5046351e7bd13e1e55

8/12/2019 5046351e7bd13e1e55

1/11

Journal of VLSI Signal Processing 41, 293303, 2005c 2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

DOI: 10.1007/s11265-005-4153-1

Data Masking: A New Approach for Steganography?

REGUNATHAN RADHAKRISHNAN, MEHDI KHARRAZI AND NASIR MEMON Polytechnic University, Brooklyn, NY 11201, USA

Abstract. It is well known that encryption provides secure channels for communicating entities. However,due to lack of covertness on these channels, an eavesdropper can identify encrypted streams through statisti-cal tests and capture them for further cryptanalysis. Hence, the communicating entities can use steganographyto achieve covertness. In this paper we propose a new form of multimedia steganography called data mask-ing . Instead of embedding a secret message into a multimedia object, as in traditional multimedia steganog-raphy, we process the entire secret message to make it appear statistically similar to a multimedia objectitself. Thereby we foil an eavesdropper who is primarily applying statistical tests to detect encrypted com-munication channels. We show that our approach can potentially give a covert channel capacity which isan order of magnitude higher than traditional steganography. Our experiments also show that steganalyzerstrained with stego objects of known steganographic have a low detection accuracy for datamasked multimediaobjects.

1. Introduction

Rapid developments in cryptography, cryptographicstandards, and relaxed export controls on encryptionschemes have made cryptographic software widelyavailable, user friendly and in some cases transpar-ent to end-users. The relative ease with which se-cure channels of communication can be created be-tween two parties using cryptography, coupled withthe explosion in network trafc, has posed a problem

to the surveillance operations typically carried out bylaw enforcement and intelligence agencies. In order tokeep up with the growth in communication trafc andsecure channels, massive eavesdropping campaigns,such as Carnivore and Echelon , have been developed.Carnivore, for example, is a combination of hardwareand software, which collects all electronic packets thataresent to andfrom theInternet Service Provider whereit is installed. Following this collection procedure,several lters are applied to isolate emails with certainkeywords which could be further analyzed off-line.These lters could potentially include the performance

of statistical tests to isolate encrypted emails for off-line cryptanalysis.

Pervasive eavesdropping like that performed byCarnivore leads to the need for covert channels, inwhich communicating entities hide their communica-tion channel itself from eavesdroppers. Steganographyrefers to thescience of invisible communication. Un-like cryptography, where the goal is to secure commu-nications from an eavesdropper, steganographic tech-niques strive to hide the very presence of the message

itself from an observer [ 1]. As broadband technolo-gies improve bandwidth at the last-mile, multimediacontent, such as still images, audio and video, havegained increasing popularity on the Internet. Given thehigh degree of redundancy present in a digital repre-sentationof multimediacontent (despite compression),there has been an increased interest in using multime-dia content for the purpose of steganography. Indeedmany such techniques have been devised in the pastfew years. For an excellent survey of such techniques,the reader is referred to [ 2]. Briey, the general modelfor steganography, illustrated in Fig. 1, where we have

8/12/2019 5046351e7bd13e1e55

2/11

294 Radhakrishnan, Kharrazi and Memon

Figure 1. Framework for Secret Key Passive Warden Steganography. Alice embeds secret message in cover image ( left ). Wendy the wardenchecks if Alices image is a stego-image ( center ). If she cannot determine it to be so, she passes it on to Bob who retrieves the hidden messagebased on secret key ( right ) he shares with Alice.

Alice wishing to send a secret message m to Bob. Inorder to do so, she embeds m into a cover-object c ,to obtain the stego-object s . The stego-object s is thensent through the public channel. Wendy who examinesall messages in the channel, should not be able to dis-tinguish in any sense between cover-objects (objectsnot containing any secret message) and stego-objects(objects containing a secret message). In this context,steganalysis refers to the body of techniques that aidWendy in distinguishing between cover-objects andstego-objects. It should be noted that Wendy has tomake this distinction without any knowledge of thesecret key which Alice and Bob may be sharing andsometimes even without any knowledge of the specicalgorithm that they might be using for embedding thesecret message.

