IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 5, NO.
2, JUNE 2010
Correspondence
SVD-Based Universal Spatial Domain Image Steganalysis
Gokhan Gul and Fatih Kurugollu
Abstract—This paper is concerned with the universal (blind) image
steganalysis problem and introduces a novel method to detect especially
spatial domain steganographic methods. The proposed steganalyzer
models linear dependencies of image rows/columns in local neighborhoods
using singular value decomposition transform and employs content inde-
pendency provided by a Wiener filtering process. Experimental results
show that the novel method has superior performance when compared
with its counterparts in terms of spatial domain steganography. Exper-
iments also demonstrate the reasonable ability of the method to detect
discrete cosine transform-based steganography as well as the perturbation
quantization method.
Index Terms—Classification, singular value decomposition (SVD), ste-
ganalysis, Wiener filtering.
I. INTRODUCTION
Steganography is the clandestine communication over public chan-
nels. The ultimate goal of steganography is to conceal the presence
of communication by embedding secret messages into innocuous dig-
ital media in a way that is statistically undetectable. In this context,
steganography is considered secure if there is no attack that can deter-
mine the existence of the secret message with a success rate better than
a random guess [1]. On the contrary, steganalysis is the art of discov-
ering the very presence of hidden data in cover objects. Steganalysis
can be broadly classified into two groups: algorithm specific (targeted)
methods [2] and universal (blind) methods [3], [4]. Algorithm specific
steganalysis assumes that the steganographic method is known by the
attacker. The attacker takes advantage of this prior knowledge to design
methods to reveal the existence of the hidden data. The shortcoming
of this type of steganalysis is that their satisfactory performance is re-
stricted to a specific steganography. Universal steganalysis methods
aim to overcome this problem. Instead of using any a priori informa-
tion, they take into account all available steganography methods to de-
vise a single steganalysis framework. It is supposed that a blind ste-
ganalysis method can detect any steganography if sufficient numbers
of cover and stego images have been taken into account during the de-
sign process [5].
All universal steganalysis methods assume that data-hiding destroys
the underlined statistics of natural images. Therefore, a common char-
acterization should be possible if the features incorporated to the classi-
fication process are sensitive to the embedding noise and insensitive to
the image content. Universal steganalysis schemes can be divided into
two categories: spatial domain and discrete cosine transform (DCT)
domain. The methods belonging to the former one extract the features
Manuscript received July 20, 2009; revised October 21, 2009; accepted De-
cember 19, 2009. Date of publication February 17, 2010; date of current version
May 14, 2010. The associate editor coordinating the review of this manuscript
and approving it for publication was Dr. Mark (Hong-Yuan) Liao.
The authors are with the School of Electronics, Electrical Engineering and
Computer Science, Queen’s University, Belfast, U.K. (e-mail: gokhangul@hot-
mail.com; f.kurugollu@qub.ac.uk).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TIFS.2010.2041826
1556-6013/$26.00 © 2010 IEEE
349
from the pixel information [4], while the latter ones attack on the DCT
coefficients [6].
In this paper, we propose a novel steganalyzer which derives the
features from the spatial domain and performs detection without re-
gard to which domain the secret message is embedded. The proposed
method models linear dependencies of image rows/columns using sin-
gular value decomposition (SVD) and employs content independency
via Wiener filtering [7]. The contributions of this paper over our pre-
vious work [8] are summarized as follows:
1) New features are derived from content independency process.
2) Rate independent comparative experiments are provided.
3) Steganalysis issue for the perturbation quantization (PQ) method
is presented by taking singular value-based steganalysis (SVBS),
which is the most successful steganalyzer for PQ, as a benchmark
algorithm [2].
In the experiments, we consider three publicly available spatial do-
main (LSB, LSB6, and Steghide) and five DCT domain (F5, Outguess,
JP Hide&Seek, MB1, and MB2) steganography algorithms as well as
the perturbed quantization (PQ) method. The LSB algorithm, which is
one of the most simple and widely used methods in steganography, sub-
stitutes the least significant bits of the pseudorandomly chosen cover
image pixels with the binary message bits. LSB6, on the other hand,
increments or decrements the value of the chosen pixels according to
a random variable with uniform distribution on f01; +1g if the LSBs
of these pixels are not the same with the corresponding message bits
[9]. Another variety of this paradigm is Steghide algorithm which ex-
changes the pixel values rather than changing their LSBs to preserve
the first-order statistics [10]. In F5 algorithm, the absolute values of
nonzero ac DCT coefficients are changed using a matrix embedding
strategy to minimize the embedding distortion [11]. The Outguess al-
gorithm tries to preserve the DCT histogram with a two-pass strategy.
