458

JOURNAL OF MULTIMEDIA, VOL. 5, NO. 5, OCTOBER 2010

Improved LSB-matching Steganography for Preserving Second-order Statistics

Guangjie Liu, Zhan Zhang and Yuewei Dai
School of Automation, Nanjing University of Science and Technology, Nanjing, China Email: gjieliu@gmail.com

Shiguo Lian
Orange Labs Beijing, Beijing, China Email: shiguo.lian@ieee.org

AbstractIn the paper, for enhancing the security of the traditional LSB matching, two improved LSB-matching methods are proposed. In the steganograhical procedure, the Markov chain distance based on the second-order statistics is chosen as the security metric to control the modification directions of 1 embedding. The first method is based on stochastic modification, which directly determines the modification directions by the empirical Markov transition matrix of a cover image and the pseudorandom number generated by a pseudorandom number generator. The second one is based on genetic algorithm, which is used to find the optimum matching vector to make the security metric as small as possible. Experiments show the proposed algorithms outperform LSB matching and LSB replacement in a sense of the firstorder and second-order security metrics. And the adjacent calibrated COM-HCF steganalytical tests also show that the two algorithms are more secure than the traditional ones. Index Termssteganography, LSB matching, stochastic modification, genetic algorithm, Markov chain

I. INTRODUCTION Communication security has become more and more important because of the demand of privacy. If people transmit information through insecure channel, they have to face to the risk that the information may be stolen by the illegal third part. There are two kinds of methods to protect the confidentiality of information. One solution depends on encryption. Its drawback is that nonsenselooked message may inspire the unauthorized part to block even decrypt the encrypted content. The other solution is steganography[1], which aims to embed secret information into a cover object, such as digital audio, image, video, software, database, network traffic and so on without arousing suspicions. For an image steganalysitical system, there are two factors, capacity and invisibility, should be carefully considered. The capacity, how much bits can be embedded in a cover image, is easy to be understood. The invisibility against perceptual analysis means the embedding cannot introduce perceptual distortion to arise the analyzers suspicion, and that against

steganalysis means the message embedding can not introduce the detectable alteration on statistics of a cover image. Usually, the first invisibility is defined as imperceptibility, and the second is defined as security. In most cases, how to deal with the trade-off between capacity and security is core issue to design a steganalysitical algorithm. In all embedding schemes, LSB steganography is most popular. There are two types of LSB steganography: LSB replacement (LSBR) and LSB matching (LSBM). In LSBR, least significant bits of cover pixels are simply replaced by the secret message bits. In LSBM, the pixel with LSB not matching with the corresponding message bit will be randomly added or subtracted by one. So LSBM is commonly called 1 embedding. In both methods, it is easy for the recipient to retrieve the hidden message bits by reading LSBs. Unlike LSBR and LSBM, which deal with the pixel values independently, LSB matching revisited (LSBMR)[2] takes two pixels as a group to perform embedding. The LSB of one pixel is used to carry the first bit, the odd-even relationship of the two pixels is used to carry the other. Through the mechanism, the change rate can decrease from 0.5 to 0.375 bits/pixel. In [3], Omoomi et al. proposed to use more than two pixels to realize the 1 data embedding scheme and further decrease the change rate. There are also some methods based on syndrome coding such as [4,5], those perform 1 data embedding before encoding the message bits to the possible parity changes on some pixels, they can achieve lower change rates while not exploit all cover pixels. In the paper, our attention is paid on full embedding, that all cover pixels are used to carry message bits. For full embedding, tuning modification directions can decrease the change rate like [2] and [3]. It can also be used to improve the security. For example, Zhang et al. [6] proposed an embedding method with histogram unchanged, and performed it on BMP images. Lu et al [7] proposed a 1 steganographical method by minimizing the distortion of first order statistics. In the method, the authors used the Euclidian distance of histograms of cover and stego images to depict the security distance and used the quadratic programming to resolve the corresponding optimization problem. In [8],