There are two problems with multimedia steganog-raphy that seem to have been ignored in the literature.Firstly, in multimedia steganography, the size of thecovert message is relatively much smaller than the sizeof the multimedia object (the stego object) that car-ries the covert message. In fact recent developments

in steganalysis techniques have demonstrated that em-bedding a secret message in a multimedia object, likesay a still image, causes certain statistically discernibleartifacts that reveal the presence of the secret messagethereby exposing the covert communication channel[3 6]. Such developments further limit the number of bits that can safely be hidden in a multimedia objectfrom sophisticated steganalysis techniques. This leadsto a cover object thatis one ormoreordersof magnitudelarger than the secret message, reducing the effectivebandwidth of covert communication channels.

The second problem is that the common interpre-tation of the warden-based framework when used for multimedia steganography described has been that thewarden examines the cover-object (or stego-object)perceptually and statistically. However, perceptualtests, which may have been practical on low volumecommunication channels, would not be practical withhigh volume communication channels. Such channelsmust rely on automated statistical tests for detectingpotential stego objects and only then may apply (if atall) some form of perceptual tests to the selected can-didates. Indeed, large scale eavesdropping operationslike Carnivore rely rst on statistical analysis to iso-late interesting messages. Hence, if Alice and Bobcan evade statistical tests performed by Wendy, theymay not have to undergo any perceptual test at all! Sofor example, if the stego object is statistically indistin-guishable from an image, its perceptual quality doesnot matter at all. Can this fact be exploited to improvecovert communications in any sense? In this paper weshow that this is indeed possible.

Specically, we propose a novel technique, which

we call Data Masking , for multimedia steganography.We again look at the problem of Alice and Bob wishingto communicate covertly, but we assume that the war-den Wendy is an entity like Carnivore and will rst ex-aminemessagesstatistically. Only if the statistical testsindicate a possibility of covert communication, wouldshe then examine the message further. Hence instead of selecting a cover object and embedding a covert mes-sage in it like traditional steganography, we mask thesecret message to make it appear like a normal multi-media object under typical statistical examinations.

8/12/2019 5046351e7bd13e1e55

3/11

8/12/2019 5046351e7bd13e1e55

4/11


Figure 3. Audio datamasking using LPC analysis/synthesis.

such that E(( X i X i )2) is minimized. Here X i isthe predicted value of the current sample based on N previous samples in the reference audio and is denedas ( N 1)k = 0 hk X k .

Using the orthogonality principle (Hilbert spaceprojection theorem), N equations (called Yule-Walker equations) can be set up to solve for the optimal lter coefcients in the MMSE (Minimum Mean Square Er-ror) sense. Then, the inverse of the LPC analysis lter so designed, can be used to lter the noise-like cipher stream to remove randomness from cipher stream andtransform it into a reference audio-like waveform thathas more correlation between samples.

With the knowledge of lter coefcients the receiver can reconstruct the cipher stream from the referenceaudio, as in the Inverse Wiener ltered cipher stream.

2.2. Image Datamasking

Since the source model for speech is well understood,LPC Analysis-synthesis could be used for datamask-ing. The all-pole synthesis lter is known to model the

human vocal tract lter. However, the source modelfor images is not as well understood and hence wepropose a different scheme for image datamasking asshown in Fig. 3.

As in audio datamasking, we use the prediction er-ror to mask the encrypted stream. Each pixels in-tensity value is predicted based on its causal neigh-borhood using a MED predictor that is employed inthe JPEG-LS lossless compression standard [ 8]. Thecausal neighborhood consists of W , NW , N , NE (west,northwest, north, northeast) pixels. According to theMED predictor, the predicted value of current pixel is

min ( W , N ) if N W max( W , N ) or it is max( W , N )if N W min( W , N ) or it is W + N NW otherwise.

A Huffman codebook, that is constructed for theLaplacian prediction error probability density function(PDF), is used to decode the input random encryptedstream. Then the PDF of Huffman decoded encryptedsamples would also have a Laplacian PDF. Hence, onecanmodify each pixel so that theprediction error at thatpixel is matched with the Huffman decoded encryptedstream sample.

