1: Iterative hash function structure

3.2 Compression Hash Function Algorithm

The core of the compression function is a random binary matrix H of size r×n. The parameters for the hash function are n the number of columns of H, r the number of rows of H and the size in bits of the function output, and w the number of columns of H added at each round.

4. System Model The system model concerned with attacks on the

extension that originate on malicious phishing sites. Password hashing is computed using a Pseudo Random Function (PRF) as follows:

hash(pwd,dom) = PRFpwd(dom)

where the user’s password pwd is used as the PRF key and the remote site’s domain name dom or some variant is used as the input to the PRF. The hash value is then encoded as a string that satisfies the site’s password encoding rules, under control of a configuration file used by the browser extension.

Password hashing is implemented naively inside a browser with rudimentary knowledge of HTML form components. Forms begin with a tag <form action=URL> that tells the browser where the form is to be submitted, and HTML password fields are tagged using <input type=“password”>. The naive browser extension listens for blur events, which fire when focus leaves a field. When the blur event occurs, the extension replaces the contents of the field with the hashed value, using the form action attribute. Thus, after the user enters a password into a form, the clear text password is replaced by a hashed version.

The goal, however, is to defend against web

scripting attacks with minimal change to the user experience. For this leverage the browser extension as a protective but largely transparent intermediary between the user and the web application. All input can be first monitored and secured by the browser extension before the web application is aware that the user is interacting with it. This requires a mechanism by which users can notify password hash browser extension that they are about to enter a password. Password hash can then take steps to protect the password as it is being entered. A distributed hash table is introduced to handle the browser utility replicas of the multiple users across hash authentication mode.

4.1 Distribute Hash Table

The distributed hash table provides incremental scalability of throughput and data capacity as more nodes are added to the cluster. To achieve this, we horizontally partition tables to spread operations and data across bricks. Each brick thus stores some number of partitions of each table in the system, and when new nodes are added to the cluster, this partitioning is altered so that data is spread onto the new node. Because of our workload assumptions, this horizontal partitioning evenly spreads both load and data across the cluster.

Given that the data in the hash table is spread

across multiple nodes, if any of those nodes fail, then a portion of the hash table will become unavailable. For this reason, each partition in the hash table is replicated on more than one cluster node. The set of replicas for a partition form a replica group; all replicas in the group are kept strictly coherent with each other. Any replica can be used to service a get(), but all replicas must be updated during a put() or remove(). If a node fails, the data from its partitions is available on the surviving members of the partitions' replica groups. Replica group membership is thus dynamic; when a node fails, all of its replicas are removed from their replica groups. When a node joins the cluster, it may be added to the replica groups of some partitions.

The illustration below describe the steps taken to

discover the set of replica groups which serve as the backing store for a specific hash table key. The key is used to traverse the DP map tries and retrieve the name of the key's replica group. The replica group name is then used looked up in the RG map to find the group's current membership.


75

We do have a checkpoint mechanism in our distributed hash table that allows us to force the on-disk image of all partitions to be consistent, the disk images can then be backed up for disaster recovery. This checkpoint mechanism is extremely heavy weight, however; during the check pointing of a hash table, no state-changing operations are allowed. We currently rely on system administrators to decide when to initiate checkpoints.

5. Experimental Result and Discussion 5.1 Experimental Implementation

In the proposed hash based password

authentication, a user can change her password at a given site without changing her password at other sites. In fact, the recommended method for using password hash is to choose a small number of strong, distinct passwords, one for every security level (e.g. one password for all financial sites, one password for all news sites, etc). The password hash extension ensures that a break-in at one financial site will not expose the user’s password at all other banks.

The system implemented the prototype as a browser helper object for Internet Explorer. The extension registers three new objects i.e., an entry in the Tools menu (to access extension options), an optional new toolbar, and the password protection service itself. Internet Explorer support COM event sinks that enable Browser Helper Objects to react to website events. Use these sinks to detect focus entering and leaving password fields, drag and drop events, paste events and double click events. The DHTML event model used by Internet Explorer allows page elements to react to these events before they “bubble” up to the extension at the top level. Since extension must handle keystroke events before scripts on the page, we intercept keystrokes using a low-level Windows keyboard hook.

When the password-key or password-prefix is detected, the browser extension determines whether the active element is a password field. If it is not a password field, the user is warned that it is not safe to enter his password. If it is a password field, the extension intercepts all keystrokes of printable characters until the focus leaves the field. The keystrokes are canceled and replaced with simulated keystrokes corresponding to the “mask” characters. The system implementation of secured hash password authentication is accomplished through following process. The client utility is in a web browser, generating hash password as shown in Fig. 2. 5.2 Result and Discussion

The proposed hash based multi-site pseudo random

password mechanism shows proposal considers N number of times that the user U might authenticate before re-registration is required. This suggests that high values of N are desirable. The host H has to store R hash function values at the server. This implies that to reduce the storage requirements, it is desirable to have a low value of R. However, N/2R is the average number of hash function computations that U has to do for every authentication session. Thus, it is desirable to have a high value of R. The parameter R therefore represents a tradeoff between computational requirements of the user U and the storage requirements of the host H. This implies that the value of N and R are best selected by the system administrator keeping in mind the system requirements. We believe that given the current state of storage technologies, the storage requirement is significantly less important than the computational requirement. Major improvement over the previous cryptographic method is the significant reduction in computational requirements per authentication session and increase in the number of logins before re-initialization.

Fig 2: Client side hash password generation


76

Table 1: Effectiveness of Proposed Hash Based Pseudorandom Password Authentication Over Existing Cryptographic Password

Resistance to eaves dropping

Web browser compatibility

Web browser compatibility no. of rounds for authentication

Computational efficiency

Storage capacity

Communication effectiveness

Existing cryptographic password authentication

Feasible False Low Low High False

Proposed hash based pseudorandom password authentication

Highly feasible

True High High Low true

Regarding the computation evaluation the host

verifies the proposed hash password sent by user by computing just a single hash function and one comparison with the stored last one time password. For the investigation of communication factor the host sends the user a hash value and an integer t. The user returns only a single hash value. The resultant of the proposed hash based pseudo random password authentication and cryptographic password authentication are listed in the below Table 1.

6. Conclusion

The paper proposed a provably secure hash

functions based password authentication scheme. This construction provides features such as both the block size of the hash function and the output size are completely scalable. The password hashing method is extremely simple, rather than send the user’s clear text password to a remote site, it sends a hash value derived from the user’s password, and the site domain name. Password Hash captures all user input to a password field and sends hash (pwd, dom) to the remote site. The hash is implemented using a Pseudo Random Function keyed by the password. Since the hash output is tailored to meet server password requirements, the resulting hashed password is handled normally at the server; no server modifications are required. This technique deters password phishing since the password received at a phishing site is not useful at any other domain. The proposed model implements the password hashing as a secure and transparent extension to modern browsers.

References

[1] N. Chou, R. Ledesma, Y. Teraguchi, and J. Mitchell, “ Client-side defense against web based identity theft “, In Proceedings of Network and Distributed Systems Security (NDSS), 2004. [2] J. A. Halderman, B.Waters, and E. Felten “A convenient method for securely managing passwords” To appear in Proceedings of the 14th International World Wide Web Conference (WWW 2005), 2005. [3] F. Hao, P. Zielinski, “A 2-round anonymous veto protocol,” Proceedings of the 14th International Workshop on Security Protocols, SPW’06, Cambridge, UK, May 2006.

[4] Muxiang Zhang, “Analysis of the SPEKE password-authenticated key exchange protocol,” IEEE Communications Letters, Vol. 8, No. 1, pp. 63-65, January 2004. [5] Z. Zhao, Z. Dong, Y. Wang, “Security analysis of a password-based authentication protocol proposed to IEEE 1363,” Theoretical Computer Science, Vol. 352, No. 1, pp. 280–287, 2006. [6] R. Sekar, V. N. Venkatakrishnan, S. Basu, S. Bhatkar, and D. C. DuVarney. Model carrying code: A practical approach for safe execution of untrusted applications. In ACM Symposium on Operating Systems Principles (SOSP), 2003. [7] P. Vogt, F. Nentwich, N. Jovanovic, E. Kirda, C. Kruegel, , and G. Vigna. Cross site scripting prevention with dynamic data tainting and static analysis. In Network and Distributed System Security Symposium (NDSS), San Diego 2007. [8] A. Perrig, R. Canetti, D. Song, and D. Tygar, "Eficient Authentication and Signing of Multicast Streams over Lossy Channels," Proc. of IEEE Security and Privacy Symposium S & P 2000, May 2000. [9] O. Hallaraker and G. Vigna. Detecting Malicious JavaScript Code in Mozilla. In Proceedings of the IEEE International Conference on Engineering of Complex Computer Systems (ICECCS), pages 85–94, Shanghai, China, June 2005. [10] Antoine Joux. Multicollisions in iterated hash functions. Application to cascaded construction. In Advances in Cryptology - CRYPTO '04 Proceedings, Lecture Notes in Computer Science, Vol. 3152, M. Franklin, ed, Springer-Verlag, 2004, pp. 306-316. [11] Yevgeniy Dodis, Thomas Ristenpart, Thomas Shrimpton. Salvaging Merkle Damgård for Practical Applications. Preliminary version in Advances in Cryptology - EUROCRYPT '09 Proceedings, Lecture Notes in Computer Science Vol. 5479, A. Joux, ed, Springer-Verlag, 2009, pp. 371-388.


77

[12] R.C. Merkle. A Certified Digital Signature. In Advances in Cryptology - CRYPTO '89 Proceedings, Lecture Notes in Computer Science Vol. 435, G. Brassard, ed, Springer-Verlag, 1989, pp. 218-238. [13] I. Damgård. A Design Principle for Hash Functions. In Advances in Cryptology - CRYPTO '89 Proceedings, Lecture Notes in Computer Science Vol. 435, G. Brassard, ed, Springer-Verlag, 1989, pp. 416-427. Authors Profile

T.S.Thangavel received the Bsc degree in Computer Science (Bharathiyar University) in 1991 and the Msc degree in computer science (Bharathidasan University) in 1993 and the Mphil degree in Computer Science (Bharathidasan University) in 2003. He is pursuing the PhD degree in department of science and humanities (Anna University).

He is working as an assistant professor in MCA department at K. S. Rangasamy College of Technology, Tiruchengode

Dr. A. Krishnan received his Ph.D degree in Electrical Engineering from IIT, Kanpur. He is now working as an Academic Dean at K. S. Rangasamy College of Technology, Tiruchengode and research guide at Anna University

Chennai. His research interest includes Control system, Digital Filters, Power Electronics, Digital Signal processing, Communication Networks. He has been published more than 176 technical papers at various National/ International Conference and Journals.


78

Detection of Masses in Digital Mammograms

J.Subhash Chandra bose1, Marcus Karnan2, and R.Sivakumar2

1Research Scholar, Anna University, Coimbatore, India

2Tamilnadu College of Engineering Coimbatore, India

[email protected]

Abstract: Mammography is at present the best available technique for early detection of breast cancer. The most common breast abnormalities that may indicate breast cancer are masses and calcifications. The challenge is to quickly and accurately overcome the development of breast cancer which affects more and more women through the world. Microcalcifications appear in a mammogram as fine, granular clusters, which are often difficult to identify in a raw mammogram. Mammogram is one of the best technologies currently being used for diagnosing breast cancer. Breast cancer is diagnosed at advanced stages with the help of the mammogram image. In this paper an intelligent system is designed to diagnose breast cancer through mammograms, using image processing techniques along with intelligent optimization tools such as GA and PSO. The suspicious region is extracted or segmented using two different approaches such as asymmetry approach and Markov Random Field (MRF) hybrid with Particle Swarm Optimization (PSO) algorithm. 161 pairs of digitized mammograms obtained from the Mammography Image Analysis Society (MIAS) database are used to design the proposed diagnosing system.

Keywords: Breast boarder, nipple identification, Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Asymmetry, Texture Segmentation, Receiver Operating Characteristics (ROC).

1. Introduction In the clinical practice of reading and interpreting medical images, radiologists often refer to and compare the similar cases with verified diagnostic results in their decision making of detecting and diagnosing suspicious lesions or diseases. Microcalcification is one of the keys for early detection of breast cancer. Cancer involves the uncontrolled growth of abnormal cells that have mutated from normal tissues. This growth can kill when these cells prevent the normal functioning of vital organs or spread throughout the body damaging essential systems. The term benign refers to a condition, tumor or growth that is not cancerous. This means that it does not spread to other parts of the body or invade and destroy nearby tissue. Benign tumors usually grow slowly. In general, benign tumor or condition is not harmful. However, this is not always the case. If a benign tumor is big enough, its size and weight can press on nearby blood vessels, nerves, organs or otherwise cause problems. Breast cancer, also known as carcinoma, is a malignant growth that begins in the tissues of the breast. There are several types of breast cancer. Ductal carcinoma begins in the cells lining the ducts that bring milk to the

nipple and accounts for more than 75% of breast cancers 20% of lobular carcinoma begins in the milk-secreting glands of the breast but otherwise fairly similar in its behavior to ductal carcinoma; 5% of other varieties of breast cancer can arise from the skin, fat, connective tissues and other cells present in the breast.

2. Over View of CAD System Detection of microcalcification is performed in two steps: preprocessing and segmentation, the global appearance (brightness, contrast, etc.) of the mammogram images may differ, usually due to variations in the recording procedure [1, 3, 15]. Initially the film artifacts and x-ray labels are removed from the mammogram images and median filter is applied to remove the high frequency components (i.e. noise) from the image. Then the mammogram images are normalized to avoid differences in brightness between the mammograms caused by the recording procedure. And to increase the reliability of segmentation, the pectoral muscle region is removed from the breast region. The enhanced images are used for segmentation. In this paper the suspicious region is extracted or segmented using two different approaches such as asymmetry approach [11] and Markov Random Field (MRF) hybrid with Particle Swarm Optimization (PSO) algorithm. In case of asymmetry approach, the suspicious regions on digital mammograms are segmented based on the asymmetries between corresponding regions in the left and right breast images. Due to the recording procedure the size and shape of the corresponding mammograms does not match. So the mammogram images must be aligned prior to subtraction. The breast border and the nipple points can be used as reference points for alignment of mammograms. In this paper the breast border is detected using Genetic Algorithm and the nipple position is identified using a novel method called Particle Swarm Optimization (PSO) algorithm. Using the border points and nipple position as references the mammogram images are aligned and subtracted to extract the suspicious region [10,11]. In the next Texture segmentation technique, a novel method, Markov Random Field (MRF) hybrid with Particle Swarm Optimization (PSO) algorithm is used to segment the microcalcifications from the mammogram image. Initially, a unique label is assigned for similar patterns in the mammogram image. The MRF based image


79

segmentation method is a process seeking the optimal labeling of the image pixels. The optimum label is which minimizes the MAP estimate. To optimize this MRF based segmentation, Particle Swarm Optimization (PSO) algorithm is implemented to compute the adaptive optimal threshold value. [12, 13, 14]. .

Figure 1. Flow diagram for mammogram image Preprocessing and Segmentation. A Receiver Operating Characteristics (ROC) analysis is performed to evaluate the classification performances of the proposed approaches [5]. The area under the ROC curve Az value is used as a measure of the classification performance. A higher Az indicates better classification performance because a larger value of True Positive (TP) is achieved at each value of False Positive (FP). The proposed algorithms and the techniques are tested on 161 pairs of digitized mammograms from Mammography Image Analysis Society (MIAS) database. The Figs. 1 shows the overview of the work. 2.1 Mammography

Medical imaging in general and mammography in particular, lets physicians evaluate a specific area of the body of a patient which might be externally visible. Medical imaging as one of the most important medical developments of the past thousand years, basically due to the fact that it provides physicians with physiology and functionality of organs and cells inside human bodies. Among the different imaging modalities used for breast cancer detection, mammography remains the key screening tool for the detection of breast abnormalities. In a recent study, the proportion of breast tumors that were detected in Vermont (US) by screening mammography increased from 2% during 1974 - 1984 to 36% during 1995 - 1999. However, it is also well known that expert radiologists can miss a significant portion of abnormalities. In addition, a large number of mammographic abnormalities turn out to be benign after biopsy. Mammograms capture the low energy X-rays which passes through a compressed

breast. Fig 2 shows directions of mammogram capturing to X-ray. Depending on the viewpoint of the X-rays, the images are classified into different categories, Cranio-Caudal view and Medio-Lateral Oblique view. Fig 3 and 4 show view of the images.

Figure 2. physically the viewpoints’ directions of mammogram capturing to X-ray

Figure 3. Cranio-Caudal view

Figure 4. Medio-Lateral Oblique view. It is important to notice that in the MLO views there is one region corresponding to a portion of the pectoral muscle which may be present in the left or the right upper corner of


80

the image. Moreover, some annotations and labels can appear in the images. Figure 5 and 6 shows the normal left and right mammogram image.

Figure 5. Right Mammogram of a woman

Figure 6. Left Mammogram of a woman 2.2 Mammographic Abnormalities There is a large number of types of abnormalities that can be present in a breast. Among those, signs of breast cancer are normally associated with: ü Asymmetry between images of left and right breasts.

ü Distortion of the normal architecture of the breast

tissue.

ü Presence of micro-calcifications in the breast.

ü Presence of masses in the breast.

It is generally accepted that the essential characteristic of a high-quality mammogram is the ability to visualize these four features. Both breasts are usually considered as almost symmetric structures. While exact mirror images are not to be expected when comparing them (usually the first practice of expert physicians), the tissue patterns within each breast should be similarly distributed. An asymmetric area may be indicative of a developing mass or a variation of normal breast tissue.

2.3 Mass Shapes Mass Margins The shape and margin of a mass are strong signs of their malignancy or benignancy degree. A distortion in the normal breast architecture (architectural distortion) refers to a derangement or disruption of the normal arrangement of the tissue strands of the breast resulting in a radiating or haphazard pattern without an associated visible centre. This includes speculations radiating from a point, and focal retraction or distortion of the edge of the parenchyma. Micro-calcifications are tiny calcifications that range from 50 to several hundred microns in diameter, which usually appear in clusters. In these cases, they are analyzed according to their size, shape, number, and distribution. The general rule is that larger, round or oval shaped calcifications with uniform size have a higher probability of being associated with a benign process, whereas smaller, irregular, polymorphic, branching calcifications heterogeneous in size and morphology are more often associated with a malignant process. A breast mass, on the other hand, is a localized swelling, protuberance, or lump in the breast, which usually is described by its location, size, shape, margin characteristics, and any other associated findings (i.e. architectural distortion, X-ray attenuation). Depending on morphologic criteria, the likelihood of malignancy can be established. Normally, a benign process is associated with the presence of circular or oval shapes, while, in contrast, speculated masses are more probable to Circular Shape Lobular Shape Speculated Shape Circumscribed Margin Well Defined Margin Ill Defined Margin. The last one has an increased probability to be malignant be the sign of a malign process. The margin refers to the border of a mass, and it should be examined carefully because it is one of the most important criteria in determining whether the mass is the result of a benign or malign process. Radiologists classify the margin among five classes: • Circumscribed margins, which are well defined and sharply demarcated with an abrupt transition between the lesion and the surrounding tissue. • Obscured margins, which are hidden by superimposed or adjacent normal tissue. • Micro-lobulated margins, which have small undulating circles along the edge of the mass. • Ill-defined margins, which are poorly defined and scattered. • Speculated margins, which are marked by radiating thin lines. The probability to find a malignancy mass is normally ordered according to this classification. The more ill-defined and speculated the margin, the higher the probability to be associated with a malignant process. It should be clear that


81

these morphological aspects can be very subtle and difficult to diagnose, even for an expert radiologist. 3. Digital Mammography As a consequence of the actual digital revolution, traditional film-based hospitals are themselves converting to digital hospitals, where patient medical records, chart information, and test results are easily available electronically for physicians from anywhere in the hospital. In that sense, full-digital mammography is gaining importance compared to the nowadays still conventional film-screen mammography, due to the fact that digital acquisition, digital storage, and digital display processes may be separated and individually optimized. 3.1 Image Acquisition Mammographic Image Analysis Society (MIAS) All mammograms used in this work are from a mini Mammographic database provided by Mammographic Image Analysis Society (MIAS), which includes 23 cases with 28 MCs . The Mammographic Image Analysis Society (MIAS) Mini Mammographic Database from the Royal Marsden Hospital in London was used in this study. It contains 322 images (Medio-Lateral Oblique (MLO)) representing 161 bilateral pairs. The database is divided into seven categories. These include normal image pairs and abnormal pairs containing microcalcifications, circumscribed masses, spiculated lesions, ill-defined masses, architectural distortion and asymmetric densities. Each mammogram from the database is a 1024 x1024 pixels and with a spatial resolution of 200µm/pixel. The odd number cases represent the left breast mammogram while the even number cases represent the corresponding right breast mammogram. The database lists the film and provides appropriate details as follows: ü 1st column: MIAS database reference number.

ü 2nd column : Character of background tissue (Fatty,

Fatty-glandular, or Dense- glandular) ü 3rd column : Class of abnormality present

(Calcification, Well-defined/circumscribed ü masses, Spiculated masses, Other/ill-defined masses,

Architectural distortion, Asymmetry, or Normal) ü 4th column: Severity of abnormality (Benign or

Malignant) ü 5th and 6th columns : x,y image coordinates of

center of abnormality. ü 7th column ; Approximate radius (in pixels) of a

circle enclosing the abnormality.

4. Need for Computer Aided Detection Breast Cancer

The idea of computer systems aiding radiologists to detect breast cancer is not recent. However, the nowadays rapid development of full digital mammographic systems has being accompanied by the natural increase of such systems. A Computer-Aided System (CAD) is a set of automatic or semiautomatic tools developed to assist radiologists in the detection and evaluation of mammographic images. The Need for Computer Aided Detection Breast cancer is the most common cancer among women in the United States, other than skin cancer. It is the second leading cause of cancer death in women, after lung cancer. The American Cancer Society estimates that 182,460 women in the United States will be found to have invasive breast cancer in 2008. About 40,480 women will die from the disease this year. In the US, breast cancer is the most common form of cancer among women and is the second leading cause of cancer deaths, after lung cancer. Women in the U.S. have about a 1 in 8 lifetime risk of developing invasive breast cancer. Incidence of breast cancer in India is on the rise and is rapidly becoming the number one cancer in females pushing the cervical cancer to the second spot. The seriousness of the situation is apparent after going through recent data from Indian Council of Medical Research (ICMR). The rise is being documented mainly in the metros, but it can be safely said that many cases in rural areas go unnoticed. It is reported that one in 22 women in India is likely to suffer from breast cancer during her lifetime, while the figure is definitely more in America with one in eight being a victim of this deadly cancer. The problem with preventing breast cancer is that there is no one cause that can be pinpointed as being the culprit. Of course screening for the presence of BRCA1 and BRCA2 mutations is available though it must be admitted of being of little use in the Indian context. It is here that the task of spreading the awareness of the prevalence of this cancer and advising women on undertaking self-breast examination comes into the picture. Health officials must try and talk about this condition so that women have a say in their own health. Finally, there are procedures like mammography and Fine Needle Aspiration Cytology (FNAC) and biopsy that need to be widely publicized so that women are aware of exactly what they are letting themselves in for. Early detection of breast cancer increases the survival rate and increases the treatment options. Screening mammography, or x-ray imaging of the breast, is currently the most effective tool for early detection of breast cancer [6, 7]. Screening mammography examinations are performed on asymptomatic woman to detect early, clinically unsuspected breast cancer. Radiologists visually search mammograms for specific abnormalities. Some of the important signs of breast cancer that radiologists look for are clusters of micro calcifications, masses, and architectural distortions.


82

5. Pre processing and Segmentation Computer-Aided Detection (CAD) systems have been developed to aid radiologists in detecting mammographic lesions that may indicate the presence of breast cancer. These systems act only as a second reader and the final decision is made by the radiologist. Recent studies have also shown that CAD systems, when used as an aid, have improved radiologists accuracy of detection of breast cancer [3, 7, 23, 24]. Computer-Aided Diagnosis (CAD) systems for aiding in the decision between follow-up and biopsy are still in development. It is important to realize that mammographic image analysis is an extremely challenging task for a number of reasons. The mammogram images were taken from the MIAS database. Initially the X-ray labels and the film artifacts are removed from the mammogram images using gradient-based tracking algorithm. And the median filter is applied to remove the noise from the mammogram images [4, 8, 9, 18]. To apply median filter, the intensity value of every pixel is replaced with the median value of the neighborhood pixels with the window size of 3×3. Due to the recording procedure the brightness between the mammograms may vary. In order to reduce the variation, and achieve computational consistency, the images are normalized, by mapping all mammograms into a fixed intensities range. In the next step, the pectoral region is removed from the breast region to increase the reliability of the segmentation. It is achieved by applying histogram-based thresholding. The enhanced mammogram images are evaluated using signal-to-noise ratio [21,22]. In the next step, the suspicious regions are extracted from the enhanced mammogram image using two different approaches such as asymmetry and texture segmentation. Asymmetry is based on the asymmetry between the corresponding left and right image. In this paper, the breast border and the nipple position are considered as reference points for mammogram alignment. The Genetic Algorithm (GA) is used for breast border detection and for nipple identification Particle Swarm Optimization (PSO) algorithm is applied. Once the mammogram images are aligned they can be subtracted to extract the suspicious region. In case of texture segmentation, Markov Random Field (MRF) is applied to label the image pixels. For labeling, a kernel is extracted for each pixel, kernel is a window of neighborhood pixels with the size of 5×5. A unique label is assigned to the kernels having similar patterns [17]. A pattern matrix is maintained to store the dissimilar patterns in the image. For each patterns in the pattern matrix, the posterior energy function value is calculated. The challenge of finding the MAP estimate of the segmentation is search for the optimum label, which minimizes the posterior energy function. In this paper a new effective approach, PSO is applied for the minimization of the energy function. The segmentation from both the methods is compared with the MIAS information, by adaptive thresholding the segmented image using various operating points. The statistical results show that the MRF-PSO performs better

than asymmetry method and existing techniques. Performance of each test is characterized in terms of its ability to identify true positives while rejecting false positives using Receiver Operating Characteristic (ROC) Analysis [4]. The area under the ROC curve is an important criterion for evaluating diagnostic performance. Usually it is referred as the AZ index. The AZ value of ROC curve is just the area under the ROC curve [19, 20].

6. Experiments and Results Ultimately, the effectiveness of the proposed technique is determined by the extent to which potential abnormalities can be extracted from corresponding mammograms based on analysis of their asymmetry image. The Mammographic Image Analysis Society (MIAS) Database is used to evaluate the technique. All 161 MIAS image pairs were used in this paper. A randomly selected set of 20 bilateral pairs drawn from the pairs with spiculated and circumscribed lesions was used for developing the algorithm and for guiding parameter setting. One of the training circumscribed cases also had an asymmetric density. The remaining abnormal and the normal image pairs were used to measure performance. The true positive detection rate and the number of false positive detection rate at various thresholds of the asymmetry images are used to measure the algorithm’s performance. These rates are represented using Receiver Operating Characteristic (ROC) curves. True Positive (TP) and False Positive (FP) rates are calculated at 20 different thresholds selected on asymmetry image pixels to generate an ROC curve. A region extracted in the asymmetry image, which overlaps with a true abnormality as provided in the ground truth of the image, is called a true positive detection. An overlap means that at least 80% of the region extracted lies within the circle indicating a true abnormality as determined by MIAS database. For example, the mammogram mdb239.pgm, the spatial coordinate position of the suspicious region, x and y are 567, 808 respectively, and the radius is 25 pixels. The resultant asymmetry image contains the suspicious region], with the radius of 25 pixels. Compared to the MIAS information on mdb239.pgm, results from the proposed method overlaps 99% of the specified region and this image is classified as true positive image. Suppose the overlap is less than 80% of the specified region, and then the image is considered as false positive image.


83

Figure 7. Truth Image of mammogram mdb239.pgm

In the previous methods such as Sallam and Bowyer, Lau and Bischof have taken the overlap region of only 40% are considered as true positive[11, 16]. But in this paper, the true positive is considered only at 80% of overlap occurs. All other regions extracted by the algorithm are labeled as false positives. Figure 5,6,7 shows the ROC curves generated on the full test set, using 20 operating points. In general, it is expected that the true positive detection rate in an ROC curve will continue to increase or remain constant as the number of false positives increase. In this case the true positive rate actually drops at certain points. If the threshold value is low true detections may become merged with false positive regions. Fig 7 shows the Truth Image of mammogram mdb239.pgm. Figure 8 shows the Asymmetry Image using PSO. Figure 9 shows the extracting suspicious region from background tissue using MRF – PSO technique.

Figure 8. Asymmetry Image using PSO

Original image segmented image

Figure 9. Segmented Image using MRF – PSO Detection Ratio: The area under the ROC curve (Az value) is an important criterion for evaluating diagnostic performance . The AZ value of ROC curve should be computed by normalizing the area under the ROC curve by the range of the abscissa. The value of AZ is 1.0 when the diagnostic detection has perfect performance, which means that TP rate is 100% and FP rate is 0%. The Az value for the proposed MRF PSO algorithm is 0.983. The Table 1 shows the comparison of classification rate between the previous works and the proposed method. Figure 10, 11 and 12: shows the ROC curve – GA, ROC curve – PSO and ROC curve – GA and PSO. Fig 13 shows the bar chart of Comparison of Classification rate.

Figure 10. shows the ROC curve – GA


84

Figure 11. shows the ROC curve – PSO

Figure 12. shows the ROC curves generated on the full test set for GA and PSO

Table 1: Comparison of Classification Rate

Figure 13. Comparison of Classification Rate

7. Conclusion In this paper the suspicious region is extracted or segmented using two different approaches such as asymmetry approach and Markov Random Field (MRF) hybrid with Particle Swarm Optimization (PSO) algorithm. In case of asymmetry approach, the suspicious regions on digital mammograms are segmented based on the asymmetries between corresponding regions in the left and right breast images. The breast border is detected using Genetic Algorithm and the nipple position is identified using a method called Particle Swarm Optimization (PSO) algorithm. In the texture segmentation technique, Markov Random Field (MRF) hybrid with Particle Swarm Optimization (PSO) algorithm is used to segment the microcalcifications from the mammogram image. To optimize this MRF based segmentation, Particle Swarm Optimization (PSO) algorithm is implemented. A Receiver Operating Characteristics (ROC) analysis is performed to evaluate the classification performances of the proposed approaches. The approach using MRF-PSO based segmentation was superior to the other methods. The overall performance and the results show that the particle Swarm Optimization algorithm performs better than other methods comparatively. Fig 14 shows the Snapshoot for the detailed result of the Detection of Microcalcification in mammograms using MATLAB 7.

Sl.No.

Authors Methods Classification Rate

1 Lau and Bischof, 1991 [11]

Asymmetry Measures

85.00%

2 Sallam and Bowyer,1999 [16]

Unwarping Technique

86.60%

3 Ferrari and Rangayyan, 2001 [2]

Directional Filtering with Gabor wavelets

74.40%

4 Thangavel and Karnan,2005 [24]

MRF-ACO 94.80%

5 The proposed Metaheuristic Approach

Bilateral Subtraction using PSO

94.60%

6 MRF-PSO Segmentation

98.30%


85

Figure 14. Snapshoot for the Detection of Masses in mammograms

References [1] Dhawan,P, Buelloni,G and Gordon,R. “Enhancement of

mammgraphic features by optimal adaptive neighborhood image processing,” IEEEtans. Acoust, Speech, Signal processing.

[2] Ferrari, R.J., Rangayyan, R.M., Desautels, J.E.L., Borges, R.A., and Frere, A.F.: “Analysis of Asymmetry in Mammograms via Directional Filtering With Gabor Wavelets,” IEEE Transactions on Medical Imaging, vol. 20, no. 9, pp: 953–964, 2001.