In the first pass, the message bits are embedded into the randomly se-
lected LSBs of DCT coefficients excluding those whose values are 0, 1,
or 01. In the second pass, an adjustment process on DCT coefficients
is carried out so that the DCT coefficient histograms of cover and stego
images are the same [12]. JP Hide&Seek embeds a content which is
encrypted using the Blowfish algorithm as a pseudorandom number
generator into DCT coefficients [13]. MB1 and MB2 are model-based
methods which try to model and to preserve the statistical properties of
DCT coefficients during the embedding process [14]. Finally, the PQ
algorithm, which is the state-of-the-art steganographic method, embeds
the message bits in the JPEG compression stage by perturbing the quan-
tization steps [15].
We compare the proposed method with four other state-of-the-art
universal (blind) steganalysis methods; Farid’s method (Farid), Binary
Similarity Measures (BSM), Wavelet Absolute Moments (WAM), and
Xuan et al. Farid uses the first four moments of the wavelet coeffi-
cients over N subbands and three orientations as well as of the norms
of the optimal linear predictor residuals [3]. BSM takes into account
two LSB planes and calculates several binary similarity features re-
lated to the correlation between these bit planes as well as the binary
texture characteristic within the bit planes [4]. WAM uses higher order
absolute moments of the noise residual to calculate the features because
the noise residual increases the features’ sensitivity to embedding [9].
Finally, Xuan et al. uses an image co-occurrence matrix to derive the
features which are clustered using class-wise nonprincipal components
analysis (CNPCA) [16].
350 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 5, NO. 2, JUNE 2010
The rest of the paper is organized as follows. SVD and the proposed
features are explained in Section II. The image sets used in the experi-
ments, the embedding processes related to the steganographic methods,
and the experimental results are presented in Section III. Finally, the
conclusions are drawn in Section IV.
II. STEGANALYZER DESIGN
A. Singular Value Decomposition
SVD is an essential matrix factorization method which is widely
used in signal processing. It decomposes a matrix A 2 m2n into
the product of two orthonormal matrices U 2 m2m , V 2 n2n and
a diagonal matrix S 2 m2n as follows:
A = USV T :
(1)
The diagonal elements of matrix S are non-negative and are sorted in
decreasing order 1  2  1 1 1 min(m;n) , where m and n are the
dimensions of A. These sorted elements produce a vector named as
“singular value vector:”
Sv = Diag(S): (2)
B. Feature Extraction
The proposed method attacks steganographic content using the fea-
tures derived from singular values. Let us assume a full rank matrix.
If any two rows or columns of this matrix are modified so that they
become linearly dependent, it can be observed that the lowest singular
value vanishes. If this process is repeated using the next row or column,
the second lowest singular value becomes zero. This observation comes
up with two main ideas which are the pillars of the proposed method.
First, the reaction to the changes on the matrix content starts from the
lowest singular value. Second, the lower valued singular values’ close-
ness to zero indicates a group of vector’s closeness to the linear de-
pendency. This observation can be used to model the soft relationship
between the image rows and columns which will be disturbed by the
embedding process.
Due to the aforementioned unequal effect of embedding noise on the
singular values, it is necessary to adopt a function which can intensify
the lower valued singular values for a powerful steganalysis and can at-
tenuate the higher valued ones to normalize the different energy levels
of different images. For this purpose, a function comprising the loga-
rithm of the inverse power of singular values log(x01 ) is devised to
derive features for the steganalysis. Since spatial domain represents a
strong dependency between the pixels in the local neighborhoods the
features are extracted from the subblocks representing the locality in
the spatial domain rather than the whole image. It is obvious that when
the block size increases the number of examined blocks decreases. This
results in decreasing the dependencies which are considered. To alle-
viate this problem, the subblocks are overlapped proportionally to the
block size in order to be able to take into account the correlations within
and among subblocks. Consequently, the feature extraction algorithm
is described as follows.
Step 1: Divide image I , into subblocks of size W 2 W , where W =
3; 4; . . . ; 27 according to the following overlapping rules:
If W < 8; no overlapping
If 8  W  13; 50% over lapping
If W > 13; 75% overlapping:
The methodology used in this overlapping strategy is ex-
plained with experiments in Section III-B.