2010 ACADEMY PUBLISHER doi:10.4304/jmm.5.5.458-463

JOURNAL OF MULTIMEDIA, VOL. 5, NO. 5, OCTOBER 2010

459

we also proposed a method based on combination of several quantization-based embedders. Through the optimal choice of combination function, the KullbackLeibler distance can be keep lower. While, the above methods can only preserve the first-order statistics, which just provides limited security. So in this paper, we begin from the second-order statistics and propose two methods based on 1 embedding to lower the secondorder security distance near to zero. The remainder of the paper is organized as following. Section 2 introduces the related work and discusses the concepts of the second-order statistical security. Section 3 describes the new two algorithms in detail. Section 4 gives experiments and the corresponding analysis. Section 5 concludes the whole paper. II. RELATED WORK A. Steganalysis against LSB steganography Although LSBR does not introduce visible artifacts on images, it reveals clear artificial effects on the histogram. The early steganalytical methods are all against LSB replacement[9]. LSBM has no obvious effects on the histogram of the image, which makes it harder be detected reliably. So in recent years, steganalysis against LSBM has attracted many researchers. Initially, Harmsen proposed to detect LSB matching by modeling data as additive noise [10]. He observed that the center of mass (COM) of the histogram characteristic function (HCF) decreases. Therefore, COM is extracted as a simple feature to make detection. The significant weakness of Harmsens method is that the detector is just effective for color images. This method was later improved by Ker[11], who called the two proposed methods as calibrated HCF-COM and calibrated adjacency HCF-COM. The experiments demonstrated a sound performance for LSB-matching steganography in grayscale images. In [12], Kers calibration technique was improved again by Li et al. via performing the features extraction on the difference image. In [13], Goljan et al. proposed the WAM (wavelet absolute moment) blind detector, which estimated the stego noise by applying some denoising techniques and extracted the features from the noise. In [14], based on the statistics of ALE (amplitude of local extrema) of the histogram, Zhang et al. presented a detector for LSB matching. Cancelli et al. extended Zhang et al.s work to the two dimensional case [15]. In [16], Gao et al. suggested using the calibration technique, the amplitude of local extrema and two dimension histograms to future improve the detection accuracy. In [17], Pevn et al. proposed to use the subsets of sample transition probability matrices as features to implement a detector which was reported to outperform ALE and WAM methods. B. First-order security metric (K-L distance) In [18], Cachin proposed an information-theoretic model that can be used to quantify the security of steganography in terms of the decision error probabilities

of hypothesis testing. In the model, the cover signal X is supposed to be composed by i.i.d. components x1 , x2 ,..., xL , and xi belongs to the limited value domain space S with M elements, and the stego signal y is also supposed to be composed by i.i.d components y1 , y2 ,..., yL . Assume that the statistical distribution of cover signal X is Pc = [ pc ( s1 ), pc ( s2 ),..., pc ( sM )] with regard to the value domain space S = {s1 , s2 ,..., sM } , and that of stego signal Y is Ps = [ ps ( s1 ), ps ( s2 ),..., ps ( sM )] . The security measure of Cachins security is based on the Kullback-Leibler distance (K-L distance) between Pc and Ps, and it is defined as (1). p ( x) D ( Pc || Ps ) = Pc ( x) log c (1) ps ( x ) xS The value of K-L distance is always non-negative and is zero i.f.f. Pc ( x) = Ps ( x) . But, as we all know, most covers have stronger dependence among cover signals, such as images, audios and videos. So it is not suitable to directly apply the Cachins security metric on common cover signal. B. Second-order security metric (M-C distance) Considering the limitation of Cachins theory, Sullivan [19] used the higher statistics, i.e. Markov chain, to measure the security of steganalytical system. According to a given scanning order, from cover and stego images, two Markov chains can be obtained as X and Y . Suppose that the value set of cover and stego images is . The divergence to quantify the statistical changes is essential the Markov Chain distance (M-C distance) between the empirical matrices M c and M s , which is defined as (2).
D(Mc , Ms ) = M c (i, j ) j M s (i, j ) c (2) M ( i , j ) log M c (i, j ) M s (i, j ) i , j j

