A Novel Method of Image Steganography in DWT Domain

Vladimír BÁNOCI, Gabriel BUGÁR, Dušan LEVICKÝ

Dept. of Electronics and Multimedia Communications, Faculty of Electrical Engineering and Informatics, technical University of Košice, Letná 9, 042 00 Košice, Slovak Republic

vladimir.banoci@tuke.sk, gabriel.bugar@tuke.sk, dusan.levicky@tuke.sk

Abstract. In this paper, we present a novel steganographic method for embedding of secret data in still grayscale image. In order to provide large capacity of the secret data while maintaining good visual quality of stego-image, the embedding process is performed in transform domain of Discrete Wavelet transform (DWT) by modifying of transform coefficients in an appropriate manner. In addition, the proposed method do not require original image for successful extraction of the secret information. The experimental results show that the proposed method provides good capacity and excellent image quality.

Keywords Steganography, Haar-DWT transform.

1. Introduction Information hiding, steganography, and watermarking

are three closely related fields that have a great deal of overlap and share many technical approaches. However, there are fundamental philosophical differences that affect the requirements, and thus the design, of a technical solution. Digital watermarking is mainly used in copyright protection, while steganography is a method of embedding the secret message into a camouflage media to ensure that unintended recipients will not be aware of the existence of the embedded secret data in cover media [2]. Thus, whole system can be considered as secret communication. However, steganography is different from cryptography as it is a part of secret communication where its techniques may fail since a cipher text has meaningless form and thus easily arouses the curiosity of malicious attackers who are willing to consume the substantial amount of time and energy to recover or destroy data. Unlike cryptography, steganography conceals the fact that there is secret communication going on and still image may be represented as well suited camouflage media for embedding of secrete data. Even more, the advantage of cryptography techniques are executed on secret data before embedding into still image; to strengthen security level and also to suppress the energy compaction of secret data. Steganalysis on the other side is the science of detecting

hidden information. The main objective of steganalysis is to break steganography system and that condition is met if an algorithm can judge whether a given image contains a secret message. There are three main types of steganalysis.

Firstly, visual attacks try to reveal the presence of hidden information through inspection with naked eye or with the assistance of a computer, which can separate the image into bit planes for further analysis. Secondly, statistical attacks are more powerful and successful, because they reveal the smallest alterations in an image´s statistical behaviour. These attacks can be further classified as (i) Passive attack and (ii) Active attack. Passive attacks deal with identifying the presence or absence of a covert message or the embedding algorithm used etc. whereas the goal of active attacks is to estimate the embedded message length or the locations of the hidden message or the secret key used in embedding. Thirdly, structural attacks are based on fact that format of the data files often changes as the data to be hidden are embedded; on identifying these characteristic structure changes can detect the existence of image.

Another criterion of steganalysis categorization is given by Xiang-Yang et al. [10] , who distinguish two categories: steganalysis for specific embedding and universal blind steganalysis. First category deals about algorithms designed for revealing the presence of secret information embedded by specific steganographic algorithms, whereas the universal blind steganalysis can detect the secret message independently to embedding algorithms, hence it has broaden practical applications.

In the last decade, several steganographic schemes have been developed to solve the privacy problem. Among these schemes is an approach that hides a secret message in the spatial domain of the cover image [3][5][6].

In Lee and Chen’s method [3], the least significant bit (LSB) of each pixel in the cover image is modified to embed a secret message. In Wang et al.’s method [5], the optimal substitution of LSB is exploited. These schemes have a higher quality stego-image but are sensitive to modification. In Chung et al.’s method [6], singular value decomposition (SVD) based hiding scheme is proposed. In Tsai et al.’s scheme [4], the bit plane of each block truncation coding (BTC) block is exploited to embed a

978-1-61284-324-7/11/$26.00 ©2011 IEEE

secret message. For such reasons, many imagery steganographic methods have been invented. Here we briefly review research carried out particularly in the Discrete Cosine Transform (DCT) domain as JSteg method that hides information sequentially in LSBs of the quantized DCT coefficients (qDCTCs) while skipping 0’s and 1’s; OutGuess method scatters information into the LSB of qDCTCs [7]. Another method employs the technique of matrix encoding to hold secret information using LSB of qDCTCs in F5. Others researches of using DCT transform are mentioned in [8][9].

