International Journal of Computer Applications Technology and Research

Volume 3 Issue 7, 439 - 445, 2014

www.ijcat.com 439
A Review on Classification Based Approaches for STE-
Ganalysis Detection

Anjani Kumar Verma
SPM College, Department Of
Computer Science,
University Of Delhi
New Delhi, India

Abstract: This paper presents two scenarios of image steganalysis, in first scenario, an alternative feature set for steganalysis based on
rate-distortion characteristics of images. Here features are based on two key observations: i) Data embedding typically increases the
image entropy in order to encode the hidden messages; ii) Data embedding methods are limited to the set of small, imperceptible dis-
tortions. The proposed feature set is used as the basis of a steganalysis algorithm and its performance is investigated using different
data hiding methods. In second scenario, a new blind approach of image Steganalysis based on contourlet transform and nonlinear
support vector machine. Properties of Contourlet transform are used to extract features of images, the important aspect of this paper is
that, it uses the minimum number of features in the transform domain and gives a better accuracy than many of the existing steganan-
lysis methods. The efficiency of the proposed method is demonstrated through experimental results. Also its performance is compared
with the Contourlet based steganalyzer (WBS). Finally, the results show that the proposed method is very efficient in terms of its de-
tection accuracy and computational cost.

Keywords: Steganography, Steganalysis, Contourlet transform, Structural similarity measure, Non linear support vector Machine,
MAE, MSE, wMSE, Bayesian

1. INTRODUCTION
Steganography refers to techniques that establish a covert
(subliminal) communications channel within regular, innocu-
ous message traffic [1-3]. Steganography is the art and science
of hiding secret messages by embedding them into digital
media while steganalysis is the art and science of detecting the
hidden messages. The goal of a high quality steganography is
hiding information imperceptibly not only to human eyes but
also to computer analysis.
The obvious purpose of steganalysis is to collect sufficient
evidence about the presence of embedded message and to
break the security of the carrier. Steganalysis can be seen as a
pattern recognition problem also since based on whether an
image contains hidden data or not, images can be classified
into Stego or Cover image classes.
Steganalysis is broadly classified into two categories. One is
meant for breaking a specific steganography. The other one is
universal steganalysis, which can detect the existence of hid-
den message without knowing the details of steganography
algorithms used. Universal steganalysis is also known as blind
steganalysis and it is more applicable and practicable [4, 5]
than the specific steganalysis. Based on the methods used,
steganalysis techniques are broadly classified into two classes;
signature based steganalysis and statistical based steganalysis.
Specific signature based steganalysis are simple, give promis-
ing results when message is embedded sequentially, but hard
to automatize and their reliability is highly questionable [6, 7].
The first blind steganalysis algorithm to detect embedded
messages in images through a proper selection of image quali-
ty metrics and multivariate regression analysis was proposed
by Avcibas et al. [8, 9]. In universal steganalysis, using statis-
tical methods and identifying the difference of some statistical
characteristic between the cover and stego image becomes a
challenge. Due to the tremendous increase in steganography,
there is a need for powerful blind steganalyzers which are
capable of identifying stego images.

In this paper, the focuses are on image steganalysis problem
and develop new algorithms based on rate-distortion concepts.
In particular, it has been observe the effect of different ste-
ganographic methods on image rate-distortion characteristics
and construct detectors to separate innocuous cover images
from message bearing stego images. This paper proposes a
new approach to blind steganalysis does not need any
knowledge of the embedding mechanism. This approach uti-
lizes contourlet transform to represent the images. A Gaussian
distribution is used to model the contourlet subband coeffi-
cients and since skewness and kurtosis of a distribution could
be analyzed using the first four moments, the first four nor-
malized statistical moments are considered as the features
along with the similarity measure among the medium fre-
quency bands. The experimental results show the efficiency of
our approach when analyzed with various steganography
methods.
The rest of the paper is organized as below. Sec-
tion 2 discusses the proposed methods. Section 3 gives Exper-
imental evaluation of the proposed Methods or Steganalyzer
with the actual results. At the end, section 5 concludes this
paper.
2. PROPOSED METHODS
2.1 METHOD 1
It has been propose novel steganalysis algorithms based on
the effect of data hididng process on image rate-distortion
characteristics. In particular, we make the following assump-
tions/observations about the data embedding process:
International Journal of Computer Applications Technology and Research
Volume 3 Issue 7, 439 - 445, 2014
www.ijcat.com 440
1. Data embedding typically increases the image
entropy: In order to encode the hidden messages, ste-
ganography methods modify parts of the image data.
These modifications typically do not conform with the
existing image statistics and therefore result in a net in-
crease in image entropy.