Here M c / s (i, j ) is ratio between the numbers of the occurrence that j follows i in the cover or stego chain and the length of the chain minus one. So, it is not hard

to deduce that

M c (i , j ) j M c (i , j )

and

M s (i , j ) j M s (i , j )

are the transition

probabilities of the value i to j for the cover and the stego image separately. According to the information theory, Theorem 1 can be proved. Theorem 1 D ( M c , M s ) 0 , and D ( M c , M s ) = 0 if and only if M c = M s .
Proof: Let A = {(i, j ) M c / s (i, j ) > 0} as the support set

of M c / s (i, j ) , according to (2)

M c (i, j ) j M s (i, j ) D ( M c , M s ) = M c (i, j ) log M c (i, j ) M s (i, j ) i , j j =
i , j A

j M c (i, j ) M s (i, j ) (i, j ) log M c (i, j ) M s (i, j ) j

Because log is strictly convex function, according to the Jensen inequality [20], we have

2010 ACADEMY PUBLISHER

460

JOURNAL OF MULTIMEDIA, VOL. 5, NO. 5, OCTOBER 2010

D ( M c , M s )

j M s (i, j ) s M (i, j ) = log c i , j A j M (i , j ) j M c (i, j ) = log j M s (i, j ) s i A j M (i , j ) = log i , jA M c (i, j ) log i , j M c (i, j ) = log1 = 0
i , j A

M log

(i, j )

j M c (i, j ) M s (i, j ) M c (i, j ) j M s (i, j )

Markov chain from a two-dimension image matrix, the row-by-row scanning mode is adopted. For an input image I, A serialized one-dimension pixel sequence X is obtained, and the empirical Markov transition matrix M c is also obtained. For a pixel x , suppose the forward pixel value is y (it may have been embedded), and the current message bit b , the embedding rule is defined as (6).
mod( x, 2) = b x c ( y , x +1) x = x + 1 mod( x, 2) b & r [0, M c ( y ,M ) (6). x +1) + M c ( y , x 1) c x 1 mod( x, 2) b & r [ c M ( y , x +c1) ,1] M ( y , x +1) + M ( y , x 1)

According to the Jenson inequality[20], the necessary and sufficient condition for the equality is j M c (i, j ) M s (i, j ) =1 i, j , M c (i, j ) j M s (i, j ) The above result means that M c (i, j ) = M s (i, j ) for any i, j . The conclusion will be drawn that if the second-order statistics is preserved, the first-order one is also be invariable. Equation (1) and (2) give the theoretical definitions about the two distances, while to obtain a stable result in the practical computation, some artifices are needed. Reviewing (1) and (2), it can be found that when pc ( x) = 0 , ps ( x) = 0 , M c (i, j ) = 0 or M s (i, j ) = 0 , the division and logarithm operations are undefined. So some additional promise should be made. The promise is based on the observation that the statistical changes after data hiding are not drastic. Despite the computed results are not very precise, the K-L and M-C distances are calculated by the two below equations. p ( x) (3) D ( Pc || Ps ) = pc ( x) log c ps ( x ) xS , pc ( x ) 0, ps ( x ) 0
D(Mc, Ms ) = M c (i, j ) j M s (i, j ) c (4) M ij log M c (i, j ) M s (i, j ) c s , M ij 0 i , j , M ij j

In the new embedding rule, when b equals to the LSB of x , x is unchanged, otherwise, the stego pixel x is changed to x + 1 when a random number r obeying the uniform distribution on [0,1] belongs to the
( y , x +1) range [0, M c ( y ,M ) , or to x 1 when r belongs to x +1) + M c ( y , x 1)
c

( y , x +1) ,1] . It implies that x will be the range [ M c ( y ,M x +1) + M c ( y , x 1)

c

changed

to

x +1

or
s

x 1
M c ( y , x 1)

by

the .
c

probability So in the

M c ( y , x +1) M c ( y , x +1) + M c ( y , x 1)

or

M c ( y , x +1) + M c ( y , x 1)