In this paper we propose a novel method, which allows hiding a secret information in still grey scale image while maintaining good visual quality and high capacity of secret information. Advantage of this method is full reconstruction of secret data without having original image present on the recipient side. The organization of the paper is as follows. In the next section, we describe objective visual quality measurements to simulate human perception which are PSNR[dB] and PSNRCFS [dB] modified by Contrast Sensitivity Function to better reflection of Human Visual System (HVS) model. In Section 3 we give overview of DWT which coefficients were used for embedding with specific way that is described later on in Section 4. In Section 5 are given results of proposed method in manner of PSNR value using cover image with different detail levels. Some open problems of image steganography related to transform domain and some interesting direction that may be worth future research are discussed in Section 5. Finally, we conclude our paper in Section 6 with discussion and contribution of our proposed method to the image steganography.

2. Visual quality of stego-image The basic principle of embedding secret object is

defined in sense of similarities stego and cover object and its measure can be expressed by function of conformity. However, this function of perceptibility does not have wide practical use. Hence, the objective measures, which stem from statistical approach, are used in general. Their application is based on measuring of the distortion between the cover data f (i, j) with resolving capacity M x N and modified data f (i, j) with same resolving capacity. To the most frequent used criterions for evaluating the quality of reconstructed information are Mean Square Error (MSE), Signal Noise Ratio (SNR) and Peak Signal Noise Ratio (PSNR). However, the subjective measures have to be still considered in application of image steganography by reason of better reflection of the human perception.

Subjective measures are based on human examination (cluster of observers), which results are categorized to multi-level hierarchy. The measurement of subjective measures is hard to conduct and replicate with identical values even though its results converging to ideal state in case of examination of stego image by human observer. Contrariwise, the objective measures are based on the mathematical apparatus and therefore are independent to

the human examination what ergo means isolating of uncertain human factor during examination phase of the stego object. In Table I. are stated the most used objective measures. The PSNR value is used in all our practical examples as mathematical simulation model of the human visual perception.

Parameter

Mean squared error

[ ]2

1 1

),(),(.1

= =

−=M

i

N

j

jifjifNM

MSE

Peak signal to noise ratio

( ) [ ]dBMSE

PSNRn 2

10

12log10 −=

PSNR weighted by CSF ( )( )

[ ]dBgII

NN

PSNR N

i

N

jCSF

WCSF

= =

∗−=

1

1

2

1

2

21

2

10

.1

255log10

Tab. 1. Objective visual quality measures.

The Contrast Sensitivity Function (CSF) denotes sensitivity of the different frequency of HVS model to the different spatial frequencies. The algorithm calculates the peak signal to noise ration weighted by CSF function. The aim of this objective measure was to approximate to ideal subjective measures. The Human Visual System model (HVS) is part of PSNRCFS calculation, where its frequency sensitivity is represented by 2D FIR filter. The coefficients for HVS are derived from CSF function. The weighting of correspondent spatial frequencies is attained by application of the 2D FIR filter to the error image.

3. Discrete Wavelet Transform with Haar wavelet Multi-resolution analysis based on two-dimensional

wavelet transform is the extension of one-dimensional analysis. If �(x) and (x) stand for one dimension scale function and wavelet function respectively, the following one two-dimensional scale function and three two-dimensional wavelet functions constitute the foundation of two-dimensional wavelet transform.

( ) ( ) ( )yxyx φφφ =, (1)

( ) ( ) ( )( ) ( ) ( )( ) ( ) ( )=

==

yxyxyxyxyxyx

D

H

V

φψψφψψ

ψφψ

,,,

(2)

After L-level decomposition of the image f(x,y), we have approximation and three detail transform coefficients

( ) ( ) ( )yxyxfyxfA LL ,,,, φ= (5)

( ) ( ) ( )yxyxfyxfD VL

VL ,,,, ψ= (6)

( ) ( ) ( )yxyxfyxfD HL

HL ,,,, ψ= (7)

( ) ( ) ( )yxyxfyxfD DL

DL ,,,, ψ= (8)

The cover image decomposition represented by approximation and detail coefficients is depicted on Fig. 1.