2. Data embedding methods are limited to the set of
small, imperceptible distortions: Typical steganography
methods make only small modifications to ensure per-
ceptual transparency. Perceptually significant parts of the
image remain intact.
Stochastic embedding does not result in artifacts with known
structures; therefore it is used to develop a generalized meth-
od to detect the changes in the rate- distortion characteristics.
Since real rate-distortion points for signals with unknown
probability distributions such as natural images cannot
reliably calculated, the data rates achieved by lossy compres-
sion scheme. The flowchart of the detection process is seen in
Figure. 1.
An image feature extraction phase is followed by a classifier
that is trained on relevant data sets. As image features we use
the distortion values at different rate points. Mean square error
(MSE), mean absolute error (MAE) and weighted mean
square error (wMSE) are used as distortion metrics. Here,
Bayesian classifier preceded by a KL transform, which reduc-
es the dimensionality of the feature vector.

Image

Feature Vector



Feature Vector




Figure 1. Flow chart describing detection of stochastic em-
bedding

2.1.1 Classifier
Let us denoting w
i
as different classes, where each corre-
sponds to a different stego method. This is assuming that 1 i
M that M such classes exist. This denotes the L dimensional
feature vector by x.


p(x|w
i
) = 1/(2)
1/2
|
i
|
1/2
exp(-1/2( x-
i
)

T

i
-1

( x-
i
)) (1)


where bf
i
= E[x]is the mean value of the w
i
class and
i
is
the covariance matrix defined as

i =
E [(x-
i
) (x-
i
)
T
] (2)

and |
i
| denotes the determinant of
i
and E[.] denotes the
expected value.
It is also define the discriminant function in the logarithmic
form as

g
i
(x) = ln(p(x|w
i
)P(w
i
)) (3)

= -1/2 ( x-
i
)

T

i
-1
( x-
i
)) + ln(P(w
i
)) 1/2ln(2)
1/2ln(|
i
|) (4)

Assuming equiprobable classes and eliminating constant
terms, Eqn.3 can be reduced to

g
i
(x )= (x-
i
)
T

i
-1
(x-
i
)+ln(|
i
|) (5)

i
and
i
are estimated from the training samples for each
class during the training phase.

When the classifier has to operate on a limited number of
training samples with relatively small number of classes, the
high dimensionality of the problem adversely affects the clas-
sifier performance. In particular, the covariance matrix be-
comes nearly singular and classifications results become sen-
sitive to acquisition noise. A method of reducing the dimen-
sionality of the classification problem while keeping the dis-
criminatory power of the feature vector is to project the fea-
ture vector onto a proper subspace.

Let us define the within class and between class scatter matri-
ces, S
w
and S
b
as,


S
w
=
M
i=1
P
i
E[(x-
i
)(x-
i
)
T
] (6)




S
b
=
M
i=1
P
i
(x-
0
)(x-
0
)
T
(7)



where
0
is the global mean vector


0
=
M
i=1
p
i

i
(8)

This is further define the scattering matrix criterion J
3
as

J
3
=trace{ S-
1
w
S
b
} (9)

It can now define a linear projection from the L dimensional
feature space to N dimensional sub-space.

=C
T
x (10)

The optimal projection matix w.r.t. the scattering matrix crite-
rion J
3
is the eigenvectors corresponding to the largest eigen-
values of the system S-
1
w
S
b
. As the individual scatter matri-
ces S
i
, the within class scatter matrix may also be ill condi-
tioned. Therefore, in practice it has been used the pseudo-
inverse of S
w
in the calculations.