Note that in the proposed system for image data-masking, there are two parameters that the communi-cating parties should exchange beforehand. The rstis a Huffman codebook designed for coding predictionerrors with a Laplacian PDF and the second is a thresh-old to select the range prediction error values to bemapped to encrypted stream samples. For instance, if the chosen threshold in M , there would 2 M + 1 huffmancodewords mapping { M . . . 0 . . . M }. This thresholdcontrols the tradeoff between datamasking capacityand quality of the data masked image. By choosingthe threshold small, one would select only smooth re-gions of the image to hide information and hence the

quality of the datamasked image would be better. Withthe knowledge of these two parameters, the receiver would map the prediction error at each pixel to therandom encrypted stream.

In the following section, we present some of our experimental results to evaluate the performance of audio and image data masking.

3. Experimental Results

We chose an AES (Advanced Encryption Standard)Advanced Encryption Standard [ 7], encrypted stream

8/12/2019 5046351e7bd13e1e55

5/11

Data Masking: A New Approach for Steganography? 297

Figure 4. Proposed system for image datamasking.

Figure5. Comparison of reference audioclip and encrypted streamin time domain.

as input to our system. The systems desired output isto generate a multimedia object from this stream. Weuse any audio clip or image as a reference to guide thesystem in the process.

3.1. Results on Audio Datamasking

We picked a speech clip of duration 2 seconds as our reference audio, A. We read every seven bits fromthe encrypted stream and append a random LSB andinterpret it as a sample from a random process, E.Figures 5 and 6 compare the encrypted stream andreference audio clip in time domain and frequencydomain respectively.

The reference audio is then segmented into framesof length 1024 samples, such that the time duration isshort enough for the stationarity assumption to hold.Then for each frame of audio, an equal number of bytes

Figure 6. Comparison of power spectral densities (PSDs) of ref-erence audio clip and encrypted stream.

from the encrypted stream are read. A LPC Analysislter of order 31 was designed, the inverse of whichwas used to lter the encrypted stream. We thus have atime-varying inverse Wiener lter shaping every 1024samples of the encrypted stream to match the reference

audio frame.Figure 7 shows the results of inverse Wiener l-

tering on encrypted stream corresponding to 2 s of reference audio clip and its PSD. It is clear that the l-tered signal has more correlation than the encryptedsignal and would potentially defeat most statisticaltests for randomness. However, the ltered signal doesnot match the reference audio waveform exactly andhence the resulting waveform was noisier when lis-tened to (perceptual test). If the shaping of the en-crypted stream perfectly matched that of reference au-dio for each frame, then the inverse Wiener ltered

8/12/2019 5046351e7bd13e1e55

6/11


Figure 7. Inverse wiener ltered encrypted stream & its PSD.

encrypted stream would also sound like the referenceaudio.

In order to reconstruct the encrypted stream at thereceiver, knowledge of LPC analysis lter coefcientsfor the current frame is essential. Therefore, the lter coefcients for the current frame are appended to theencryptedstreamthat was shaped in the previous frameand the coefcients corresponding to the rst framewere included in the header of the .wav le. Wiener ltering can then be performed for each frame by re-trieving the ltering coefcients reconstructed fromthe previous frame to obtain the encrypted stream for the current frame. Figure 8 shows the reconstructed en-crypted streamand itsdifference from the original. Themaximum difference between the samples of originaland reconstructed encrypted stream is within 0.015.Since each sample was interpreted as a byte from theencrypted stream and scaled by (1/64) and shifted by 1, a bit ip in the encrypted stream can occur only if the error is greater than (1/64). Therefore, if the error iswithin (1/64) or (0.0156), it is possible to reconstruct

the encrypted stream without any bit error which iscritical for proper decryption.