Fig. 1. Image decomposition by DWT

4. Implementation of proposed method in DWT domain The proposed method processes grey scale images as

cover object for creating subliminal channel and it utilizes transform coefficients of 2-Dimensional Discrete Transform (2-DWT) for embedding process. The secret object could consist of any kind of binary data file. The elected cover image is decomposed by DWT transform. According previous chapter describing 2-DWT, this transform provides an approximation and three detail coefficients (horizontal, vertical and diagonal) on each decomposition level. The main objective is fulfil the condition of stealth communication as well as imperceptibility of embedded secret message in cover object and that is in the case of image steganography scrutinized by human observer or any statistical methods. Therefore, the detail coefficients are the most convenient area for secret message embedding. Another important requirement which has to be followed is information capacity of the method. Because the amount of embedded data and total distortion of cover object caused by embedding are reciprocally related, the apt practice for comparison of methods could be specified by payload in relation to PSNR value [dB]. The payload can be expressed in number of bit embedded per pixel (bit/pixel). The proposed method uses the 1-level decomposition of DWT transform by using Haar's wavelet specifically. An application of Haar mother wavelet offers specific attributes of detail coefficients, on which relays proposed method. Generally, approximation coefficients are not suitable for embedding because they carry the most information content of the whole cover image. Therefore, details coefficient are the most convenient area for embedding a secret data.

The mentioned method is based on three modes, where each mode defines the robustness level of the method. The modes are based on fact that the same data are replicated to other details areas of DWT coefficients, where the mode-1 embeds secret data to one area of detail coefficients. The mode-2 inserting the same secret data to another area of details coefficients and mode-3 uses all three details areas VL, DL, HL respectively. Application of

higher mode alternates more details coefficients without contribution to transfer capacity. However, by this approach is guaranteed the robustness of the secret data what ergo means better extraction. The coefficients suitable for embedding that are elected from detail areas (VL, DL, HL) are also denoted as C'={cij | 1 i Mc, 1 i Nc} and secret message as S={si | 1 i n, si �{0,1}} that is resized to same dimension as C'. The process of embedding consists of comparing the secret message Sij with detail coefficients ijC ′ in form of following sequence:

Embedding of secret data

if 01mod0 ≠′∧== ijíj CS then ijij CC ′=*

if 01mod0 =′∧== ijíj CS then ( ) ( )21sgn* −′⋅′= ijijij CCC

if 01mod1 ≠′∧== ijíj CS then ( ) ( )21sgn* −′⋅′= ijijij CCC

if 01mod1 =′∧== ijíj CS then ijij CC ′=*

where *ijC are modified coefficients of DWT detail area.

After embedding all secret data and performing inverse DWT (IDWT) on *

ijC , the stego image is obtained. The stego-image with secret message embedded is then ready for transmission. The extraction process consists of simple modulo operation of stego image coefficients *ˆ

ijC in following way:

Extraction of secret data

≠=

=.0modˆ101modˆ0ˆ

*

*

ij

ijij Cif

CifS (9)

The advantage of this method is that the original cover image does not have to be present on the receiver side. Therefore, the risk of disclosure of secret communication is lower. This fact is more critical in steganalysis research of area, where its methods could easily find the patterns of artificial changes by comparing the original and stego object.

5. Experimental Results The capacity of the method remains the same and it is

represented by 1/4 of cover image size for 1-level decomposition of the cover image. The payload is 0.25 bit/pixel in case of using the maximum capacity and it also varies depending on numbers of detail coefficient are used during the embedding phase. The proposed algorithm employs 1-Level L decomposition of the image hence the total capacity (in bits) is represented by 1/4 of image size number of DWT detail coefficient, which are altered.

The tests were performed with a gray standard testing image Lena and other cover image depicted in Fig. 2 with the same size 256x256 pixels (containing 64 kB). The

secret message what represented by (containing 2048B).

Fig. 2. Cover images a - Lena, b - Motorbike, c

In case of mode-1, the rate reconstruction is around 49% what could blow. However, the visual quality of surpassing with PSNR value approximately3 depicts the dependence between usedmethod in process of embedding and qualitthat is expressed by PSNR value in successful extraction rate in [%] for LePSNR value remains almost same for otheResults for those images are additionally sand the results vary in extraction rate due tlevels, which they carry.

Fig. 3. Quality of stego image expressed by extraction rate in %.