2.2 METHOD 2
The objective of the proposed scheme is to select the most
relevant features using statistical characteristics of the sub-
band coefficients, thus reduce the dimensionality of feature
set and increase the accuracy of detection. In this paper, the
first four normalized moments of high frequency, low fre-
quency subband coefficients and structural similarity measure
Lossy Com-
pression
Distortion
Metrics
KLT
Classifier
International Journal of Computer Applications Technology and Research
Volume 3 Issue 7, 439 - 445, 2014
www.ijcat.com 441
of medium frequency sub band coefficients are taken as the
feature set. With these five features, a Non linear Support
Vector Machine is trained for further classification. The block
diagram of the proposed model is given in Figure 2. The fol-
lowing sub sections briefly explain contourlet transformation
and how the feature set is extracted from images.

Figure 2. Block diagram for proposed scheme

2.2.1 Contourlet Transform
The Contourlet transform is a two-dimensional extension of
the wavelet transform proposed by Do and Vetterli [11, 12]
using multiscale and directional filter banks. The contourlet
expansion is composed of basis images oriented at various
directions in multiple scales with flexible aspect ratio that
could effectively capture smooth contours of all images. The
contourlet employs an efficient tree structured implementa-
tion, which is an iterated combination of Laplacian Pyramid
(LP) [13] for capturing the point discontinuities, and the Di-
rectional Filter Bank (DFB) [14] to gather nearby basis func-
tions and link point discontinuities into linear structures. Con-
tourlet transform is more powerful than the wavelet transform
in characterizing images rich of directional details and smooth
contours [15, 16].

Let the image be a real-valued function I(t) defined on the
integer valued Cartesian grid [2
l
, 2
l
]. The Discrete Contourlet
Transform with scale j, direction k and level n of I(t) is de-
fined as follows [17,19]:


where d
k
(m) is the directional coefficient and

where ( t) is the scaling function and f
i
is the spatial domain
function.
Furthermore, the current existing steganalysis algorithms are
limited to the domain of wavelet and DCT transforms. There-
fore, identifying stego (constructed by embedding data into
their contourlet coefficients) and cover image from the image
data set is not easy by these steganalysis algorithms. This fact
motivates us to develop efficient steganalysis algorithm in
contourlet domain. In this paper, contourlet subband based
features are used for steganalysis. Sub-band Coefficient Mod-
elling The coefficients in the produced sub bands of contour-
let transformed image are very appropriate to obtain the tex-
ture feature due to coarse to fine directional details of the
image in these sub-bands. Besides, the distribution of the sub-
bands coefficients is symmetric and unimodal with mean
skewness approximately near to zero, though they have not
exactly Gaussian distribution [18]. These special characteris-
tics of subband coefficients make them suitable for modelling
by Gaussian distribution with density function.


where and are the mean and standard deviation of all the
coefficients of sub-bands.

2.2.2 Feature Extraction
There are various methods in the literature to extract the rele-
vant features of digital images based on different transforms
or filtering techniques. Even though the accuracy of classifiers
is based on the number of suitable features, higher the number
of features slower will be the classification. So identifying a
minimum number of features which can produce efficient
classification is a challenge. In this paper, only 5 features have
been used which is very less compared to the number of fea-
tures used in the existing steganalysis. Contourlet transform is
more sparser than wavelet as the majority of the coefficients
have amplitudes close to zero. Also the moments of contourlet
coefficients are more sensitive to the process of information
hiding. The first four normalized moments of the high fre-
quency and low frequency subband coefficients are more
sensitive to the process of steganography. Since these mo-
ments could be a good measure for skewness and kurtosis due
to information hiding, the first four normalized moments are
extracted as features. Moments are computed as below:

k=1,2,3 and 4.

where X represents the coefficients of contourlet sub bands.
Since these moments alone are not sufficient to detect the
changes in the medium frequency sub-bands, another feature
namely structural similarity measure (SSIM) is also included.
For estimating SSIM, medium frequency band is split into two
equal number of subband groups X and Y respectively. SSIM
includes three parts: Luminance Comparison (LC), Contrast
Comparison (CC) and Structural Comparison (SC) and they
are defined as below [20, 21, 22]:




The similarity of the whole image (I) is

International Journal of Computer Applications Technology and Research
Volume 3 Issue 7, 439 - 445, 2014
www.ijcat.com 442


where n is the number of middle frequency sub bands in the
image. Feature set consists of the first four normalized mo-
ments m
k
(k=1,2,3,4) and the similarity measure SSIM(I).