Now that we areconvinced that theproposed schemewould mask encrypted data from statistical tests, let usevaluate the performance of this scheme as a audiodata hiding technique. If we assume that each lter coefcient occupies a word, we have an overhead of 124 bytes (4 31) for every 1024 bytes of encryptedstream. The effective payload, (1024 124 128) =772 bytes, can be thought of as embedding capacity for this scheme per frame of audio. It has been observedthat the quantization error introduced by a precision

Figure 8. Reconstructed encrypted stream & its difference fromoriginal encrypted stream.

of 8 bits per sample of the ltered encrypted streamwould result in an error greater than (1/64). Therefore,we use 16 bit precision for each sample in the lteredencrypted stream and assume that the transmitted in-verse Wiener ltered stream was received without anydistortion. We can compare our schemes embeddingcapacity to that of LSB embedding, in which one bitis embedded in the LSB per sample of audio. The pro-posed scheme hides 6176 bits (8 772) per 1024 bytesof audio data whereas LSB embedding (which also haslow robustness) can hide 1024 bits per 1024 bytes of audio data. Therefore, the embedding capacity of pro-posed scheme is 6.0313 times that of the capacity of LSB embedding. In order to achieve an embeddingcapacity of 6176 bits per 1024 bytes using LSB em-bedding technique, it would require changing at least 6bits of the 8 bits in each sample! Note that in practice,steganalysis techniques given in [ 4 6] can detect thepresence of LSB embedding even when 10% of thebits in the cover message have been ipped (in somecases this can be as low as 2%;). So we potentially have

an order of magnitude larger capacity as compared totraditional multimedia steganography. Of course, wegain this by assuming that warden does not examinethe message perceptually, but only statistically.

In order to see what kind of statistical tests candetect the datamasked audio signals, we tested themwith a number of existing audio steganalysis tools[9]. Table 1 summarizes the performance of somesteganalysis techniques on 66 datamasked audiostreams of total duration 29 min. It can be observedthat all of the detectors trained with stego objects of known steganographic techniques have a low detection

8/12/2019 5046351e7bd13e1e55

7/11


Table 1. Detection performance of classiers trained with stegoaudio les of specic embedding techniques with embedding capac-ity of 6.03bits per audio sample ; [A]: Number of actual datamaskedaudio streams; [B]: Number of audio streams detected by the ste-ganalyzer; [C]: Number of false alarms; Precision: ([ B] [C ])[ B] ; Recall:([ B] [C ])

[ A] .

[A] [B] [C]Precision

(%)Recall

(%)

DSSS 66 100 46 54.0 81.82

FHSSS 66 38 12 68.42 39.39

Echo 66 36 7 80.56 43.94

DctWHas 66 20 9 55 16.67

Steganos 66 17 4 76.47 25.76

S-tools 66 54 15 72.22 59.09

rate for the datamasked audio streams. This suggeststhat the artifacts introduced to the cover audio signalby the datamasking process has a different signaturethan that introduced by existing steganographictechniques. Hence, the classiers trained with stegoaudio les of specic embedding techniques were lesssuccessful with a dataset consisting datamasked audioles.

3.2. Results on Image Datamasking

A reference image was chosen and a Huffman code-book was designed for the prediction error PDF. Thereference images gray scale values were re-mapped

to a smaller range so that the modied pixels value

is still within [0 255]. Figure 9 shows the originalreference image and the PDF of the prediction error.We used a threshold of 50 on the absolute value of prediction error and hence there were 101 huffmancodewords to map the encrypted stream to predictionerrors. Figure 10 shows the datamasked image and itsprediction error PDF.

The embeddingcapacity of thisdatamasking schemefor the chosen image was found to be 5.47 bits per byteof image stream. Figure 11 shows the tradeoff betweenembedding capacity and quality of the datamasked im-age for two different values of M.

For all of the chosen M values, the embedding ca-pacity is high compared to many of the multimediasteganographic schemes. With such a high embeddingcapacity, would the datamasked image escape statisti-cal tests performed by existing steganalyzers?

This motivated us to test a number of datamaskedimages with existing steganalyzers for known stegano-graphic techniques. Tables 1 and 2 present the detec-tion performance on two sets of datamasked images of a linear classier trained with stego images of specicembedding techniques. Dataset 1 contains 100 data-masked images with embedding capacity of 4.5 bpp& their corresponding 100 original images. Dataset 2

contains 71 datamasked images with the same embed-ding capacity & their corresponding original images.The original images in the Dataset 2 were chosen to bealready noisy images.

Figure 9. Original reference image and its prediction error histogram.

8/12/2019 5046351e7bd13e1e55

8/11


Figure 10. Datamasked image and its prediction error histogram.