Step 2: For each particular W , calculate the singular value vector
Sv of each subblock j
SVD(Subblockj ) 0! Svj = [1j ; 2j ; . . . ; Wj ]: (3)
Step 3: Calculate the natural logarithm of the inverse power of each
singular value and add the singular values up with respect
to the related subblock j
W
SvBj = log ij01 ; j = 1; 2; . . . ; TW ij 6= 0 (4)
i=1
where TW is the total number of subblocks sizes of W 2W .
Step 4: Sum the final results obtained in Step 3 and normalize them
with the number of total subblocks
T
FW = T1 SvBj ; W = 3; . . . ; 27: (5)
W j=1
Using this algorithm, we obtain 25 dimensional (25D) features for each
image.
One of the crucial obstacles for an efficient steganalysis is the image
content. To alleviate this obstacle, we drew an analogy between the de-
tection problem for steganalysis and the classical dc signal detection in
additive white Gaussian noise (AWGN) channel [17, p. 67, Example
3.2]. According to the latter one, an increase in the noise variance re-
sults in a decrease in the detection performance. For image steganal-
ysis, assuming that the embedding noise is the dc value which is aimed
to be detected and AWGN as the image content, an increase in the dy-
namic range corresponds to an increase in the variance of AWGN. Ac-
cordingly, any attempt to decrease the dynamic range can improve the
detection performance. In respect of this analogy, we need to expect
the pixel values of bitmap images to be in the interval of [0 f ], where
f < 255 before the data embedding. This is obviously not possible.
However, it is possible to estimate the cover image with a certain pre-
cision. The difference image between the image under consideration
and the estimated image can be viewed as an estimation of the em-
bedding noise. This estimation provides diverse information from the
original data. Therefore, we resort to extract other 25 features from the
difference image yielding 50 features in total. To estimate the original
image, we use Wiener filtering with a 3 2 3 window. In Section III-B,
we justify the use of this filtering and we show that the merged 50 fea-
tures (WFLogSv) are superior to both 25D LogSv features [8] and 25D
Wiener filtering process features (WF). These 50 features are used in
the classification process to make a decision on the image under test,
whether it is cover or stego.
III. EXPERIMENTAL RESULTS
A. Image Set
Greenspun’s image database consisting of 1800 natural color im-
ages with the quality factor equals to 100 is used in the experiments
[18]. These images are converted to grayscale and black borders around
them are then cropped resulting in images of size 480 2 480 pixels. We
never compressed the images since it is harder to detect spatial-domain
steganography in raw images than the preprocessed ones [9]. All im-
ages in the data set are embedded with relative payloads of 0.4, 0.2,
0.1, and 0.05 bit per pixel (bpp) for spatial domain steganographic al-
gorithms or 0.4, 0.2, 0.1, and 0.05 bit per nonzero DCT coefficient (bpc)
for DCT domain ones in order to obtain the stego image set. Note that
due to the limited capacity of the Outguess algorithm [8], the maximum
embedding rate is considered instead of 0.4 bpc.
IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 5, NO. 2, JUNE 2010
TABLE I
DETECTION PERFORMANCES OF VARIOUS OVERLAPPING STRATEGIES
For the PQ method, 1000 images from the last experiment of [2] are
combined with the Greenspun’s image set. There are 2603 images in
this combined set since PQ was not able to embed 0.4 bpc to 197 images
in the Greenspun’s set. Only a 0.4-bpc rate is considered for PQ due to
its very poor detectibility for lower embedding rates [15]. JPEG quality
Q
( ) factor is determined by the steganographic algorithms as 75 and
80 for Outguess and F5 methods, respectively. For these two methods,
the same JPEG quality factor used to compress the stego images is
exploited to get the cover images. For PQ, because of the nature of the
method, recompressed images are used as cover images via doubling
Q
the first quantization matrix ( 1 = 85) [2].
B. The Overlapping Rule and the Content Independency Process
The design of the proportionally increasing overlapping rule relies
on the trade-off between the accuracy of the steganalyzer and the time
necessary to extract the features. The greater the number of overlap-
pings, the higher the amount of collected statistics is. However, after
exceeding a number of overlappings, in terms of detection performance
we do not gain much. Therefore, our strategy was to achieve an ac-
ceptable detection performance within a considerable time slot. Ran-
domly chosen 800 images from the Greenspun image set and the corre-
sponding 0.05 rate embedded LSB images are used in the experiments.