[3] Guido M. te Brake and Nico Karssemeijer,”Single and multiscale detection of masses in digital mammograms., IEEE Transactions on Medical Imaging, vol. 18, No. 7, July 1999, pp. 628-638.

[4] Gonzalez, R.C and Wintz,P, Digital Image Processing(Add-Wesley, Reading, 1987.

[5] J. A. Hanley, and B. J. McNeil. The meaning and use of the area under a receiver operating characteristic (ROC), curve. Radiology, 143:29–36, 1982.

[6] Harvey.J.E, Fajardo.L.L, & Inis.G.A, ”Previous mammograms in patients with impalpable breast carcinoma: Rctrospective vs. blinded interpretation. AJR, vol. 161, PP. 1167-1172, 1993

[7] Hult.I.W, Astley.S.M, & Boggis.C.R.M,. Prompting as an aid to diagnosis in mammography; in Digital Mammography,. A. G. Gala, S. M. Astley, D. R. Dance, & A. Y. Cairns, Eds. Amsterdam. The Netherlands : Elsevier, 1994, pp. 389-398.

[8] Jain, A.K,” fundamentals of digital image processing. 1995.

[9] Jain, A.K, Duin, R.P.W, and Mao, J. Statistical pattern recognition: A review. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(1):4–37, 2000.

[10] Jong kook kim and Hyun wook park,”Statistical textural features for detection of micro calcification in digitized mammogram”,IEEE trans.on medi. Imaging, vol.no.18, no.3, mar 1999.

[11] Lau, T.K and Bischof,W., “Automated detection of breast tumors using the asymmetry approach,” Comput. Biomed. Res. 24, 273-295(1991).

[12] M. Karnan, R. Sivakumar, M. Almelumangai, K. Selvanayagi and T. Logeswari,” Hybrid Particle Swarm Optimization for Automatically Detect the Breast Border and Nipple position to Identify the Suspicious Regions on Digital Mammograms Based on Asymmetries”, International Journal of Soft Computing 3 (3): 220-223, 2008

[13] Karnan and K.Thangavel, Automatic Detection of the Breast Border and Nipple Position on Digital Mammograms Using Genetic Algorithm, International Journal on Computer Methods and Programs in Biomedicine (Elsvier). VOL 87, pp.12-20 2007

[14] M.Karnan, K.Thangavel, “Weight Updating in BPN Network Using Ant Colony Optimization for Classification of Microcalcifications in Mammograms, International Journal of Computing and Applications, Vol:2,no.2, pp 95-109, 2007

[15] Naga R. Mudigonda, Rangaraj M Rangayyan and Leo Desautel.J.E, ”Gradient & Texture analysis for the classification of Mammographic Masses. IEEE trans. on MI, vol. 19, no. 10, Oct. 2000 pp. 1032 ‘ 1042.

[16] Sallam, M.Y., and Bowyer, K.W.: “Registration and difference analysis of corresponding mammogram images,” Medical Image Analysis, vol. 3, no. 2, pp: 103-118, 1999

[17] Stephan Olariu, Albert Y. Zomaya, “Handbook of Bioinspired Algorithms and Applications” CHAPMAN & HALL/CRC COMPUTER and INFORMATION SCIENCE SERIES, 2006

[18] K.Thangavel and M.Karnan, “CAD system for Preprocessing and Enhancement of Digital Mammograms,” International Journal on Graphics Vision and Image Processing, vol. 9, no. 9, pp: 69-74, 2006.

[19] K.Thangavel, M.Karnan, P.Jaganathan, A. Pethalakshmi, R. Sivakumar, “Computer-Aided Diagnosis: Automatic detection of microcalcifications in Mammography Images using Soft Computing”, Lecturer Notes in Engineering and Computer Science IMECS Hong Kongm PP 280-286 June, 2006.

[20] M.Karnan, K.Thangavel, R.Sivakumar, “Ant Colony optimization algorithm for Feature Selection and Classification of Microcalcifications in Mammograms, IEEE International Conference on Advanced Computing and Communications, 2006. ,IEEE press, pp: 298-303,2006

[21] K.Thangavel, M.Karnan, R. Siva Kumar, and A. Kaja Mohideen. “Automatic Detection of Microcalcification in Mammograms-A Review,” International Journal on Graphics Vision and Image Processing, vol. 5, no. 5, pp: 31-61, 2005.

[22] K.Thangavel and M.Karnan. “Computer Aided Diagnosis in Digital Mammograms: Detection of Microcalcifications by Meta Heuristic Algorithms,” International Journal on Graphics Vision and Image Processing, vol. 7, no. 7, pp: 41-55, 2005.


86

[23] K.Thangavel and M.Karnan, “Automatic Detection of Asymmetries in Mammograms Using Genetic Algorithm,” International Journal on Artificial Intelligence and Machine Learning, vol. 5, no. 3, pp: 55-62, 2005.

[24] K.Thangavel, M.Karnan, R.Siva Kumar and A.Kajamohideen, “Segmentation and Classification of Microcalcification in Mammograms Using the Ant Colony System,” International Journal on Artificial Intelligence and Machine Learning, vol. 5, no. 3, pp: 29-40, 2005.

Authors Profile

J.Subashchandra bose Received the Master of Computer Science and Engineering Degree from Anna University Chennai, Tamil Nadu, India, Bachelor of Electrical and Electronics Engineering Degree from Anna University Chennai, Tamil Nadu, India, Currently he is working as Assistant Professor,

Department of Computer Science & Engineering, Hindusthan College of Engineering and Technology, Tamil Nadu, India, and doing part-time paper in the Department of computer Science and Engineering, Anna University- Coimbatore, Tamil Nadu, India. His area of interests includes medical image processing, artificial intelligence, neural network, and fuzzy logic

Marcus Karnan received the BE Degree in Electrical and Electronics Engineering from Government College of Technology,Bharathiar University, India. Received the ME Degree in Computer Science and Engineering from Government College of

Engineering ,Manonmaniam Sundaranar University in 2000. Received the PhD degree in CSE from Gandhigram Rural University, India in 2007, Currently he is working as Professor, Department of Computer Science & Engineering Department, Tamilnadu College of Engineering, India. He has been in teaching since 1998 and has more than eleven years in industrial and paper experience. His area of interests includes medical image processing, artificial intelligence, neural network, genetic algorithm, pattern recognition and fuzzy logic

Sivakumar Ramakrishnan, Received the Master of Computer Science and Engineering Degree from Computer Science and Engineering Department, from Government College of Engineering, Manonmaniam Sundaranar University, Tamil Nadu, India, in 2000. Currently he is working

as Assistant Professor, Department of Computer Science & Engineering, Tamilnadu College of Engineering, Tamil Nadu, India. And doing part-time paper in the Department of computer Science, Bharathiar University, Tamil Nadu,

India.His area of interests includes medical image processing, artificial intelligence, neural network, and fuzzy logic


87

A Dynamic Trust-Based Context-Aware Authentication Framework with Privacy Preserving

Abdallah MHAMED1, Pierre E. ABI-CHAR1, Bachar EL-HASSAN2 Mounir MOKHTARI1

1Laboratory of Reseaux et Service De Telecommunication

(RST), Telecom SudParis (ex. INT), 9 Rue Charles Fourier, Evry, France {Pierre.abi_char, abdallah.mhamed, mounir.mokhtari}@it-sudparis.eu

2Laboratory of Electronic Systems, Telecommunication and

Networking (LASTRE), Faculty of Engineering, Branch 1, Al Arz Street, El Kobbeh, Tripoli, Lebanon.

[email protected]

Abstract: As ubiquitous technologies ingrain themselves further into our lives, rapid progress has been made in context-aware computing. Context-aware environments are set to become a reality. However, major challenges remain to be addressed including privacy, authentication, access control, and trust. These security challenges have to be non-intrusive, intelligent, and able to adapt to the rapidly changing contexts of users. Context-aware environments are expected to make these challenges more accurate and to consider them in place from the start, so that a mutual trust relationship can be formed between entities. It is therefore, a key challenge in a ubiquitous network society to design an effective privacy preserving authentication and access control framework that adequately meet security requirements posed by the context-aware service paradigm in pervasive computing environment. In this paper, we propose a security framework that integrates context-awareness to perform authentication and access control approach in a very flexible and scalable model that is both context-aware and privacy preserving. Moreover, we show how our framework can be integrated with trust management. In this paper, we focus on introducing an anonymous authentication and access control scheme to secure interactions between users and services in ubiquitous environments. The architecture focuses on the authentication of users who request access to the resources of smart environment system through static devices (i.e. smart card, RFID, etc.), or dynamic devices (i.e. PDA, mobile phones, etc.).

Keywords: Context-Aware, Authentication, Access Control, Smart Spaces, Privacy Control, Fuzzy Logic, Trust Management, Risk Assessment, Quality of Privacy.

1. Introduction The growing evolution of Information and Communication Technology (ICT) systems towards more pervasive and ubiquitous infrastructures contribute significantly to the deployment of services anywhere, at anytime and for anyone. To provide personalized services in such infrastructures, we should consider both user's privacy and security requirements and context-awareness environment. Security, Privacy and Trust in pervasive computing are currently hot issues in digital information technology area. Security is used to describe techniques that control who may use or modify private data and context information, privacy is viewed as the ability of an entity to determine whether, when, and to whom information is to be released and finally

trust denotes the grounds for confidence that a system will meet its security objectives. The development of mobile communications technologies and ubiquitous computing paradigm and the convergence of m-healthcare, m-business, m-entertainment and m-education services have raised the urgency of dealing with privacy threats (i.e. personal information, etc.). These threats are caused by the detection of personal sensitive information such as location, preferences, and activities about individuals through sensors or invisible computing devices gathering collating data and deriving user context. Moreover, the ubiquitous computing environment is characterized by people constantly moving, and engaged in numerous activities simultaneously. Therefore, we proposed an authentication and access control agent framework for context-aware services. Our framework’s objectives are to provide the most suitable security scheme on the basis of context, such as users' location and profiles, and to protect personal information such as user location, user's ID, etc. This paper provides us a scheme to protect privacy of users and to maintain the flexibility for users while using available service in ubiquitous environments. The ultimate goal is anonymity which keeps the users anonymously interacting with the services, through that, preserving context privacy of users. And also it keeps confidentiality and integrity on communication channels. The proposed schemes is at application level without relying on any underlying system infrastructure such as “light house” or “Mist router” in [6]. This scheme possesses many desirable security properties, such as anonymity, nonlinkability, trust management, etc.

The rest of this paper is as follows. Context-aware definition and usage, authentication and access control characteristics and their privacy effects, and trust management definition and properties are outlined in Section 2. Section 3 provides an outline for the mathematical backgrounds needed for our protocol process. Section 4 provides a summary regarding relevant related work. Our proposed agent framework, its process descriptions, and security discussion are introduced in Section 5, 6, and 7 respectively. Finally, the paper future work and conclusion are described in Section 8.


88

2. Pervasive Computing Paradigm In this section we briefly introduce some assumptions, concepts, and values that constitute a real way for viewing the necessity to a novel scheme.

2.1 Context-Aware Context-Aware computing is an emerging computing paradigm that tries to exploit information about the context of its users to provide new or improved services. [2] Have defined context as: any information that can be used to characterize the situation of an entity. An entity is a person, place, or object that is considered relevant to the interaction between a user and an application, including the user and applications themselves. The use of context information gives a number of advantages in communication as motivated before. Context-awareness is an enabling technology to build helpers that are disappearing from the users’ perception. This allows creating new classes of services. The combination of several context values provides a very powerful mechanism to determine the current situation.

2.2 Authentication and Access Control The title Authentication systems are used for security purposes to verify the authenticity of one or more parties or entities during a transaction. Most traditional authentication methods either do not scale well in massively distributed environments, with hundreds or thousands of embedded devices like smart spaces, or they are inconvenient for users roaming around within smart environments. In addition, authentication in smart environments can not use a one-size-fits-all approach, as authentication requirements differ greatly among different spaces and different applications and contexts within the same smart space. In general, users must be able to authenticate with other entities with a varied level of confidence, in a transparent, convenient, and private manner. The concept of context-aware authentication and access control is: (1) Collect and recognize the user’s current context, and (2) Generate and control a secure user environment based on the current context. Generally, the context includes user’s location and services, present information, environmental information (temperature, loudness, and brightness), terminal attributes, and network status (QoS), etc

2.2.1 Privacy Effects An inherent tension exists between authentication and privacy because the act of authentication often involves some disclosure or confirmation of personal information. System designers sometimes fail to consider the myriad impact that authentication affects privacy. When designing an authentication system, selecting one for use, or developing policies for one, we should authenticate only for necessary (well-defined purposes), minimize the scope of the data collected, articulate what entities will have access to the collected data, articulate what kinds of access to and use of the data will be allowed, and finally provide means for

individuals to check on and correct any information held about them for use in authentication. Context-aware services should be able to trust context data provided to them from these various sources and to respond to changes. The dynamic nature of a context-aware environment necessitates the need for a very active and flexible authentication mechanism that allows members across different domains to identify and communicate with each other with a reasonable level of trust. More generally, systems architects' developers should focus more on reconciling authentication and privacy goals when designing, developing, and deploying systems. Understanding security needs and developing appropriate threat models are keys for determining whether and what authentication are necessary and what kind is needed. According to [1], [3] the context-aware authentication service has to hold the following distinguishing properties:

Context-Awareness: A context-aware service has to use context data to provide relevant services to users. The security system adapts itself to match with the dynamism of context information. It also has to be able to prune its services accordingly to changes in context data, such as changes in time, location, activity, etc. Therefore, it is critical to check the authenticity and integrity of the context data from context-providers.

Autonomy: The context-aware service should involve the last human intervention possible. The security may improvise new policies based on the available or new context data.

Scalability: The authentication service has to be capable of bootstrapping trust and authentication across heterogeneous domains.

Flexibility: In an open, massively distributed, pervasive computing system, using different means of authentication should be made possible, and it does not have to be constrained to a specific format. Therefore, the system has to be able to provide a great level of customization to each individual.

Privacy-Preserving: In a context-aware environment, there will be thousands of sensors recording every type of important information about users. They will silently track user's location, preferences, and activities in the environment. Therefore, protecting privacy of the user is important, and there has to be a provision to protect it against abuse.

Anonymity: The real identity of a user should never be revealed from the communications exchanged between the user and a server unless it is intentionally disclosed by the user. Different communication sessions between the same user and service should not be linkable. Different devices of user should not be linkable.

Context privacy: Except users want to disclose their context information (location, time, preference, name of services, etc), no one should know about such information even system administrator or service providers they interact with.

Confidentiality and integrity: System should provide protection measures on the communication channels while


89

users are interacting with services in order to protect sensitive information from eavesdroppers.

Nonlinkability: Ideally, nonlinkability means that, for both insiders (i.e., service) and outsiders, 1) neither of them could ascribe any session to a particular user, and 2) neither of them could link two different sessions to the same user.

2.3 Trust Management Trust in pervasive computing is a complex subject relating to belief in the honesty, trustfulness, competence, and reliability of an entity. In the context of pervasive computing, trust is usually specified in terms of a relationship between a resource or service requester and a resource or service provider [1]. To trust pervasive computing systems, we must be able to manage the privacy, confidentiality, availability, and controlled access to digital information as it flows through the system. Trust forms the basis for allowing a requester to use services or manipulate resources owned by a service provider. Also, it may influence a requester's decision to use a service or resource from a provider. So trust is an important factor in the decision-making process.

For trust establishment in the pervasive computing environments, the mobility and uncertainty of the systems and clients need more dynamic and flexible trust strategies. In addition to the traditional trust strategies such as access control and PKI, other trust strategies are proposed and used for trust establishment and management in pervasive computing environments [1]. These trust strategies are:

Trust Negotiation: Is needed when system does not have the client information and there is no third party to consult with on the trustworthiness of the client. In this case, it is only reasonable and practical for the client and system to build their trust relationship by disclosing their credentials gradually to meet the access control policies of each other.

Trust Delegation: Is needed when one entity in the system trusts the client and can assign its rights to the clients.

Trust Based on Reputation: Is used when the system can derive the clients' trustworthiness from the client's behavior records. Because the system may need to collect the clients' reputation from other peer systems, the trust level of the network and the peers systems are taken into account when deciding the trust reputation of the clients.

Trust Based on Context and Ontology: Can be use when clients and the systems may have the smart sensing devices. This ontology information can help the system to determine the trust levels of its clients or assign them trust rights in the given context.

3. Mathematical Backgrounds: In this section we briefly introduce some mathematical backgrounds necessary for the description of our scheme.

3.3 Elliptic Curve Cryptography, (ECC): Many researchers have examined elliptic curve cryptosystems, which were firstly proposed by Miller [18]

and Koblitz [19]. The elliptic curves which are based on the elliptic curve discrete logarithm problem over a finite field have some advantages than other systems: the key size can be much smaller than the other schemes since only exponential-time attacks have been known so far if the curve is carefully chosen [20], and the elliptic curve discrete logarithms might be still intractable even if factoring and the multiplicative group discrete logarithm are broken. In this paper we use an elliptic curve E defined over a finite field pF . The elliptic curve parameters to be selected

[21,22] are: 1 -Two field elements a and pFb ∈ , which define the

equation of the elliptic curve E over pF

(i.e., baxxy ++= 32 ) in case 4≥p , where

0274 23 ≠+ ba . 2 -Two field elements px and py in pF which define a

finite point ),( pp yxP of prime order in )( pFE ( P is not

equal to O , where O denotes the point at infinity). 3 -The order n of the point P . The Elliptic Curve domain parameter can be verified to meet the following requirements [21] and [22]. In order to avoid the Pollard-rho [23] and Pohling-Hellman algorithms for the elliptic curve discrete logarithm problem, it is necessary that the number of pF -rational points on E ,

denoted by )( pFE# , be divisible by a sufficiently large

prime n . To avoid the reduction algorithms of Menezes, Okamoto [29] and Vanstone [24] and Frey and Ruck [25], the curve should be non-supersingular (i.e., p should not

devide ))(1( pFEp −#+ . To avoid the attack of Semaev

[26] on pF -anomalous curves, the curve should not be pF -

anomalous (i.e., pFE p ≠# )( ).

3.4 Bilinear Pairing: This section briefly describes the bilinear pairing, the BDHP and CDHP assumptions. Let 1G and 2G denote two groups

of prime q , where 1G is an additive group that consists of

points on an elliptic curve, and 2G is a multiplicative group of a finite field. A bilinear pairing is a computable bilinear map between two groups, which could be the modified Weil pairing or the modified Tate pairing [27]-[28]. For our proposed architecture within this paper, we let e denote a general bilinear map 221: GGGe →× which has the following four properties:

1-Bilinear: if 1,, GRQP ∈ and *qZa ∈ ,

),().,(),( RQeRPeRQPe =+),().,(),( RPeQPeRQPe =+ and

aQPeaQPeQaPe ),(),(),( == .


90

2 -Non-degenerate: There exists 1, GQP ∈ , such that

1),( ≠QPe . 3 -Computability: There exist efficient algorithms to

compute ),( QPe for all 1, GQP ∈ .

4 -Alternative: 1),(),( −= PQeQPe . Definition 1 -The bilinear Diffie-Hellman problem

(BHDP) for a bilinear pairing is defined as follows: Given

1,,, GcPbPaPP ∈ , where cba ,, are random numbers

from *qZ , compute 1),( GPPe abc ∈ . BDHP assumption:

The BDHP problem is assumed to be hard, that is, there is no polynomial time algorithm to solve BDHP problem with non-negligible probability. Definition 2 -The computational Diffie-Hellman problem (CDHP) is defined as follows: Given 1,, GbPaPP ∈ ,

where ba, are random numbers from *qZ , compute

1GabP ∈ . CDHP assumption: There exists no algorithm running in polynomial time, which can solve the CDHP problem with non-negligible probability.

4. Related Work: Basic Related Work: Recently, many papers have been

published to address mechanisms designed against security, privacy threats, and trust in pervasive computing environments. However, most of these designs fall in the scope of establishing a general security framework identifying general security and privacy requirements. Some of these efforts focused on designing security infrastructures to protect users' personal information such as Mix-Network architecture, Mist system, Aware Home Architecture, Solar, etc. Others focused on designing identity management approach. Some efforts focused on providing privacy control through integrating privacy preferences (P3P), policies and context-aware systems. Various trust management strategies including, trust negotiations and trust establishments, have been proposed to prevent unauthorized disclosure of any relevant information that can be used for inferring sensitive credentials. Based on most important relevant schemes, a full exhaustive comparison study of the most important features is provided in [1] and summarized in Table 1.

Table 1: Protocols Security Features Comparison (P: Partially, H: High, M: Medium, N.A: Not Available)

MA UCP NL LA DCI DS CA TM RA Mist[6] P. N.A Yes H. Yes No No No No

Aware H[9] Yes Yes N.A N.A Yes No No No No Solar[8] N.A No N.A N.A N.A N.A No No No

PawS[10] P. Yes N.A H. Yes No Yes No No Jend02[7] No No No M. No Yes No No No He04[5] Yes Yes No M. No No Yes No No

Ren05[48] Yes Yes P. H. Yes Yes Yes No No Ren06[49] Yes Yes P. H. No Yes Yes No No Kim07[4] Yes Yes Yes H. Yes Yes Yes No No Ren07[50] Yes Yes Yes H. Yes Yes Yes No No

Fire04[47] No Yes N.A N.A N.A No Yes Yes No Dim04[45] No N.A N.A N.A Yes No Yes Yes Yes Dim05[46] No No N.A N.A N.A No No Yes Yes

Yuan06[43] No Yes N.A N.A N.A N.A Yes Yes No Yuan06[44] No Yes N.A N.A N.A N.A Yes Yes No Ries07[42] No Yes N.A N.A N.A N.A Yes Yes Yes Xu07[41] No No N.A N.A N.A N.A No Yes Yes

Uddin08[39] No Yes N.A N.A N.A N.A Yes Yes No Mohan08[40] No Yes N.A N.A N.A N.A Yes Yes No

The comparison is done based on privacy and security

related features. The following comparison cover these features includes Trust Management (TM), Context-Awareness (CA), Mutual Authentication (MA), User Context Privacy (UCP), Non-Linkability (NL), Data Confidentiality and Integrity (DCI), Differentiated Service Access Control (DS), Level of Anonymity (LA), Quality of Privacy (QoP), and Risk Awareness (RA).

Closely Related Work: Authors, in [11], have defined a

model that uses contextual attributes to achieve an approach to authentication that is better suited for dynamic, mobile computing environments. They examined the use of trusted platforms to provide assurances for these contextual attributes. Although authors claimed that their model provides a seamless and flexible user experience that can protect privacy and reduce administrative overhead, it does not provides trust and reasoning and there no mention about how to protect privacy (i.e, user, attributes, and data privacy). Marc Langheinrich [10], introduces a privacy awareness system that allows data collectors to both announce and implement data usage policies. The announced data collections of each services and their policies is delegated by a mobile privacy assistant to a personal privacy proxy residing on the platform, which interact with corresponding service proxies and inquires their privacy policies (Privacy Beacon). Corner et al. [12] describe Transient Authentication as a means of authenticating users with devices through a small, short-ranged wireless communications token. This research is limited to the use of location-based context (i.e., proximity) as an attribute in authentication.

A similar approach is taken by Glynos et al. [14] where they combined traditional authentication with a limited set of contextual information used to identify users. Other similar approaches were taken by [52]-[53] where they also have used a limited set of attributes to perform authentication process. However, we have presented a more generic approach that allows any attributes to be used for authentication. Creese et al. [13] present a general overview of security requirements for authentication in pervasive computing and discuss how traditional authentication does not fit these requirements. Although they discuss authentication of entities using attributes, they did not present a framework for authentication as we have done. In [15], Authors present a service provision mechanism which can enable effective service provision based on semantic similarity measure with the combination of user profiles and situation context in WLAN enabled environment. The paper suggests the combination of user profiles and contextual


91

information to provide a more pervasive service experience in smart assistive environments with mobile device. Behzad et al. [16] propose a framework to construct a context-aware authentication system. Although the framework is flexible and privacy preserving, it is not context-aware user authentication and does not support user trustworthiness evaluation neither user role assignment. Moreover, the framework is designed to be applicable to Ad-Hoc network, does not provide users a way to control attributes, and not suitable for static environments where users may be holding RFID tags only. In [17], authors propose an authentication scheme for a mobile ubiquitous environment, in which the trustworthiness of a user’s device is authenticated anonymously to a remote Service Provider (verifier), during the service discovery process. However, the scheme does not provide support for contextual information, and does not support fuzzy private matching.

5. Toward A New Solution: Here, we outline our proposed authentication-based privacy enhancing infrastructure. Our framework is based on a privacy control layer, a context-aware authentication broker, a context-aware Access Control broker and the use of attributes-based private set intersection and trust evaluation engines. Our framework is a layered architecture that discriminates service providers (context consumers), privacy control process, authentication process, access control process, service receivers (context producers) and the borders that separate these layers. The figure below (Figure 1) shows the process of granting access to resources with the help of user and attributes. Attributes can contain identity and other contextual information (i.e user's profile).

Figure 1. Context-Aware Framework

In our framework, we design an integration scenario

where mobile subjects (i.e users) carrying embedded devices (i.e., smart phones, PDA, etc.) receive pervasive services according to their identity and real-time context information environments. The cornerstone of our framework is the flexibility to provide authentication and access control for independent and dependent (With a special need) people both at context level and where privacy is preserved. Moreover, our framework provides a distributed infrastructure that allows the tracking of the context in a real-time manner. In the following sections, we detail the functionality of these components and describe how they

interact with one another. A high-level overview of these logical components and how they interact is given in following figure 2.

Figure 2. A High Overview of the Framework Our model is based on contextual information obtained

from a distributed network of sensors. In the following we will detail the functionality of these components.

6. Context-Based Authentication Scheme The dynamic nature of a context-aware environment necessitates the need for a very active, flexible authentication mechanism that allows users to securely authenticate and access services with a reasonable level of trust and while privacy is preserved. Our framework consists of the following layers: A Privacy Control Layer (Layer 1) for providing users a way for controlling privacy over the reveal of their personal and contextual information. An access layer, (Layer 2) which combines authentication process (SubLayer 2.1) and access control (SubLayer 2.2) both at context-aware level. The authentication process (Figure 3) contains a trust process (Figure 4 ) where the trustworthiness parameters value are computed in order to provide access to users, and contains a private set intersection process (PSI). In the following sections, we detail the functionality of these processes and describe how they interact with one another. In this section, we present the access process architecture scheme. Figure 3 shows the authentication process architecture. The purpose of access process is to provide authentication and access control according to user's profile and environment (attributes-based authentication and access-control) and then to establish a secure communication link between entities, whilst preserving the privacy of users. Moreover, we will introduce context-aware based user trustworthiness and role's required trustworthiness and show how to improve user assignment and role activation.


92

Our framework is composed of various mechanisms that altogether yield a flexible, scalable context-aware based authentication. In our model, confidence and trust are defined based on each user's contextual information. First, we introduce the system parameters initialization used for the protocol process. Next, we state the different phases upon which the scheme is based. Finally, we describe the operation of the architecture.

Figure 3. The Authentication Architecture Process

6.1 The Scheme: Our infrastructure involves a context-based authentication process, a context-based access control process, a Trusted Key Generation Center (TKGC ), embedded devices EDs , Service Providers ( SP ), Inference engines IEs ,

and users denoted by ( iU ). The TKGC chooses two primes

order group 1G and 2G of prime order q . q is a prime which is large enough to make solving discrete logarithm problem in 1G and 2G infeasible. The TKGC chooses

G as a generator of 1G , chooses Map-To-Point/Curve

function H and chooses e where e is the bilinear pairing map. The TKGC compute GsPTKGC .= , where *

qZs ∈

is the TKGC 's private master key and keep s secret. We define each user as ),( rai AKIDU = , where ID is a user

identity information and raAK is a set of assigned keys corresponding to the roles assigned to the user defined as

{ }nIDrIDrra KKAK ,....,

1= . For each user iU to be

registered, the TKGC calculates iQ , where iQ is user's

partial public key with )( ii IDHQ = , determines iU 's

partial private key ii QsS .= and calculates SPQ , PSIQ

and TEQ which are the framework entities' partial public keys. Moreover, the TKGC calculates a user's or an entity's

public key [30] as GsxPxP uPubuU ... == , where *qu Zx ∈

is generated on user's or entity's behavior. In addition, we define a role as a set of pair of public and private keys belonging to the role. Each role is represented as ),( privpub rrr = . When a role ir is added to the system,

the TKGC picks a random irpk as ir 's private key and

sets GrpkRPK ii .= as ir 's public key. To assign the role

ir to a user with an identity ID , the TKGC check the

user ID , computes )(IDHQID = , and generates the

user's assigned key riIDK corresponding to ir with

)(. IDQrpkK iriID = and where irpk is the ir 's private key. Finally, the TKGC sends iS , iP , Z and the set of

{ }TEPSISP QQQQ ,,= to the user via a secure channel.

The User-Based Authentication Engine UBAE manages an stores, for each user iU with an ED , a record pair

consisting of 21,,, ssSQ ii , where ),( 21 ss are the

prover’s secret. The Table 2, below, shows the mathematical parameters that are used in our proposed framework.

Table 2: Mathematical notations Index Explanation TKGC The trusted key generation center

1G An additive group with prime order q

2G A multiplicative group with prime order q

G A generator of 1G

PubP The public key of TKGC, GsPPub .= s Chosen from *

qZ by TKGC, s and kept secret

iID The identity of the user i, { }*1,0∈iID

iS The long term private key of user i, ni ≤≤1

iQ The long term public key of user i, )( ii IDHQ = where H is a Map function

21, HH Hash functions

H A map to curve algorithm where an ID is mapped into a point on 1G

e A bilinear pairing map qp, Large prime numbers, where 1.2 += qp

QP, Random points over elliptic curve

ba , Random generated private keys

E Non-super singular elliptic curve

B )( pFE with order q

)(Qx x coordinate of point Q


93

In the following, we will propose our model to achieve attribute-based authentication. In our architecture, end-users can interact with the infrastructure (e.g. walking into a room, entering the subway system using smart phone, PDA, etc). The infrastructure provides a set of resources generally tied to different geographical areas, such as printers, surveillance cameras, campus-based location tracking functionality, and so on. These resources are all modeled as services that can be automatically discovered based on different relevant mechanisms which are out of our band. Our Authentication scheme involves two distinct phases: the Join Phase, and the Mutual Authentication Phase. We will describe the various interactions that take place between the entities described in our logical system model. We refer our readers to (Figure 2) for a comprehensive high level overview of our framework model.

Join Phase: The purpose of this phase is to automatically provide services to users through a context-based provision process. In our attributes-based authentication, we aim to have a service provision framework that combines user's profiles and contextual information to select appropriate services to the end users from thousands of desultory services. In order to achieve our contributions, we firstly have adopted the framework proposed by Qin et al. [31] that automatically provide appropriate services to the right person with the right form with the relevant consideration of contextual information in smart environment. Moreover, we took the assumption that the proposed protocol in [31] is extended to add two new context type fields which will be executed during the provision process. The first context type is related to users with special needs equipped with a body network sensor. This context type is collected by a BNS adapter and translated to the provision protocol in order to be proceeded. The second context type is related to a Meta classification process which will be helping in well selecting services. Once the service provider (SP) has initiated the context-aware service provision process, we can go a step forward to start the Authentication Phase.

Authentication Phase: Service discovery typically

involves the exchange of service advertisement and service reply messages between the user and service provider. To avoid increasing the communication overheads, we incorporate our extended previous authentication mechanism into these messages. In other words, service discovery and authentication can take place concurrently. We now examine how these messages are constructed to achieve our aim of attributes-based authentication.