Figure 11. Datamasked images with M = 25 and M = 15 with corresponding embedding capacities of 4.56 & 3.65 bpp.

The results from tables 1 and 2 show that mostof the steganalyzers have a low detection rate for both datasets. Also, the false alarm rate is consistentlyhigher for dataset 2 compared to dataset 1. Since theimages in dataset 2 were already noisy before embed-ding, the detectors found it difcult to differentiate be-tween the noise introduced by the embedding processand the noise already present, thereby increasing thefalse alarm rate. Since the choice of the cover imagefor steganography is upto the communicating parties,

the noisy images like those in dataset 2 would be moredesirable for datamasking.

In order to see the effect of embedding capacity onthese two classiers, a similar experiment was per-formed on the same datasets but with a lower embed-ding capacity of 2.8 bpp. Tables 3 and 7 present theresults for the testing dataset 1 with SVM classierswith two different embedding capacities.

Tables 4 and 8 present the results of SVM classierson dataset 2 with two different embedding capacities.

8/12/2019 5046351e7bd13e1e55

9/11


Table 2. Top Table: Detection performance of linear classierstrained with stego images of specic embedding techniques usingBinary Similarity Metric (BSM) as features for Dataset 1 with em-bedding capacity of 4.5 bpp; [A]: Number of actual datamaskedimage streams; [B]: Number of image streams detected by the ste-ganalyzer; [C]: Number of false alarms; Precision: ([ B] [C ])[ B] ; Recall:([ B] [C ])

[ A] .


(%)Recall

(%)

F5r12 100 111 28 74.77 83 . 0

F5r11 100 49 15 69 . 38 34 . 0

Out 100 164 64 60 . 97 100 . 0

Out + 100 29 29 0 . 0 0. 0

Lsb 100 52 52 0 . 0 0. 0

Table 3. Detection performance of linear classiers trained withstegoimages of specic embedding techniquesusing BinarySimilar-ity Metric(BSM) as features for Dataset 2 with embedding capacityof 4.5 bpp ; [A]: Number of actual datamasked image streams; [B]:Number of image streams detected by the steganalyzer; [C]: Number of false alarms; Precision: ([ B] [C ])[ B] ; Recall:

([ B] [C ])[ A] .


(%)Recall

(%)

F5r12 71 89 28 68 . 54 85 . 92

F5r11 71 56 20 64 . 29 50 . 7

Out 71 111 40 63 . 96 100 . 0Out + 71 43 43 0 . 0 0. 0

Lsb 71 48 48 0 . 0 0. 0

Table 5 presents the results of a linear classier ondataset 1 with embeddingcapacityof 2.8bpp. Comparethis with Table 1 which presents the detection resultsfor the same classier and dataset but with higher em-bedding capacity of 4.5 bpp.

Table 6 presents the results of a linear classier ondataset 2 with embeddingcapacityof 2.8bpp. Compare

this with Table 2 which presents the detection resultsfor the same classier and dataset but with higher em-bedding capacity of 4.5 bpp.

From all the experiments we consistently observethat, the detection rate of classiers decreases as onedecreases the embedding capacity of datamaskingalgorithm. Also, the reduction in detection rate wasonly marginal.

While these experiments were performed with ste-ganalyzers trained with stego images of specic em-bedding techniques, we also performed an experimentwith a generic steganalyzer trained with stego im-

Table 4. Detection performance of SVM classiers trained withstego imagesof specic embedding techniques usingBinary Similar-ity Metric(BSM) as features for Dataset 1 with embedding capacityof 4.5 bpp ; [A]: Number of actual datamasked image streams; [B]:Number of image streams detected by thesteganalyzer; [C]:Number of false alarms; Precision: ([ B] [C ])[ B] ; Recall:

([ B] [C ])[ A] .