The detection rates are calculated as the average of the maximum ac-
) are
curacy over 100 random trainings and tests using an SVM classifier
with optimum parameters (see Section III-C for details). We compared
the detection performance of the proposed overlapping algorithm to no
overlapping, 50% and 75% overlappings. The detection performance
of 25D LogSv features for various training and testing sets and for the
considered overlapping rules are listed in Table I. It is clear that using
this strategy we only reduce the computational time; the performance
loss compared to 75% overlapping is negligable.
The design parameter of a Wiener filter is the window size. Due to the
stochastic nature of the distribution of the natural images, it is not pos-
sible to provide a good window size by means of detection performance
with theoretical analyses. Besides, ideally, an infinite number of window
sizes limits the number of practical experiments. We tested the window
sizes 2 2 2, 3 2 3, 4 2 4, and 5 2 5 with the same experimental setup.
Out of 800 images, 500 images are used for training and the rest of the
images for testing. With 25D features from the Wiener filtering process,
we obtained detection performances; 86.7%, 87.1%, 86.6%, and 88.0%
for 2 2 2, 3 2 3, 4 2 4, and 5 2 5 window sizes, respectively. In case we
consider 50D features (25D LogSv and 25D Wiener filtering features),
we obtained 91.3%, 91.8%, 88.6%, and 89.3% detection performances
for 2 2 2, 3 2 3, 4 2 4, and 5 2 5 window sizes, respectively. Since the
final detection performance is of our interest, we decided for a Wiener fil-
tering with a 3 2 3 window. In the next experiment, we investigated the
performance improvement with additional 3 2 3 Wiener filtering fea-
tures for all considered steganography algorithms. The detection per-
formance of the proposed 50D WFLogSv features, 25D LogSv features
[8], and 25D Wiener filtering features are given in Table II. It can easily
be seen that WFLogSv leads to at least about 1% and at most about 7%
performance gain for all steganography algorithms. This clearly indi-
cates one of the main contributions of this paper.
351
TABLE II
DETECTION PERFORMANCES OF LogSv, WF AND WFLogSv
C. Spatial Domain Steganalysis
Besides the proposed approach, we have implemented the CNPCA
classifier as well as the high-dimensional features of Xuan et al. [16]
and BSM [4]. The source code of Farid’s steganalysis method [3] is
obtained from the Internet [19] while that of WAM is provided by its
authors [9]. For a fair comparison, each steganalyzer is associated with
a classifier which provides the best performance on average. Three clas-
sifiers are considered: Support Vector Machine (SVM), Fisher Linear
Discriminator (FLD), and CNPCA. FLD for WAM and Farid, CNPCA
for Xuan and SVM for BSM, SVBS and WFLogSv have been chosen
empirically according to the best detection performance for the related
staganalysis method, except for Xuan. Due to high feature dimension-
ality CNPCA is adopted inevitably for the Xuan method. CNPCA and
R C
SVM have some free parameters, and ( , ), respectively. Optimum
parameters for the steganalyzers are determined before the steganal-
ysis depending on solely the steganographic method by the use of a
grid search algorithm having a step size of one for CNPCA and two
with five fold cross validation for SVM. Pair of parameters ( C;
found as (217 2017 ), (215 21 ), and (215 203 ) for WFLogSv, BSM,
; ; ;
and SVBS, respectively, for all steganographic methods. The param-
R
eter found for Xuan is = 523.
Using the optimum parameters, the classifiers are trained by the ran-
domly chosen 1200 cover and the corresponding 1200 stego images
(300 images from each embedding rate) in the training phase. The rest
of images are used in the testing phase. These images intersect neither
within nor among the training and testing sets (as it is true in all exper-
iments). Note that in [8] the classifiers are trained only with the images
of a particular embedding rate according to Kerckhoffs’ principle (em-
bedding rate is assumed to be known by the steganalyzer) whereas here
we consider all embedding rates for the construction of the training set.
Because in the steganalysis community this principle has been found
unrealistic by asserting that in a realistic scenario the embedding rate
may not be known [20]. Moreover, except for PQ, the experiments
are iterated 100 times for all steganograpic methods in order to pro-
vide reliable results. For PQ, we iterated the training and the testing
phases for 10 times. The detection performances of the state-of-the-art
steganalyzers are determined by averaging the maximum accuracy of
:
the classifiers, which is given as 1 0 0 5 2 (false alarm percentage +
miss percentage), over all iterations. From Tables III–V, we give the
detection performances of steganalyzers performing on spatial domain
steganography. It can be easily observed that the proposed method out-
performs the prior arts for all spatial domain steganographic methods
with a significant margin. Notice that for very low embedding rates this
divergence is more significant.