):(From RoundFirst TheWithin EDSP →⇒ :

Our Attributes-based authentication model will start with a service provider engine advertising available context-aware services to the end user, clients iC , as indicated in (1).

SP Services AwareContext Advertise iC (1) For example, a location-based service allow providers to

advertise its services to any user within a certain acceptable proximity. The advertised service announcement contains the following: A Universal Resource Locator (URL), that could allow a client to locate and access the advertised access. Authentication Requirements ( AR ), allowing clients to package their access request with the necessary authentication credentials and contextual information. The exchange of traffic between the Service Provider ( SP ), the user iU , and inference engines is based on an extension for

our previous work [32]. For the SP to construct and send the authenticated services advertisement message, he will be performing the following: The SP starts the protocol by

generating two fresh random nonce 1r and nZr ∈2 , then

he calculates the point X where 2211 PrPrX ×+×= .

Next, SP constructs the service advertisement message as in (2):

)),,....,,(,( 21 XsrvsrvsrvQAdv nsp= (2)

Where { }isrvsrvsrv ,...,, 21 represent the set of available suitable context-aware services defined in the first phase (Join~Phase). Finally, the service provider encrypts and sends the Adv message to the embedded device ED , as given in (3):

),),,....,,(,( 21 URLXsrvsrvsrvQE nspKe (3)

In our framework and hereafter, any two entities denoted by X and Y, can directly compute a partial private shared key between them without exchanging any previous message. Based on the one's own partial private key and the other party's partial public key, they can directly compute the share key as follows. We denote their partial private key/public key by xx QsS .= , where )(1 xx IDHQ = and

by yy QsS .= , where )(1 yy IDHQ = . The nodes X and

Y then compute ),(/ yxyx QSeK = and

),(/ yxxy SQeK = , respectively. And finally the private

shared key will be eK where:

[ ]syxyxYX QQeHKHK ),()( 2/2/ ==

eXY KKH == )( /2 This approach is very efficient in terms of communication and computation and this feature makes it very attractive to the environments where the entities capabilities are limited.

):(From Round Second TheWithin SPED →⇒ :

After receiving the advertised service announcement, the client iC decrypt the message and retrieve the credentials. Suppose that the client is interested in an advertised service

isrv , (i.e, request access to perform an operation O on

service isrv from the service provider), he will be

performing the following: As isrv is a context-based


94

resource, iC is promoted to present not only identity credentials but also all the required contextual information and bundle them with the access request that is sent to iSP . In our attribute-based authentication model, authentication requirements are dynamic and can vary dramatically from one access to the next. Moreover, we must expect that some attributes will be generated by the user while others by the platform. Our model provides the client with a full control over the reveal of the personal information. The option of collecting contextual information attributes from the platform is done by using a Privacy Control Layer PCL . In order to retrieve needed attributes to fulfill the access request, the user issues a service request which is handled by the user agent. The user agent does not directly invoke the service. Instead, it retrieves the privacy policy of the service, without revealing any information about the user. The user agent compares the service's policy to the user's preferences and performs the following:

Based on the user’s preferences: 1-If there is a preference rule that accepts the privacy

policy, then: - Extract the context-dependent preferences from the user’s extended preferences document. - Store an association between the user, the service and the user’s context-dependent privacy preferences in the platform and finally a request for contextual information is issued to the PCL .

2-If there is no accepting rule, or there is a rule that

indicates that the user should be alerted, and then the service will not be invoked. The user is prompted for further evaluation of the policy.

Whenever a request for contextual information arrives at

the privacy control layer, the PCLshould perform the following actions:

1- Check for an association record between the service

that is requesting the contextual information and the user about whom information is requested. If this association does not exist, try to contact the user’s agent and ask it to store an association record.

2-Retrieve and evaluate the context-dependent

preferences referenced in the association: a: If the context -dependent preferences evaluate to ‘true’,

then retrieve the requested information from the context interpreter and return the information to the user agent.

b: If the context-dependent preferences evaluate to ‘false’, then refuse the request for contextual information.

When the PCL is introduced in the infrastructure, the

access request itself is altered to include information that was provided by the PCL . Context-Aware providers will publicize to their users information such as positions, roles,

activities, etc. The validity of these data could be verified by introducing Context Trustworthy Engine, CTE in the framework. This is the role of the authentication broker, using the CTE , to validate these data before starting the authentication process. After receiving relevant reply message from the PCL , the user agent retrieves the set of contextual information received from the attribute provider(s) through the PCL , and performs the following: The queried ED selects the role or the corresponding set of roles denoted by { }hrrrSR ,....,, 21= . Generates the

message Q and calculates the signature QSig on Q with

erpi SRSQ = and where erp is the permission that the

user wants to enforce. The QSig is denoted by VU , . In

addition, ED generates two fresh random nonces f and a ,

where tR Zf 2∈ and *

qZa ∈ , and calculates EDT , where

GaTED .= . For a static context-less system, the user

computes ),( xx TR , where ),( xx TR is the signature pair

over the user's private key iS . This ),( xx TR will be

replacing the couple VU , in equation (6) for the protocol

process run. Finally, the client will package all the collected attributes encrypted (i.e., user's profile and environment's attributes) with needed information in order to be sent to the service provider for authentication process. Let assume that a user iU has received the request set of context-data D

from the PCL . Therefore, the set A, given in (5), denotes

all the attributes that user aU may present to set her rules in the authentication process.

{ } { }jiAPC bbaaAAAii

,...,,,...,, 11==

{ } Dcacacaca jii ⊆= + ,...,,,..., 11 (5) D is the reference set that contains all the attributes a user may hold or the context data received and ca represent the context data collected. Finally, the client packages the final required set of context and attributes that the service provider may use for authentication process and construct the message as described in (6):

iC VUfTAEservQE EDKiCK PSIUie,,,)),(),(,(

/ iSP (6)

Hereafter, these attributes are mapped into integer numbers

ica for li ,....,3,2,1= that is 1ca is a number

representing Name, 2ca is a number representing Location, and so on. Our model is very flexible in that the service provider engine may accept or refuse a subset of attributes in Acorresponding to different level of confidence. If the user

can present all attributes in A required by service provider in order to access for identification, a full confidence will be achieved, otherwise the confidence level will be depending both on PSI 's reasoning process and on the user's requirements by computing user's trustworthiness and role's


95

required trustworthiness. In the next section, we will demonstrate how our scheme could be combined with Timed Fuzzy Logic [33] in order to set a threshold under uncertainty and to account for changes in context-data.

):(From Round Third TheWithin IEsSP →⇒ :

The service provider now has an authentication package, containing the requested context attributes that was provided by the client. The first step requires the SP to decrypt the encrypted message and retrieve the data in order to determine the source and authenticity of these attributes provided by both iU and AP , and later on to complete authentication process. Once the service provider has retrieved the data set from equation (6), the authentication process will be performed as follows: The service provider send the encrypted set A where { }jcacacaA ,...,, 21= to

both PSIE and TE engines, and send { }fVUB ,,= to

UBAE engine. The service provider's platform is composed of the two

main brokers. The authentication process and the access control process. Each of these brokers contains different relevant engines that they interacted altogether provide a flexible, a scalable context-aware authentication framework. For the authentication broker, we have the following engines: a Private Set Interaction Engine ( PSIE ), a Trust Engine (TE ), and a User-Based Authentication Engine (UBAE ). We also have an Identity Based Encryption Engine ( IBEE ) that will be responsible for setting a shared secret key for secure future communications. This IBE protocol will be interacting with the PSI in order to calculate the shared secret key. Moreover, the PSI engine will be interacting with the CTE engine to accomplish the attributes verification process. Therefore, our authentication process decision will be based on the output of these several engines. The description of these engines and their interacting process will be explained in the coming section.

):(From RoundFourth TheWithin SPIEs →⇒ :

Upon receiving the encrypted messages from the service provider, the PSI start the attributes verification process. To verify the source of AP ’s attributes, we have introduced the Context Trustworthy Engine ( CTE ) which is responsible for verifying all attributes provided by APs and other contextual information provided by the client (i.e., case of an RFID or a smart card and a client with special need). The interactions (7) and (8) show the PSI requesting the CTE to verify the validity of the

attributes ica .

PSI )),...,,(,( 21 jSPK cacacaQEie iCTE (7)

PSI sultsonVerificatiQQCTE Re, iCTE (8)

Once PSI determines the verification process of these attributes provided on behalf of the client, it passes the authentication credentials and attributes to the relevant engines that will complete the processing of the client's access request. Each engine will start it's own process as follows:

Description of the PSI Engine: One new component that

will be added to our architecture is the Private Set Intersection Engine ( PSIE ). PSI are cryptographic techniques allowing two or more parties, each holding a set of inputs, to jointly identify the intersection of their inputs sets (i.e, shared context), without leaking any information about credentials that each entity might have. Nevertheless, both entities, the prover and the verifier, need to protect their credentials from each other. Moreover, any entity waiting to be authenticated by a server has to establish enough confidence in it and be able to present the required attributes. Therefore, the conditions that the server sets for authentication become extremely valuable, as they determine the reasoning mechanisms in the authentication protocol. To keep a high level of security, the server needs to keep those attributes private. For this purpose, we make use of the Private Set Intersection ( PSI ). Once the PSIE receives and extract/decrypt the set A of attributes and upon the sender request's selected isrv , the PSIE will

initializes a PSI protocol over the two sets AandisS .

{ }jsrvsrvsrvs iiiiSSSS ,...,, 21= represent the needed set of

contextual information defined by the service isrv

administrator deployment. The isS set reside on a Services

Proxy Server ( SPS ), and the PSI protocol will be initialized between PSI engine and SPS . There are many PSI protocols in the literature. We can adopt the one that was chosen by [33], [34] since it has a provision for approximate matching, referred to as Fuzzy Private Match. The PSI engine performs two kinds of tasks: First, it gives a level of confidence when a user is on an authentication process. It makes use of authentication contextual information to assign the confidence level. Second, it evaluates a Fuzzy Logic Matching protocol queries from applications about whether a certain entity is allowed to access a certain resources. It makes use of applications specific contextual information, the credentials of the entity, and entity's contextual information to decide whether an entity is authenticated and has access to resources. For convenient readerships, we urge our readers who want to go deeper in the theory of Fuzzy Private Matching Protocol and getting acquainted with the principles of PSI theory to refer to [3], [33]. Moreover, the PSI engine will be also interacting with the Identity Based Encryption protocol to calculate the secret shared key. This step will be discussed


96

in the following paragraph: Description of The Identity-Based Encryption Protocol

The IBE removes the need to set and exchange certificates as the message can be encrypted based on the identity of the entities. The identity can be defined as a location, name, email address, time, etc. or a combination of them. The combination of them could be refereed to the context data. For convenient readership, we urge our readers to refer to [35]. In the following, we will describe the details of how PSI interacts with IBE protocol in order to calculate the shared secret key. From the PSI, let

iSSAI be the

intersection set of A and iSS defined above:

{ }iS ddddSAi

,....,,, 321=I (9)

where id denotes the context that are shared between the user and the service provider. Finally, the IBE will calculate and send SPT to PSI engine with GdT iSP ).(∑= .

Description of the Trust Engine: Another new

component that will be added to our architecture is the Trust Engine TE . To trust pervasive computing, we must be able to manage privacy, confidentiality, availability, and controlled access to digital information as it flows through the systems. In the following, we will describe the Trust process design architecture shown in Figure 4. Our ultimate goal is to provide a trust model that is flexible and adaptive for various application scenarios and environments. This approach could be solved using the concept of fuzzy-based trustworthiness.

In this section, a dynamic trust model is formally introduced to incorporate trust strategies in order to first build up the user's and role's required trustworthiness level and than the User Assignment UA trustworthy value. There are several ways and approaches to design trust models. The component-based approach is chosen for our model design because it can be implemented in a distributed way and be extended easily and transparently (i.e, to include later the Risk Assessment Engine). During a real-time trust management process in pervasive computing environments, the trust information may be from different resources at any time. Therefore, our adopted trust model is designed to be able to evaluate the trust information concurrently. Using this approach, the trust engine derives the level trustworthiness of a user UT and role's required trustworthiness RT by using user’s attributes and roles permission, respectively.

Figure 4. The Trust/Risk Process Architecture

The user assignment UA level is performed based on the trust level UT in comparison with the trust level RT . However, our trust model is based on the trust policies, the environment contextual information, and the user’s roles permissions. As a cognitive process, trust is complex and fuzzy. That is, for a special context, we can not easily make a decision about whether to trust an entity or distrust it. Therefore, Our Trust evaluation engine is adopted as a combination from [36] [37] where trust model is provided by integrating trust into a fuzzy logic-based trusted decision upon building the trustworthiness's prediction. For convenient readership of this work, we will briefly describe the trust model process here: Trust establishment can be thought as a process that identifies or verifies the principal's claim against the trust evidence. Trust evidence,

{ }brcecev TTTT ,,= , are further classified into the following categories: credentials, the context of the environments, and behavior records. We define UT as the trust level of the user by using a function F , as given in (10): ),,,(),( brcecpevpres TTTTFTTFUT == (10)

where resF is the function of the trust level of the client to

access the resource and pT is the set of trust policies for the

resources. In our definition, the trust level of the user, UT , for accessing the resource in the system is determined by evaluating the trust evidence against the trust policies for the resource and the user assignment, UA, is evaluated based on UT in comparison with RT . For simplicity, we will consider aattributesev TTT == and finally

)( aTFUT = . F , UT and RT could be calculated using the formal mathematical equations from [36]. Once these parameters are calculated, the trust decision modular will evaluate the user assignment UA based on UT in comparison with RT , and will package the final result in order to be sent to the Authentication Process Decision.

Description of the UBA Engine: Moreover, upon

receiving the encrypted signature pair message VUEeK ,

from the service provider, the UBA engine will decrypt the message, and then verify the signature pair. If it is valid, then the UBA engine accepts, and the pair

),( 21 ss associated with the authenticated ED is extracted from the database server, and encrypted using the Weil-Pairing-based encryption algorithm. Finally, the user based authentication engine packages the encrypted message

),( 21 ssEeK with the evaluated result in order to be sent

to the authentication process decision broker. The authentication process decision will take the decision

based on its different engines evaluation and package the final output result and send it encrypted to the service provider.


97

):(From RoundFifth TheWithin EDSP →⇒ :

Upon receiving the message from the authentication process decision broker, the service provider first decrypts the message and then evaluates the output. If the result is false, he will deny access request to resources, otherwise, if true (i.e., the user is authenticated, the user trustworthiness parameters are acceptable, and the confidence level is acceptable), the service provider extracts the pair ),( 21 ssEKe and then computes:

)))(mod(( nsfry iii ×+= (11)

For 2,1=i and starting packaging the following data

( iy with 2,1=i ) in order to be sent later to the ED . Meanwhile, as the final decision will be evaluated based on both authentication and access control process decision brokers, the user's access request is also subject to context-aware access control rules which will be discussed in the following:

Context-Based Access Control Process: A key challenge

in ubiquitous environment is the design of an effective active access control schemes [36] that can adequately meet the security challenges represented by the system's ability to capture security relevant contextual information, such as time, location, user's profile, or environmental state available at the time the access request are made and to incorporate these information in its access control process. We specify and integrate our own context-aware access control rules definitions to further enhance the security of our proposed authentication-based framework scheme. Moreover, the context directly affects the level of trust associated with a user, and hence the authorizations granted to him. Therefore, we introduce the user trustworthiness and role's required trustworthiness parameters into the design the context-based access control by incorporating them within the development of the context constraints. Conditions on the access control to solve the semantic problem is to check the trust engine parameters UT and UA if they satisfy the condition, the user will be subject to authorization rules and policies based on the available presented attributes. We believe that the introduction for the rules definitions is necessary for providing an adequate authorization decision for any Service Access Request and to accomplish a secure authentication process. Figure 5 shows our extended access control scheme with the rules definitions.

Figure 5. The Extended Access Control Process

In the following, we describe needed rules definitions for

a dynamic context-aware access control infrastructure to fulfill the framework's security requirements:

Rule Definition 1: Dynamic Adjustment. In our

approach, we believe that any pervasive model should dynamically adjust role assignments and permission assignments based on presented context information. Therefore, we consider DRBAC concept [38] where each user is assigned a set of roles and the context information is used to decide which role is active at a time. User will access the resource with the active role. Moreover, each role is assigned a set of permission, where the context information will be used to decide which permission is active for that role. The systems-based context for resources should be taken into consideration, and the security policy for the resources should be able to define a permission transition for a current role.

Rule Definition 2: Context Type. A context type is

defined as a property related to every participant in a service. In simple scenario, context type may be a concrete property familiar in everyday life, such as time or location, etc. However, in a more complex scenario, we believe that context type should be extended to describe more attributes such as user's capability and/or willingness (i.e, case of People with special need equipped with a hidden body network sensor). We define such context type by cCT .

Therefore, based on a complete users' context type iCT , we

can define that each resource ir has its own context set

irCS which is defined as follow:

{ }ncr CTCTCTCTCSi

,....,,....,, 21= (12)

In any access control design to be integrated within our framework, we define two sets of context types, passive and active sets. While the authentication process will be subject to only the active set, the access control decision will be


98

subject to the two sets. Rule Definition 3: Context Constraint. We define our context constraint as a regular expression that is capable of specifying any complex context related constraint to introduce all kinds of security requirements. In general a context set is defined as: Context Constraint iClauseClauseClauseCC ∪∪== .....:: 21 where

jconditionconditionConditionClause ∩∩= ...: 21

and VALUEOPCTCondition =: ,

where CSCT ∈ , OP is a logical operator in the set { }=≠≥>≤< ,,,,, , and VALUE is a specific value of CT . Therefore, we suggest that should be extended to accommodate user trustworthiness UT and user assignment trustworthiness UAas a new clause. The new context constraint will be: newClauseClauseClauseCC ∪∪= .....: 21 (13)

Where ))()(( VALUEUAVALUEUTClausenew ≥∩≥= (14)

As an illustration, suppose we have a context set =CS Time, Location, Authentication Level and we have a

partial security rule such as a patient data can be accessed from within the hospital between 8am and 5pm with a trust level of a password; otherwise a higher level of trust is required.

Rule Definition 4: Authorization Policy. We define an

authorization policy as a quadruple CCOPSAP ,,,=

where S is the subject in this policy, which could be a user or a set of roles, P the mode of operation defined by {READ, APPEND, DELETE, UPDATE, WRITE}, O is a data object and CC is a context constraint defined according to Definition 3.

Rule Definition 5: Resource_Access_Request, denoted

by RAR, is defined as a quadruple

iiii RCOPURAR ,,,= where UserSetU i ∈ ,

SeTPermissionPi ∈ , O is the data object requested, and

context RC is a runtime context set of values for every context type in the context set CS . iRC is defined according to Definition 2 and captured dynamically at the time of the access request.

Dynamic Context Evaluation: Finally, the access control

decision for any service access request

iiii RCOPURAR ,,,= is granted only if there exists

an authorization policy CCOPSAP ,,,= , such that

SU i ∈ , OPOP ii ,, = , and CC evaluated to true

under iRC (that is, when all CTs in constraint CC are

replaced with their available presented values in iRC , then the resulted Boolean expression is true).

⇒ Finally, the service provider evaluate the final access request decision (the one from authentication broker and the other from the access control broker) and packages the results with the relevant data ( SPT , iy for 2,1=i ) and send

it to the user with the embedded device. The ED computes: ))(( ZfPy ii ×+×∑ (15)

and then checks that if ))(( ZfPy ii ×+×∑ is equals to

X , if so the ED accepts and extract the shared secret key in order to be used for encrypting future communications, else rejects.

After the above messages, EDT and SPT are exchanged,

the service provider and the user can agree and compute the secret shared key:

),,(.),(/ EDspspb

EDEDEDSP TSxePQeK = (16)

and ),,(.),(/ SPeded

aSPSPSPED TSxePQeK = (17)

respectively. We denote by SPEDEDSP KKK // == , the key shared between the entities. To ensure forward security, we can use the new shared key hK after applying a hash

function to K . Once the protocol run completes

successfully, both parties may use hK to encrypt subsequent session traffic in order to create a confidential communication channel. In the following we will present a brief verification regarding the similarity of the shared key equations:

),,(.),(/ EDspspb

EDEDEDSP TSxePQeK =

).,,(.)..,( GaSxeGsxQe spspb

edED=

).,.,(.).,..( GaQsxeGbQsxe SPspEDed= a

SPSPSPeded PQeTSxe ),(.),.(=

SPEDK /=

7. Protocol Analysis: Security Analysis: Our proposed architecture is

considered to provide privacy and anonymity for users. In the following, we evaluate our architecture regarding the security and privacy requirements.

Mutual Authentication: Considering the fact that the

digital signature pair ),( VU , created by the ED , is verified

by the Back-end server. Considering that the pair ),( 21 ss ,

sent by the back-end server UBAE , is recalculated by the


99

service provider under ),( 21 yy and verified by the ED . Therefore, our proposed architecture guarantees the secure mutual authentication between the embedded device ED and the back-end server.

Passive attack: Suppose an attacker performs a passive

attack, then the session will terminate with both legitimates parties accepting. That is, the two parties successfully identify themselves to each other. And regarding the fact that the exchanges messages between the service provider and the ED are generated from random nonce which are generated with every new session, so it is infeasible that an attacker computes any useful information including the iID

of a user iU . Therefore the architecture resists against the passive attack.

Man in the middle attack (or active attack): Suppose that

an attacker intercepts X and replaces it with 'X , the attacker then receives f and ),( VU from the ED . He

would like to replace the pair with )','( VU , as before. However, and unfortunately for the attacker, he can not compute the value of the new pair because he does not know the users credentials and parameters and because the transmitted messages are meaningless. Therefore the proposed scheme thwarts the man in-the-middle attack.

Perfect forward secrecy: Each run of the protocol

computes a unique X , a unique Signature pair ),( VU and

a unique pair ),( 21 yy . In addition the transmitted messages are meaningless as they are generated for each new session using new random nonce. Thus, the architecture is secure against perfect forward secrecy.

Data Confidentiality: Since our architecture provides

secure mutual authentication between the ED and the system and since the information transmitted between the EDand system is meaningless, thus, our architecture provide data confidentiality and the user privacy on data is strongly protected.

ED Anonymity and Location Privacy: During the

authentication processes, a signature algorithm is used to produce the signature pair ),( VU . The pair ),( VU and f

that are transmitted between the ED and SP are randomized and anonymous since they are updated for each read attempt. Thus, our architecture provides user anonymity and location privacy is not compromised.

Unauthorized SP Detection: Our Proposed architecture is

based on the insecure communication channel between SP and UBA engine. The unauthorized 'SP is detected and prevented by the back-end server IDUBAE using the weil pairing based encryption algorithm between the service

provider and the back-end server, and by verifying the pair ),( 21 yy by the legitimate user or ED . Thus, our scheme

protects against unauthorized service provider. Protocol Correctness: We can choose one of many

identity-based signature scheme to compute the QSig .

Therefore, we will adopt the signature scheme that was used by [51]. To compute the QSig , the user selects random

*qZr ∈ , computes IDQrU .= , computes ),( UQHh = ,

,computes )( 1 iIDrhiSR KK =∑= , and finally computes

SRKhrV )( += . The validity of QSig can be

accomplished by verifying if ),(?),( IDAR hQUPeVPe += . The Proof is given below:

),(),( 1 IDIDi

kiIDAR hQrQPehQUPe +∑=+ =

))(,.( 1 IDki QhrPse +∑= =

))(,( 1 IDiki QshrPe =∑+=

))(,( IDARShrPe += ),( VPe=

As a summary, a context-based access is developed and it

can thus be granted to both known and unknown agents. The integration for the IEs engines, the extension for the context-based access control definitions, and the development of IBE engine form the core of our context-based authentication framework where every request is authenticated and filtered in order to remove any unauthorized actions. After filtration the service provider can evaluate the request and create an appropriate response depending on the contextual information.

8. Conclusion: In this paper, we have proposed a usable dynamic

authentication framework for pervasive computing environments. We hope that our work can raise interests on the design of authenticated-base framework dedicated for areas with specials needs. From the above arguments, we can see that our context-aware access control rules provide an effective second-line defense against attacks on authentication. We have identified security and privacy threats that may arise during access services in a pervasive computing environment; we also derived corresponding security and privacy requirements. We have presented our attributes-based authentication scheme, using Trusted Computing Functionality, which preserves user privacy. The scheme also satisfies all the identified security requirements. To a user and service provider, security and privacy are both desirable. However, they are potentially conflicting requirements, and it is challenging to achieve them both. However, this is achieved by our attributes-based authentication scheme presented here, enabling secure


100

access to services while privacy is preserved. In the coming future work, we are working on introducing risk assessment, quality of privacy within our framework. Moreover, we are working on an implementation for the proposed model into the platform at the lab of Telecom SudParis (ex. INT).

9. Acknowledgment: The authors would like to thank Telecom SudParis and

Lebanese University Laboratories’ staff for their contributions and support. I would like to thank everyone for his help, guidance, advice as well as his enthusiasm and many valuable contributions to this work. Their suggestions and observations were extremely helpful throughout this paper. Without their input, I would not have been able to complete this work.

References [1] P. ABI-CHAR, A. Mhamed, B. EL-Hassan and M.

Mokhtari, “Controlling Trust and Privacy in Context-Aware Environments, State of Art and Future Directions”, In Trust Modeling and management in Digital Environments: From Social Concept to System Development. Book Edited by Nokia Research Center, Finland, January 2010.

[2] K. Dey and D. Abowd, “Towards a better understanding of context and context-awareness. The CHI 2000 Workshop on the What, Who, when, and How of Context-Awareness”, The Hague, Netherlands, Apr. 2000,ftp://ftp.cc.gatech.edu/pub/gvu/tr/1999/92-22.pdf.

[3] B. Malek, A. Miri, and A. Karmouch, “A Framework for Context-Aware Authentication”, In 2008 IET 4th International Conference on Intelligent Environments. IEEE Computer Society Press, pp. 1-8, 2008.

[4] J., Kim, Z. Kim, and K. Kim, “A lightweight Privacy Preserving Authentication and Access Control Scheme for Ubiquitous Environment”, In Proceeding of the 10th international Conference on Information Security and Cryptography, ICISC07, pp. 37-48, 2007.

[5] Q. He, L. Wu, and P. Khosla, “for Personal Control over Mobile Location Privacy”, In IEEE Commun. Mag. vol. 42, no. 5, pp. 130-136, 2004.

[6] J. Al-Muhtadi, R. Campell, A. Kapadia, M. Mickunas, and S.Yi, “Routing Through the Mist: Privacy Preserving Communication In Ubiquitous Computing Environments”. In Proceedings of the International Conference of Distributed Computing Systems (ICDCS02), pp. 65-74, 2002.

[7] U. Jendricke, M. Kreutzer, and A. Zugenmair, “Pervasive Privacy with Identity Management”. In Proceeding of. 1st Workshop Security, UbiComp 2002.

[8] K., Minami, K., and D., Kotz, “Controlling Access to Pervasive Information in the "Solar" system”.

Dartmouth Computer Science Technical Report TR2002-422, February 28, 2002.

[9] M., Covington, W., Long, S., Srinivasan, A., Dey, M., Ahamad and D.G., Abowd, “Securing Context-Aware Applications Using Environments Roles”. In Proceeding of the sixth ACM Symposium on Access Controls Models and Technologies, pp. 10-20, 2001.

[10] L. Marc, “A Privacy Awareness System for Ubiquitous Computing Environments”. In Proceeding of the 4th International Conference on Ubiquitous Computing, (UbiComp) pp. 237-245, 2002.

[11] J.M. Covington, M. Sastry, and D.J. Manohar, “Attribute-Based Authentication Model for Dynamic Mobile Environments”. In Proceeding of the Third International conference on Security in Pervasive Computing (SPC), York, UK, pp. 227-242, 2006.

[12] M.D. Corner, and B.D. Noble, “Protecting Applications with Transient Authentication”. In Proceeding of the First International Conference on Mobile Systems, Applications and Services, pp. 57-70, 2003.

[13] S.J. Creese, M.H. Goldsmith, and B.R. Zakiuddin, “Authentication in Pervasive Computing”. In Proceeding of the First International Conference on Security in Pervasive Computing (SPC), 2003.

[14] D. Glynos, P. Kotzanikolaou, and C. Douligeris, “Preventing Impersonation Attacks in MANET with Multi-Factor Authentication”. In Proceeding of the Third International Symposium on Modeling and Optimization in Mobile Ad-hoc, And Wireless Networks, pp. 59-64, 2005.

[15] W. Qin, Z. Daqing, M. Mounir, S. Yuanchun, and D. Kejun, “Combining User Profiles and Situation Context for Spontaneous Service Provision in Smart Assistive Environments”. In Proceeding of the 5th international conference on Ubiquitous Intelligence and Computing, pp. 187-200, 2008.

[16] M. Behzad, M. Ali, and K. Ahmed, “A Framework for Context-Aware Authentication”. In Proceeding of 2008 IET 4th International Conference on Intelligent Environments, pp. 1-8, 2008.

[17] L. Adrian, and M.J. Chris, “Ninja: Non Identity Based, Privacy Preserving Authentication for Ubiquitous Environments”. In Proceeding of 9th International Conference on Ubiquitous Computing, (UbiComp2007), Springer LNCS 4717, pp. 73-90, 2007.

[18] V. Miller, “Uses of elliptic curves in cryptography”. In Proceeding of Crypto '85, Santa Barbara, pp. 417 - 426. 1986.

[19] N. Koblitz, “Elliptic Curve cryptosystems”. Mathematics of Computation, vol. 48. pp. 203 - 209, 1987.

[20] N. Koblitz, “CM-Curves with Good Cryptography Properties”. In Proc. of Crypto' 91, Santa Barbara, USA, 1992.

[21] L. Law, A. Menezes, M. Qu, J. Solinas and S. Vanstane, “An efficient Protocol for Authenticated Key Agreement”. Technical report CORR98-05, Department of CO, University of Waterloo, 1998.


101

[22] L. Law, A. Menezes, M. Qu, J. Solinas and S. Vanstane, “An efficient Protocol for Authenticated Key Agreement”. Designs, Codes and Cryptography, vol. 28, pp. 119-134, 2003.

[23] J. Pollard, “Monte Carlo methods for index computation mod p”. Mathematics of Computation, vol. 32, pp. 918-924, 1978.

[24] A. Menezes, T. Okamoto and S. Vanstane, “Reducing elliptic curve logarithms in a finite field”. IEEE Transactions on Information Theory, vol. 39, pp. 1639-1646, 1993.

[25] G. Frey and H. Ruck, “A remark concerning m-divisibility and the discrete logarithm in the divisor class group of curves}, Mathematics of Computation, vol. 62, pp. 865-874, 1994.

[26] I. Semaev, “Evaluation of Discrete logarithms in a group of p-torsion points of an elliptic curve in Characteristic p”. Mathematics of Computation, vol. 67, pp. 353-356, 1998.

[27] D. Boneh and M. Franklin, “Identity-based encryption from the Weil Pairing”. Advanced in CRYPTO2001, LNCS 2139, pp. 213-229, 2001.

[28] G. Frey, M. Muller and H. Ruck, “The Tate Pairing and the discrete logarithm applied to elliptic curve cryptosystem”. IEEE Transaction on Information Theory, Vol. 45, No.5, pp. 1717-1719, 1999.

[29] D. R. Stinson, “Cryptography Theory and Practice”. In Proceeding of Chapman and Hall/CRC, Third Edition, pages: 353-438, 2006.

[30] S. Wang, Z. Cao, and H. Bao, “Efficient Certificateless Authentication and Key Agreement (CL-AK) for Grid Computing”. In Proceeding of the International Journal of Network Security, vol.7, No.3, pp. 342-347, 2008.