2.2.3 Classification
A three back propagation Neural Network (NN) is used as a
classifier for identifying stego images as well as images [10].
The power of back propagation is that it enables us to com-
pute an effective error for each hidden unit, and thus derive a
learning rule for the input to hidden weights. Non linear Sup-
port Vector Machine (NSVM) classifier is used for effective
classification of stego images and cover images in this work.
3. EXPERIMENTAL RESULTS
3.1 Method 1
The database collection is done through 108 images in Kodak
PhotoCD format and spans a large variety of image subjects.
In fact, the collection even includes some digitally manipulat-
ed images. It is assume that these images have not been modi-
fied by steganography software and hence represent the set of
cover images.
All images are converted from their original photoCD format
to RGB TIFF images at the base resolution of 768X512 pix-
els. As proposed methods operate on mono-chrome images,
only green channel is used. Furthermore, the image is crop to
remove black boundary regions (30 and 20 pixels).
3.1.1 Detection of stochastic embedding
The process can be modeled by noise addition, without loss of
generality. Although, the method allows for alternative noise
statistics, this uses a Gaussian noise in this experiment. It uses
two different embedding strengths at
2
=3 and
2
=9, corre-
sponding to PSNR of 41dB and 38dB, and embedding rates of
0.84bpp and 0.91bpp, respectively.
For each image, mean square error, mean absolute error, and
weighted mean square error between the image and com-
pressed version are computed. Compression is performed with
JPEG2000 at 95, 90, 85, 80, 70, 60, and 50% of the lossless
rates.
During training, feature vectors are processed to obtain an
optimal projection onto a two dimensional feature space. Then
a Bayesian classifier is trained on the reduced features using
three classes (namely no embedding, low embedding, and
high embedding). In the test phase, the projection matrix ob-
tained in the training phase is used to reduce the feature vector
dimensions. Afterwards, classification is performed using the
previously learned parameters.
In the whole scenario, 9 cover images out of 54 are mis-
labeled as stego-image, while 13 stego-images are mis-labeled
as a cover image. Corresponding false alarm and miss rates
are 16.7% and 12% respectively.
3.2 Method 2
The proposed steganalysis is implemented using MATLAB
7.6.0 with MATLAB scripts. The experiments are conducted
on a personal computer with a 1 GB RAM and P-IV proces-
sor. For training we have used 12,200 images from Computer
Vision image dataset and INRIA image dataset. It contains
5,500 cover images and 6,700 stego images which are gener-
ated by different embedding algorithms like LSB, F5,
Contsteg, and YASS. Washington image dataset [22] is used
for testing the proposed steganalytic method. 100 images are
used to test the proposed scheme, with 60 cover images and
40 stego images.
In order to analyze the proposed method, four typical ste-
ganography methods are used. Table 1 gives a comparison of
the average detection accuracy between NN classification and
non linear support vector classification with same feature set.
From this table, one can see that Non linear Support Vector
Machine classifies stego images and cover images more accu-
rately. Figure 3. Depicts the performance comparison of NN
classifier and NSVM classifier in classifying stego images.

Table 1. Average correct detection rates for natural images
and stego images
Ste-
ganog-
raphy
Methods
Clas-
sifier
Average correct detection rates
Embedding
rates
Different image size
10
0%
50
%
25
%
512
X51
2
256
X25
6
128
X12
8
64
X6
4
LSB NSV
M
.94
5
.9
22
.9
04
.952 .976 .937 .90
2
NN .92
2
.9
12
.8
94
.975 .973 .895 .87
0
F5 NSV
M
.95
0
.9
71
.9
77
.957 .951 .931 .89
4
NN .91
0
.9
69
.8
98
.985 .942 .973 .76
3
ConSteg NSV
M
.92
8
.9
17
.9
08
.905 .957 .903 .82
3
NN .92
1
.8
97
.8
78
.870 .856 .831 .72
3
YASS NSV
M
.96
6
.9
36
.9
14
.941 .987 .912 .85
3
NN .95
6
.9
06
.8
44
.901 .892 .879 .73
3
International Journal of Computer Applications Technology and Research
Volume 3 Issue 7, 439 - 445, 2014
www.ijcat.com 443