Cover image Mode-1 (D1)

Mode-2(H1, D1

Lena 256x256

PSNR [dB] 63,174 60,164PSNRCSF [dB] 311,286 106,647Extraction [%] 49,908 99,975

Motorbike 256x256

PSNR [dB] 63,144 60,134PSNRCSF [dB] 311,662 95,787Extraction [%] 49,56 99,926

Apple 256x256

PSNR [dB] 63,096 60,085PSNRCSF [dB] 98,957 74,574Extraction [%] 48,999 99,975

Tab. 2. Experimental result of extraction rate anfor different cover images and numberused during embedding.

6. Discussion and ConcluExperimental results show that pr

posses with very good visual quality of the

stream of bits

- Apple

of successful be considered as stego image is y 63dB. The Fig. d modes of the ty of stego image

[dB] and also ena image. The er cover images. howed in Tab. 2 o different detail

PSNR[dB] and

2 1)

Mode-3 (H1,V1, D1)

4 58,4037 98,777

5 1004 58,3737 93,6686 1005 58,3254 72,5115 100

nd PSNR values r of detail areas

usion roposed method e stego-image, in

all cases higher than 58dB anvariety in implementation to acqufault tolerance. As it could be smode-2, the extraction rate is nexcellent visual quality of stegowith 100% extraction rate. Fproposed method could be improSequence Spread Spectrum theoHence, this also could significanof the first variance of algorithdetail coefficients are used. Theapplied to second variation of prthe algorithm allows combinatcoefficients. Proposed method implementation and as it wassecurity level as the original extraction.

Acknowledgements The paper was supported by MSlovak Republic VEGA Grant also the result of the project DevInformation and CommunicaKnowledge Systems (project supported by the Research & Program funded by the ERDF (40

References [1] BÁNOCI, V., BUGÁR, G., LEVI

using by CDMA techniques, Radio183 - 186.

[2] COX, I. J., MILLER, M. L., BKALKER, T. Digital watermarkiBurlington: Morgan Kaufmann, 200

[3] LEE, Y.K., CHEN, L. H. High capVision, Image Signal Process., 2000

[4] TSAI, P., HU, Y. C., CHANG, Cusing block truncation coding. ProDigital Steganogarphy. Kitakyaushu

[5] WANG, R. Z., LIN, C. F., LIN, J. substitution and genetic algorithm. 3, p. 671 - 683.

[6] CHUNG, K. L., SHEN, C. H., CHABased Image Hiding Scheme. Patteno. 9, p. 1051 - 1058.

[7] PROVOS, N. Defending against sta10th USENIX Security Symposium,

[8] SALLE, P. Model based steganogron Digital Watermarking, 2003, p. 1

[9] MIYAKE, K., IWATA, M., SHIOutilizing features of JPEG images. IElectronics, Communications and C4, p. 929 – 936.

[10] XIANG-YANG, L., DAO-SHUN, review on blind detection for image2008, Vol. 88, no. 9, p. 2138 - 2157

nd algorithm also allows uire desired robustness and seen from table results, in early 100%. Results show

o image no less than 58dB For further research, the oved by applying of Direct-ory DSSS for instance [1]. ntly enhance extraction rate hm where just one area of e same deduction could be oposed method. Moreover,

tions of any three details allows great variety in

s mentioned also higher image in not needed for

Ministry of Education of No. 1/0065/10 (60%) and velopment of the Center of ation Technologies for

number: 26220120030) Development Operational 0%).

ICKÝ, D. Steganography system oelectronics conference, 2009, p.

BLOOM, J. A., FRIDRICH, J., ing and steganography. 2nd ed. 08.

pacity image model. IEEE Proc. - 0, Vol. 147, no.3, p. 288 - 294.

C. C. An image hiding technique oc. of Pacific Rim Workshop on u (Japan), 2002, p. 54 - 64.

C. Image hiding by optimal LSB Pattern recog., 2001, Vol. 34, no.

ANG, L.C. A Novel SVD- and VQ-ern recog., letters, 2001, Vol. 22,

atistical steganalysis. Proc. of the Washington, 2001, p. 323 – 336.

aphy, in: International Workshop 154 – 167.

OZAKI, A. digital steganography IEICE Trans. on Fundamentals of Computer, 2004, Vol. E87-A, no.

W., PING, W., FEN-LIN, L., A. e steganography, Signal Process.,

7.