[31] W. Qin, Z. Daqing, M. Mounir, S. Yuanchun, and D. Kejun, “Combining User Profiles and Situation Context for Spontaneous Service Provision in Smart Assistive Environments”. In Proceeding of the 5th international conference on Ubiquitous Intelligence and Computing, pp. 187-200, 2008.

[32] P. ABI-CHAR, M. Mokhtari, A. Mhamed and B.~EL -Hassan, “Towards a Robust Privacy and Anonymity Preserving Architecture for Ubiquitous Computing”. In Proc. of the Third International Conference on Risks and Security of Internet and Systems (CRISIS08). Tozeur, Tunisia, IEEE Computer Society Press, October 28-30, pp. 125-132, 2008.

[33] J.F. Michael, N. Kobbi, P. Benny, “Efficient Private Matching and Set Intersection”. In Proceeding of the Advances in Cryptology-Eurocrypt'04', vol. 3027 of lectures Notes in Computer Science, Springer-Verlag, pages 1-19, 2004.

[34] M. Behzad, M. Ali, and K. Ahmed, “A Framework for Context-Aware Authentication”. In Proceeding of 2008 IET 4th International Conference on Intelligent Environments, pp. 1-8, 2008.

[35] B. Dan, and F. Matthew, “Identity-Based Encryption from the Weil Pairing”. In Proceeding of the SIAM Journal of Computing, 32(3):586-615, 2003.

[36] T. Hassan, A. Morteza, and J. Rasool, “Trust-Based User-Role Assignment in Role-Based Access

Control”. In Proceeding of the IEEE/ACS International Conference on Computer Systems and Applications (AICCSA), pp. 807-814, 2007.

[37] D., Xiu, and Z., Liu, “A Dynamic Trust Model for Pervasive Computing Environments. A Research Paper}, Research Supported by the NSF0406325, coitweb.uncc.edu/~zhliu/Research/Papers/asc.pdf, 2004.

[38] G. Zhang, and P. Manish, “Context-Aware Dynamic Access Control for Pervasive Applications”. In Proceedings of the Communication Networks and Distributed Systems Modeling and Simulation Conference, (CNDS'04), USA, 2004.

[39] M. Uddin, M. Zulkernine, and S. Ahamed, “CAT: a context-aware trust model for open and dynamic systems}, In Proceedings of the 2008 ACM symposium on Applied computing, pp. 2024-2029, 2008.

[40] A., Mohan, and M., Blough, “Attribute Trust-a Framework for Evaluating Trust in Aggregated Attributes via a Reputation System”. In proceeding of the 6th Annual Conference on Privacy, Security and Trust. IEEE Computer Society Press, pp. 201-212, 2008.

[41] W., Xu, T., Xin, and G., Lu, “A Trust Framework for Pervasive Computing Environments”. In the proceeding of the International Conference on Wireless Communications, Networking and Mobile Computing (WiCom07), IEEE Society Press, pp. 2222-2225, 2007.

[42] S., Ries, “Certain Trust: A Trust Model for Users and Agents”. In Proceeding of the 22nd Annual ACM Symposium on Applied Computing, ACM Press, pp. 1599-1604, 2007.

[43] W., Yuan, D., Guan, S., Lee, Y., Lee, and H., Lee, “Filtering out unfair recommendations Finding for trust model in ubiquitous environments”. In proceeding of the second International Conference on Information Systems Security (ICISS06), pp. 258-263, 2006.

[44] W., Yuan, D., Guan, S., Lee, Y., Lee, and A., Gavrilov, “Finding Reliable Recommendations for trust model”. In proceeding of the 7th International Conference on Web Information Systems Engineering (WISE06), pp. 375-386, 2006.

[45] N., Dimmock, A., Belokosztolski, D., Eyers, J., Bacon, D., Ingram, and K., Moody, “Using Trust and Risk in Role-Based Access Control Policies”. In Proceedings of the 9th ACM Symposium on Access Control Models and technologies. USA, ACM Press, pp. 156-162, 2004

[46] N., Dimmock, J., Bacon, D., Ingram, and Moody, “Risk Models for Trust-Based Access Control (TBAC)”. In Proceedings of the 3rd Annual Conference on Trust Management, France, pp. 364-371, 2005.

[47] T., Huynh, N., Jennings, and N., Shadbolt, “FIRE: An integrated trust and reputation model for open multi-agent systems”. In Proceeding of the 16th European Conference on Artificial Intelligence, Spain, pp. 18-22, 2004.


102

[48] K. Ren, and W. Lou, “Privacy Enhanced Access Control in Ubiquitous Computing Environments”. In Proceeding of the 2nd International Conference of Broadband Networks, Vol. 1, pp. 356-365, 2005.

[49] K., Ren, and W., Lou, “Privacy-Enhanced, Attack-Resilient Access Control in Pervasive Computing Environments with Optional Context Authentication Capability” In Springer Science LLC, Mobile Netw Appl-12:79-92, 2007.

[50] K., Ren, W., Lou, K., Kim, and R., Deng, “A Novel Privacy Preserving Authentication and Access Control Scheme for Pervasive Computing Environments”. In IEEE Transactions on Vehicular Technology, Vol. 55, no. 4, pp. 1373-1384, 2007.

[51] J. Wang, J. Yu, D. Li, X. Bai, and Z. Jia, “Combining User Authentication With Role-Based Authorization Based on Identity-Based Signature”. In Proceeding of International Conference on Computational Intelligence and Security, CIS, pp.847-857, 2006.

[52] B. Jakob, K. Rasmus, P. Michael, “Context-aware user authentication: Supporting proximity-based login in pervasive computing”. In Proceeding of International conference on ubiquitous computing No5, Seattle WA, vol. 2864, pp. 107-123, 2003.

[53] C. Wang, L. Fang, and Q. Wang, “Zero-Knowledge-Based User Authentication Technique in Context-aware System”. In Proceeding of the International Conference on Multimedia and Ubiquitous Engineering, pp. 874-879, 2007.


103

Integration of Metaheuristic Algorithms for Minimum Spanning Tree

R.Thamilselvan1, Dr.P.Balasubramanie2

1Assistant Professor, Department of Computer Science and Engineering,

Kongu Engineering College, Perundurai, Erode 638 052, Tamilnadu, India [email protected]

2Professor, Department of Computer Science and Engineering,

Kongu Engineering College, Perundurai, Erode 638 052, Tamilnadu, India [email protected]

Abstract: This paper presents a new algorithm based on integrating Tabu Search (TS), Simulated Annealing (SA) and Genetic Algorithms (GA) for solving minimum spanning tree problem. This proposed integration method is a general kind of solution method that orchestrates the interaction between local improvement procedures and higher level strategies to create a process that is capable of escaping from local optima and performing a robust search of a feasible region. This paper applies Tabu Search, Simulated Annealing and Genetic algorithm for minimum spanning tree problem and compares the results obtained by each. With the implementation of our approach the minimum spanning tree problem is solved with feasible solution.

Keywords: Tabu Search, Simulated Annealing, Genetic Algorithm, Minimum Spanning Tree.

1. Introduction There are two cases in minimum spanning tree [12], an

undirected and connected network is being considered, where the given information includes some measure of the positive length (distance, cost, time etc.) associated with each link. The problem has to choosing a set of links that have the shortest total length among all sets of links that satisfy a certain property. For the minimum spanning tree problem, the required property is that the chosen links must provide a path between each pair of nodes. The minimum spanning tree problem can be summarized as follows.

i. There are set of nodes of a network but not the links. Instead, you are given the potential links and the positive length for each if it is inserted into the network.

ii. To design the network by inserting enough links to satisfy the requirement that there be a path between every pair of nodes.

iii. The objective is to satisfy this requirement in a way that minimizes the total length of the links inserted into the network.

A network with n nodes required (n-1) links to provide a path between each pair of nodes. No extra links should be used, since this would needlessly increase the total length of the chosen links. The links need to be chosen in such a way that the resulting network forms a spanning tree. The basic

minimum spanning tree algorithm [10] is shown in table 1 7 2 2 5 4 5 4 1 3 1 7 4

Figure 1. Network with nodes are connected

Table 1: Algorithm for Minimum Spanning Tree

To apply the algorithm in Table 1 to the Network in Figure 1. Keep the arbitrarily select node O to start. The unconnected node closest to node O is node A. Connect node A to node O. Repeat the algorithm, then we will get the spanning tree in Figure 2.

2 2 5 1 3 1

Figure 2. Minimum Spanning Tree

The minimum spanning tree problem is the one problem that falls into the broad category of network design.

1. Select a node arbitrarily, and then connect it. 2. Identify the unconnected node that is closest to a

connected node, and then connect these two nodes. Repeat this step until all nodes have been connected.

3. Tie breaking: Ties for the nearest distinct node or the closest unconnected node may be broken arbitrarily, and the algorithm must still yield an optimal solution.

A

O B

C

D

E

T

A

O B

C

D

E

T


104

2. Related Work 2.1 Tabu Search

Tabu Search [1] [2] is a widely used metaheuristic that uses some common sense ideas to enable the search process to escape from a local optimum. Any application of tabu search includes as a subroutine a local search procedure that seems appropriate for the problem being addressed. The process begins by using this procedure as a local improvement procedure in the usual way to find a local optimum. A key strategy of tabu search is that is then continues the search by allowing non improving moves to the best solutions in the neighborhood of the local optimum. Once the point is reached where better solutions can be found in the neighborhood of the current trial solution, the local improvement procedure is reapplied to find a new local optimum. The danger with this approach is that after moving away from a local optimum, the process will cycle right back to the same local optimum. A tabu list records these forbidden moves, which are referred to as tabu moves.

2.1.1 Basic Tabu Search Algorithm Initialization: Start with a feasible initial trial solution

Iteration: Use an appropriate local search procedure to define the feasible moves into the local neighborhood [5] of the current trial solution. Eliminate from consideration any move on the current tabu list unless that move would result in a better solution than the best trial solution found so far. Determine which of the remaining moves provides the best solution. Adopt this solution as the next trial solution, regardless of whether it is better or worse than the current trial solution. Update the tabu list to forbid cycling back to what has been the current trial solution. If the tabu list already had been full, delete the oldest member of the tabu list to provide more flexibility for future moves.

Stopping rule: Use some stopping criterion, such as a fixed number of iterations [7], a fixed amount of CPU time, or a fixed number of consecutive iterations without an improvement in the best objective function value. Also stop at any iteration where there are no feasible moves into the neighborhood of the current trial solution. Accept the best trial solution found on any iteration as the final solution.

The basic idea behind tabu search is that, adding short-term memory to local search, improves its ability to locate optimal solutions. It is an is an iterative search that starts from some initial feasible solution [8] and attempts to determine the best solution in the manner of a hill-climbing algorithm.

2.1.2 A Minimum Spanning Tree Problem with Constraints

Figure 3 shows a network with five nodes, where the dashed lines represent the potential links that could be inserted into the network and the number next to each dashed line represents the cost associated with inserting that particular link. Thus, the problem is to determine which four of these links should be inserted into the network to minimize the total cost of these links. The following

constraints must be observed when choosing the links to include in the network.

Constraint 1: Link AD can be included only if link DE also included.

Constraint 2: At most one of the three links AD, CD and AB can be included.

To take the constraint into account is to charge a huge penalty, such as charge a penalty of 100 if constraint 1 is violated and charge a penalty of 100 if two of the three links specified in constraint 2 are included. 30 30 10 5 25

15 40

Figure 3. Data for a Minimum Spanning Tree before choosing the links to be included in the network

30 30 10 5 25

15 40

Figure 4. Optimal solution for a Minimum Spanning Tree

The Figure 4 shows the desired minimum spanning tree, where the dark lines represent the links that have been inserted into the network with a total cost of 50. This optimal solution [8] is obtained by the greedy algorithm. The method to conduct tabu search is follows.

Local Search Procedure: At each iteration, choose the best immediate neighbor [6] of the current trial solution that is not ruled out by its tabu status.

Neighborhood structure: An immediate neighbor of the current trial solution is one that is reached by adding a single link and then deleting one of the other links in the cycle that is formed by the addition of this link.

Form of tabu moves: List the links that should not be deleted

Addition of tabu move: At each iteration, after choosing the link to be added to the network, also add this link to the tabu list.

A

D

B

C E

A

D

B

C E


105

Maximum size of tabu list: Two

Stooping rule: Stop after three consecutive iterations without an improvement in the best objective function value.

In Figure 4 both constraints are violates, so that penalties of 100 need to be imposed twice. Therefore the total cost of this solution is

Cost = 20+10+5+15+200 = 250

Iteration 1: There are three options for adding a link to the network in Figure 4 are BE, CD and DE. If BE were to be chosen, the cycle formed, so the three options for deleting a link would be CE, AC and AB. If CE were to be deleted, the change in the cost would be 30-5=25 with no change in the constraint penalties, so the total cost would increase from 250 to 275. So that iteration 1 is followed as per Table 2.

Table 2: The options for adding a link and deleting another link in iteration 1

Add Delete Cost

BE BE BE

CE AC AB

75+200=275 70+200=270 60+100=160

CD CD

AD AC

60+100=160 65+300=365

DE DE DE

CE AC AD

85+100=185 80+100=180

75+0=75 (Minimum)

The final output of the iteration 1 in a network is shown in Figure 5. 30 30 10 5 25

15 40

Figure 5. Modified network after Iteration 1 with new cost=75

Iteration 2: In the second iteration the following decisions are made. 30 30 10 5 25

15 40


Add link BE to the network shown in Figure 5, the automatically place this added link on the tabu list. Delete link AB from the network shown in Figure 5. Various options for adding a link and deleting a link in iteration 2 are shown in Table 3.


Add Delete Cost

AD AD AD

DE CD AC

(Tabu move) 85+100=185 80+100=180

BE BE BE

CE AC AB

100+0=100 95+0=95

85+0=85 (Minimum) CD CD

DE DE

60+100=160 95+100=195

Iteration 3: The following are the decisions are made in this iteration.

Add a link CD to the network shown in Figure 6, and then automatically place this added link on the tabu list. Delete link DE from the network shown in Figure 6. Various options for adding a link and deleting a link in iteration 3 are shown in Table 4. 30 30 10 5 25

15 40

Figure 7. Modified network after Iteration 3 with optimal solution cost=70


Add Delete Cost

AB AB AB

BE CE AC

(Tabu move) 100+0=100

95+0=95 AD AD AD

DE CE AC

60+100=160 95+0=95 90+0=90

CD CD

DE CE

70+0=70 (optimal solution 105+0=105

With a well designed tabu search algorithm, the best trial solution found in Figure 7 after the algorithm has run a modest number of iterations is likely to be a good feasible solution. It might even be an optimal solution [1], [2] but no such guarantee can be given. Selecting a stopping rule that provides a relatively long run of the algorithm increases the chance of reaching the global optimum.

A

D

B

C E

A

D

B

C E

A

D

B

C E


106

2.2 Simulated Annealing Simulated annealing [1]-[5] is another widely used metaheuritic that enables the search process to escape from a local optimum. Each iteration of the simulated annealing search process moves from the current trial solution to an immediate neighbor in the local neighborhood of this solution, just as for tabu search. However, the difference from tabu search lies in how an immediate neighbor is selected to be the next trial solution. Let

Zc = objective function value for the current trial solution Zn = objective function value for the current candidate to be the next trial solution. T = a parameter that measures the tendency to accept the current candidate to be the next trial solution if this candidate is not an improvement on the current trial solution.

The rule for selecting which immediate neighbor will be the next trial solution is the following.

Move Selection rule: Among all the immediate neighbors of the current trial solution [9], select one randomly to become the current candidate to be the next trial solution. Assuming the objective is maximization of the objective function, accept or reject this candidate to be the next trial solution as follows:

If Zn ≤ Zc, always accept this candidate

If Zn < Zc, accept the candidate with the following probability

Prof {acceptance} = ex where T

ZcZnx −=

2.2.1 Basic Simulated Annealing Algorithm Initialization: Start with a feasible initial trial solution.

Iteration: Use to move selection rule to select the next trial solution.

Check the temperature schedule: When the desired number of iterations has been performed at the current value of T, decrease T to the next value in the temperature schedule and resume performing iterations at this next value.

Stopping rule: When the desired number of iterations have been performed at the smallest value of T in the temperature schedule as final solution.

2.3 Genetic Algorithms Genetic algorithms [1]-[3] provide a third type of metaheuristic that is quite different from first two. This type tends to be particularly effective at exploring various parts of the feasible region and gradually evolving toward the best feasible solutions. The modern field of genetics provides a further explanation of this process of evolution and the natural selection involved in the survival of the fittest. In any species that reproduces by sexual reproduction, each offspring inherits some of the chromosomes from each of the two parents, where the genes within the chromosomes determine the individual features of the child. A child who

happens to inherit the better features of the parents is slightly more likely to survive into adulthood and then become a parent who passes on some of these features to the next generation. The population tends to improve slowly over time by this process. A second factor that contributes to this process is a random, low-level mutation rate in the DNA of the chromosomes. Thus, a mutation occasionally occurs that changes the features of a chromosome that a child inherits from a parent.

2.3.1 Basic Genetic Algorithm Initialization: Start with an initial population of feasible trial solutions, perhaps by generation then randomly.

Iteration: Use a random process [5] that is biased toward the more fit members of the current population to select some of the members to become parents. Pair up the parents randomly and then have each pair of parents give birth to two children whose features are a random mixture [10] of the features of the parents, except for occasional mutations. Evaluate the fitness for each new member in the new population.

Stopping rule: Use some stopping rule, such as a fixed number of iterations, a fixed amount of CPU time, or a fixed number of consecutive iterations without any improvement in the best trial solution found so far. Use the best trial solution fond on any iteration as the final solution.

3. Proposed Work The new algorithm based on integrating genetic

algorithms, tabu search and simulated annealing methods to solve the minimum spanning tree problem. The core of the proposed algorithm is based on tabu search algorithms. Genetic algorithm is used to generate new population members in the reproduction phase of the tabu search algorithm. Simulated annealing method is used to accelerate the convergence of the tabu search algorithm by applying the simulated annealing test for all the population members. A new implementation of the tabu search algorithm is introduced. In the genetic algorithm part of the proposed algorithm, a simple short-term memory procedure is used to counter the danger of entrapment at a local optimum, and the premature convergence of the tabu search algorithm. A simple cooling schedule has been implemented to apply the simulated annealing test in the algorithm [3]-[5].

3.1 Basic Integration Algorithm

Initialization: Start with a feasible trial solution.

Iteration: Use an appropriate local search procedure to define the feasible moves into the local neighborhood [6] of the current trial solution. Determine which of the remaining moves provides the best solution. At the same time Use the move selection rule to select the next trial solution. For each iteration invoke the algorithm to test the cycles in a graph.

Stopping rule: If a pre-defined termination condition is satisfied, output Best Solution and exit.

3.1.1 Algorithm for testing Cycles


107

Step 1: From the combination of edges and incidence matrix, obtain degree of each node contributed by the edges under consideration. Step 2: Test whether at least two nodes of degree one? If not, go to step 6. Otherwise continue.

Step 3: Test whether at least three nodes of degree more than one? If not go to step 5. Step 4: Delete pendant edges, if exists of n-1 edges and modify the degree of the modes accordingly and go to step 2. Otherwise go to step 6.

Table 6: Comparisons of TS, GA, SA and Integration Algorithm for Spanning Tree

Graph size (No.

of Nodes)

TS GTA SA Integration Algorithm

Time(s) Cost Iteration Time(s) Cost Iteration Time(s) Cost Iteration Time(s) Cost Iteration

5 50 75 3 52 75 3 55 75 4 50 75 2 7 55 80 3 54 80 4 55 80 4 48 80 2 9 58 95 3 57 90 3 60 90 4 52 90 2

11 65 100 4 68 100 3 75 110 5 60 95 1 13 70 120 4 75 120 4 78 120 5 60 115 1

Figure 8. Comparisons of TS, GA, SA and Integration Algorithm for Spanning Tree Step 5: Edge combinations are tree. Step 6: Stop.

Iteration 1: Instead of charge a penalty for the constraint, the constraint should be verified for every change in the graph. There are three options for adding a link to the network in Figure 4 are CD, DE and DE. If CD were to be chosen, the cycle formed, so the two options for deleting a link would be AD and AC. If AD is added then AB should be deleted according to the second constraint. So that for connecting B, BE should be added. So that iteration 1 is followed as per Table 5.


Constraints violated.

Constraints violated.

The final output of the iteration 1 in a network is shown in Figure 9. Well defined tabu search algorithm [2], the best trial solution found after the algorithm has run a modest number of iterations is likely to be a good feasible solution [8]. But the integration algorithm found the feasible solution in only little iteration from the starting. 30 30 10 5 25

15 40


4. Experiments and Results Five testing graphs of the size 5, 7,9,11 and 13 nodes. All

the five graphs are generating a parse tree using tabu search, genetic algorithm, simulated annealing and integration algorithm. All the algorithm were stopped when the solutions

Add Delete Cost

BE BE BE

CE AC AB

75 70 60

CD CD CD

AD AC AB

60 65 70

DE DE

CE AC

85 80

A

D

B

C E


108

did not change over a fixed number of iterations. The relative performance between TS, GTA, SA and Integration algorithm for the minimum spanning tree problem is given in Table 6 and Figure 8. Five trials were taken on each graph and the average value of the cost and time in milliseconds are tabulated in each case. It is evident from the table that Integration algorithm performed very well as the problem size increased and showed superliner speedups at higher problem sizes. 5. Conclusion and Future Work

In this paper, we have implemented an integration of all the three metaheuristics [5] algorithms for minimum spanning tree problem. In this approach, a pre-defined number of starting solutions are chosen from widely separated regions in the sample space, and used in local search procedures to obtain a set of locally optimal solutions. These locally optimal solutions [9] [10] are then examined to provide an idea about the probability of each vertex being included in an optimal solution to the instance. Using these ideas, the neighborhood of each solution is searched in a probabilistic manner, such that the moves excluding vertices that have a high probability of being in an optimal solution or of including vertices that have a low probability of being in an optimal solution are discouraged. Experiment result shows that the integration of metaheuristics provides a feasible solution [11] if the numbers of nodes are increased in a graph. In future based on this result this may be extended for feasible dynamic routing in a network.

References [1] R.Thamilselvan and Dr.P. Balasubramanie, “A Genetic

Algorithm with a Tabu Search (GTA) for Travelling Salesman Problem”. International Journal of Recent Trends in Engineering, Issue. 1, Vol. 1, pp. 607-610, June 2009.

[2] R.Thamilselvan and Dr.P. Balasubramanie, “Integrating Genetic Algorithm, Tabu Search Approach for Job Shop Scheduling” International Journal of Computer Science and Information Security, Vol. 2, No.1, pp. 134-139, 2009

[3] Jingjing Zhang, Yaohui Jin and Wsisheng Hu, “A Genetic Algorithm for Smooth Scheduling in Slotted WDM Network”, International Journal of Information Technology, Vo. 12, No. 6, pp.26-34, 2006.

[4] S.Jayalakshmi and S.P. Rajagopalan, “Modular Simulated Annealing in Classical Job Shop Scheduling”, International Journal of Information Technology, Vol. 6, No.2, pp. 222-226, 2007.

[5] N.K.Cauvery and Dr. K.V Viswanatha, “Routing in Dynamic Network using Ants and Genetic Algorithm”, International Journal of Computer Science and Network Security, Vol. 9, No. 3, pp. 194-200, March 2009.

[6] Van Laarhoven, P.J.M., E.H.L., AArts, and Jan Karel Lenstra, “Job Shop Scheduling by Simulated Annealing”, Operation Research, Vol. 40, pp. 113-125, 1992.

[7] K. Aggarwal and R. D. Kent, “An Adaptive Generalized Scheduler for Grid Applications”, in Proc. of the 19th Annual International Symposium on High Performance Computing Systems and Applications (HPCS’05), pp.15-18, Guelph, Ontario Canada, May 2005.

[8] Anant Oonsivilai, Wichai Srisuruk, Boonuruang Marungsri and Thanatchai Kulworawanichpong, “Tabu Search Approach to Solve Routing Issues in Communication Networks”, World Academy of Science, Engineering and Technology, pp. 1174-1177, 2009.

[9] D.Janaki Ram, T.H.Steenivas and K.Ganapathy Subramaniam, “Parallel Simulated Annealing Algorithms”, International journal of Parallel and Distributed Computing, Vol.37,pp. 207-212, 1996.

[10] D.Janakiram, “Grid Computing”, A Research Monograph, Tata McGraw-Hill Publishing Company Limited, 2005.

[11] Bernard Chazelle, “A Minimum Spanning Tree Algorithm with Inverse Ackermann Type Complexity”, Journal of the ACM, 47(6), pp. 1028-1047, 2000.

[12] Sanjay Kumar Pal, “Renovation of Minimum Spanning Tree Algorithms of Weighted Graph”, Journal of ACM Ubiquity, Vol.9, Issue 7, pp. 1-5, 2008.

[13] Greening, Daniel R., “Parallel Simulated Annealing Techniques”, Physica D, Vol.42, pp. 293-306,1990.

[14] Van Laarhoven, P.J.M., E.H.L., AArts, and Jan Karel Lenstra, “Job Shop Scheduling by Simulated Annealing”, Operation Research, Vol. 40, pp. 113-125, 1992.

Authors Profile

R.Thamilselvan is an Assistant Professor in the department of computer Science and Engineering, Kongu Engineering College, Perundurai, Tamilnadu India. He has completed his M.E Computer Science and Engineering in 2005 under Anna University Chennai. He has completed 8 years of teaching service. He has published 3 papers in national conference and 2 paper in International Journal. He was the recipient of the Best Faculty award during the year 2007-2008. His area of interest includes Grid Computing, Parallel Processing, and Distributed Computing. He has organized 2 national level seminar sponsored by AICTE, New Delhi.

Dr.P.Balasubramanie is a Professor in the department of computer Science and Engineering, Kongu Engineering College, Perundurai, Tamilnadu India. He was awarded junior research Fellowship by Council of Scientific and Industrial Research (CSIR) in 1990 and he has completed his Ph.D degree in 1990 under Anna University in 1996. He has also qualified for the state Level Eligibility test for Lectureship in 1990. He has completed 13 years of teaching service. He has published more than 60 articles in International/National Journals. He has authored six books with the reputed publishers. He was the recipient of the Best Faculty award for consecutively two years. He is also the recipient of the CTS Best Faculty Award- 2008. He has guided 3 Ph.D scholars and 23 research scholars are working under his guidance. His area of interest includes Image processing, data mining, networking and so on. He is a member of Board of studies, Anna University Coimbatore. He has organized several seminar/workshops.


109

Forward Collision and Delay Avoidance Using Energy Efficient Collision Sets in WSN

Bhushan N. Mahajan 1, Dr. R. V. Dharaskar 2 and Dr. V. M. Thakare 3

1 Computer Science Department

GHRCE, Nagpur [email protected]

2 HOD , Computer Science Department

GHRCE, Nagpur. [email protected]


SGB Amaravati University. [email protected]

Abstract: This paper present battery-efficient system design, and outlines collision and delay avoidance using scheduling method. Batteries power electronic systems and maximizing battery life needs understanding of the capabilities and limitations of the batteries. The amount of energy that can be supplied by a given battery varies significantly. It depend on how the energy is drawn. Researchers are attempting to develop new battery driven approaches to system design.

Keywords: Collision , Delay , Energy level , power , Battery power , Sleep mode , Ideal mode.

1. Introduction Battery powered electronic systems, and the integrated circuits within them, account for a large and rapidly growing revenue segment for the computer, electronics, and semiconductor industries. We know “battery gap”, between trends in processor power consumption , and improvements in battery capacity . Bridging this gap is a challenge that system designers must face for the future. Low power design techniques are successful in reducing the energy that is drawn from the battery, and hence improve battery life to some extent. But, truly maximizing battery life requires an understanding of both the source of energy and the system that consumes it. [1] 2. Related work This energy efficient routing protocol is modified version of the Ad-hoc on demand distance vector (AODV) protocol by taking into consideration the results of the pre-simulation, the existing feature of AODV for implementation of the design . An Ad Hoc network Networks that do not require a pre-established infrastructure . An Ad Hoc network does not guarantee that a mobile node can directly communicate with destinations all the time. So, there is independence of any fixed infrastructure or centralized administration. [2]An Ad Hoc network is capable of operating autonomously. It is completely self-organizing and self-configuring. It has multi-hop capability. A mobile node, which lies outside the transmission of its specific destination, would need to relay

its information flow through other mobile nodes. This implies that mobile nodes in Ad Hoc networks bear routing functionality so that they can act both as routers and hosts. These networks provide mobile users with everywhere communication capacity and information access regardless of location. The Ad Hoc networks can be seen in to two categories whether dynamically changing their position or not, once create communication link. These are wireless sensor networks and Mobile Ad Hoc networks (MANETs). Wireless sensor networks’ mobile nodes are deployed in large number on small area. Once the nodes are deployed, they are static. In Mobile Ad Hoc networks the nodes can dynamically change their position . An Ad Hoc network can be used in an area where infrastructures for mobile communication are not available, probably due to high deployment costs or disaster destruction. The typical application of Ad Hoc networks includes battle field communication, emergency relief and extension of the coverage area of cellular networks. Ad-hoc routing algorithms broadly can be categorized into pro-active and on-demand routing . The on-demand routing algorithms initiate to find out the suitable route when a route is requested . The pro-active routing algorithm exchanges routing information periodically and generates the routing table in advance of route request . [3] These protocols select the routes based on the metrics of minimum hop count. Ability to forward packets is depend on Battery power . Overall network lifetime is depend on battery power Design and manufacturing of less energy consume components of mobile nodes such processors, memory and OS power management strategies is used to reduce non-communication energy consumption. During communication, energy is consumed in either inactive state of communication or active communication states. The energy consumption of active communication is more significant than the others for high traffic environment. Energy efficient routing protocols can be designed to formulate energy efficient active communications . Energy efficient virtual topology can be designed to formulate


110

energy efficient active communications . Minimum hops , Minimum message exchange , Minimum collision , Minimum retransmission strategies can be designed to formulate energy efficient active communications . Result is Energy efficient active communications. It will prolong the network life time . The network life time is defined as the time when a node runs out of its own battery power for the first time . [4] Active communication energy efficient routing protocols can be categorized into two categories: • Transmission power control approach • Load distribution approach. For protocols that belong to the former category, the active communication energy can be reduced by adjusting each node’s radio power just enough to reach the receiving node but not more than that. This transmission power control approach can be extended to determine the optimal routing path that minimizes the total transmission energy required to deliver data packets to the destination. The specific goal of the load distribution approach is to balance the energy.

How active transmission energy can be reduced? . Answer is by adjusting each node’s radio power just enough to reach the receiving node but not more than that , by determining the optimal routing path , by determining the task distribution approach to balance the energy usage of all mobile nodes . by Selecting a route with underutilized nodes rather than the shortest route, by Selecting energy levels to minimizes transmission energy to deliver data packets to the destination. All above approaches prolong the network life time in different ways. [5] Two approaches are mutually contradicting each other with some extent. Advantages- • Increased the life time of the node. • Increased packet delivery fraction • Reduced the Variance of the nodes residual battery Energy. • Minimized energy consumed per packet. The purpose of energy-aware routing protocols is to maximize the network lifetime. The energy-efficient routing protocols should consider energy consumption from the viewpoints of both the network and the node level. Why task distribution approach is required from node point of view ?

Nodes have limited remaining battery energy. Nodes can perform every kind of task and while working they consume energy. But equal consumption should occur among all nodes. So, work should not be assigned to same set of nodes again and again. It may lead to over consumption of those nodes battery power . Constant amount of power of all nodes should get utilized. It is called as balancing .The way for achieving balancing is –

• Equal distribution of work.

• Selecting different routing path at different time instance.

• Changing routing path frequently.

• Same routing path will not be repeatedly used again and again.