Figure 3. Performance comparison of Neural Network
and Non linear Support vector machine based classifiers

The relevancy of the extracted features used in this steganaly-
sis is evaluated using error estimation. Table 2 and
Figure 4 display the sample Median Absolute Error (MAE)
which exhibits a higher error than bias for all the embedding
algorithms. So it is clear that, with this minimum dimensional
feature set, proposed method can able to detect the stego im-
age.

Table 2. Median absolute error and bias for the proposed
method
Algorithm MAE Bias
LSB 5.91X10
-3
-1.70X10
-4

F5 6.63 X10
-3
-3.78X10
-4

YASS 4.19X10
-3
1.87X10
-4

ContSteg 3.25X10
-3
0.58X10
-4



Figure 4. Median Absolute Error (MAE) and Bias of proposed
steganalyzer, with respect to embedding rates
The proposed work is compared with the Contourlet-Based
Steganalysis (CBS) [23] methods and the results show signifi-
cant improvement and they are tabulated in Table 3. The Data
set used in the proposed scheme for comparison is the Wash-
ington dataset which is used in ContSteg [24] and CBS [23].

Table 3. Accuracy of CBS and proposed steganalysis methods
on detection of stego-image produced by ContSteg

Secret Data
Size (bits)
Steganalysis
Method
Average Detec-
tion Accuracy
(%)
5000 CBS 59
Proposed Method 77
10,000 CBS 63
Proposed Method 89
15,000 CBS 68
Proposed Method 93

The correct detection rate is improved in the proposed method
compared to existing steganalysis schemes [23]. Especially
proposed scheme is independent of file formats and image
types. The new method based on statistical steganalysis utiliz-
es fewer features than rest of the methods. Hence, it is fast and
the computational cost of the new method in extracting the
features and detecting the stego image are much less than that
of the methods based on feature extraction.
International Journal of Computer Applications Technology and Research
Volume 3 Issue 7, 439 - 445, 2014
www.ijcat.com 444
4. CONCLUSIONS
In this paper, it has been proposed new steganalysis tech-
niques based on rate-distortion arguments. These techniques
are base on the observation that the steganographic algorithms
invariably disturb the underlying statistics therefore change in
rate-distortion characteristics of the signals. This is demon-
strated the effectiveness of the proposed approach against the
stochastic embedding algorithms with varying degrees of
success.

On the other hand another approach has been proposed a ste-
ganalysis blind detection method based on contourlet trans-
form and non linear support vector machine. This method
extracts the statistical moments and structural similarity of the
contourlet coefficients as the feature set. The performance of
the proposed scheme is illustrated using various testing met-
rics. The average correct detection rate is improved, at the
same time the dimension of the feature set and the average run
time is reduced in this proposed scheme. Furthermore, the
method proposed here is an universal blind scheme, which is
independent of image type and file format.
5. ACKNOWLEDGMENTS
I would like to thanks the experts who are involved in this
area for so long time, as well as their valuable resources and
also like to indebtness towards my colleagues of their valua-
ble help to write this paper.

6. REFERENCES
[1] R.J. A nderson and F.A. Petitcolas, On the limits of ste-
ganography, IEEE Journal of selected Areas in Communica-
tions 16, pp. 474-481, May 1998. Special issue on copyright
& privacy protection.

[2] I.J. Cox, M.L. Miller, and J.A. Bloom, Digital Watermark-
ing, Morgan Kaufmann Publishers, San Francisco, CA, USA,
2002.
S.Katzenbeisser and F.A.P. Petitcolas, eds., Information Hid-
ing: techniques for steganography and digital watermarking,
Artech House, Boston, MA, 2000.

[3] J. Fridrich and M. Goljan, Digital image steganography
using stochastic modulation, in Proc. SPIE: Security and
Watermarking of Multimedia Contents V, E.J. Delp and P.W.
Wong , eds., E123, pp. 191-202, Jan. 2003.