D. DCT Domain Steganalysis
In order to detect DCT domain steganography, there exist quite pow-
erful DCT-based steganalysis algorithms, for example [6] and [16].
352 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 5, NO. 2, JUNE 2010

TABLE III TABLE VI

DETECTION PERFORMANCES FOR LSB DETECTION PERFORMANCES FOR F5

TABLE IV TABLE VII

DETECTION PERFORMANCES FOR LSB 6 DETECTION PERFORMANCES FOR JP HIDE&SEEK

TABLE V TABLE VIII

DETECTION PERFORMANCES FOR STEGHIDE DETECTION PERFORMANCES FOR OUTGUESS

TABLE IX
However, until recently, the performance of spatial domain steganal- DETECTION PERFORMANCES FOR MB1
ysis algorithms to detect DCT domain steganography has not yet been
investigated. This kind of steganalytic attack can especially be useful
if the secret information is embedded in the DCT domain and the stego
image is advertised as a bitmap image by adopting some error correc-
tion codes and decompressing the image to the spatial domain. Actu-
ally a similar scenario has been introduced recently [21] in the YASS
method. YASS embeds information in the spatial domain but the stego
image is advertised in the DCT domain (as a JPEG image) by em-
ploying some error correcting codes. Our scenario considered in the
DCT tests is the reverse version of the one provided in YASS.
TABLE X
Therefore, at one hand, DCT domain steganalysis algorithms should DETECTION PERFORMANCES FOR MB2
have a good reason to attack bitmap images since it is always assumed
that if the image is represented in a particular domain, the embedding
should have been performed on this domain as well [6], [16]. On the
other hand, DCT domain steganalysis algorithms can experience se-
vere performance degradation due to the masking of decompression
on the embedding noise. Another reason for attacking DCT domain
steganography using spatial domain steganalysis is that, although de-
tection can intuitively be estimated to be poor due to the decompres-
sion, spatial domain features are likely to provide diverse information
from the DCT-based ones which is useful for some fusion schemes.
From Tables VI–X, the detection performances of the state-of-the-art of the steganalysis methods over all DCT stegonagraphic methods, the
spatial domain steganalysis methods on F5, JP Hide&Seek, Outguess, proposed method (WFLogSV) is the second best after WAM.
MB1, and MB2 steganography methods are tabulated. The proposed
method shows reasonable performance in the DCT domain. It is the
E. Steganalysis of PQ
second best for JP Hide&Seek, MB1, and MB2 and the third best for
F5 and Outguess on average. It is also the second best for F5 and Out- For the sake of completeness, we have investigated the potential of
guess for the lowest embedding rate (0.05). If we rank the performances the steganalyzers on PQ steganography. As a benchmark, we adopt the
IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 5, NO. 2, JUNE 2010
Fig. 1. ROC curves of universal steganalysis schemes and SVBS.
SVBS steganalyzer with type-I features using the aferomentioned sub-
block scheme resulting in 350 features [2]. Two thousand randomly
chosen images from the combined image set are used for the training
and the rest for the testing. In Fig. 1, receiver operating characteristic
(ROC) curves of steganalyzers averaged over all ten random tests for
0.4 bpc are depicted. Area under the curve (AUC) values of the stegan-
alyzers are calculated as 0.68, 0.71, 0.62, 0.56, 0.57, and 0.76 for the
proposed method, Xuan, WAM, Farid, BSM, and SVBS, respectively.
Please note that SVBS is a targeted steganalyzer for PQ and its perfor-
mance is given here for a comparison. It can be seen that the proposed
method is the second best universal method after Xuan et al. while the
other steganalyzers provide poor detection performance.
IV. CONCLUSION
In this paper a novel spatial domain steganalysis method was pro-
posed. Its performance on detecting spatial domain and DCT domain
steganographic methods as well as PQ algorithm were investigated. The
results were compared with well-known spatial domain universal ste-
ganalyzers including WAM, Farid, BSM, and Xuan. The experiments
reveal that the proposed algorithm is superior to the prior arts for all
spatial domain steganographic algorithms. It shows reasonable perfor-
mance for DCT domain methods (ranked as the second best stegana-
lyzer according to overall rankings). The poor performance of the spa-
tial domain steganalyzers on the DCT domain can be related to the ef-
fect of decompression on the embedding noise. However, the features
extracted from the spatial domain can be used to increase the accuracy
of DCT-based steganalyzers. We also conclude that PQ can be detected
relatively accurately for the 0.4 embedding rate. The best performance
is obtained by Xuan’s method and the proposed algorithm is ranked as
the second best one.