III. IMPROVED LSB-MATCHING METHODS The data embedding rule of LSB matching is well known as mod( x, 2) = b x x = (5). x 1 mod( x, 2) b Here, x is the pixel value, b is the message bit. Equation (5) implies that there are two modification directions (+ or - 1) to change a pixel to embed a message bit. The core motivation of this paper is to dynamically and optimally determine the modification directions to make M s approximate M c to preserve the so-called second-order statistics. In this paper, two improved LSB matching methods are proposed to preserve the second-order statistics. They are both applied to gray-level BMP images. Of cause, after some small modifications, they also can be used for color BMP or other formatted images. A. LSB matching based on stochastic modification The first method is based on stochastic modification. We name it as SM-LSBM. To obtain the one-dimension

probability meanings, M should near to M . The SM-LSBM data embedding process can be illustrated by Figure 1. In Figure 1, I denotes the covet image, which is input into the row-by-row scanning module to generate the one-dimension vector X. The Mc is obtained via the empirical Markov transition matrix module according to X. The notation (n by 1) denotes to divide an n-length vector into 1-length fragments. For example, assume that the current image pixel equals to 123, the forward image pixel value is 65, c M c (65,122) = 1135 / 65534 and M (65,124) = 970 / 65534 . If current message bit is 0 , 123 is modified to 122 or 124 to carry message bit 0. If PRNG outputs 0.3431, the image pixel will be changed to 122.

Figure 1. SM-LSBM embedding process

After all pixels are embedded, X is inverted to produce the stego image I. To extract the message bits, the receiver just needs to rescan the stego image to extract LSBs of all pixels.

2010 ACADEMY PUBLISHER

JOURNAL OF MULTIMEDIA, VOL. 5, NO. 5, OCTOBER 2010

461

B. LSB matching based on genetic algorithm Notice that SM-LSBM cant ensure M approximates M c as near as possible. In the subsection, we propose another method based on optimization. For the cover image pixel sequence X = [ x1 , x2 ,..., xn ] , according to (5), the stego ,..., xn ] can be obtained by pixel sequence X = [ x1, x2
X = X +T s.t. mod( X ) = B
s

next (1 )Q chromosomes are reproduced by the crossover and mutation. Crossover: Randomly select (1 )Q pair chromosomes according to the selection probability F( g ) P( gi ) = Q F(i g ) and the roulette wheel rule. For the any
i =1 i

(7)

Here, T = [t1 , t2 ,..., tn ], ti {1, 0,1} and B = [b1 , b2 ,..., bn ],

bi {0,1} . The secure steganography can be converted to

the below optimization problem.

T * = arg min ti {1,1,0} D( M c , M s ( X , B, T ))

(8)

Here, D(, ) is M-C distance function and T is called the matching vector. For X, it is hard to solve problem (8) to get the optimum matching vector because the dimension of the variable is too larger. For example, the length of the matching vector of a cover image with 256 256 pixels is as long as 65535. So we suggest to divide the image pixel sequence X and message B into [n/m] m-length fragments. For the fragment i, the data hiding is to find the matching vector Ti = [ti1 , ti 2 ,..., tim ] by (9). Ti = arg min tij {1,1,0} D( M c , M s ( X i , Bi , Ti )) (9) In (9), tij = 0 when mod( xij , 2) = bij , assume that other tij s compose a L-length vector g with L less than m, the following question is to solve (9) to obtain g . Because the object function is non-linear, it is very hard to obtain the formulated solution. In the paper, we give the genetic algorithm to get the approximately optimal solution. The genetic algorithm (GA) [21] can be regarded as a randomized search procedure that is commonly used to solve the optimization problems. In general, GA is mainly comprised of the following three operators, namely, reproduction, crossover and mutation. Here, a chromosome consisting of L genes is described by (10). g = ( g1 , g 2 ,..., g L ) (10) Here, the value of each gene belongs to {1,1} . The genetic algorithm can be described as follows. Initialization: Randomly generate Q individual chromosomes g1 , g 2 ,...., gQ to form an initial population.
Selection: During each successive generation, a proportion of the existing population is selected to breed a new generation through a fitness function as (11), where is a parameter to control the sensitivity of the fitness function. F( g ) = ( D( M c , M s ( X i , Bi , Ti ( g ))) + ) 1 (11) In the phase, the aQ chromosomes with best fitness values are selected. Reproduction: The selected aQ chromosomes are chosen as individuals of the successive population. The