[4] Fridrich J, Goljan M.(2002) Practical: Steganalysis of
digital images- state of the art. In: Proceedings of SPIE, Se-
curity and Watermarking Multimedia Content IV.Vol. 4675.
New York: SPIE, pp 1-13.

[5] McBride B T, Peterson G L, Gustafson S C.(2005) A new
blind method for detecting novel steganography. Digit In-
vest, 2: 50-70.

[6] Johnson.N.F, Jajodia.S.:( 1998) Steganalysis: the investi-
gation of hidden information, In: Proc. IEEE Information
Technology Conference, Syracuse, NY.

[7] Fridrich.J, Goljan.M.: Practical steganalysis of digital
images state of the art, in: Proc. SPICE Photonics West, Elec-
tronic Imaging (2002), Security and watermarking of multi-
media contents, San Jose, CA, vol. 4675, January 2002, pp 1-
13.

[8] Avcibas I, Memon N D, Sankur B.:(2001) Steganalysis of
watermarking techniques using image quality metrics In:
Proceedings of SPIE, Security and Watermarking of Multi-
media Content III, vol. 4314. New York: SPIE, 2001. 523-
531.

[9] Avcibas I, Memon N, Sankur B. (2003): Steganalysis
using image quality metrics. IEEE Trans Image Process, 12:
221- 229.

[10] V. Natarajan and R. Anitha,(2012) Universal Ste-
ganalysis Using Contourlet Transform, Advances in Intel-
ligent and Soft Computing, Springer Verlag, Volume
167/2012, 727-735.

[11] Do, M.N., Vetterli, M.: Contourlets (2002) A direc-
tional multiresolution image representation, Proc. of
IEEE Int. Conf. on Image Process., Piscataway, NJ, pp.
357-360.

[12] Minh N.Do, Martin Vertterli. (2006) The Contourlet
Transform: An Efficient Directional Multiresolution Image
Representation, IEEE Transaction on Image Processing,
vol.14 no.12,pp.2091-2106.

[13] Burt P.J and Adelson E.H.(1983) The Laplacian pyr-
amid as a compact image code, IEEE Trans.Commun,vol
31,no.4, ppl 532-540.

[14] Bamberger R.H and Smith M.J.T.(1992) A filter
bank for the directional decomposition of images: theory
and design, IEEE Trans.Signal Process.
Vol.40,n0.4,pp.882-893.

[15] `Yazdi, M., Mahyari, A.G. (2010) A new 2D fractal
dimension estimation based on contourlet transform for
texture segmentation, The Arabian Journal for Science
and Engineering, vol. 35, No. 13, pp.293-317.

[16] Ali Mosleh, Farzad Zargari, Reza Azizi. (2009) Tex-
ture Image Retrieval Using Contourlet Transfrom, Inter-
national Symposium on Signal, Circuits and Systems.

[17] M.N.Do, and Vetterli.M. (2006) Directional mul-
tiscale modeling of images using contourlet transform,
IEEE Transactions on Image Processing, Vol.15,
no.6,pp.1610-1620.

[18] Chun Ling Yang, Fan Wang, Dongqin Xiao.(2009)
Contourlet Transform based Structural Similarity for
image quality assessment, Intelligent computing and in-
telligent systems.

[19]
http://www.cs.washington.edu/research/imagedatabase.
[20] Kodak PhotoCD images.
ftp://ftp.kodak.com/www/images/pcd.

[21] HP Labs LOCO-I/JPEG-LS Home Page.
http://www.hpl.hp.com/loco/.
International Journal of Computer Applications Technology and Research
Volume 3 Issue 7, 439 - 445, 2014
www.ijcat.com 445

[22] Mathworks (MATLAB) http://www.mathworks.com

[23] Hedieh Sajedi, Mansour Jamzad.(2008) A Steganal-
ysis method based on contourlet transform coefficients,
International Conference of Intelligent Information Hiding
and Multimedia Signal Processing.

[24] Sajedi H., and Jamzad M. ContSteg: Contourlet-
Based Steganography Method, Wireless Sensor Network,
Scientific Research Publishing (SRP) in California
(US),1(3),163-170.