• Same set of nodes will not get over utilized again and again

Existing wake-up schemes Nodes have to remain awaken when they are not receiving any data. Nodes have to listen idle channel . Network nodes cannot easily know exactly when events happen. The delay exist between receiving wake up message and actual waking up of node. That delay is called as wake-up latency. Node wakes up itself when it senses an communication from other nodes. Then, it wakes up the MCU and RF transceiver. Time-based wake-up mechanisms It require each node wake up periodically to listen to the radio channel .It is time-based wake-up mechanisms. Low power sleeping nodes wake up at the same time periodically to communicate. Data are transferred from sensor nodes to the sink through a multi-hop communication paradigm [3]. By choosing a good time synchronized sleep and wake-up mechanism the network may save much power consumption. Drawbacks to this mechanism is the high quality clock synchronization requirement. It makes it hard to be implemented in large WSN. [6 , 11] If the packet arrives at node at the end of the listening cycle, then wake up call will be delayed while the node sleeps. It is wake up latency . Asynchronous wake-up mechanisms Each node follows its own wake-up schedule in idle states, as long as the wake-up intervals among neighbors overlap. It do not require time synchronization among the different nodes in the network. Energy harvesting for sensor is still in its early stages, and is gaining momentum in the research community [11]. Drawbacks to this mechanism are this strategy can lead to large packet delays within the network .It cannot achieve the same level of power savings that synchronous approach can. On-demand wake-up mechanism Possible solution is to design an on-demand wake-up In this mechanism , Here out-band signal is used to wake up sleeping nodes in an on-demand manner. For example, a transceiver is woken up by a special wake-up signal from a terminal that wants to communicate with a sleeping transceiver. By this way, a transceiver is woken up on demand instead of periodical wake-up so that the power consumption can be dramatically reduced . Span [11] is a connectivity-driven protocol that adaptively elects “coordinators” of all nodes in the network. Coordinators stay awake continuously and perform multihop routing, while the other nodes stay in sleeping mode and periodically check if there is a need to wake up and become a coordinator. The protocol achieves the following four goals. First, it ensures that there is always a sufficient number of coordinators so that every node is in the transmission range of at least one coordinator. Second, to spread energy consumption as uniformly as possible among network nodes Span rotates the coordinators. Third, it tries to minimize the number of coordinators (to increase the network lifetime) while avoiding a performance degradation in terms of network capacity and message latency. Fourth, it elect coordinators in a decentralized way by using only local information. STEM (Sparse Topology and Energy Management) [11]


111

It uses two different radios for wakeup signals and data packet transmissions, respectively. The wakeup radio is not a low power radio (to avoid problems associated with different transmission ranges). Therefore, an asynchronous duty cycle scheme is used on the wakeup radio as well. Each node periodically turns on its wakeup radio for Tactive every T duration. When a source node (initiator) has to communicate with a neighboring node (target), it sends a stream of periodic beacons on the wakeup channel. As soon as the target node receives a beacon it sends back a wakeup acknowledgement, and turns on its data radio. [7] If we select longer-hop routes, which spend more energy end to end.

Minimum total power routing (MTPR): S is the set containing all the possible routes. Rth route is selected . Transmission power for route R is P

R .

This approach may select the route that includes one or more mobile node with least energy level or min-hop routing . it makes no effort to use energy evenly among the nodes . This leads to “die” of the first node sooner and it causes partition of the network early. Power aware localized routing: It is assumed that the power needed for transmission and reception is a linear function of d

α where d is distance

between the two neighboring nodes and α a parameter that depends on the physical environment. [8] The authors make use of GPS position information to transmit packets with the minimum required transmit energy. Two drawbacks are,GPS cannot provide useful information about the physical environment and the second is that the power dissipation overhead of the GPS device is an additional power draw on the battery source of the mobile node. Minimum Battery Cost Routing (MBCR): It tries to use battery power evenly by using a cost function which is inversely proportional to residual battery power. One possible choice for the cost function of a node i is given as

bi is the residual battery energy of node i. the total cost of the route is defined as the sum of costs of nodes that are the components of the route, and MBCR selects a route with minimum total cost. [9] The Min-Max Battery Cost Routing (MMBCR):

It selects the route with the minimum path cost among possible routes. Because this metric takes into account the remaining energy level of individual nodes instead of the total energy, the energy of each node can be evenly used. The limitation of this algorithm is that since there is no guarantee that paths with the minimum hop-count or with the minimum total power are selected. It can select paths that results in much higher power dissipation in order to send traffic from a source to destination nodes. This feature actually leads to in shorter network lifetime because in essence the average energy consumption per delivered packet of user data has been increased. [10]

Conditional Max-Min Battery Capacity Routing (CMMBCR): If there are nodes that have more battery power than threshold power, it applies MTPR to the nodes. Otherwise, it mimics MMBCR. When battery power is plentiful, it minimizes the total energy consumption like MTPR, and the other case it considers the nodes with lower energy like MMBCR. When the current drawn is sufficiently large, the rate of diffusion fails to keep up with the rate at which ions are consumed at the cathode. As a result, the concentration of positively charged ions decreases near the cathode and increases near the anode, degrading the battery’s output voltage. However, if the battery is allowed to idle for a period of time, the concentration gradient decreases (due to diffusion), and charge recovery takes place at the cathode. As a result, the capacity and lifetime of the battery increase.. 3. Algorithm Nodelist={all nodes as per pre-defined index}; Hierarchy _ I = {set of nodes in level I} ; Hierarchy _II = {set of nodes in level II} ; Hierarchy _III = {set of nodes in level III} ; Hierarchy _IV = {set of nodes in level IV} ; Threshold = x ; // pre-set Send_Nodelist = nodelist; Rev_Nodelist = nodelist; SendDoneNodelist={set all zero} ; RecvDoneNodelist={set all zero} ; do { slot = new Slot(); // set flags here . // flag may be Fwd or Rev check_direction (node); Fwd_send_collision_set = null; Fwd_recv_collision_set = null; Rev_send_collision_set = null; Rev_recv_collision_set = null; for each node in nodelist { node->u = check_ Hierarchy (node) ; node->v = check_ Hierarchy _Energy_Level (node) ; node->w = check_incoming_packet_seq_no (node) ; node->x = check_rec_slots(node) ; node->y = check_send_slots(node) ; node->z = check_traffic_over (node) ; if ( (node not in Fwd_send_collision_set) & & (node->dest not in Fwd_recv_collision_set) & & ( node->z not “over crowding” )

& & ( node->y not “0” ) & & ( node->x == “0” ) & & ( node->w == “No duplication” )


112

& & ( node->w == “Above_threshold” ) ) { // 7à 6 . 7=send , 6 =Recv slot[node] = ‘‘SEND’’; slot[node?dest] = ‘‘RECV’’; // Nodes those will cause collision Fwd_send_collision_set.add( node->dest->neighbors.flag Fwd ); Fwd_send_collision_set.add( node->dest.flag.Fwd); // Nodes those will face collision Fwd_recv_collision_set.add( Node.flag Fwd); Fwd_recv_collision_set.add( node->neighbors.flag Fwd); // Nodes those will cause collision Rev_send_collision_set.add( node->dest->neighbors.flag Rev ); Rev_send_collision_set.add( node->dest.flag.Rev); // Nodes those will face collision Rev_recv_collision_set.add( node->neighbors.flag Rev); change_priority_algo(node->dest->recv); SendDoneNodelist.addtoLastPos(node); RecvDoneNodelist.addtoLastPos(node?Dest); Send_Nodelist.Remove(node); Recv_Nodelist.Remove(node?Dest); If ( decide_(collision , threshold ) > 1 ) { Node_List = Recv_Nodelist. } else { Node_List = Send_Nodelist. } } // if } // for frame.add(slot); } while (slot.length > 0) // do We will see why task distribution approach from network point of view. The best route in network is that one which require minimum total transmission power . The network lifetime is the time when a node runs out of its own battery

power for the first time. If a node stops its operation, it can result in network partitioning and interrupt communication. We will see why transmission power control approach from node point of view .The nodes with lower power should avoid transmission up to large distance. The node should balance the energy usage among the neighbor nodes . It minimizes the total transmission energy required to deliver data packet to the destination .[15] We will see why transmission power control approach at network level . The best route in network is that one which require minimum total transmission power . The network lifetime is the time when a node runs out of its own battery power for the first time. [16] If a node stops its operation, it can result in network partitioning and interrupt communication. Minimizing the total energy consumption tends to favor the route that consumes minimum energy .

4. Experiment Now we will concentrate on delay avoidance and collision avoidance . Here is a topology and related routing table . Node 0 is a cluster head .Routing table contain list of all neighbors and their send slot and receive slots . we will define a schedule for each node so that path wake up timing will get decided [11].

Figure 1

Node Neighbors 6 2, 5, 7, 8 7 6, 9 8 0, 6, 10 9 7, 11 Our aim is to reach up to node 0 .

Node 1-hop Neighbors Dest Hops upto node O 6 2, 5, 7, 8 8 2 7 6, 9 6 3 8 0, 6, 10 0 1 9 7, 11 11 3

Node Neighbors Dest Hops Recv Send 6 2, 5, 7, 8 8 2 1 1 7 6, 9 6 3 0 1 8 0, 6, 10 0 1 1 1 9 7, 11 11 3 0 1 Now algorithm will decide slots for receive and send . Node/ slot 1 2 3 4 5 6 7

4 8

11 9

7 6

2

5

14

16 15

10 0 1 3

13

12


113

6 R S 7 S The more improved topology construction is suggested below . It will include delay avoidance strategy . Few other strategies that can be adopted are also suggested below . Three levels of power should be decided at each node. Level 1 = Nodes those are at one hop . Level 2 = Nodes those are at two hop . Level 3 = Nodes those are at three hop . As shown in figure 2, we can create topology. Node 0 is currently acting as cluster head . Its task can be handed over to any other node which is at one hop distance from node 0 . Node in square boundary can act as cluster heads of their own clusters. They can communicate to other cluster head by using node 0 using two hops . Node 3,5,1,16 are still available for forming new cluster .

Figure 2.

Cluster head can communicate with node 0 in two hops. Scheduling information can be stored at node 0. Task of node is to dispatch packet up to cluster head in two hops. It is now node0 task to forward packet up to destination in two hops only . Hierarchical addressing can be also used . Probability of send , probability of receive , probability of idle , probability of Sleep are depend on four important factors .Data forward rate , data arrival rate , delay or waiting time , time spend in wake up .

Division of task can be designed at different level Level 0 = Scheduling information Level 1 = Routing information . Level 3= Energy, collision set information

Figure 3. Five levels of power should be decided at each node. Level 1 = sleep mode Level 2 = forward data Level 3 = receive data Level 4 = send data to neighbor. Level 5 = send data up to cluster head . Level 6 = send busy tone . Three types of sleep periods should be created Level 1 = data to be send after x unit of time . Level 2 = data to be stored after x unit of time . Level 3 = data to be forwarded after x unit of time . Node list = { 17,18,19,20,21,22,23,24,25,26,27,28, 29,30,31,32,33,34,35,36,37,38,39,40 } Start from 17 . 17->41 , [ 17,1] = ”S” [ 41,1] = ”R” Node One hop neighbors 16 0.7,15,17,18,19,20 17 41,42,16 Node Choose Dest Hops Send Recv 16 0 1 4 4 17 16 2 4 4 Divide transmission of 17 in two plus two slots by blocking alternate reverse paths . Fwd_send_collision_set = { 16, 0.7,15 } // new nodes // They cause collision // 16 and its neighbor Fwd_recv_collision_set = { 17,41,42,16 } // existing nodes // They face collision // 17 and its neighbor Rev_send_collision_set = { 17,18,19,20 } // new nodes // They cause collision // 16 and its neighbor Rev_recv_collision_set = { 17,41,42,16 } // existing nodes // They face collision // 17 and its neighbor Now we will assign a last priority to 7 in SendDoneNodelist . we will assign a last priority to 6 in RecvDoneNodelist. So , the collisions those were likely to occur will get avoided .and superior priority will be given to rest of the nodes of Fwd_send_collision set nodes.

5. Power related parameters


114

We can use multi channel sender initiated random access mac layer protocols - Divide available bandwidth into multiple channel .Instead of switch off mode , use low power switch mode . They can be turned on using timer Or using wake-up packets. Mac layer can decide turn off/on of module. [12]

Various advantages of using this scheme are –

• Minimum delay • Energy level selection • Schedule distribution • Collision level • Selection of bandwidth and speed. • Fast input and fast output • Allocate function to individual nodes • Achieve division of work. When a new packet is generated, the sensor node must wait until the next TDMA frame. [13] It transmits the Path-WU message to wake-up all the nodes. Since the packet is generated randomly, the average time until the next frame is Tf/2. [11]

Here , Power as a function of the arrival rate for the same delay. Delay as a function of the arrival rate, for the same power consumption. The minimum values for Ti and Tslot are limited by the available hardware and by the reception sensing time. we consider Ti = Tslot. [11]

L = average size of packets .

. = data packet arrival rate. . = data packet forward rate.

n = the number of timeslots the sensor

node listens

Tf = the length of the total period Tslot = the length of each TDMA timeslot. n * (Tslot/Tf) = The time a sensor node spends in periodical

wakeups . LWait. = The average of waiting time Based on the calculation , the balance energy level is calculated . The balance energy is calculated and it is used to decide the remaining lifetime of battery of node . 6. Delay related parameter

The end-to-end delay, is the sum of the transmission delay, the access delay, the queuing delay , the propagation delay. The delay is proportional to the number of hops N. Delay is affected by number of hops times the period Tf. The transmission at each hop is delayed until the next receiving sensor node wakes up. [14 ]Small distance between the wireless sensor nodes makes the propagation delay small. Large distance between the wireless sensor nodes makes the propagation delay large. Say (N - 1)intermediate forwarders are there between sender node and receiver node .[11] Average delay

The delay in the S-MAC protocol

The delay in the adaptive listening

The delay in TDMA

tcs = Access time ttx = Transmission delay. N = No. of nodes. Tf = No. of hops times. Tf/2. = Average time until the next frame is generated. When Tf increase , end to end delay increase . When Tf decrease , nodes must become active more often . It result in increased power consumption. End to end delay is affected by term which is proportional to Tf . This is because, transmission at each hop is delayed until the next receiving sensor node wakes up . Transmission must wait until the next wake up time . [11]

7 . Conclusion The study of scheduling algorithm helps to achieve the balance between energy saving and delay. By sensing energy from the radio signals, the node provides a interrupt to the MCU. The MCU detects the preamble and wakes up RF transceiver when wake-up signal is indeed to wake itself up. The analysis results show that, based on the scheduling, nodes can be woken up timely when it is necessary to wake up with the aid of extremely low power.

Reference [1] V. Raghunathan, C. Schurghers, S. Park, M. Srivastava, “Energy-aware Wireless Microsensor Networks”, IEEE

Signal Processing Magazine, March 2002, pp. 40-50. [2] G. Pottie, W. Kaiser, “Wireless Integrated Network

Sensors, Communication of ACM, Vol. 43, N. 5, pp. 51- 58, May 2000.


115

[3] I. Akyildiz, W. Su, Y. Sankarasubramaniam and E. Cayirci, “Wireless Sensor Networks: a Survey”, Computer Networks , Volume 38, N. 4, March 2002.

[4] IEEE Pervasive Computing, “Energy Harvesting and Conservation”, Vol. 4, Issue 1, Jan-Mar. 2005.

[5] A. Kansal, J. Hsu, V. Raghunathan, J. Friedman, and M.B. Srivastava, UCLA Energy-Harvesting Project, http://nesl.ee.ucla.edu/projects/heliomote.

[6] L. Doherty, B.A. Warneke, B.E. Boser, K.S.J. Pister, “Energy and Performance Considerations for Smart Dust,” International Journal of Parallel Distributed Systems and Networks, Volume 4, Number 3, 2001, pp. 121-133.

[7] J.A. Paradiso and T. Starnar, “Energy Scavenging for Mobile and Wireless Electronics”, IEEE Pervasive Computing, Vol. 4, Issue 1, pp.18-26, Jan-Mar. 2005.

[8] J. Krikke, “Sunrise for Energy Harvesting Products”, IEEE Pervasive Computing, Vol. 4, Issue 1, pp. 4-8, Jan-Mar. 2005.

[9] M. Philipose, J.R. Smith, B. Jiang, A. Mamishev, S. Roy, and K. Sundara-Rajan, “Battery-Free Wireless Identification and Sensing”, IEEE Pervasive Computing, Vol. 4, Issue 1, pp. 37-45, Jan-Mar. 2005.

[10] C. Schurgers, V. Tsiatsis, M. B. Srivastava, “STEM: Topology Management for Energy Efficient Sensor

Networks”, IEEE Aerospace Conference '02, Big Sky, MT, March 10-15, 2002. [11] Energy efficiency in wireless sensor networks using sleep mode TDMA scheduling Nikolaos A. Pantazis , Dimitrios J. Vergados , Dimitrios D. Vergados , Christos Douligeris .

[12] Y. Xu, J. Heidemann, D. Estrin, “Geography-informed Energy Conservation for Ad Hoc”, Proc. ACM MobiCom 2001, pp. 70 – 84. Rome, 2001.

[13] P. Santi, “Topology Control in Wireless Ad Hoc and Sensor Networks, ACM Computing Survey, Vol. 37, n. 2, p. 164-194, June 2005.

[14] A. Cerpa, D. Estrin, “Ascent: Adaptive Self-Configuring Sensor Network Topologies, Proc. IEEE INFOCOM 2002. [15] B. Chen, K. Jamieson, H. Balakrishnan, R. Morris. “Span: An Energy-Efficient Coordination Algorithm for Topology Maintenance in Ad Hoc Wireless Networks”, ACM Wireless Networks, Vol. 8, N. 5, September 2002.

[16] F. Koushanfar, N. Taft and M. Potkonjak, “Sleeping Coordination for Comprehensive Sensing Using Isotonic Regression and Domatic Partitions”, Proc. of Infocom 2006, Barcelona, Spain, April 2006.

Authors Profile

Bhushan N. Mahajan received Diploma in Mechanical Engineering [DME] in 1999 , A.M.I.E. [CSE] Engineering degree in 2007 , BCA degree in 2007 , qualified GATE 2008 and MCA degree in 2009 . He is perusing Master of Engineering degree [M.E.] in WCC in Computer Science department at

GHRCE , Nagpur university , India . He is a professional software developer . He is now working on energy and power management topics and various schedule development strategies in WSN . He has a special interest in topology modeling of ad-hoc network i.e. wireless sensor network , wireless mesh network and MANET . He has simulated various network scenario using ns2 network simulator software and other programming languages.

Dr. Rajiv Dharaskar is presently working as Professor at PG Department of Computer Science and Engineering, G.H. Raisoni College of Engineering, Nagpur. He is Ph.D. in Computer Science & Engineering in the Faculty of Engineering & Technology, M.Tech. in Computers, P.G. Dip., M.Phil., and M.Sc. He is having 24

years of teaching and 18 years of R&D experience in the field of Computers & IT. He is approved PhD guide for Computer Engineering and Science for Nagpur and Amravati University and 22 research scholars are perusing Ph.D. degree under his guidance. He is an author of number books on Programming Languages..

Dr V M Thakare is Professor and Head of PG department of computer Science and Engg in SGB Amravati University Amravati, Maharastra (India) and has completed ME in Advance Electronics and Ph.D. In computer Science/Engg. His Area of Research are Robotics and Artificial

Intelligence, Information Technology. He is Recognized Giude for computer science and computer engineering in this University and In other universities also. He has also received received national level Award for excellent paper award. More than 10 candidates are working for Ph D Under his supervision. He has Published and presented more than 115 papers at National and international level. He has worked on various national level bodies like AICTE/UGC and also worked on various bodies of other universities. He is presently member of BOS, RRC, BUTR of this university and also chairman and Member of various committees of this university .


116

Efficient Code Dissemination

Reprogramming Protocol for WSN R.P.Shaikh 1, Dr. V.M. Thakare2, Dr. R.V. Dharaskar3

1 Computer Science Department GHRCE, Nagpur

[email protected]

2 HOD , Computer Science Department GHRCE, Nagpur.

[email protected]


SGB Amaravati University. [email protected]

Abstract: Network reprogramming is a way of reprogramming wireless sensor nodes by disseminating the program code over radio for uploading new code or for changing the functionality of existing code.. Existing reprogramming protocols, such as Deluge,achieve this by bundling the reprogramming protocol and the application as one program image, hence it increases the overall size of the image which is transferred through the network. This increases both time and energy required for network reprogramming. A proposed protocol divides the code image into application and reprogramming support. It pre-installs the reprogramming protocol as one image and the application program equipped with the ability to listen to new code updates as the second image that mitigates the above problem. Keyword : Wireless Sensor Network, Sensor, Wireless reprogramming, code dissemination. 1. Introduction A wireless sensor network is expected to consist of a potentially large number of low-cost, low-power, and multifunctional sensor nodes that communicate over short distances through wireless links. Due to their potential to provide fine-grained sensing and actuation at a reasonable cost, wireless sensor networks are considered ideal candidates for a wide range of applications, such as industry monitoring, data acquisition in hazardous environments, and military operations. It is desirable and sometimes necessary to reprogram sensor nodes through wireless links after they are deployed, due to, for example, the need of removing bugs and adding new functionalities. The process of propagating a new code image to the nodes in a network is commonly referred to as code dissemination.

Traditionally, reprogramming was done manually. Therefore, nodes were reprogrammed one by one. However, as the size of sensor nodes becomes larger and larger this technique is not very efficient. What is more, it might be impossible to collect all the nodes from the field and then to reprogram them. Hence, reprogramming needs to be accomplished without physical contact with the nodes. A reliable data dissemination protocol is to be implemented which takes under consideration the previous factors and

can disseminate efficiently a large data object from one node to many other nodes over an entire wireless sensor network. The three important steps for code dissemination protocols is: advertisement of available software, selection of a source, and reliable download to the target,which may then become a source in turn (Figure1).

Figure 1. Three Way handshake for code distribution

Thus, reprogramming sensor nodes, i.e. changing the software running on sensor nodes after deployment, is necessary for sensor networks. A scheme is required to wirelessly reprogram the nodes The scenario poses many challenges, of them being energy, bandwidth and reprogramming. Requirements and Properties of Code Distribution are : 1. The complete image, starting from specific points in the network, must reach all the nodes. This is a requirement. We do not consider the extended problem of reaching only a subset of the nodes. 2. If the image cannot fit into a single packet, it must be

placed in stable storage until the transfer is complete, at

which point the node can be safely reprogrammed. This is

also a required property.

3. The lifetime of the network should not be severely affected by the distribution operation. This is a desirable property.


117

4. The memory and storage requirements of the mechanism should not be very high since that would limit the available space for the normal application. This property is also desirable. 2. Relatedwork Code dissemination protocols have been developed to propagate new code images using the wireless network formed by the sensor nodes. Data dissemination in wireless networks, retransmission of broadcasts can lead to the broadcast storm problem, where redundancy, contention, and collisions impair performance and reliability. Scalable Reliable Multicast (SRM) is a reliable multicast mechanism built for wired networks [15], using communication suppression techniques to minimize network congestion and request implosion at the server. SPIN-RL is an epidemic algorithm designed for broadcast networks that makes use of a three phase (advertisement-request-data) handshaking protocol between nodes to disseminate data [16]. The epidemic property is important since WSNs experience high loss rates, asymmetric connectivity, and transient links due to node failures and repopulation. However, their results show control message redundancy at over 95% as it only considers the suppression of redundant request messages, and SPIN-RL does not perform as well as naive flooding for lossy network models. The earliest network reprogramming protocol XNP[17] only operated over a single hop and did not provide incremental updates of the code image. . A special Boot Loader must be resident in a reserved section of program memory, and the xnp protocol module must be wired into an application (to allow for subsequent XNP updates). A host PC application xnp loads the image, via a base station mote running TOSBase (this acts as a serial-to-wireless bridge) to one (mote-id specific) or many (group-id specific) nodes within direct radio range of the base. The image is sent in capsules, one per packet; there is a fixed time delay between packet transmissions. In unicast mode, XNP checks delivery for each capsule; in broadcast mode, missing packets are handled, after the full image download has completed, using a follow-up query request (nodes respond with a list of missing capsules). The program is loaded into external (nonprogram) memory. Applications are halted during the program download. When a reboot command is issued (via the xnp host program), then the boot loader is called: this copies the program from external to program memory, and then jumps to the start of the new program. MOAP [4] is a multi-hop, over-the-air code distribution mechanism. It uses store-and-forward, providing a ‘ripple’ pattern of updates; lost segments are identified by the receiver using a sliding window, and are re-requested using a unicast message to prevent duplication; a keep alive timer is used to recover from unanswered unicast retransmission requests – when it expires a broadcast request is sent. The basestation broadcasts publish messages advertising the version number of the new code. Receiving nodes check this against their own version number, and can request the update with subscribe messages. A link-statistics mechanism is used to try to avoid unreliable links. After waiting a period to receive all subscriptions, the sender

then starts the data transfer. Missing segments are requested directly from the sender, which prioritises these over further data transmissions. Once a node has received an entire image, it becomes a sender in turn. If a sender receives no subscribe messages, it transfers the new image to program memory from EPROM, and reboots with the new code. Sliding window acknowledgements reduce power consumption (reduced EEPROM reads) at the cost of reduced out-of-order message tolerance. There is no support for rate control, or suppressing multiple senders (apart from link statistics). Trickle[6] runs under TinyOS/Mate – it acts as a service to continuously propagate code updates throughout the network. Periodically (gossiping interval τ) using the maintenance algorithm every node broadcasts a code summary (‘metadata’) if it has not overheard a certain number of neighbours transmit the same information. If a recipient detects the need for an update (either in the sender or in the receiver) then it brings everyone nearby up to date by broadcasting the needed code. Trickle dynamically regulates the per-node, Trickle-related traffic to a particular rate (rx + tx), thus adjusting automatically to the local network density. This scales well, even with packet loss taken into account. A listen-only period is used to minimise the short-listen problem (where de synchronised nodes may cause redundant transmissions due to a shift in their timer phases). The CSMA hidden-terminal problem does not lead to excessive misbehaviour by Trickle, as long as the traffic rate is kept low. By dynamically changing the gossip interval, Trickle can propagate changes rapidly, while using less network bandwidth when there are no known changes. Programs fit into a single TinyOS packet. 3. System Models The conventional reprogramming protocol system model for sensor networks is depicted in figure 2, in which the code images are propagated from base station to every sensor node in the network

Figure 2. Reprogramming model for sensor network 4. The three substantially more sophisticated protocols :


118

4.1 MNP The design goal of MNP[8] is to choose a local source of the code which can satisfy the maximum number of nodes. They provide energy savings by turning off the radio of non-sender nodes. MNP is targeted at MICA2 motes running TinyOS and uses the XNP boot loader along with a dedicated network protocol to provide multi-hop, in-network programming. The MNP protocol operates in 4 phases: 1. Advertisement/Request, where sources advertise the new version of the code, and all interested nodes make requests. Nodes listen to both advertisements and requests, and decide whether to start forwarding code or not (this acts as a suppression scheme to avoid network overload); 2. Forward/Download, where a source broadcasts a StartDownload message to prepare the receivers, and then sends the program code a packet at a time (in packet-sized segments) to the receivers to be stored in external memory (EEPROM) – there is no ack, the receiver keeps a linked-list of missing segments in EEPROM to save RAM space; 3. Query/Update, where the source broadcasts a Query to all its receivers, which respond by unicast by asking for the missing packets (segments) – these are then rebroadcast by the source node, and then another Query is broadcast until there are no requests for missing packets. The receivers, having received the full image, now become source nodes and start advertising the new program; 4. Reboot, entered when a source received no requests in response to an advertisement, where the new program image is transferred to program memory, and the node reboots with the new code. A node sends a download request to all senders, this assists in sender selection, and also allows the hidden terminal effect to be reduced (as other potential senders can overhead this request). The sender selection algorithm attempts to allow only one active sender in a particular neighborhood. 4.2 FRESHET Freshet[10] is different in aggressively optimizing the energy consumption for reprogramming. It introduces a new phase called blitzkrieg when the code update is started from the base node. During the blitzkrieg phase, information about the code and topology (primarily the number of hops a node is away from the wave front where the code is at) propagates through the network rapidly. Using the topology information each node estimates when the code will arrive in its vicinity and the three way handshake will be initiated – the distribution phase. Each node can go to sleep in between the blitzkrieg phase and the distribution phase thereby saving energy. Freshet also optimizes the energy consumption by exponentially reducing the meta-data rate during conditions of stability in the network when no new code is being introduced, called the quiescent phase.

4.3 DELUGE Deluge[6] is a density-aware protocol with epidemic behavior that can help propagate the code reliable over unpredictable network conditions. It represents the data

object as a set of fixed-size pages, a key feature needed for spatial multiplexing. Deluge is based on protocol Trickle , a protocol designed for manipulating code updates in sensor networks. Deluge's basinality (borrowed from Trickle) is the suppression and dynamic adjustment of the broadcast rate so as to limit the transmitted messages among n. Deluge uses an epidemic protocol for efficient advertisement of code meta data and spatial multiplexing for efficient propagation of code images. Deluge is generally accepted as the state of the art for code dissemination in wireless sensor networks, and has been included in recent TinyOS distributions Deluge is a data dissemination protocol and algorithm for propagating large amounts of data throughout a WSN using incremental upgrades for enhanced performance. It is particularly aimed at disseminating software image updates, identified by incremental version numbers, for network reprogramming. The program image is split into fixed size pages that can be ‘reasonably’ buffered in RAM, and each page is split into fixed size packets so that a packet can be sent without fragmentation by the TinyOS network stack. A bit vector of pages received can be sent in a single TinyOS network packet. Nodes broadcast advertisements containing a version number and a bit vector of the associated pages received, using a variable period based on updating activity. If a node determines that it needs to upgrade part of its image to match a newer version, then, after listening to further advertisements for a time, it sends a request to the selected neighbour for the lowest page number required, and the packets required within that page. After listening for further requests, the sender selects a page, and broadcasts every requested packet in that page. When a node receives the last packet required to complete a page, it broadcasts an advertisement before requesting further pages – this enhances parallelisation (‘spatial multiplexing’) of the update within the network (as the node can now issue further requests in parallel with responding to requests from other nodes). The protocol keeps the state data to a fixed size, independent of the number of neighbours. There are no ACK’s or NACK’s – requesters either request new pages, or re-request missing packets from a previous page. There is no global co-ordination to select senders; heuristics are used to try and elect relatively remote senders in order to minimise radio network contention. Incremental updating is supported through the use of Complete Advertisements which indicate which pages in an image have changed since the previous version; requesters can then request just the changed pages. Future versions of Deluge are expected to address the following issues: control message suppression, running updates concurrently with applications, explicitly reducing energy consumption, and support for multiple types and versions of images.