selected two chromosomes g and g , the crossover result g is defined as ,..., g L ) , gi= gi gi, i {1, 2,..., L} g = ( g1 , g2 (6) Mutation: Randomly select Q chromosomes from the new individuals to make the mutation. For a chromosome g = ( g1 , g 2 ,..., g n ) , the mutation is performed by switching two randomly selected genes. The operation can be repeated several times to produce the ultimate chromosome. Termination: The generational process is repeated until a fixed number of generations reaches. We call the scheme as GA-LSBM. The main process of the algorithm is described as Figure 2.

Figure 2. GA-LSBM embedding process

In Figure 2, The notation (n by m) denotes to divide an n-length vector into m-length fragments. Here X i = ( xi1 , xi 2 ,...., xim ) , Bi = (bi1 , bi 2 ,...., bim ) and X is the stego image pixel sequence. After all fragment are embedded, X is inverted to produce the stego image I. The extraction is same as LSB matching. IV. EXPERIMENTAL RESULTS The embedding rate in our experiments is set to 1bpp. The secret messages were generated by a pseudorandom number generator with identical probabilities for bits 1 and 0. The experiments contain two main parts: security metrics and steganalysis resistance tests. In the first experiment, five standard test images, Lena, Peppers, Boat, Plane and Baboon, all with size equal to 512512, are used. The experiment of GA-LSBM is performed under P = 100, = 0.1, = 0.05, = 0.05, = 50, n = 256 . In the mutation phase, the switching is done 5 times. Table 1 gives the experimental results including K-L and M-C distance. For comparison, the LSB replacement(LSBR) and LSB matching(LSBM) are also tested.

2010 ACADEMY PUBLISHER

462

JOURNAL OF MULTIMEDIA, VOL. 5, NO. 5, OCTOBER 2010

TABLE I. . K-L AND M-C DISTANCES FOR FIVE IMAGES Test images Lena (gray,512 512) Peppers (gray,512 512) Boat (gray,512 512) Plane (gray,512 512) Baboon (gray,512 512) algorithm LSBR LSBM SM-LSBM GA-LSBM LSBR LSBM SM-LSBM GA-LSBM LSBR LSBM SM-LSBM GA-LSBM LSBR LSBM SM-LSBM GA-LSBM LSBR LSBM SM-LSBM GA-LSBM K-L distance(10-3) 1.2982 0.6546 0.5302 0.4463 1.5657 0.8643 0.4854 0.3962 11.9681 6.5396 3.6250 3.3865 5.4208 1.3623 0.6724 0.5572 0.8041 0.5933 0.5065 0.4785 M-C distance(10-1) 0.5785 0.4268 0.3825 0.3356 0.5222 0.4237 0.3803 0.3402 0.8187 0.6379 0.5465 0.4680 1.0245 0.6268 0.4573 0.3777 0.7097 0.6468 0.6221 0.6174

more reliable than for uncompressed imags. The reason is that pixels of JPEG compressed images have stronger correlation than uncompressed images.