(%)Recall

(%)

F5r12 100 23 23 0 . 0 0. 0

F5r11 100 66 20 69 . 69 46 . 0

Out 100 30 25 16 . 66 5 . 0

Out + 100 21 20 4 . 76 1 . 0

Lsb 100 121 89 26 . 44 32

Table 5. Detection performance of SVM classiers trained withstego imagesof specic embedding techniques usingBinary Similar-ity Metric(BSM) as features for Dataset 1 with embedding capacityof 2.8 bpp ; [A]: Number of actual datamasked image streams; [B]:Number of image streams detected by thesteganalyzer; [C]:Number of false alarms; Precision: ([ B] [C ])[ B] ; Recall:

([ B] [C ])[ A] .


(%)Recall

(%)

F5r12 100 23 23 0 . 0 0. 0

F5r11 100 58 20 65 . 52 38 . 0

Out 100 29 25 13 . 79 4 . 0

Out + 100 20 20 0 . 0 0. 0Lsb 100 106 89 16 . 04 17 . 0

Table 6. Detection performance of SVM classiers trained withstego imagesof specic embedding techniques usingBinary Similar-ity Metric(BSM) as features for Dataset 2 with embedding capacityof 4.5 bpp ; [A]: Number of actual datamasked image streams; [B]:Number of image streams detected by thesteganalyzer; [C]:Number of false alarms; Precision: ([ B] [C ])[ B] ; Recall

([ B] [C ])[ A] .


(%)Recall

(%)

F5r12 71 15 14 6 . 67 1 . 41

F5r11 71 13 9 30 . 77 5 . 63

Out 71 23 18 21 . 74 7 . 04

Out + 71 6 6 0. 0 0. 0

Lsb 71 75 66 12 . 00 12 . 68

ages of different embedding schemes. The detectionperformance of the generic steganalyzer on the sameset of datamasked images was found to be 2.81%at 4.5 bpp and 1.31% at 2.8 bpp. Also, the em-bedded message length estimate using RS steganal-

8/12/2019 5046351e7bd13e1e55

10/11


Table 7. Detection performance of SVM classiers trained withstegoimages of specic embedding techniquesusing BinarySimilar-ity Metric(BSM) as features for Dataset 2 with embedding capacityof 2.8 bpp ; [A]: Number of actual datamasked image streams; [B]:Numberof image streams detected by the steganalyzer; [C]: Number of false alarms; Precision: ([ B] [C ])[ B] ; Recall

([ B] [C ])[ A] .


(%)Recall

(%)

F5r12 71 14 14 0 . 0 0. 0

F5r11 71 11 9 18 . 18 5 . 63

Out 71 22 18 18 . 18 5 . 63

Out + 71 6 6 0. 0 0. 0

Lsb 71 83 66 20 . 48 23 . 94

Table 8. Detection performance of linear classiers trained withstegoimages of specic embedding techniquesusing BinarySimilar-ity Metric(BSM) as features for Dataset 1 with embedding capacityof 2.8 bpp ; [A]: Number of actual datamasked image streams; [B]:Number of image streams detected by the steganalyzer; [C]: Number of false alarms; Precision: ([ B] [C ])[ B] ; Recall

([ B] [C ])[ A] .


(%)Recall

(%)

F5r12 100 109 28 74 . 31 81 . 0

F5r11 100 53 15 71 . 7 38 . 0

Out 100 164 64 60 . 97 100 . 0

Out + 100 29 29 0 . 0 0. 0Lsb 100 52 52 0 . 0 0. 0

Table 9. Detection performance of linear classiers trained withstegoimages of specic embedding techniquesusing BinarySimilar-ity Metric(BSM) as features for Dataset 2 with embedding capacityof 2.8 bpp ; [A]: Number of actual datamasked image streams; [B]:Number of image streams detected by the steganalyzer; [C]: Number of false alarms; Precision: ([ B] [C ])[ B] ; Recall

([ B] [C ])[ A] .


(%)Recall

(%)

F5r12 71 91 28 69 . 23 88 . 71

F5r11 71 44 20 54 . 55 33 . 80

Out 71 114 40 62 . 28 100 . 0

Out + 71 43 43 0 . 0 0. 0

Lsb 71 48 48 0 . 0 0. 0

ysis was only 25% of the actual embedded messagelength. All of these experiments suggests that thep-roposed datamasking scheme has a different signa-ture from that of existing steganographic schemes.Hence, oneneeds to train a new steganalyzer with data-

masked images to learn the statistics of the datamask-

ing scheme. In order to verify that we also trained asteganalyzer using datamasked images and have foundthat both linear & SVM classiers have a precisionof above 87% while maintaining a recall rate above97%.