REFERENCES
[1] R. Chandramouli, M. Kharrazi, and N. Memon, “Image steganography
and steganalysis: Concepts and practice,” in Proc. 2nd Int. Workshop
on Digital Watermarking (LNCS 2939), Seoul, Korea, Oct./Nov. 2004,
pp. 35–49.
[2] G. Gul, A. E. Dirik, and I. Avcibas, “Steganalytic features for JPEG
compression based perturbed quantization,” IEEE Signal Process.
Lett., vol. 14, no. 3, pp. 205–208, Mar. 2007.
353
[3] S. Lyu and H. Farid, “Detecting hidden messages using higher-order
statistics and support vector machines,” in Lecture Notes in Computer
Science. New York: Springer-Verlag, 2002, vol. 2578, pp. 340–354.
[4] I. Avcibas, M. Kharrazi, N. Memon, and B. Sankur, “Image ste-
ganalysis with binary similarity measures,” EURASIP J. Appl. Signal
Process., vol. 17, pp. 2749–2757, 2005.
[5] J. Fridrich, “Steganalysis,” in Multimedia Security technologies
for Digital Rights Management, W. Zeng, H. Yu, and C. Y. Lin,
Eds. New York: Academic, 2006.
[6] T. Pevny and J. Fridrich, “Merging marcov and DCT features for multi-
class JPEG steganalysis,” in SPIE 2007 Security Steganography and
Watermarking of Multimedia Contents IX, San Jose, CA, Jan. 2007.
[7] I. Avcibas, “Audio steganalysis with content-independent distortion
measures,” IEEE Signal Process. Lett., vol. 13, no. 2, pp. 92–95, Feb.
2006.
[8] G. Gul and F. Kurugollu, “A novel universal steganalyser design:
Logsv,” in IEEE Int. Conf. Image Processing (ICIP 2009), Cairo,
Egypt, 2009.
[9] M. Goljan, J. Fridrich, and T. Holotyak, “New blind steganalysis and its
implications,” Proc. SPIE Security, Steganography, and Watermarking
of Multimedia Contents VIII, vol. 6072, pp. 607 201–1, 2006.
[10] S. Hetzl and P. Mutzel, “A graph-theoretic approach to steganography,”
Lecture Notes in Computer Science, vol. 3677, pp. 119–128, 2005.
[11] Software [Online]. Available: http://wwwrn.inf.tu-dresden.de/~west-
feld/f5.html.
[12] Software [Online]. Available: http://www.outguess.org/.
[13] Software [Online]. Available: http://linux01.gwdg.de/~alatham/stego.
html.
[14] Software [Online]. Available: http://www.philsallee.com/mb-
steg/index.html.
[15] J. Fridrich, M. Goljan, and D. Soukal, “Perturbed quantization
steganography with wet paper codes,” in Proc. ACM Multimedia and
Security Workshop, Magdeburg, Germany, Sep. 20–21, 2004, pp.
4–15.
[16] G. Xuan et al., “Steganalysis using high-dimensional features derived
from co-occurrence matrix and class-wise non-principal components
analysis (CNPCA),” in Int. Workshop on Digital Watermarking, Korea,
Nov. 8–10, 2006.
[17] S. M. Kay, Fundamentals of Statistical Signal Processing: Detection
Theory. Englewood Cliffs: Prentice-Hall, 1998.
[18] Image set [Online]. Available: http://philip.greenspun.com.
[19] Software [Online]. Available: http://www.cs.dartmouth.edu/farid/re-
search/steg.m.
[20] L. Marvel, B. Henz, and C. Boncelet, “A performance study of
6 1 steganalysis employing a realistic operating scenario,” in IEEE
Military Communications Conf. (MILCOM 2007), Orlando, FL,
2007, pp. 1–7.
[21] K. Solanki, A. Sarkar, and B. S. Manjunath, “YASS: Yet another
steganographic scheme that resists blind steganalysis,” in 9th Int.
Workshop on Information Hiding, Saint Malo, France, Jun. 2007.