5. Proposed Protocol Each protocol discuss above transfers the image of the entire reprogramming protocol together with the minimally necessary part.The researchers have found that it is difficult to improve over Deluge the rate of transfer of data over the wireless link. Hence to optimize what needs to be transferred, keeping the basic mode of transfer the same as


119

in Deluge ,transfer just what is needed, in other words, the application code (or the code of the updates to the application).This idea gives rise to our proposed protocol.It transfers close to the minimally required image size by segmenting the total program image into an application image and the reprogramming image.(Application image refer to the user application , reprogramming image refer to protocol component for protocol, such as MNP ,Deluge or Freshet) The benefit of our protocol shows up in fewer number of bytes transferred over the wireless medium leading to increased energy savings and reduced delay for reprogramming

5.1 Protocol Description An application is modified by linking it to a small component called Application Support (AS) while Reprogramming Support (RS) is pre-installed in each node. Overall, design principle is to limit the size of AS and providing it the facility to switch to RS when triggered by a code update related message. Consider that initially all nodes have RS as image 0 and the application with AS as image 1 Each node is executing the image 1 code. The node that initiates the reprogramming is attached to a computer through the serial port and is called the base node. Following is the description of how Stream works when a new user application, again with the Stream-AS component added to it, has to be injected into the network. 1.Reboot procedure takes place as follows: a. The base node executing image 1 initiates the process by generating a command to reboot from image 0. It broadcasts the reboot command to its one hop neighbors and itself reboots from image 0. b. When a node running the user application receives the reboot command, it rebroadcasts the reboot command and itself reboots from image 0. 2.When all nodes receives reboot command they all start running RS. Then the new user application is injected into the network using RS. 3.Reprogramming of entire network stars using three way handshake as discussed above. Each node maintains a set S containing the node ids of the nodes from which it has received the requests for code. 4. Once the node downloads the new code completely, it performs a single-hop broadcast of an ACK indicating it has completed downloading. 5. When a node receives the ACK from a node, it removes the id of from its set S. 6. When the set S is empty and all the images are complete and after sometime entire network is reprogrammed and nodes will reboot from apllcation support.

5.2 Advantages : •Reduce transmitted bit over wireless medium leading to increased energy savings and reduced delay for reprogramming •Reduce programming time, energy costs and program memory

•Improve the protocol for a new node to get image from network •It optimizes the steady state energy expenditure by switching from a push-based mechanism where periodically node sends advertisements to pull based mechanism where newly inserted node request for the code. •In Freshet to save energy the sleeping time of node is to be estimated prior and this estimation if often found inaccurate due to variability of the wireless channel however stream protocol achieve this goal by rebooting the node from Stream-RS only when new node arrives at one of its neighbors thus the user application running on the node can put the node to sleep till the time to reboot comes. This opportunistic sleeping feature conserve energy in resource constrained sensored network. •In Deluge, once a node’s reprogramming is over, it keeps on advertising the code image it has hence radio resources are continuously used in the steady state but in stream ,Stream-AS does not advertise the data it has .

5.3 Evaluation Results

With the help of reprogramming using the ns-2 simulator we have to evaluate the message inter-arrival period and compared it with the total energy consumption of the sensor nodes. Indeed our aim is to compare our proposed protocol with the known Deluge protocol [6] for wireless sensor network and obtain the result and graph as displayed in Table I.and fig 3. Main objective is observe that the energy consumption has also been reduced because of the reduction in the overall size of the program image that needs to be transferred over the wireless medium which may increase the time and energy required for reprogramming the sensor nodes. Thus fewer number of bytes transferred over the wireless medium leading to increased energy savings and reduced delay for reprogramming

Table 1: Time Taken for Download Phase

Code Size Download time

Case 1 45.2 KB 112 sec Case 2 54.3 KB 120 sec Case 3 67.8 KB 135 sec Case 4 75.7 KB 139 sec Case 5 80.2 KB 141 sec


120

Figure 3. Message inter arrival period (SEC) Other Parameters of implementation on ns-2 is as shown in in Table II

Table 2: Parameters for Implementation on ns2

There are a number of open research problems common to all the classes: 1.Before initiating an update, it would be invaluable to be able run a model (such as a simulation) to determine/analyze reprogramming time and the energy cost of different update options on the current network configuration, and thus allow the user to make informed tradeoffs against energy. For example: which is the best injection strategy for the current configuration? What size reduction technique will result in the quickest update? etc. 2. There is not yet a definitive answer to the best way to reduce the size of updates, with modular, differences-based, and compression all showing promise in different circumstances. 3. There are a number of different injection strategies in use: simulating a base station, sending the update to a base station for dissemination, sending the update to a number of seed nodes for further dissemination, and sending the update individually to each node. 4. It is likely that an energy aware approach will have to be taken in order to respond to current energy patterns in a

sensor net (ref. energy-aware MAC layers, and energy-aware Routing). 5.In order to support the various possible patterns in which software updates may be received, and to support any requirements for backwards and forwards version compatibility, tighter control over the order of node activation will be required. 6. There are a number of aspects of this which are not directly related to software updating, but the key ones which are related are: checking the downloaded software before activation (integrity, version mismatches, platform mismatches) and dynamically checking the operation of the downloaded software after is has been activated. It is likely that further advances will be necessary in this area, probably using techniques from autonomic computing, to increase the robustness of software updates.

7. There is a need for tools to monitor the ‘version’ state of a WSN and report status and problems to an operator/user. These will be able to use existing techniques for fusing data to reduce the overhead, and for tracking update-related faults. 8. The normal issues of: key-distribution, authentication, secrecy, integrity, and authorization needing to be addressed. Results from existing WSN security research will be needed, along with other work specific to the software update problem. 9. The protocols used need to be energy-aware, so that the current energy-state of both individual nodes and the entire network can be taken into account during an update. 10. Recovering from faulty updates methods are required before execution and during execution has been done 6. Conclusion This paper examines the challenges of incorporating scalable bandwidth management scheme and reducing the reprogramming time, the number of bytes transferred, the energy expended, and the usage of program memory for wireless reprogramming in WSN environment with brief description of some existing proposals that specifically address this problem . In future analysis of parameters as shown in table I & tableII by Simulation experiments to show the increasing advantages of proposed protocol over Deluge with larger network sizes. Certain issues were not addressed in this work, like the security issue, reliability etc. If an acknowledgement/code segment lost in a particular hop of multihop network due wireless medium constraints, then the nodes which are in that hop have to take some necessary actions to achieve reliability. References [1] P. Levis, N. Patel, S. Shenker, D. Culler, “Trickle: a

selfregulating algorithm for code propagation & maintenance in wireless sensor network,” in: Proceedings of the First USENIX/ACM Symposium on Networked Systems Design and Implementation “ (NSDI 2004) .


121

[2] “Remote Incremental Linking for Energy-Efficient Reprogramming of Sensor Networks”,Joel Koshy & Raju Pandey University of California, Davis, California 95616, USA.

[3] J.W. Hui, D. Culler, The dynamic behavior of a data dissemination protocol for network programming at scale, in: The Proceedings of the Second International Conference on Embedded Networked Sensor Systems, Baltimore, MD,USA, 2004, pp. 81–94.

[4] T. Stathopoulos, J. Heidemann, D. Estrin, A remote code update mechanism for wireless sensor networks, Technical Report CENS Technical Report 30, 2003.

[5] Synapse: A Network Reprogramming Protocol for Wireless Sensor Networks using Fountain Codes” Michele Rossi, Giovanni Zanca, Luca Stabellini Riccardo Crepaldi, Albert F. Harris III and Michele Zorzi Dept. of Information Engineering, University of Padova, 35131 Padova, Italy.

[6] R.K. Panta, I. Khalil, S. Bagchi, “Stream: low overhead wireless reprogramming for sensor networks, in: Proceedings of the 26th IEEE International Conference on Computer Communications (INFOCOM), May 2007, pp. 928–936.”

[7] M. D. Krasniewski, S. Bagchi, C-L. Yang, W. J. Chappell, “Energy efficient, On-demand Reprogramming of Large-scale Sensor Networks,” Submitted to IEEE Transactions on Mobile Computing (TMC). Available as Purdue ECE Technical Report TR-ECE-06-02, 2006.

[8] S.S.Kulkarni and L. Wang, “MNP: Multihop Network Reprogramming Service for Sensor Networks,” in IEEE ICDCS, Columbus, Ohio, USA, Jun. 2005.

[9] Efficient wireless reprogramming through reduced bandwidth usage and opportunistic sleeping,” Rajesh Krishna Panta , Saurabh Bagchi , Issa M. Khalil a Dependable Computing Systems Lab, School of Electrical and Computer Engineering, Purdue University

[10] N. Reijers, K. Langendoen, “Efficient code distribution in wireless sensor networks, in:” Proceedings of the Second ACM International Conference on Wireless Sensor Networks and Applications (WSNA), 2003, pp. 60–67.

[11] J. Koshy, R. Pandey, “Remote incremental linking for energy efficient reprogramming of sensor networks” in: Proceedings of the Second European Workshop on Wireless Sensor Networks (EWSN), 2005, pp. 354–365

[12] “Updating Software in Wireless Sensor Networks: A Survey” S. Brown, Dept. of Computer Science, National University of Ireland, Maynooth C.J. Sreenan, Mobile & Internet Systems Laboratory, Dept. of Computer Science, University College Cork, Ireland Technical Report UCCCS- 2006-13-07

[13] Shen,, Srisathapornphat, Jaikaeo: “Sensor Information Networking Architecture and Applications”. In: Proc. of the International Workshop on Pervasive Computing, Toronto, Canada, August. IEEE(2004) 52-59

[14] Stann, F., Heidemann, “RMST: Reliable Data Transport in Sensor Networks”. In: Proc. of the 1st IEEE Intl. Workshop on Sensor Network Applications and Protocols. IEEE (2003) 102-112

[15] Beutel, J., Dyer, M., Meier, Ringwald, Thiele: Next-Generation Deployment Support for Sensor Networks. TIK-Report No: 207. Computer Engineering and Networks Lab, Swiss Federal Institute of Technology(ETH), Zurich (2004)

[16] S. K. Kasera, G. Hj´almt´ysson, D. F. Towsley, and J. F.Kurose. “Scalable reliable multicast using multiple multicast channels. IEEE/ACM Transactions on Networking, 8(3):294–310, 2000.”

[17] J.Kulik, W. R. Heinzelman, and H.Balakrishnan.” Negotiation-based protocols for disseminating information in wireless sensor networks. Wireless Networks,8(2-3):169–185, 2002.”

[18] Crossbow Tech Inc.,” Mote In-Network Programming” User Ref, http://www.tinyos.net/tinyos 1.x/doc/Xnp.pdf, 2003

[19] Q.Wang, Y.Y. Zhu., L. Cheng, “Reprogramming wireless sensor networks: challenges and approaches”, IEEE Networks, 2006, 20(3): 48.

[20] A. Chlipala, J. Hui, and G. Tolle. Deluge: Data dissemination for network reprogramming at scale.

Authors Profile

Dr V M Thakare is Professor and Head of PG department of computer Science and Engg in SGB Amravati University Amravati, Maharastra (India) and has completed ME in Advance Electronics and Ph.D. In computer Science/Engg. His Area of

Research are Robotics and Artificial Intelligence, Information Technology. He is Recognized Giude for computer science and computer engineering in this University and In other universities also. He has also received received national level Award for excellent paper award. More than 10 candidates are working for Ph D Under his supervision. He has Published and presented more than 115 papers at National and international level. He has worked on various national level bodies like AICTE/UGC and also worked on various bodies of other universities. He is presently member of BOS, RRC, BUTR of this university and also chairman and Member of various committees of this university .

Dr. Rajiv Dharaskar is presently working as Professor at PG Department of Computer Science and Engineering, GH Raisoni College of Engineering, Nagpur. He is Ph.D. in Computer Science & Engineering in the Faculty of Engineering & Technology, M.Tech. in Computers, P.G. Dip., M.Phil., and M.Sc. He is having 24 years of teaching and 18

years of R&D experience in the field of Computers & IT. He is approved PhD guide for Computer Engineering and Science for Nagpur and Amravati University and 22


122

research scholars are perusing Ph.D. degree under his guidance. He is an author of number books on Programming Languages.

Riyaz Shaikh received B.E degree in Computer Technology from Nagpur University.Joined as MIS INCHARGE in Govt Project at ZP.Before that she also worked as Lecturer in Polytechnic and MCA college. Presently perusing Master of Engineering degree in Wireless communication and

Computing branch under Computer Science department at GHRCE , Nagpur university , India . Her area of interest are wireless adhoc and sensor network.


123

Performance analysis of Dynamic load Balancing Techniques for Parallel and Distributed Systems

Z.Khan1, R.Singh2, J. Alam3 and R.Kumar3

1 Department of Computer Science and Engineering, IIET, Bareilly India

[email protected]

2 Department of Computer Science & IT, IET, R.U. Bareilly India

[email protected]

3 University Women’s Polytechnic, Faculty of Engineering & Technology, AMU Aligarh India [email protected]

SMCA, Thapar University- Patiala India

[email protected]

Abstract: The DAANNS (Diffusion algorithm asynchronous nearest neighborhood) proposed in this paper is an asynchronous algorithm which finds the unbalanced nodes in a network of processors automatically, and makes them balanced in a way that load difference among the processors is 0.5 units. In this paper the DAANNS algorithm has been evaluated by comparing it with RID (Receiver Initiated Diffusion) algorithm across a range of network topologies including ring, hypercube, and torus where the no of nodes have been varied from 8 to 128. All the experiments were performed on Intel parallel compiler. After simulation we have noticed that the DAANNS performed very well as compared to RID.

Keywords: RID (receiver initial diffusion), DAANNS, Load balancing, IPC.

1. Introduction The load balancing problem in parallel and distributed systems deals with how to distribute the work load (computational tasks) among the available processors so that each processor has equal or nearly equal load. Load balancing is of two types - static load balancing and dynamic load balancing. In static load balancing the load balancing is done prior to execution and is done only once. Static load balancing is also called static mapping and it is quite effective for those computations which are of predictive runtime behavior. For this type of load balancing the computational time is non deterministic i.e. the load balancing problem is NP Complete. Performing load balancing in the beginning of execution is not sufficient for certain types of problems like parallel data applications. In such cases load balancing is performed again and again or periodically during run time when the resources required by a particular part of a problem are available in the system. Other examples of using this type of load balancing include protein structure prediction, seismological data analysis, meteorological data analysis and high quality animation generation. A task or load is a set of modules and module execute on one of the processing node and communicate with other

process with inter process communication (IPC). The load allocation and then balancing it among the processor is an essential phase in parallel and distributed systems. This paper addresses the problem of load balancing in parallel and distributed systems. Here each processor has its own identity and processor pass messages through inter process communication between the tasks they are running. A peer to peer interconnection network is used for communication. Many commercial parallel computers fall under this category including Intel Paragon, The Thinking Machine, CM-5, IBMSP2, Origin 2000, and CrayT3D/T3E. In these systems to achieve load balancing a processor relies on nearest neighborhood methods in which the processor in question tries to transfer (accept) load (tasks) to (from) all of its neighbors simultaneously. The strategies proposed in [2],[3],[4],[5],[6] assume that the load of a processor is a real number. Such assumption may be successful for parallel programs which enjoy large grain parallelism. Under the more realistic cases when parallel programs have medium or small grain parallelism these strategies may fail. This paper proposes an algorithm called Diffusion Algorithm Asynchronous Nearest Neighborhood Strategy (DAANNS) which is capable of handling the indivisible tasks and therefore targeted for the parallel programs which exploit medium or small grain parallelism. The rest of the paper is organized as fallows. Section 2, 3 and 4 describe the DAANNS (diffusion algorithm asynchronous nearest neighborhood strategy) and the methodology used. In section 5 simulation results have been presented and section 6 concludes the paper.

2. The DAANNS strategy: DAANNS is a diffusion algorithm asynchronous nearest Neighborhood strategy based on RID (receiver initiated diffusion) proposed by Willebeck Le Mairet et. al. in [5]. RID is a highly distributed local approach which makes use of near neighbor load information in which under loaded processor request load from heavily loaded near neighbors. The balancing process is initiated by any processor whose


124

load drops below a pre specified threshold (LLOW). Second upon receipt of a load request a processor will full fill the request only up to an amount equal to half of its current load (this reduces the effect of aging of data upon which the request was based). Finally in the receiver initiated approach the under loaded processor in the system take on the majority of the load balancing over head. The RID algorithm can be adapted to the integer work load model using floor and ceiling functions. However this integer approach may fail to guaranteed a global balanced situation although the load of each processor may differ by only one unit at most from that of its neighbor. The global load balance may be very poor because of overhead. The DAANNS is developed to solve this problem detecting the unbalanced domain and it performs local exchange of load among the processors to achieve a global balancing state (where the maximum load difference between two processor is 0.5 units i.e. the threshold value). The DAANNS algorithm is explained in table 1 where load of a processor P at a moment (m) is defined as and

represent the

global load vector at a time m. Each processor executes the same group of operations, at every iteration, of the load balancing. Now each processor sends its load information to all its neighbors and receives the load information from all its neighbors. Then first it computes the average load in the domain.

(1)

It also computes its load value . If the processor

load is below the average load by more than a threshold value i.e. 0.5 units or further, it proceeds to load balancing

(2)

Under load That is a processor is under loaded processor when the

value is non negative ( otherwise the process

is overloaded and the value will be negative value .

An under loaded processor performs load

balancing by requesting proportionate amount of load from over loaded neighbors. And each neighbor k is assign a weight .

if (3)

otherwise (4) These weights are summed to determine the total surplus

(5)

The amount of load requested by processor from neighbor

is computed as

(6)

3. Algorithm:

Table 1: Algorithm DAANNS (Diffusion algorithm asynchronous nearest neighborhood)

4. Methodology: In this section we have discussed an important mechanism used in most dynamic load balancing schemes the load update strategy. Many dynamic load balancing strategies like the one proposed here make the balancing decisions based on the load levels of a subset of processor in the system. This subset may include anything from a single neighbor to all processors in the system. The degree of knowledge may vary from one strategy to another. This quality of information governs the intelligence of the load balancing decisions. The quality of information depends on three primary factors: a. The accuracy of processor load estimates b. The aging of information due to communication latency

of inter connection network. And the destination of load information and finally

c. The frequency of load messages updates.

The first factor is application dependent and may involve a tradeoff between the quality of the estimate and the complexity of estimation process. The second factor is dependent on machine architecture and the load balancing


125

strategy. The third factor is the interval between updated messages. It may be computed as a function of time or function of load level. The differentiation of performance between RID and DAANNS is because of following reasons: 4.1 Implementation: In theory both approaches should yield similar results, practical implementation issues however distinguish these approaches from one another. The RID strategy receive local update message from neighbor. While DAANNS is a superset of RID and performs well. 4.2 Stability: both schemes RID and DAANNS strategies make load balancing decision based on load status of their neighbors. This load information suffers from aging processes. 4.3 Overhead: In minimizing total execution time it is beneficial to spare overloaded processor burden of load balancing responsibilities. The extent of the overhead is dependent on the task granularity, and may become significant if tasks are small.

5. Simulation/results and analysis: In this section the simulation of both algorithms is presented. We have compared RID and DAANNS algorithm with respect to their stability and efficiency. The stability (or balance quality) measures the ability of an algorithm to coerce any initial load distribution into an equilibrium state, i.e., to reach the global uniform distribution state. The efficiency measure is reflected by the time incurred in the load communication steps and the number of balancing steps required by the algorithm to drive an initial workload distribution into a stable distribution. To check the aspects we implemented these algorithms on the hypercube, torus and ring processor topologies. The communication network sizes in terms of number of processors were 8, 32, 64 and 128. The initial load distribution is denoted as and the total work load is denoted as L. So we can evaluate priori on the expected final load at each processor i.e. average load └L/n┘ or ┌L/n┐, where n is the size of topology. In our experiment the problem size was L=2520 and initial load distribution were likely distribution and pathological distribution. The parameters used in likely distribution were as fallow.

• Varying 25% from global load average for all P




Here 25% variation patterns indicate that processors have equal load. Otherwise 100% variation shows that there is differences of load among the processors while 50% and 75% refer to intermediate situation among processors. Now the pathological distribution is classified as fallows

• All load on single processor • 25% of idle processor • 50% of idle processor • 75% of idle processor

Simulation runs till termination and we get the variance with respect to load average obtained by both strategies for all distribution used in our experiment. Results of experiment have shown in table 2 and table 3, and graphically in fig 1 and fig 2 respectively.

Table2: Variance obtain on an average by DAANNS Likely

distribution(Variance) Pathological

distribution(Variance) No. of Proc.

Hypercube

Torus

Ring Hypercube

Torus Ring

8 0.02 0.10 0.25 0.04 0.24 0.36 16 0.25 0.23 1.00 0.25 0.25 1.00 32 0.25 0.24 5.21 0.23 0.26 5.21 64 0.25 0.35 19.20 0.33 0.45 23.04 128 0.25 0.41 24 0.33 0.79 37.24

Table 2: Variance obtain on an average by RID

Likely distribution(Variance) Pathological distribution(Variance)

No. of Proc.

Hypercube

Torus Ring Hypercube

Torus Ring

8 7.24 5.76 28.09 1.44 4.41 23.04 16 26.01 32.39 31.36 22.09 19.36 161.29 32 33.64 33.64 118.8

1 70.56 127.69 524.41

64 37.21 54.75 342.25

70.12 184.96 1011.24

128 40.96 67.24 482.41

84.64 207.36 1755.61


126

Fig 5.1Vafriance of Torus and hypercube Topology by DAANNS

Fig 5.2 Variance of Ring Topology by DAANNS

Fig 5.3Variance of ring torus and hypercube by RID

6. Analysis: As we know that stability reflects the ability of an algorithm in bounding any initial load distribution into an equilibrium state i.e. to reach the global uniform distribution state. The variance of load with respect to the average load has been obtained by both the strategies for all the distribution used in our experiment. As it can be seen DAANNS achieve very low variations for all topologies (always less than 0.5 for hypercube and torus) and less than 25 in ring. In contrast RID exhibits a high variation for all cases. Now we can conclude that on one hand DAANNS achieve a small maximum difference than RID on average on the other hand all the processor have a better final state i.e. system is in its optimal state . 7. Conclusion: In this paper we have compared two algorithms namely DAANNS and RID for dynamic load balancing in parallel systems. This comparison has been carried out by considering a large set of load distribution which includes different amount of load imbalances. These distributions have been applied on ring, hypercube and torus topologies and the no of processor ranges from 8 to 128. The experiment was conduct on INTEL parallel computer for analyzing the balance degree achieved by both algorithms.

From our experiment we have noticed that the DAANNS performed better than RID. References [1] M. Willebeek-LeMair, A. P. Reeves, “Strategies for

Dynamic Load Balancing on Highly Parallel Computers”, IEEE Transactions on Parallel and Distributed Systems, vol. 4, No. 9, September 1993, pp. 979-993

[2] R. Subramain, I. D. Scherson, “An Analysis of Diffusive Load-Balancing”, In Proceedings of 6th ACM Symposium on Parallel Algorithms and Architectures, 1994

[3] V. Kumar, A. Y. Grama and N. R. Vempaty, “Scalable load balancing techniques for parallel computers”, J. of Par. And Distrib. Comput., 22(1), 1994, pp. 60-79.

[4] S. H. Hosseini, B. Litow, M. Malkawi, J. McPherson, and K. Vairavan, “Analysis of a Graph Coloring Based Distributed Load Balancing Algorithm”, Journal of Parallel and Distributed Computing 10, 1990, pp. 160-166.

[5] G.C.Fox, M.A. Johnson, G.A. Lyzenga,S.W. Otto, J.K.Salmon and D.W. Walkeer, Solving Problems on Concurrent Processors, vol. 1, Prentice-Hall, 1998.

[6] A. Cortés, A. Ripoll, M.A.Senar , F. Cedó and E. Luque, “On the convergence of SID and DASUD load-balancing algorithms”, Technical Report, UAB, 1998

[7] A. Cortés, A. Ripoll, M. A. Senar and E. Luque, “Dynamic Load Balancing Strategy for Scalable Parallel Systems”, PARCO’97, 1997.

[8] C. Z. Xu and F. C. M. Lau, Load Balancing Parallel Computers - Theory and Practice, Kluwer Academic

Publishers, 1997 Authors Profile

Zubair Khan received his Bachelor Degree in Science and Master of Computer Application Degree from MJP Rohilkhand University Bareilly, India in 1996 and 2001 respectively and also Master in Technology (M.Tech) Degree in Computer Science and

Engineering from Uttar Pardesh Technical University in the year 2008. He is currently pursuing his P.hd in computer science and Information Technology from MJP Rohilkhand University Bareilly, UP India. He also worked as a senior lecturer in JAZAN University Kingdom of Saudi Arbia . He is also servicing as Reader in the Department Of Computer Science and Engineering Invertis Institute of Technology Bareilly, India. His area of interest include data mining and warehousing, parallel systems and computer communication networks. He is an author/ co-author of more than 15 international and national publication in journals and conference proceedings.

R. Singh received the B. Engg. Degree in Electronics Engineering from M.I.T. Aurangabad, India in 1991 and the Ph. D. in Computer Science and Engineering from Lucknow University/Institute of Engineering & Technology Lucknow, India. He is a doctoral

investigator at MJP Rohilkhand University, Bareilly and U.P. Technical University, Lucknow and a visiting Associate Professor at various Technical Collages/Universities in India. After a


127

number of years with the research and development wing of different industries at various positions, he joined as a faculty for the Department of Computer Science & Information Technology, M0-JP Rohilkhand University, Bareilly (India) in Dec 1997 and is currently working as a Head of the Department, CS & IT. . He had also worked as a Professor & Director at R. B. Institute of Engineering & Technology, Bareilly (India). His Research interests are in the area of Routing issues in the Wired and Wireless Network, QoS provisioning for Service Level Agreements in IP networks, Software Architectures with admission control schemes for Real time communication over the Internet, Acoustic communications in Under-water Sensor Network and issues related to Data mining Techniques.

Jahangir Alam graduated in science from Meerut University, Meerut in the year 1992 and received Master degree in Computer Science and Applications from Aligarh Muslim University, Aligarh in the year 1995 and Master in Technology (M.Tech) Degree in Computer Engineering from

JRN Rajasthan Vidyapeeth University, Udaipur, Rajasthan, in the year 2008. He is currently working towards his Ph.D. in Computer Engineering from Thapar University, Patiala. He is also serving as an Assistant Professor in University Women’s Polytechnic, Faculty of Engineering and Technology at Aligarh Muslim University, Aligarh. His areas of interest include Interconnection Networks, Scheduling and Load Balancing, Computer Communication Networks and Databases. He has authored/ co-authored over 10 publications in journals/conference proceedings.

Rajesh Kumar received his bachelor degree in science from Gurukul Kangri University, Haridwar, Utrakhand India and his Master of Science in Mathematics, Master of Philosophy in Computer Science and Ph.D degrees from IIT Roorkee in the year 1988,1990 and 1995 respectively. He has also

done several skill oriented practical courses in computer science from different Institutes/ Universities. Presently he is serving as an Associate .Professor in School of Mathematics and Computer Application at Thapar University Patiala, Punjab India. He is also discharging the duties of Head Computer Centre at ThaparUniversity,Patiala. He has published several papers in National and International Journals/Conferences and has guided a large no of students on their MCA projects. His areas of interest include fracture mechanics, image processing, S/W Engineering and parallel systems. A number of Ph.D theses are being guided by him at Thapar University and at some other Indian universities.


128

Effect of error concealment methods on UMTS radio network capacity and coverage planning

Dr. Bhumin H. Pathak1, Dr. Geoff Childs2 and Dr. Maaruf Ali3 1Airvana Inc., Chelmsford, USA, [email protected]

2School of Technology at Oxford Brooks University, Oxford, UK, [email protected] 3School of Technology at Oxford Brooks University, Oxford, UK, [email protected]

Abstract: The radio spectrum is a precious resource and careful utilization of it requires optimization of all the processes involved in data delivery. It is understood that achieving maximum coverage and capacity with the required QoS is of the utmost importance for network operators to maximize revenue generation. Whilst current methods of video compression accelerate transmission by reducing the number of bits to be transmitted over the network, they have the unfortunate trade-off of increasing signal sensitivity to transmission errors. In this paper we present an approach to transmit MPEG-4 coded video in a way that optimizes the radio resources while maintaining the required received video quality. This approach thereby provides increased capacity for network operators. Two different methods used for this purpose are selective retransmission of erroneous parts in frames and dynamic changes in the reference frames for predictive coding. Network operators can still provide the required video quality with implementation of these two methods despite the low signal-to-interference ratios and can therefore increase the number of users in the network. We show here with the help of simulation results the performance enhancements these methods can provide, for various channel propagation environments. Comparison is performed using a standard PSNR metric and also using the VQMp metric for subjective evaluation.

Keywords: MPEG-4, UMTS, NEWPRED, UEP, network coverage and optimization.

1. Introduction

Future wireless mobile communication is expected to provide a wide range of services including real-time video with acceptable quality. Unlike wired networks, wireless networks struggle to provide high bandwidth for such services which therefore require highly compressed video transmission. With high compression comes high sensitivity to transmission errors. Highly compressed video bitstreams like MPEG-4 Error! Reference source not found. can lose a considerable amount of information with the introduction of just a few errors. In this context, it is important to have a video codec with a repertoire of efficient error-resilience tools. Transmission errors in a mobile environment can vary from single bit errors to large burst errors or even intermittent loss of the connection. The widely varying nature of the wireless channel limits the use of classical forward error correction (FEC) methods which may require a large amount of redundant data to overcome bursty errors.

In this case as an alternative to FEC, an optimized automatic repeat request (ARQ) scheme with selective repeat and dynamic frame referencing provides better performance. Correcting or compensating for errors in a compressed video stream is complicated by the fact that bit-energy concentration in the compressed video is not uniformly distributed, especially with most motion-compensated interframe predictive codecs like MPEG-4. The encoded bit-stream has a high degree of hierarchical structure and dependencies which impact on error correction techniques. The problem is compounded by the fact that it is also not feasible to apply a generalized error-resilience scheme to the video stream as this may impact on standardized parameters in the different layers of the UMTS protocol stack. For example when overhead or redundancy is added to the existing video standard for FEC implementation, the modified video stream can become incompatible with the standard and the subsequent data format may not be decoded by all standard video decoders. It is therefore important that any overhead or redundancy be added in a way which does not make the modified video bit-stream incompatible with the standard. With reference to the above mentioned constraints it is vital to design an error-protection mechanism that exploits the hierarchical structure of the inter-frame video compression algorithm, is compatible with the standardized wireless communication system, and preserves the integration of the standard bitstream.. In this paper we present one such scheme which uses unequal error protection (UEP) in conjunction with ARQ to exploit different modes at the radio link layer (RLC) Error! Reference source not found. of the UMTS architecture. This scheme is coupled with a dynamic frame referencing (NEWPRED) Error! Reference source not found. scheme which exploits frame dependencies and interframe intervals. The techniques discussed in this paper are relevant and applicable to a wide variety of interframe video coding schemes. As an illustrative example, the video compression standard, MPEG-4, is used throughout. The rest of the paper is organized into the following sections. Section-2 gives an introduction to the general


129

principles of the predictive interframe MPEG-4 codec, the frame hierarchies and dependencies involved. Section-3 discusses the general UMTS network and protocol architecture. Section-4 classifies transmission errors into different classes and introduces the UEP and NEWPRED protection schemes.. Section-5 details the UEP scheme while Section-6 details NEWPRED. In Section-7 the simulation scenarios and results are presented. All the achieved results are then discussed from the network operators’ point of view in Section-8. Finally the conclusion is drawn at the end in Section-9.

2. Motion-compensated predictive coding

2.1 General Principles

Image and video data compression refers to a process in which the amount of data used to represent images and video is reduced to meet a bit rate requirement, while the quality of the reconstructed image or video satisfies a requirement for a certain application. This needs to be undertaken while ensuring the complexity of computation involved is affordable for the application and end-devices. The statistical analysis of video signals indicates that there is a strong correlation both between successive picture frames and within the picture elements themselves. Theoretically decorrelation of these signals can lead to bandwidth compression without significantly affecting image or video resolution. Moreover, the insensitivity of the human visual system to loss of certain spatio-temporal visual information can be exploited for further bitrate reduction. Hence, subjectively lossy compression techniques can be used to reduce video bitrates while maintaining an acceptable video quality. Figure 1 shows the general block diagram of a generic interframe video codec Error! Reference source not found..