4. Conclusions

In this paper, we proposed a novel solution to removerandomness from cipher streams in a reversible man-ner. This would make difcult for an eavesdropper to distinguish encrypted streams from other networkstreams, thereby, providing covertness to secure chan-nels. We have proposed schemes to generate audiostreams and image streams from encrypted streams.Since perceptual tests are not feasible with such largevolumes of data, we relax the requirement for per-ceptual transparency to increase the steganographicstreams. However, we maintain certain statistical prop-erties in the datamasking process: the shape of power spectral density in the case of audio datamasking andthe shape of prediction error PDF in the case of imageof datamasking.

Our experimental results show that the proposed

datamasking schemes have a different signature thanexisting steganographicschemes.Eventhough existingsteganalyzers do notperform well to detectdatamaskedstreams, we have also shown that one can train a ste-ganalyzer with datamasked streams to detect them withhigh accuracy. For the existing steganalyzers, ensuringspectral shape and prediction error PDFs shape wassufcient. However, it remains to be seen if one candevise a framework for steganography which wouldescape any other higher order statistical analysis. Our experiments seem to suggest that an embedding mech-anism which induces similar noise characteristics as

the noise that is already present in the image, would behard to be detected by any statistical analysis.

Acknowledgments

We would like to thank Kulesh Shanmugasundaram(CIS department, Polytechnic University, Brooklyn,NY) for many helpful discussions on data masking andnetwork security. We are also grateful to Hamza Ozer (National Research Institute of Electronics and Cryp-tology, Marmara Research Center, Gebze, Turkey) for

8/12/2019 5046351e7bd13e1e55

11/11


helping in evaluating datamasked streams with the au-

dio steganalyzers.

References

1. G.J Simmons, Prisoners Problem and the Subliminal Channel(The), CRYPTO83Advances in Cryptology , Aug. 1984, pp.5167.

2. N.F Johnson and S. Katzenbeisser, A Survey of SteganographicTechniques, in Information Hiding S. Katzenbeisser and F. Pe-titcoals (Eds.), Artech House, Norwood, MA, 2000, pp. 4378.

3. N.F. Johnson and S. Jajodia, Steganalysis of Images CreatedUsing Current Steganography Software, in Information Hiding , David Aucsmith (Ed.), LNCS 1525, Springer-VerlagBerlin Heidelberg, 1998, pp. 3247.

4. A. Westeld and A. Ptzmann, Attacks on SteganographicSystems, in Information Hiding , LNCS 1768, Springer-VerlagHeidelberg, 1999, pp. 6176.

5. J. Fridrich, M. Goljan, and R. Du, Reliable Detection of LSBSteganography in Color and Grayscale Images, in Proc. of the

ACM Workshop on Multimedia and Security , Ottawa, CA, Oct.,2001, pp. 2730.

6. I. Avcibas, N. Memon, and B. Sankur, Steganalysis UsingImage Quality Metrics, Security and Watermarking of Multimedia Contents , SPIE, San Jose, 2001.

7. Advanced Encryption Standard, FIPS 197, 2001.8. ISO/IEC 14495-1, ITU Recommendation T.87, Information

technologyLossless and Near-Lossless Compression of Continuous-Tone Still Images, 1999.

9. H. Zer, I. Avcibas, B. Sankur, and N. Memon, Steganalysisof Audio Based on Audio Quality Metrics, submitted to SPIE2004.

10. Elke Franz et al., Steganography Preserving Statistical Prop-erties , Fifth International Workshop on Information Hiding,2002.

11. S. Lyu and H. Farid, Detecting Hidden Messages Using Higher-Order Statistics and Support Vector Machines ,Fifth International Workshop on Information Hiding,2002.

12. S. Dumitrescu, X. Wu, and Z. Wang, Detection of LSBSteganography via Sample Pair Analysis, IEEE Transactionson Signal Processing , 2003, pp. 19952007.

5046351e7bd13e1e55

Documents

Transcript of 5046351e7bd13e1e55