Figure 3. ROC curves on uncompressed images

According to Table I, It can be seen that the K-L and M-C distances of our methods are always less than those of LSBR and LSBM. It means that our two methods are more secure than LSBR and LSBM. And from Table 1, It could be found that the security metrics of GA-LSBM are better than those of SM-LSBM. And It is also found that images with more textures and edges are more suitable to be taken as cover images. The experimental results of Baboon and Plane are the good example. In the steganalysis resistance tests, we use the steganalysis technique based on calibrated adjacency HCF-COM proposed by Ker[11]. The detection is applied to both cover and stego images. The receiver operating characteristic (ROC) curves are plotted in Figure 4 and 5. The two image sets are: 1338 Uncompressed Images. The set contains 1338 uncompressed TIF images with the size of 384512 (or 512384) from UCID[22]. The color images are converted to gray-scale images in the experiments. 2000 JPEG Images. The set contains 2000 JPEG images with the size of 32642448 (or 24483264), which are all with Nikon CoolPix P5100. The color images are sized 384512 (or 512384), converted to grayscale, and stored with factor 75. The cover images are the decoded gray-value matrix of those JPEG images. The ROC curves show tradeoff between the false alarm probability and the detection probability when the threshold is varying. The cover images in the two sets are all embedded with maximal-length random messages. Note that the x-axes have been scaled to focus on regions of interest. From the two figures, we can see that the calibrated adjacency HCF-COM detector has more reliable detection results for LSBM, and the detection results of SM-LSBM are better than that of GA-LSBM. The ROC results declare that our proposed two methods are more secure than LSBM and GA-LSBM is more secure than SM-LSBM. According to Figure 4 and 5, we can also find that the detection for JPEG compressed image is

Figure 4. ROC curves on JPEG compressed images

V. CONCLUSIONS For improving the security of steganalysitical algorithms, two improved LSB-matching methods are proposed in the paper. Different from the existing schemes, we choose the second-order statistics, i.e., the Markov Chain distance as the security metric. In the first method, the modification directions are determined by the empirical Markov transition matrix of the cover image and the pseudorandom numbers generated by a PRNG. In the second method, the genetic algorithm is used to find the optimum matching vector to make the M-C distance, as small as possible. Experiments indicate that the proposed two algorithms have smaller values of the K-L distance and M-C distance than LSBF and LSBM. And the steganalytical tests also show that our new algorithms are more secure than LSB matching, and GA-LSBM is better than SM-LSBM. But, from the viewpoint of computation complexity, the SM-LSBM is better because that GA-LSBM need more time to complete the optimization process. How to improve the computation speed of GA-LSBM is our work in the future.

2010 ACADEMY PUBLISHER

JOURNAL OF MULTIMEDIA, VOL. 5, NO. 5, OCTOBER 2010

463

ACKNOWLEDGMENT The work was supported by the China Post Doctor Foundation of China (Grant No. 20070421017), NSF of Jiangsu Province(Grand No. BK2008403) and Post Doctor Foundation of Jiangsu Province(Grant No. 0801044B). REFERENCES
[1] S. Lian and Y. Zhang (Eds.). Handbook of research on secure multimedia distribution, New York: IGI Global (formerly Idea Group, Inc), March 2009. [2] J. Mielikainen, LSB matching revisited, IEEE Signal Processing Letter, vol. 13, no. 5, pp. 285-287, May 2006. [3] M. Omoomi, S. Samavi and S. Dumitrescu, An efficient high payload 1 data embedding scheme, Multimedia Tools and Applications, in press, DOI: 10.1007/s11042010-0517-z [4] W. Zhang, X. Zhang, and S. Wang, Maximizing steganalysitical embedding efficiency by combining Hamming codes and wet paper codes, In 10th IHW, vol. 5284 of LNCS, pp. 60-71, Springer-Verlag, 2008. [5] J. Fridrich, P. Lisonek; Grid Colorings in Steganography, IEEE Transactions on Information Theory, vol. 53, no. 4, pp. 1547-1549, April 2007. [6] X. Zhang, S. Wang and K. W. Zhang, Steganography with least histogram abnormality. In International Workshop on Mathematical Methods, Models, and Architectures for Computer Network Security, vol. 2776 of LNCS, pp. 395406, Springer-Verlag, 2003. [7] Y. Lu, X. Li and B. Yang, A 1-based steganography by minimizing the distortion of rst order statistics, In the Fifth International Conference on Intelligent Information Hiding and Multimedia Signal Processing, pp. 754-758, 2009 [8] G. Liu, S. Lian, Z. Ren, Y. Dai and Z. Wang, Image steganography based on quantization-embedders combination, in Proc. of ICME, pp. 1115-1118, 2007. [9] S. Dumitrescu, X. Wu, Z. Wang, Detection of LSB steganography via sample pair analysis, IEEE Transaction on Signal Processing, vol. 51, no. 7, pp. 1995-2007. July 2003. [10] J. Harmsen, Steganalysis of additive noise modelable information hiding. MS Thesis, Rensselaer Polytechnic Institute, Troy, New York, 2003. [11] A. Ker, Steganalysis of LSB matching in grayscale images. IEEE Signal Processing Letters, vol. 12, no. 6, pp. 441-444, June 2005. [12] X. Li, T. Zeng, and B. Yang, Detecting LSB matching by applying calibration technique for difference image, in Proc. of the 10th Workshop on Multimedia & Security, pp. 133-138, 2008. [13] M. Goljan, J. Fridrich, and T. Holotyak, New blind steganalysis and its implications, in Security, Steganography, and Watermarking of Multimedia Contents VIII, vol. 6072 of SPIE, pp. 1-13, 2006 [14] J. Zhang, I. J. Cox, and G. Doerr, Steganalysis for LSB matching in images with high-frequency noise, in Proc. of the IEEE MMSP, pp. 385-388, 2007 [15] G. Cancelli, G. Doerr, I. J. Cox, and M. Barni, Detection of +-1 LSB steganography based on the amplitude of [16]