Figure 1. Generic interframe video codec In interframe predictive coding the difference between pixels in the current frame and their prediction values from the previous frame are coded and transmitted. At the receiving end after decoding the error signal of each pixel, it is added to a similar prediction value to reconstruct the picture. The better the predictor, the smaller the error signal, and hence the transmission bit rate. If the video scene is static, a good prediction for the current pixel is the

same pixel in the previous frame. However, when there is a motion, assuming the movement in the picture is only a shift of object position, then a pixel in the previous frame, displaced by a motion vector, is used. Assigning a motion vector to each pixel is very costly. Instead, a group of pixels are motion compensated, such that the motion vector overhead per pixel can be very small. In a standard codec a block of 16 × 16 pixels, known as a Macroblock (MB), are motion estimated and compensated. It should be noted that motion estimation is only carried out on the luminance parts of the pictures. A scaled version of the same motion vector is used for compensation of chrominance blocks, depending on the picture format. Every MB is either interframe or intraframe coded. The decision on the type of MB depends on the coding technique. Every MB is divided into 8 × 8 luminance and chrominance pixel blocks. Each block is then transformed via the DCT – Discrete Cosine Transform. There are four luminance blocks in each MB, but the number of chrominance blocks depends on the color resolution. Then after quantization, variable length coding (entropy coding) is applied before the actual channel transmission.

2.2 Frame types

An entire video frame sequence is divided into a Group of Pictures (GOP) to assist random access into the frame sequence and to add better error-resilience. The first coded frame in the group is an I-frame. It is followed by an arrangement for P and B frames. The GOP length is normally defined as the distance between two consecutive I-frames as shown in Figure-2. The I-frame (Intra-coded) frame is coded using information only from itself. A Predictive-coded (P) frame is coded using motion compensated prediction from a past reference frame(s). While a Bidirectionally predictive-coded (B) frame is a frame which is coded using motion and texture compensated prediction from a past and future reference frames.

Figure 2. Frame structures in a generic interframe video codec. It is important to note here that the frame interval between two consecutive I-frames and between two consecutive P-frames has a significant effect on the received video quality as well as on the transmission bitrate. It is actually a trade-off between error-resilience capabilities and the required operational bitrate. A major disadvantage of this coding scheme is that transmission errors occurring in a frame which is used as a


130

reference frame for other P or B frames causes errors to propagate into the following video sequence. This propagation continues until an intra-refresh is applied. In the presented work this hierarchy of frame types is exploited. It should be obvious from the above discussion that the I-frame which acts as a reference frame for the entire GOP needs to be protected the most from transmission errors with perhaps weaker protection for P-frames. Since no other frame depends on B-frames, errors occurring in B-frames affect just a single frame and do not propagate into the video sequence. As an illustrative example, the video compression standard, MPEG-4, is used throughout. The rest of the paper is organized into the following sections. Section-2 gives introduction of general principles of predictive interframe MPEG-4 codec, frame hierarchies and dependencies involved. Section-3 discusses the general UMTS network and protocol architecture. Section-4 classifies transmission errors into different classes and introduces the UEP and NEWPRED protection schemes. Section-5 introduces UEP while Section-6 introduces NEWPRED. In Section-7 error detection method is described. Section-8 discusses video quality measurement techniques used in the paper. Section-9 provides information on header compression protocol used in UMTS architecture. Section-10 introduces various radio propagation environments used in simulations. In section-11 simulation scenarios and results are presented. All the achieved results are then discussed from the network operators’ point of view in Section-12. Finally the conclusion is drawn at the end in Section-13.

3. UMTS network and protocol architecture

3.1 Brief introduction

UMTS is a very complex communication system with a wide range of protocols working at different layers of its network architecture Error! Reference source not found.. UMTS has been designed to support a wide range of applications with different Quality of Service (QoS) profiles Error! Reference source not found.. It is essential to understand the overall architecture of the system before we consider details of transmission quality for a given application. UMTS can be briefly divided into two major functional groupings: Access Stratum (AS) and Non-Access Stratum (NAS). The AS is the functional grouping of protocols specific to the access techniques while NAS aims at different aspects for different types of network connections. The Radio Access Bearer (RAB) is a service provided by the AS to the NAS in order to transfer data between the user equipment (UE) and core network (CN). It uses different radio interface protocols at the Uu interface, those are layered into three major parts as shown in Figure 3.

The radio interface protocols are needed to set-up, reconfigure and release the RAB. The radio interface is divided into three protocol layers: Physical layer (L1) Data Link layer (L2) Network layer (L3) The data link layer is split further into the Medium Access Control (MAC) Error! Reference source not found., Radio Link Control (RLC) Error! Reference source not found., Packet Data Convergence Protocol (PDCP) Error! Reference source not found. and Broadcast/Multicast Control (BMC) Error! Reference source not found.. Layer 3 and RLC are divided into Control (C) and User (U) planes as shown in Figure-3.

control

control

RRC

RLCRLC

RLCRLC

RLCRLC

RLCRLC

PDCPPDCP

BMC

MAC

PHY

control

control

control

U-planeinformation

L3

RadioBearers

L2/PDCP

L2/BMC

L2/RLC

Logicalchannels

L2/MACTransportchannels

L1

C-planesignalling

Figure 3. Radio interface protocol structure for UMTS The service access point (SAP) between the MAC and physical layers are provided by the transport channels (TrcCHs). While the SAP between RLC and MAC sub-layers are provided by the logical channels (LcCHs). We will explain the RLC in further details as the services provided by different RLC modes are exploited in order to provide different levels of UEP for video transmission.

3.2 RLC modes

Different services provided by RLC sub-layers includes, Transparent Mode (TM) data transfer, Unacknowledged Mode (UM) data transfer, Acknowledged Mode (AM) data transfer, maintenance of QoS as defined by upper layers and notification of unrecoverable error Error! Reference source not found.. TM data transfer transmits upper layer PDUs without adding any protocol information, possibly including segmentation/reassembly functionality. It ignores any errors in received PDUs and just passes them onto the upper layer for further processing. UM data transfer transmits upper layer PDUs without guaranteeing delivery to the peer entity. PDUs which are received with errors are


131

discarded without any notice to the upper layer or to the peer entity. AM data transfer transmits upper layer PDUs with guaranteed delivery to the peer entity. Error-free delivery is ensured by means of retransmission. For this service, both in-sequence and out-of-sequence delivery are supported. As mentioned before, in the presented scheme unequal error protection is applied to the different types of video frames transmitted. The I-frame which has the most crucial data is always transmitted using the AM data transfer mode of service to ensure error-free delivery.

4. Error classification

From the discussion about the interframe predictive coding hierarchical structure of MPEG-4 it is clear that various parts of the bit stream have different levels of importance. We have classified errors into three main classes based on their impact on the coded hierarchy:

• Most critical errors – Class-A errors • Less critical errors – Class-B errors • Least critical errors – Class-C errors

Class-A errors contain errors which can significantly degrade the received video quality. This class includes errors introduced in the video sequence header, I-frame headers and I-frame data. Class-B, includes errors in P-frame headers and P-frame data while Class-C includes errors in B-frame headers and B-frame data. In the following sections we discuss two different methods to deal with Class-A and Class-B errors separately as described

4.1 Unequal Error Protection (UEP) for Class-A errors

As mentioned before, the I-frame has relatively more importance than the P and B frames. Hence an I-frame data needs to be transmitted with higher protection using the proposed UEP scheme. This is achieved by using a different mode of transmission on the RLC layer of UMTS. The I-frames are transmitted using AM while other frames are transmitted using TM. If any error is introduced during radio transmission, the specific PDU of the I-frame is retransmitted using AM. This guaranties error-free delivery of the I-frames.

4.2 NEWPRED for Class-B errors

The MPEG-4 ISO/IEC 14496 (Part-2) Error! Reference source not found. standard provides error robustness and resilience capabilities to allow accessing of image or video information over a wide range of storage and transmission media. The error resilience tools developed for this part of ISO/IEC 14496 can be divided into three major categories: synchronization, data recovery and error concealment. The NEWPRED feature falls into the category of error concealment procedures. Recovery from temporal error

propagation is an indispensable component of any error robust video communication system. Errors introduced during transmission can lead to frame mismatch between the encoder and the decoder, which can persist until the next intra refresh occurs. Where an upstream data channel exists from the decoder to the encoder, NEWPRED or demand intra refresh can be used. NEWPRED is a technique in which the reference frame for interframe coding is replaced adaptively according to the upstream messaging from the decoder. NEWPRED uses upstream messages to indicate which segments are erroneously decoded. On receipt of this upstream message the encoder will subsequently use only the correctly decoded part of the prediction in an inter-frame coding scheme. This prevents temporal error propagation without the insertion of intra coded MBs (Macro Blocks) and improves the video quality in noisy multipath environments. The following section explains the concept in more detail.

Figure 4. Error propagation due to interframe decoding dependencies When a raw video sequence is encoded utilizing MPEG-4, each of the raw video frames is categorized according to the way in which predictive encoding references are used. An Intra-coded (I) frame is coded using information only from itself. A Predictive-coded (P) frame is coded using motion compensated prediction from a past reference frame(s). While a Bidirectional predictive-coded (B) frame is a frame which is coded using motion and texture compensated prediction from a past and future reference frames. A disadvantage of this coding scheme is that transmission errors occurring in a frame which is used as a reference frame for other P or B frames, causes errors to propagate into the video sequence. This propagation continues until an intra-refresh is applied. In the example shown in Figure-4, an error occurred in frame P3 which acts as a reference frame for P4, subsequent P-frames and B-frames (B5, B6 etc), until the next intra-refresh frame (I2) occurs. Where the transmission error has damaged crucial parts of the bit-streams such as a frame header, the decoder may be unable to decode the frame which it then drops. If this dropped frame is a P-frame, none of the frames that are


132

subsequently coded with reference to this dropped P-frame can be decoded. So in effect all subsequent frames until the next intra-refresh are dropped. This situation can seriously degrade the received video quality leading to the occurrence of often long sequences of static or frozen frames. If through the use of an upstream message the encoder is made aware of errors in the particular P-frame (P3), the encoder can change the reference frame for the next P-frame (P4) to the previous one which was received correctly (P2). P-frames and B-frames after P4 then refer to the correctly decoded P4, rather than the faulty P3 frame. The technique therefore reduces error propagation and frame loss occurring from dropped P-frames. This method can significantly improve the performance of the received video quality. To implement the NEWPRED feature, both the encoder and decoder need buffer memories for the reference frames. The required buffer memory depends on the strategy of the reference frame selection by the encoder, and transmission delay between the encoder and decoder. In this paper results for two different schemes are reported.

5. Error detection

To implement NEWPRED it is important to identify errors at the frame level at the decoding end. Mapping of errors identified by the lower layer to the application layer with the precision of a single video frame or video packet often results in a complicated process consuming a considerable amount of processing resource and introduces severe processing delays. Insertion of CRC bits in the standard MPEG-4 bit-stream at frame level provides a simpler solution to this problem. With the insertion of extra bits which are not defined as part of the standard video encoded sequence, would normally result in the production of an incompatible bit-stream which standard decoders will not be able to decode. But as mentioned in Error! Reference source not found. this would not be the case if these bits are inserted at a particular place of the standard MPEG-4 bit-stream Error! Reference source not found.. While decoding, the decoder is aware of the total number of macroblocks (MB) in each frame. It starts searching for a new video frame header after decoding these macroblocks. It ignores everything between the last marcoblock of the frame and the next frame header as padding. If generated CRC bits are inserted at this place as shown in Figure-5, after the last macroblock and before the next header, this should preserve the compatibility of the bit-stream with standard MPEG-4. Such insertion of CRC bits does not affect the normal operation of any standard MPEG-4 decoder. Also because the inserted CRC only consists of the 16 bits, generated using the polynomial G16 defined for the MAC layer of the UMTS architecture, it is not possible for it to emulate any start code sequences.

Figure 5. CRC insertion This method adds an extra 16 bits of overhead to each frame but the performance improvements in video quality with NEWPRED implementation coupled with CRC error detection, justifies this overhead. The discussion in the previous sections described the proposed methods for error-protection for different parts of the video stream. The following section describes the simulation scenarios used to test the implementation of these methods. The entire end-to-end transmission and reception system scenario is developed using different simulation tools ranging from simple C programming to using SPW 4.2 by CoWare and OPNET 10.5A.

6. Video clips and video quality measurement techniques used

For evaluation purposes three standard video test sequences were used, these being: Mother-Daughter, Highway and Foreman. Each of these clips is of 650 frames in length of QCIF (176 × 144) resolution and is encoded with the standard MPEG-4 codec at 10 fps. The objective video quality is measured by the PSNR (Peak Signal to Noise Ratio) as defined by ANSI T1.801.03-1996 Error! Reference source not found.. One more sophisticated model which is developed and tested by ITS (Institute for Telecommunication Sciences) is the Video Quality Metric (VQM) Error! Reference source not found.. VQM has been extensively tested on subjective data sets and has significantly proven correlation with subjective assessment. One of the many models developed as a part of this utility is Peak-Signal-to-Noise-Ratio VQM (VQMP). This model is optimized for use on low-bit rate channels and is also used in this paper for the near subjective analysis of the received video sequence. Typical VQMP is given by:

( )6675.251701.011

−×+= PSNRe

VQMp

The higher the PSNR value the better is the objective quality whilst the lower the VQMP value the better is the subjective quality of the video sequence.


133

7. Upper layer protocol overheads and PDCP header compression

Each lower layer defined in the UMTS protocol stack provides services to the upper layer at defined Service Access Points (SAPs) Error! Reference source not found.. These protocols add a header part to the video frame payload to exchange information with peer entities. Depending upon protocol configurations and the size of the video frames, these headers can be attached to each video frame or multiple video frames can be used as a single payload as defined by RFC-3016 Error! Reference source not found.. As many of the successive headers contain a huge amount of redundant data, header compression is applied in the form of the Packet Data Convergence Protocol (PDCP) Error! Reference source not found.. With PDCP compression, higher layer protocol headers like RTP, UDP and IP headers are compressed into one single PDCP header. In the presented simulation, header attachments, compression and header removal was achieved using the C programming language. Figure-6 shows a typical structure of the video frame payload and header before it is submitted to the RLC layer for further processing.

Figure 6. PDCP header compression Once the PDCP compression is achieved this packet is then submitted to the RLC layer for further processing. The RLC layer can be configured in any one of the transparent, unacknowledged or acknowledged modes of transmission. The RLC then submits the PDU to the lower layers, where the MAC layer and physical layer procedures are applied as appropriate. In the presented simulation, each video frame coming from the application layer is mapped to the RTP layer. Each frame is mapped into a separate RTP packet after addition of the above mentioned RTP, UDP, IP headers with PDCP compression, and if required, RLC headers. For TM RLC service, the RLC SDU cannot be larger than the RLC PDU size. For this reason, if the video frame size is larger than the RLC payload size for TM, then the frame is fragmented into different RTP packets. Other protocol headers are then added to these packets separately. Each RLC SDU is then either of the same size as the RLC PDU or smaller. For AM RLC service, an entire frame can be considered as one RLC SDU regardless of the size. For this reason protocol headers are added to each video frame. This RLC SDU is then fragmented into different RLC PDUs at the RLC layer.

Once these PDUs are mapped onto different transport channels, they are transmitted using the WCDMA air interface. The physical layer of WCDMA is simulated using the SPW tool by CoWare which models an environment to generate error patterns for various types of channel propagation conditions defined by the 3GPP standards. A 64 kbps downlink data channel and 2.5 kbps control channel were used for this UMTS simulation. These two channels were multiplexed and transmitted over the WCDMA air-interface. The transmission time interval, transmission block size, transmission block set size, CRC attachment, channel coding, rate matching and inter-leaving parameters were configured for both channels compliant with the 3GPP TS 34.108 specification Error! Reference source not found.. The typical parameter set for reference RABs (Radio Access Barriers) and SABs (Signaling Access Barriers) and relevant combinations of them are presented in this standard. The different channel propagation conditions used in the simulation were static, multi-path fading, moving and birth-death propagation. These channel conditions are described in some details in the following section.

8. Propagation conditions

Four different standardized propagation conditions – static, multi-path fading, moving and birth-death were used to generate different error patterns. The typical parameter sets for conformance testing as mentioned in 3GPP TS 25.101 Error! Reference source not found. is used for the radio interface configuration. A common set of parameters for all kinds of environment is listed in Table 1, while any specific parameters to the environment are mentioned in the respective sections.

Table 3. Common set of parameters Interference -60 dB

Received signal / Noise (SNR) -3.0 dB

AWGN noise 4 *10-9 watts

Eb/No (Overall) 6.01 dB

BER (Bit Error Rate) 0.001

Data Rate (Downlink) 64 kbps

8.1 Static propagation condition

As defined in 3GPP TS 25.101 V6.3.0, the propagation condition for a static performance measurement is an Additive White Gaussian Noise (AWGN) environment. No fading and multi-paths exist for this propagation model.


134

Table 2, lists the received values of BLER and FER for this propagation condition.

Table 4. Parameters for static propagation conditions BLER (Block Error Rate) 0.034

FER (Frame Error Rate) 0.0923

8.2 Multi-path fading propagation conditions

Multi-path fading normally follows a Rayleigh fading pattern. In this simulation Case-2 as mentioned by TS 25.101 is used with frequency band-1 (2112.5 MHz) and the number of paths set to 3 with relative delay between each paths of 0, 976 and 20000 ns and a mean power of 0 dB for all three paths. The delay model used in this case is fixed. The vehicle speed is configured to be 3 km/h. The received values of BLER and FER are given in Table 3.

Table 5. Parameters for multi-path fading propagation conditions

BLER (Block Error Rate) 0.007


8.3 Moving propagation conditions

The dynamic propagation conditions for this environment for the test of the baseband performance is a non-fading channel model with two taps as described by 3GPP TS 25.101. One of the taps, Path-0 is static, and other, Path-,1is moving,. Both taps have equal strengths and phases but unequal time difference exists between them. The received values of BLER and FER are given in Table 4.

Table 6. Parameters for moving propagation conditions BLER (Block Error Rate) 0.031


8.4 Birth-Death propagation conditions

These conditions are similar to the Moving propagation condition except, in this case both taps are moving. The positions of paths appear randomly and are selected with an equal probability rate. Table 5 lists the received values of BLER and FER for this propagation condition.

Table 7. Parameters for birth-death propagation conditions BLER (Block Error Rate) 0.037


Generated error patterns are applied to the data transmitted from the RLC layer. Different RLC modes are simulated using the C programming language.

9. Simulation Results

As mentioned before VQMP is used as the quality measurement metric in this simulation. Table 6 presents the VQMP values obtained during the simulations. The following conventions are used in Table 6. Video clips names:

- Mother and Daughter – MD - Highway – HW - Foreman – FM

VQMP without UEP and without NEWPRED – Results A VQMP with UEP and without NEWPRED – Results B VQMP without UEP and with NEWPRED – Results C VQMP with UEP and with NEWPRED – Results D Note the lower the VQMP value the better the subjective image quality

Table 8. Simulation results

Video Clip

Results A

Results B

Results C

Results D

Static Environment MD 0.70 0.46 0.45 0.25 HW 0.64 0.36 0.32 0.19 FM 0.85 0.83 0.74 0.68 Multi-path Environment MD 0.27 0.18 0.18 0.16 HW 0.37 0.21 0.18 0.13 FM 0.62 0.53 0.47 0.34 Moving Environment MD 0.63 0.45 0.44 0.23 HW 0.55 0.35 0.31 0.26 FM 0.83 0.70 0.67 0.52 Birth-Death Environment MD 0.58 0.32 0.35 0.27 HW 0.49 0.40 0.31 0.22 FM 0.85 0.72 0.68 0.45

The following observations can be made on the above results. The received VQMP values for all video sequences are improved by implementation of the UEP and NEPWRED methods. It can be observed that in most cases the VQMP values show greater improvement with implementation of the NEWPRED method than the improvements obtained with the UEP method. Errors in the wireless environment occur in bursts and are random in nature. Due to the time varying nature of the wireless environment it is not possible to predict the exact location of the error. The NEWPRED method provides protection to the P-frames while UEP is aimed at protecting the I-frames. As P-frames are much more frequent than I-frames in the encoded video sequence, P-frames are more susceptible to these bursty randomly distributed errors. This explains why


135

the NEWPRED implementation gives better performance than the UEP implementation. Combined implementation of the UEP and NEWPRED methods always outperforms a single method implementation and gives the best quality of received video. The same observation is made in terms of performance assessment using the VQMP score in most cases. Objective assessment of VQMP scores is highly dependent on bit-by-bit comparison which does not take into consideration the number of frozen frames. In most cases frozen frames give better correlation with original frames compared to frames with propagation errors.

10. Network capacity and coverage improvements

As shown by the simulation results above, implementation of UEP and NEWPRED enhances the received video quality. This implies that the same video quality or QoS can be attained at lower Eb/No if the two proposed methods are utilized. A reduced Eb/No value have a direct impact on radio network coverage and capacity planning for UMTS networks Error! Reference source not found., Error! Reference source not found.. A number of variable parameters determine the resultant received video quality. The nature of the video clip, the encoding parameters, channel propagation environment and many more variables can have varying effects on the overall video quality. For these reasons it is not easy to quantify exactly how much reduction in Eb/No value would be achieved using the techniques described above. For the simulation purpose, a reduction of 1 dB is assumed. Network capacity and coverage enhancements are quantified assuming 1 dB reduction in the required Eb/No value for video service. The basic radio network used for this simulation is shown in following Figure. 10 Node-B sites each with three sectors are simulated. 2000 UEs are randomly distributed over an area of 6 × 6 km2. The UMTS pedestrian environment for the micro cell is selected as a propagation model. The mobiles are assumed to be moving at a speed of 3 km/h. Downlink bit-rate of 64 kbps is selected and other link layer parameters and results are imported from simulations discussed above. In the first run, an Eb/No value of 4.6 dB is used and the total throughput per cell is represented in graphical form as shown in Error! Reference source not found..

Figure 7. Throughput in downlink per cell at Eb/N0 – 4.6 dB

Now in second run, the Eb/No value is decreased by 1 dB to 3.6 dB and the total throughput per cell is presented in graphical form in Error! Reference source not found..

Figure 8. Throughput in downlink per cell at Eb/N0 – 3.6 dB As can be clearly compared from Error! Reference source not found. to Error! Reference source not found., the decrease in the Eb/No value by 1 dB results in a significant increase in the total throughput per cell in the downlink direction.

11. Conclusions

As can be seen from the simulation results, implementation of UEP and NEWPRED results in significant improvements on the received video quality. Improvements achieved using these methods provides extra margin for network operators to increase capacity. With implementation of these error concealment methods the same video quality can be achieved using a lower Eb/No which in turn provides flexibility for network operator to increase the number of users. This implementation obviously requires some processing overhead on both the encoder and decoder sides, but considering the increasing processing power of mobile stations, this should present a major obstacle and the error concealment methods described should provide considerable enhancements. References: [1] International Standard ISO/IEC 14496-2: Information

Technology - Coding of audio-visual objects-Part 2, Visual, International Organization for Standardization. 2001.

[2] 3GPP, Technical Specification Group Radio Access Network; Radio Link Control (RLC) protocol specification; 3GPP TS 25.322, V4.12.0.

[3] Pereira, F., Ebrahimi, T., The MPEG-4 Book; Prentice Hall PTR (July 20, 2002), ISBN-10: 0130616214.

[4] 3GPP, Technical Specification Group Radio Access Network;Radio interface protocol architecture; 3GPP TS 25.301 (2002-09), Ver 5.2.0.


136

[5] 3GPP, Technical Specification Group Services and System Aspects; Quality of Service (QoS) concept and architecture, 3GPP TS 23.107 v4.0.0.

[6] 3GPP, Technical Specification Group Radio Access Network; MAC protocol Specification; 3GPP TS 25.321, V4.0.0.

[7] 3GPP, Technical Specification Group Radio Access Network; PDCP protocol specification; 3GPP TS 25.331 (2003-12), Ver 6.0.0.

[8] 3GPP, Technical Specification Group, Radio Access Network; Broadcast/Multicast Control BMC; 3GPP TS 25.324, v4.0.0.

[9] Worrall, S., Sadka, A., Sweeney, P., Kondoz, A., Backward compatible user defined data insertion into MPEG-4 bitstream. IEE Electronics letters.

[10] ATIS Technical Report T1.TR.74-2201: Objective Video Quality Measurement using a Peak-Signal-to-Noise Ratio (PSNR) Full Reference Technique. October 2001, Alliance for Telecommunications Industry Solutions.

[11] Wolf, S., Pinson, M., Video Quality Measurement Techniques. June 2002, ITS, NTIA Report 02-392.

[12] RFC-3016, RTP Payload Format for MPEG-4 Audio/Visual Streams, November 2000.

[13] 3GPP, Technical Specification Group Terminals; Common test environment for UE conformance testing; 3GPP TS 34.108 (2003-12), Ver 4.9.0.

[14] 3GPP, Technical Specification Group Radio Access Network; UE radio transmission and reception (FDD); 3GPP TS 25.101 (2003-12), Ver 6.3.0.

[15] Laiho, J., Wacker, A., Novosad, T., Radio Network Planning and Optimization for UMTS, Wiley, 2nd edition (December 13, 2001).

[16] Holma, H., Toskala, A., WCDMA for UMTS: Radio Access for Third Generation Mobile Communications, Wiley, 3rd edition (June 21, 2000).

Dr. Bhumin Pathak received his M.Sc. and Ph.D. degree from Oxford Brookes University, Oxford, UK. He has been working as Systems Engineer at Airvana Inc., since March 2007.

Dr. Geoff Childs is Principal Lecturer at School of Technology at Oxford Brookes University, Oxford, UK. Dr. Maaruf Ali is Senior Lecturer at School of Technology at Oxford Brookes University, Oxford, UK.


137

Comprehensive analysis of UMTS Authentication and Key Agreement

Engr. Mujtaba Hassan1, Engr. Munaza Razzaq2 and Engr. Asim Shahzad3

1 Kohat University of Sciences and Technology, Kust Institute of Engineering Sciences,

Department of Computer Engineering, Kohat. Pakistan [email protected]

2 NWFP University of Engineering and Technology Peshawar, Abbottabad Campus

Near Post Graduate College No. 2 University Road Mandian, District Abbottabad, Pakistan. [email protected]

3University of Engineering and Technology Taxila, Pakistan

[email protected]

Abstract: This paper presents an analysis and evaluation

of the security of UMTS. This paper provides information on the 3rd generation mobile communication system, UMTS, its Authentication and Key Agreement (AKA) procedures and security aspects. The AKA procedure is the essence of authenticating a user to the network and vice versa. AKA procedures in UMTS have increased security compared with GSM. The new feature of two-way authentication eliminates the problem with false base stations. This is a very important security improvement. Even though the security has improved in some areas, there are still security features that should be improved. Some weaknesses are also pointed out in UMTS. One of the major weaknesses in UMTS is sending IMSI in plaintext. We have simulated this weakness pointed out in the literature survey. In this paper we have shown simulation scenarios for an attack on IMSI of MS when it sends a registration request to the serving network. Keywords: UMTS, AKA, IMSI, Security

1. Introduction The Universal Mobile Telecommunications System (UMTS) is one of the new ‘third generation’ (3G) mobile cellular communication systems being developed within the framework defined by the International Telecommunications Union (ITU) known as IMT-20001. UMTS security builds on the success of Global System for Mobile communications (GSM) by providing new and enhanced security features. UMTS aims to provide a broadband, packet-based service for transmitting video, text, digitized voice, and multimedia at data rates of up to 2 Mbps while remaining cost effective. UMTS utilizes Code Division Multiple Access (CDMA) as it is far better suited for fast data stream transfer. Although GSM security has been very successful but GSM suffers from security problems such as weak authentication and encryption algorithms, short secret key length (only 32 bits) with no network authentication. This has lead to false base station attack and lack of data integrity, allowing denial of service attacks, limited encryption scope and insecure key transmission. An objective of the UMTS security design was to address weaknesses [1] in GSM. UMTS introduces new

and enhanced security features that are designed to stop threats [2], [3], [4], [5], [15]. These include: Mutual Authentication which allows the mobile user and serving network(SN) to authenticate each other [6], Network to Network security that secure communication between serving networks which suggested the use of IP security to do so, wider security scope, secure International Mobile Subscriber identity (IMSI) usage, user to mobile station authentication where more flexibility in that security features can be extended and enhanced as required by new threats and services plus GSM compatibility.

2. UMTS Security Architecture The security architecture in UMTS is based on three security principles: Authentication, Confidentiality and Integrity

2.1 Authentication Authentication is provided to assure the claimed identity of an entity. A node that wants to authenticate itself to someone has to show its own identity. This can be done either by showing knowledge of a secret only the nodes involved knows; or by letting a third party that both nodes trusts, vouch for their identities. Authentication in UMTS is divided into two parts:

• Authentication of the user towards the network • Authentication of the network towards the user

2.2 Confidentiality Confidentiality is to keep information secured from unwanted parties. With more and more people using the terminals for both personal and business calls (e.g. online services like banking) the need for keeping the communication secure grows rapidly. Confidentiality in UMTS is achieved by ciphering communications between the subscriber and the network and by referring to the subscriber by temporary (local) identities instead of using the global identity, IMSI. The properties that should be confidential are:

• The identity of the subscriber


138

• The current location of the subscriber • User data (both voice and data communications

should be kept confidential).

2.3 Integrity Sometimes a message’s origin or contents have to be verified. Even though it might come from a previously authenticated party, the message may have been tampered with. To avoid this, integrity protection is necessary. The message itself might not even have to be confidential; the important thing is that it’s genuine. The method for integrity protection in UMTS is to generate stamps to be added to messages. The stamps can only be generated at the nodes that know the keys derivate of the pre-shared secret key, K. They are stored in the Universal Subscriber Identity module (USIM) and the Authentication Centre (AuC). It is very important to offer integrity protection, especially since the SN often is operated by another operator than the subscriber’s own operator. The property that should be integrity protected is: Signaling messages and signaling data.

3. Authentication and Key Agreement The authentication is performed by the Authentication and Key Agreement (AKA) procedure [7]. The AKA procedure is built on the RIJNDAEL block cipher [8]. In addition to authentication, AKA procedure also results in the Cipher Key (CK) and the Integrity Key (IK). In UMTS, only the encryption mode of the RIJNDAEL block cipher is used [9] as an iterated hash function [10]. The block and key length have been set to 128-bit. The USIM AKA (Fig. 1) is chosen in such a way as to achieve maximum compatibility with the current GSM/GPRS security architecture. [11], [12]. USIM AKA is a one-pass challenge response protocol [13], [14].

Figure 1. UMTS Authentication and Key Agreement [8]

3.1 When to use AKA • Registration of a user in a SN • After a service request • Location Update Request

• Attach Request • Detach request • Connection re-establishment request

Registration of a subscriber in a SN typically occurs when the user goes to another country. The first time the subscriber then connects to the SN, he gets registered in the SN. Service Request is the possibility for higher-level protocols/applications to ask for AKA to be performed. E.g. performing AKA to increase security before an online banking transaction. The terminal updates the Home Location Register (HLR) regularly with its position in Location Update Requests. Attach request and detach request are procedures to connect and disconnect the subscriber to the network. Connection re-establishment request is performed when the maximum number of local authentications has been conducted. In the following an overview of how the UMTS AKA protocol works is given:

3.2 Procedures Authentication and key agreement (Fig. 2) [16] consists of two procedures: First, the Home Environment (HE) distributes authentication information to the SN. Second, an authentication exchange is run between the user and the SN.