[17]

[18]

[19]

[20] [21] [22]

histogram local extrema, in Proc. of the IEEE ICIP, pp. 1288-1291, 2008. Y. Gao, X. Li, B. Yang and Y. Lu, Detecting LSB matching by characterizing the amplitude of histogram, in Proc. of ICASSP, pp. 1505-1508, 2009. T. Pevn, P. Bas and J. Fridrich, Steganalysis by subtractive pixel adjacency matrix, in Proc. of the 11th ACM Workshop on Multimedia and Security, pp. 75-84, 2009. C. Cachin, An information-theoretic model for steganography. Information and Computation, vol. 192, no. 1, pp. 41-56, 2004. K. Sullivan, U. Madhow, S. Chandrasekaran and B. S. Manjunath, Steganalysis for Markov cover data with applications to images, IEEE Transaction on Information Forensics and Security, vol.1, no.2, pp. 275-287, 2006. T. Cover and J. Thomas, Elements of Information Theory, NewYork: John Wiley & Sons, 1991. J. Holland, Adaptation in Natural and Artificial Systems. Ann Arbor, MI: University of Michigan Press, 1975. G. Schaefer and M. Stich, Ucid-An uncompressed colour image database. in Proc. of SPIE, Storage and Retrieval Methods and Applications for Multimedia, San Jose, CA, pp. 472-480, 2004.

Guangjie Liu received the B.E. and M.E. from Nanjing University of Science and Technology in 1998, 2002 respectively. He now is a Associate Professor with Automation School of Nanjing University of Science and Technology. He has contributed more than 40 refereed papers covering topics of network and multimedia steganography. Zhan Zhang received the M.E. from Guizhou University in 2006. He now is a PhD candidate with Automation School of Nanjing University of Science and Technology. His research interests includes steganography and steganalysis. Yuewei Dai Yuewei Dai received the M.E., andDr. Eng. degrees from Nanjing University of Science and Technology in 1987 and 2002 respectively. He now is a professor with Automation School of Nanjing University of Science and Technology. He has contributed more than 100 refereed papers including automation control, information technology and information security. He is also the director of System Engineering Academyof Jiangsu Province in China. Shiguo Lian got his Ph.D. in multimedia security from Nanjing University of Science and Technology, China, in July 2005. He was a research assistant in City University of Hong Kong in 2004. He has been with France Telecom R&D (Orange Labs) Beijing since July 2005. He is the author or coauthor of more than 70 refereed international technical papers and book chapters. He has held 13 filed patents. He is the author of the book Multimedia Content Encryption: Techniques and Applications (CRC Press, 2008) and co-editor of Handbook of Research on Secure Multimedia Distribution (IGI Global, 2009). His research interests include network and multimedia security, and intelligent services, e.g. lightweight cryptography, digital rights management (DRM), and intelligent multimedia services and security, etc.

2010 ACADEMY PUBLISHER