Figure 2. Overview of Authentication and Key Agreement

[16]

Figure 2 shows that, after receiving an authentication information request, the HE generate an ordered array of n authentication vectors. Each authentication vector (AV) consists of five components (and hence may be called a UMTS ‘quintet’ in analogy to GSM ‘triplets’): A random number RAND, an expected response XRES, a cipher key CK, an integrity key IK and an authentication token AUTN. This array of n authentication vectors is then sent from the HE to the SN. It is good for n authentication exchanges between the SN and the USIM. In an authentication exchange the SN first selects the next (the i-th) AVfrom the array and sends the parameters RAND(i) and AUTN(i) to


139

the user. The USIM checks whether AUTN(i) can be accepted and, if so, produces a response RES(i) which is sent back to the SN. AUTN(i) can only be accepted if the sequence number contained in this token is fresh. [16] The USIM also computes CK(i) and IK(i). The SN compares the received RES(i) with XRES(i). If they match, the SN considers the authentication exchange to be successfully completed. The established keys CK(i) and IK(i) will then be transferred by the USIM to the mobile equipment(ME) and by the Visitor Location Register (VLR) or Serving General packet Radio Service Support Node(SGSN) to the Radio Network Controller (RNC); the keys are then used by the ciphering and integrity functions in the Mobile Station (MS) and in the RNC.

4. AKA Algorithms The security features of UMTS are fulfilled with a set of cryptographic functions and algorithms.[17] A total of 10 functions are needed to perform all the necessary features, f0-f5, f1*, f5*, f8 and f9.

Table 1: Authentication Functions [17]

Function

Description

Output

f0

Random challenge generating function

RAND

f1

Network authentication function

MAC-A/XMAC-A

f1*

Re-synchronization message authentication function

MAC-S/XMAC-S

f2

User authentication function

RES/XRES

f3

Cipher key derivation function

CK

f4

Integrity key derivation function

IK

f5

Anonymity key derivation function

AK

f5*

Anonymity key derivation function for the resynchronization message function

AK

f8

Confidentiality key stream generating function

<Key stream block>

f9

Integrity stamp generating function

MAC-I/XMAC-I

5. Key Generation functions The functions f1-f5* are called key generating functions and are used in the initial Authentication and Key Agreement procedures.

5.1 Functions in AuC When generating a new AV the AuC reads the stored value of the sequence number, SQNHE and then generates a new SQN and a random challenge RAND [17].

Figure 3. AV and Key Generation in AuC [17]

.

5.2 Functions in USIM To generate the output keys in the USIM it has only one of the four parameters that the AuC has the pre-shared secret key (K). The rest of the parameters it has to receive from the AuC.

Figure 4. RES Generation in USIM [17]

When the USIM receives the (RAND||AUTN) pair it starts by generating the Anonymity Key (AK) by applying function f5 on the received RAND. By XOR-in the AK with the (SQN XOR AK) from the Authentication Token, the sequence number of the AuC is revealed (SQNHE). [17] The secret key K is then used with the received AMF, SQN and RAND to generate the Expected Message Authentication Code (XMAC-A). This is then compared with the MAC-A. If the X-MAC and MAC matches, the USIM have authenticated that the message (RAND||AUTN pair) is originated in its HE.

6. Authentication parameters The parameters used in the Authentication and Key Agreement procedure are: AV, AUTN, RES and XRES, MAC-A and XMAC-A, AUTS, MAC-S and XMAC-S [17].


140

Table2: Size of Authentication Parameters [17]

7. Integrity function The integrity protection of signaling messages between Mobile Equipment (ME) and RNC starts during the security mode set-up as soon as the integrity key and integrity protection algorithm is known. A MAC function is applied to each individual signaling message at the RRC layer of UMTS Terrestrial Radio Access Network (UTRAN) protocol stack [14]

Figure 5. Integrity Function [14]

Figure 5 illustrates the use of integrity algorithm f9 to authenticate the data integrity of an RRC signaling message. Input Parameters to the integrity function are COUNT, IK, FRESH and Message [17].

8. Confidentiality function In the 3G Security, user data and some signaling information elements are considered sensitive and may be confidentiality protected [14]. The need for a protected mode of transmission is fulfilled by a confidentiality function f8 as shown in Fig.6 [14]. The encryption function is applied on dedicated channels between the ME and the RNC [14].

Figure 6. Confidentiality function

Table 4: Input parameters to confidentiality function [17]

Both f8 and f9 algorithms are based on KASUMI algorithm. The block cipher KASUMI is a modification of MISTY1 [14], [18]. KASUMI has been tested by the design team and independent evaluation teams using crypt analytical methods [19]. KASUMI constructions have also been proven to provide pseudo randomness.

9. Weaknesses in UMTS security mechanisms

To sum up, the main weaknesses in UMTS security mechanism are:

• Integrity keys used between UE and RNC generated in VLR/SGSN are transmitted unencrypted to the RNC (and sometimes between RNCs).

• IMSI is transmitted in unencrypted form. • For a short time during signaling procedures,

signaling data are unprotected and hence exposed to tampering.

10. Simulation Scenarios Our simulation which identifies the problem is developed and tested on the version designed for the Microsoft Windows environment, MATLAB R2008a. Figure 7 is the screenshot for the normal AKA procedure and figure 8 shows that an intruder has captured the IMSI of MS, as it was transmitted in plain text.


141

Figure 7. AKA normal procedure

It can be seen from Figure 8 and 9 that after the intruder has captured IMSI, it will send this IMSI to VLR/SGSN and has authenticated itself as an original MS although it is an attacker that is acting as an authenticated user.

Figure 8. An intruder captures the IMSI

Figure 9. An intruder launches the attack

Rest of the five USIM functions can be generated through SIM cloning process by the attacker in order to prove that it is the real user which has requested the service. This problem can be avoided by proposing such algorithm which also encrypts the IMSI during authentication process. This problem has also been identified in [22], [23]. 11. Further developments in UMTS security Work on the next UMTS release has started. This will introduce new security features. Many of these features will be introduced to secure the new services which will be introduced, e.g. presence services, push services and multicast/broadcast services. Looking more into the future, mobile cellular systems will have to accommodate a variety of different radio access networks including short-range wireless technologies, connected to a common core network. On the user side the concept of a monolithic terminal, as we know it, is dissolving. Distributed terminal architectures are appearing whose components are interconnected by short-range radio links. These new developments represent a major challenge to the UMTS security architecture. A collaborative research project funded by the European Union and called SHAMAN (Security for Heterogeneous Access in Mobile Applications and Networks) have tackled these issues. A separate project is also underway to identify research topics in the area of mobile communications; this project is called PAMPAS (Pioneering Advanced Mobile Privacy and Security).


142

11. Conclusion

AKA procedures in UMTS have increased security compared with GSM. The new feature of two-way authentication eliminates the problem with false base stations. This is a very important security improvement. Even though the security has improved in some areas, there are still security features that should be improved. It is not sufficient to just require integrity protection on signaling messages. All messages should be integrity checked, but indirectly by requiring confidentiality protection together with integrity. AKA concept is used to perform authentication of the user and network, as opposed to 2G systems, which only authenticated users in a system. The confidentiality algorithm is stronger than its GSM predecessor. The integrity mechanism works independent of confidentiality protection and provides protection against active attacks. The design of cryptographic algorithms is open and they are extensively crypto analyzed. Moreover, the architecture is flexible and more algorithms can be added easily. Although 3G Security marks a large step forward however there are some short comings. IMSI is sent in clear text when allocating TMSI to the user. In this paper this problem is discussed in detail with the help of simulation scenarios. Some future work on UMTS security architecture is also elaborated in this paper.

References [1] “3G Security; Security principles and objectives”,

Release 4, March, 2001. [2] Johnson, M. (2002). Revenue Assurance, “Fraud and Security in 3G Telecom Services. VP Business Development Visual Wireless AB”, Journal of

Economic Management, 2002, Volume 1, Issue 2. [3] Stalling, W. Cryptography and Network Security, Principles and Practice. 3rd edition. USA, Prentice

Hall. 2003 [4] Stefan, P, and Fridrich R. (1998). “Authentication Schemes for 3G mobile radio Systems”. The Ninth

IEEE International Symposium on, 1998. [5] Zhang, M. and Fang, Y. (2005).” Security Analysis

and Enhancements of 3GPP Authentication and Key Agreement Protocol”. IEEE Transactions on wireless Communications, Vol. 4, No. 2. 2005 [6] 3GPP TS 21.133. “3GPP Security; Security

Architecture”. [7] 3GPP TS 33.102, Technical Specification Group Services and System Aspects,” 3G Security. Security Architecture”, V 4.2.0 September 2001. [8] Daemen J, Rijmen V. AES Proposal: Rijndael.

Available:

http://csrc.nist.gov/encryption/aes/round2/AESAlgs/Rijndael

[9] 3GPP TS 35.206 V4.0.0, Technical Specification Group Services and System Aspects, 3G Security,

Specification the MILENAGE Algorithm Set: An example algorithm set for the 3GPP authentication and

key generation functions f1, f1*, f2, f3, f4, f5 and f5*, Document 2: Algorithm Specification, April 2001.

[10] 3GPP TR 35.909 V4.0.0, Technical Specification Group Services and System Aspects, 3G Security, Specification of the MILENAGE. Algorithm Set: An example algorithm set for the 3GPP authentication and key generation functions f1, f1*, f2, f3, f4, f5 and f5*, design and evaluation, April 2001.

[11] C. J. Mitchell, “Security for Mobility”, Institute of Electrical Engineers, December, 2004.

[12] 3GPP TS 33.102 (5.2.0), “3G Security; Security Architecture”, Release 5, June, 2003.

[13] ISO/IEC 9798-4: "Information technology – Security techniques - Entity authentication - Part 4: Mechanisms using a cryptographic check function"

[14] Evaluation of UMTS security architecture and services 1-4244-9701-0/06/$20.00 ©2006 IEEE [15] “Extension of Authentication and Key Agreement

Protocol (AKA) for Universal Mobile Telecommunication System (UMTS)International Journal of Theoretical and Applied Computer Sciences” Volume 1 Number 1 (2006) pp. 109–118 (c) GBS Publishers and Distributors (India) http://www.gbspublisher.com/ijtacs.htm

[16] UMTS Security by K. Boman, G. Horn, P. Howard, and V. Niemi October 2002 issue of IEE Electronics & Communication Engineering Journal.

[17] UMTS Authentication and Key Agreement - A comprehensive illustration of AKA procedures within the UMTS system By Jon Robert Dohmen , Lars Sømo Olaussen, , Grimstad - Norway, May 2001

[18] M. Matsui, “Block encryption algorithm MISTY” in Proceedings of Fast Software Encryption (FSE’97), Volume 1267, Springer-Verlag, 1997.

[19] 3GPP TR 33.909 V1.0.0 (2000-12) Technical Report; 3rd Generation Partnership Project; Technical Specification Group Services and System Aspects; Report on the Evaluation of 3GPP Standard Confidentiality and Integrity Algorithms (Release 1999)

[20] S. Babbage and L. Frisch, “On MISTY1 higher order differential cryptanalysis”, in Proceeding of International Conference on Information Security and Cryptology (ICISC 2000), Lecture Notes in Computer Science Volume. 2015, Springer-Verlag, 2001.

[21] U. Kühn, “Cryptanalysis of reduced-round MISTY”, in Proceedings of Eurocrypt’01, Lecture Notes in Computer Science, Volume 2045, Springer-Verlag, 2001.

[22] G. M. Koien, “Privacy enhanced cellular access security” ACM - 2005

[23] ” S.Y. A.-R. A. Mustafa Al-Fayoumi, Shadi Nashwan, “A new hybrid approach of symmetric/asymmetric authentication protocol for future mobile networks IEEE - 2007


143

Automation of Data Warehouse, Extraction, Transformation and Loading Update Cycle

Atif Amin1 and Abdul Aziz2

1Faculty of Information Technology, University of Central Punjab, Lahore, Pakistan

atif.amin @ucp.edu.pk

2Faculty of Information Technology, University of Central Punjab, Lahore, Pakistan

[email protected] Abstract: Business enterprises invest lots of money to develop data warehouse that gives them real, constant and up to date data for decision making. To keep data warehouse update, traditionally, data warehouses are updated periodically. Periodic updates make a delay between operational data and warehouse data. These updates are triggered on time set; some may set it to evening time when there is no load of work on systems. This fixing of time does not work in every case. Many companies run day and night without any break, then in these situations periodic updates stale warehouse. This delay depends upon the periodic interval, as interval time increase the difference between operational and warehouse data also increase. The most recent data is unavailable for the analysis because it resides in operational data sources. For timely and effective decision making warehouse should be updated as soon as possible. Extraction, Transformation and Loading (ETL) are designed tools for the updating of warehouse. When warehouse is refreshed for the update purpose, it often gets stuck due to overloading on resources. Perfect time should be chosen for the updating of warehouse, so that we can utilize our resources efficiently. Warehouse is not updated once, this is cyclic process. We are introducing automation for ETL, Our proposed framework will select best time to complete the process, so that warehouse gets updated automatically as soon as resources are available without compromising on data warehouse usage.

Keywords: ETL, Updating, Loading, Data Warehouse.

1. Introduction Computers were used for data transactions and to provide information to support decision making. As early as the merit of placing information in different platform for decision making were used. This approach is for easy to access needed data, improves system response time and assures security and data integrity. These systems were pioneer to use this approach. Its end user saw many applications for example executive summary and etc for having specially prepared data.

Before two decades, organizations developed data warehouse to provide users decision support system. There are different approaches from earlier systems. One is the use of special purpose system which task was to clean data, extract useful data and loading all data into data warehouse. Depending on the application needs many software can be used to store data. Enhanced data access tools make it easy for end user to access, transform, analyze and display computed information without writing queries.

Many organizations are becoming customer focused. They are using data mining to provide information for

business advantage. In many organizations the way for users is to obtain timely information for correct decision. The fundamental role of data warehouse is to provide correct decision making system. To achieve these kinds of information an application / tools are required for its implementation. Their requirements include easy to access data, scope and accuracy. They have also on-line analytical processing (OLAP) based on relation databases.

Decision support systems (DSS), was named given in 1970s to information systems designed to help managerial staff for making decisions. Managerial problems are ranging from assigning budget and choosing correct site locations for business etc. The basic idea behind developing these systems was that mangers could create and operate these systems at their own. Therefore, in 1980s number of organizations called it executive information system (EIS) [3].

The basic idea behind EIS was the mangers needs standard information about their firms and the external environment related to business. This information includes the time history of problems and their output for predicting their future state, so that manger could instantly know what is going on. The EIS system does not have analytical advantages of DSS. Some writers say that EIS is used by senior managers and DSS is used by junior staff.

Although very useful, the EIS and DSS often lacked a strong database. Generally, information gathered for one database cannot be used for another database. Managerial decision making, required consideration of the past and the future, but not just the present. As a result, DSS and EIS have to create their own databases, an area which was not their prime expertise. This activity is demanding and time consuming.

Figure 1. ETL follow in Data Warehouse


144

Data of the interest is across the different sources, for the centralize warehouse it is collected over the heterogeneous sources. Data is firstly extracted from the desired source, then transformed and cleansed respectively according to globalize rules. In figure 1 basic cycle of data warehouse updating is shown. ETL is the process of managing operational source data uniform so that it can be loaded into the warehouse. In order to keep data warehouse up to date, operational source data is loaded into the warehouse timely. Loading into the warehouse is done by ETL tool. Warehouse should be intelligent enough to make decision for updating.

2. Common Problem Data warehouse is repository of business data that is drawn from multiple sources across the heterogeneous data stores. Since warehouse implement materialized view of data. Data in the warehouse is stored and information is extracted from the warehouse. Whenever data arrives in the warehouse, the extracted metalized information should also be implemented according to current data.

As we know that warehouse is the rising technique for the retrieval of data from the distributed and heterogeneous data sources. A data warehouse is a repository of integrated information, available for queries and analysis (e.g. decision support and data mining) [1]. When relevant information becomes the part of warehouse, the information extracted from source is translated into a common model called Relational Model, and integrated with the existing data of warehouse. Data mining and querying to data warehouse can be done quickly and efficiently because formation of data already exists due to warehouse property, and all differences found resolved in warehouse.

Figure 2. The ETL process

We can think of data warehouse as a defining and storing integrated materialized view over the multiple data sources. Figure 2 shows a cyclic process; it will be executed each time when warehouse would be refreshed. Warehouse should be refreshed with the new operational source data timely, so that analysis can be performed over it. Refresh of warehouse requires the maximum utilization of computerized resources. In business environment computing resources are being used for gaining the efficiency in the business. If in peak hours this updating starts it will acquire the use of these resources and will slow down the business system. These resources are not used all the time. Sometimes, these recourses are used at average and some

time near to no use. We have to utilize the time of below average of resources for the updating of warehouse; it will be selected on historical records of resources.

3. Related Work Many researchers [2], [3], [7] have done a lots of work on the ETL life cycle. They have purposed different techniques regarding updating time. This process stuck out the data warehouse when ETL is being performed, because in Extract-Transform-Load makes warehouse unstable. When this process finishes the warehouse is updated not only with data but also with Meta data repository as shown in figure 2. It becomes stable when complete updating has been performed over it. We review different techniques that have been used and proposed earlier. Many proposals consider that the development phase of data warehouse is totally different from the development of RDBMS. Either they include extra stages, like workload refinement [10] or the ETL process [11, 12]. Moreover, they provide different methods for the requirements analysis phase. Several authors argue the importance of metadata as an integral part of the data warehouse design process [13, 14, 15].

3.1. Offline ETL Process This technique has become obsolete; it is not being used by any modern data warehouse repository. We would like to discuss it. Data in the warehouse was loaded offline. During updating process, data warehouse was shut down from functionality. In organizations every concern authorities was well informed that updating of data warehouse is in progress and warehouse will not be working for some time interval. When data in the warehouse was updated, intimation to the all concern authorities of organization sent that they can use it.

This technique was also called manually updating the data warehouse. If any problem to the warehouse comes, all others have to suffer from using it. During maintaining phase warehouse was not capable of performing its functionality. Warehouse is not one time activity, every time it needs to be updating to capture new requirements and more data. Most of the time warehouse was not functioning due to its maintain process. This technique was very resource consuming because lots of resource were left unused most of the time due to unavailability of warehouse functionality.

3.2. Periodic ETL Process When offline ETL process was not functioning well for the goodwill of the organization, a periodic ETL process was introduced. This technique does not automate the process of updating but it somehow has semi automated this task. In this technique off-hours were used for the updating process. When there was no need of warehouse. It also stops the functionality of the warehouse while updating. It does not matter because in off-peak hours there is no need of data warehouse. Warehouse is mostly needed because when there is management staff and they have to make decision. Now problem arises how to select off-peak hours.

3.2.1 Daily Periodic Activity Daily periodic activity is best for those organizations which need up to date data for correct decision making and these


145

organizations works in one shift only. If they work in the morning time and closed their work in the evening. We can see that their all resources remain unused when their office remains closed from evening to morning up till new day starts. Office close time will be the peak hours for that organization which works for one shift only. Their warehouse administrator will chose evening time for the updating of warehouse. It will utilize this resource when these were at not in use. Our primary goal is to maximize the use of resource to save our cost. This is online updation because in this technique there is no need to shut down the activities of warehouse. Administrator is also not necessary to be there when warehouse is being updating. When warehouse is updated, it automatically sends confirmation report to the concerned person that the warehouse has been updated.

3.2.2 Weekly Periodic Activity Weekly periodic activity is best for those organizations which need up to date data for correct decision making but these organizations works in all shifts in working days i.e. they work round the clock and close their work in the weekend holidays. We can see that their all resources remain unused when their office remains closed. Warehouse administrator will chose week holidays for the updating of warehouse. It will utilize resource when these were in no use. Our always goal is to maximize the use of resource to save our cost. This same situation can be reciprocal in an organization if it works at week ends and holidays, and remains closed during week days. It depends on organization to organization according to their structural needs; their updating will be performed in those weeks holidays respectively. This is online updating because in this technique there is no need to shut down the activities of warehouse. Administrator is also not necessary to be there when warehouse is being updating. When warehouse is updated, it automatically sends confirmation report to the concerned person that warehouse has been updated.

3.2.3 Monthly Periodic Activity Monthly periodic activity is best for those organizations which need up to date data for correct decision making but these organizations works in all shift in working days as well as on holidays also. If they work in the all the shifts and does not close their work even in the weekends, holidays. Their warehouse administrator will chose closing date of the month for the updating of warehouse. It depends on organization to organization according to their structural needs; their updating will be performed in those closing month’s days when there is less use of warehouse resources, because normally these organizations does not make decisions at the closing of the months.

3.2.4 Quarterly Periodic Activity Quarterly periodic activity is best for those organizations which need up to date data for correct decision making but these organizations works in all shift in working days as well as on holidays. If they work in the all the shifts and does not close their work in the weekends, holidays. However, we require some extra time for the updating of data warehouse. In this time updating is performed offline and data warehouse is also maintained to accommodate

extra business needs. During maintenance time no user can communicate with the warehouse for getting results out of it.

3.2.5 Yearly Periodic Activity There are some organizations which update their data warehouse on annual basis. Mostly these companies make their decision and policies at the beginning of every year. Once they have decided their work plan then it will remain in effect for rest of the year. In market different organizations have different plans and policies. They use the strategy which suits them best. Moreover, it also depends on the business nature and market trends.

3.3. Online periodic Queue Activity An extensive study on the modeling of ETL jobs is published by Simitsis and Vassiliadisl [5][9]. Online periodic queue is best suited for that environment where different heterogeneous sources update the warehouse. Business situation does not remain the same for a long time, situation change quite frequently. One operation source may be in one time zone and other may be in different time zone. That is when one is using warehouse and other needs to update warehouse online because this is peak off time of that organization. In this scenario which is the best time to update warehouse.

We see that none of the given solution works for that kind of problem where operational data stores are at different places and at different time zones. Then researcher introduced online periodic queue activity. This activity allows warehouse to have queue attached with it.

WarehouseOperational Source

Operational Source

Operational Source

Operational Source

Queue

Figure 3. Online Periodic queue activity

In figure 3 it is shown that queue is attached with the warehouse. Each operation source sends its ETL output to the warehouse for the updating. It stores that into queue and update one by one. This technique allows writing as much time as operational data sources needs to do so. Warehouse keeps updating itself whenever it gets off-hours. There are some problems with these techniques that are when operational data source keeps on sending their updating request in the warehouse queue for the updating of warehouse, and warehouse does not get time for the updating. This creates problem of buffer overflow. Warehouse would not be updating as it was desired. Lastly it has another problem that is if more and more updating


146

request will be coming from the different data source for the updating of warehouse, it will keep busy warehouse mostly in the updating and it will reduces the productive work out of warehouse. This will decrease the efficiency of the warehouse because we know that updating stuck the warehouse system.

4. Proposed Solution All above techniques waits for the off-hours so that resources gets free and ETL can be performed easily. In reality this is not the case with multinational organizations because these organizations run round the clock. There are no holidays even on Christmas, Eids and Easters. Although employees get off individually but these organizations has maintained a cycle that if one employee will be on holidays other will be working for it alternatively, on top of that, customers satisfaction and quality matters to them. In this type of working environment where organizations remain busy round the clock; it is very difficult to find any off-hours and warehouse shut down also creates a problem for them. Our proposed framework will identify the times for the updating of the warehouse using prediction technique applying historical data.

It identifies all the resources that are used by the warehouse while it is being maintained and updated. It keeps the record each resource that is utilized with and without updating warehouse. It also identifies those machines that get maximum use of warehouse. We will attach histogram with each resource and machines that needs data. They will observe the utilization of the resources. It will calculate the threshold limit for the each resource and machines. This observation keeps on checking at what level of load this resource perform in timely manner and what stage it would start sticking or malfunctioning.

Once we identify these things we will notice that there will be some times where updating can be performed easily. This framework identifies that times when office is working but concerned resources are not being used. It will apply updating process at that time and will starts observing load of the overall working environments up till functionality is being performed well. When loads gets back its normal stage of threshold we stop updating the warehouse at that stage and will resume from that stage once updating will be completed.

Figure 4. Minimum and maximum of resource utilization

These are the maximum and minimum load limit where resource can work properly as shown in figure 4 by the

limits. We called the threshold that is average of maximum and minimum.

Figure 5. Threshold of Resource

Figure 4 and figure 5 explain the overall load on resource. When it is near or under the threshold more load can be applied as it is applied in the off-hours in the periodic updating activity.

Now we see resources utilization that participates in the working of organization and updating of warehouse. This resource can be any server that is designed for both work. We anticipate that resource that participates in duel actions remains overburden and causes the system malfunction from the warehouse. In figure 6 we maintain a graph that checks the load on that resource during the office hours and after the office hours. It maintains a history record set of these activities on daily basis. We have shown a history record set of day Monday in table 1. We have picked those record set time and their duration that stay below the average threshold time.

Table 1: Resources duration with their status Resources Day Time Duration Status Res 1 Mon 6:00 Am 4 hours Below Res 2 Mon 5:00 Am 4 hours Below Res 3 Mon 12:00

Am 10 hours Below

Res 5 Mon 7:00 Am 3 hours Below

We show some calculations that will get time for the warehouse updating during the office hours. We gets the overlapped time that is free for all the activities. Table 1 is the history of one day, we gets two hours free for the updating of the warehouse. This overlap time is from 7:00 Am to the 9:00 Am where all the resources are free and warehouse updating activity will not makes them in the state of the stuck system. This recording of resources duration and their status will be a continuous process. As long as this process goes old, it will start giving the true prediction of time for the warehouse updating. This philosophy is improved with the passage of time, as statistical population/data size grows prediction becomes more and more accurate.

There will be a situation where ETL update cycle will predict to update the warehouse but current status would not allow doing so because updating in this situation will cause warehouse failure. We will skip that time and wait for another time.


147

Figure 6. Resource utilization graph below and above

threshold point

Lastly, there may be a situation where current utilization of resources and predicted time for the updating of warehouse will allows to update warehouse smoothly but when we will start updating the warehouse, the utilization of the resource increase due to external use of resources. At that time we will stop the updating process and give the utilization to the system environment and we will wait for the next suitable time, and this is completely automatic, self activation and stopage.

Figure 7 gives the utilization picture of the resource that is below to its threshold time. We consider these as a free of utilization. We record the history of those resources with their reason being free, which remains below average mostly. If this case continues to be happened, we removed them from the list of resources that we inspect because these resources lie in the dual usage of system and warehouse.

Figure 7. Resource Utilization below threshold point

There may comes a situation when there does not exist any time frame when resources are not below threshold time but are not at the maximum level of resource utilization. Then we calculate the approximate load of ETL process with its duration. We predict the load on the resource after putting the ETL update cost. If this remains at normal level that is below the maximum level, we start updating the warehouse until any of the resource goes beyond the maximum load limit.

Our architecture of warehouse would also be modified with little changes. It has been shown in figure 8. It adds the queue with each operational data source. This queue will send the updating request when they will get response from the warehouse that their pervious request has been updated. This approach wills benefits in many ways. It will avoid

warehouse buffer overflow due to the limited request of update. Each data source will have a request of maximum of one job in warehouse queue. Operational data sources will contain there updating request in their queues until it does not ask.

Figure 8. ETL update cycle architecture

These queues will give equal opportunity to each operational data source. It avoids them from starvation. In previous architectures some operation data sources may be starved due to the rapid updating of another warehouse.

5. Conclusion In this paper we discussed the issue regarding ETL update cycle. We argued that update cycle method is very efficient compared to complete loading of data from operational data sources because in multinational organizations it is very difficult to find off-hours easily. However, ETL update cycle method is preferable in general because it gives maximum utilization of resources and gets maximum throughput out of it and availability of warehouse so that different users gets maximum queries from it. Since ETL update cycle requires extra checking cost for resources utilization every time that is why the cost is increased, but this becomes minimal when organizations gets more business due to availability of warehouse.

Our main contribution is to provide certain rules for the ETL update cycle that makes it fully automated process and efficient without the human interaction with the data warehouse. It selects its best time automatically for loading data into the warehouse without compromising of its functionality and response time to the users. If warehouse gets some data from their operational connected data sources, then it give first priority to users and second priority to updating otherwise our algorithms will not occupy the resources un-necessary. Warehouse does not need to update every time it gets updated call by some predefined time schedule. Our proposed mechanism saves the time and cost as compared to periodic update and other existing techniques for ETL.


148

References [1] Gupta, A., Jagadish, H. V., Mumick, "Data Integration

using Self-Maintainable Views," EDBT, pp. 140-144, 1996

[2] JÄorg, T., Dessloch, "Towards generating ETL processes for incremental loading," IDEAS, pp. 101-110, 2008

[3] JÄorg, T., Dessloch, "Formalizing ETL Jobs for Incremental Loading of DataWarehouses," BTW pp. 327-346, 2009

[4] Kimball, R., Caserta, The Data Warehouse ETL Toolkit: Practical Techniques for Extracting, leaning, Conforming, and Delivering Data, John Wiley & Sons, 2004

[5] Panos Vassiliadis, Alkis Simitsis, and Spiros Skiadopoulos, "Conceptual modeling for ETL processes," In DOLAP, pp. 14–21, 2002

[6] Alkis Simitsis, "Mapping conceptual to logical models for ETL processes," In DOLAP, pp. 67–76, 2005

[7] Alkis Simitsis, Panos Vassiliadis, and Timos K. Sellis, "Optimizing ETL Processes in Data Warehouses," In ICDE, pp. 564–575, 2005

[8] Alkis Simitsis, Panos Vassiliadis, Manolis Terrovitis, and Spiros Skiadopoulos, "Graph-Based Modeling of ETL Activities with Multi-level Transformations and Updates," In DaWaK, pp. 43–52, 2005

[9] Labio, W., Garcia-Molina,"Efficient Snapshot Differential Algorithms for Data Warehousing," VLDB, pp. 63-74, 1996

[10] M. Golfarelli and S. Rizzi. A methodological framework for data warehouse design. In I.-Y. Song and T.J. Teorey, editors, Proceedings of the 1st ACM International Workshop on Data Warehousing and OLAP, DOLAP’98, pp. 3–9. ACM Press, 1998.

[11] R. Kimball, L. Reeves, M. Ross, and W. Thornthwaite. The Data Warehouse Lifecycle Toolkit: Expert Methods for Designing, Developing, and Deploying Data Warehouses. Wiley, 1998.

[12] S. Luj´an-Mora and J. Trujillo. A comprehensive method for data warehouse design. Proceedings of the 5th International Workshop on Design and Management of Data Warehouses, DMDW’03. CEUR Workshop Proceedings, 2003.

[13] C. Ballard, D. Herreman, D. Schau, R. Bell, E. Kim, and A. Valencic. Data Modeling Techniques for Data Warehousing. IBM Redbooks SG24-2238-00, 1998.

[14] L. Carneiro and A. Brayner. X-META: A methodology for data warehouse design with metadata management. In [157], pp. 13–22.

[15] F. Paim, A. Carvalho, and J. Castro. Towards a methodology for requirements analysis of data warehouse systems. In Proceedings of the 16th Brazilian Symposium on Software Engineering, SBES’02, pp. 1–16, 2002.

Authors Profile Atif Amin received his B.S. in Computer Science degree from University of Central Punjab in 2008. He has been the winner of All Pakistan Software Competition, Softcom ‘08 in 2008 where he won first position. He has been chairman IEEE from 2007-2008. He is now doing M. S. in Computer Science from University of Central Punjab.

Prof. Dr. Abdul Aziz did his M.Sc. from University of the Punjab, Pakistan in 1989; M.Phil and Ph.D in Computer Science from University of East Anglia, UK. He secured many honors and awards during his academic career from various institutions. He is currently working as full Professor at the University of Central Punjab, Lahore, Pakistan. He is the founder and Chair of Data Mining Research Group at UCP. Dr. Aziz has delivered lectures at many universities as guest speaker. He has published text books and large number of research papers in different refereed international journals and conferences. His research interests include Knowledge Discovery in Databases (KDD) - Data Mining, Pattern Recognition, Data Warehousing and Machine Learning.

He is member of editorial board and referee for various well known international journals / conferences including IEEE publications. (e-mail: [email protected]).

vol2 no 2

Documents

Transcript of vol2 no 2