Fusion of Two Typical Quantitative Steganalysis
Based on SVR
Chunfang Yang
1
, Fenlin Liu
1
, Xiangyang Luo
1,2
, Ying Zeng
1
1
Zhengzhou Information Science and Technology Institute, Zhengzhou, China
2
State Key Laboratory of Information Security, Institute of Information Engineering, Chinese Academy of Science,
Beijing, China
Email: chunfangyang@126.com, liufenlin@vip.sina.com, xiangyangluo@126.com, zengying510@yahoo.com.cn
Abstract— For the LSB steganography, a fusion method
is proposed to fuse two typical quantitative steganalysis
methods based on support vector regression (SVR). This
paper first gives some main factors influencing the errors
of structural steganalysis and weighted stego image ste-
ganalysis, viz. the local variance and saturation. Then, the
estimated embedding ratios of above two methods, the local
variance, the histogram of local variance and saturation are
fed to the SVR to train the fusion rule which is used to
fusing these two methods. Experimental results show that
the proposed fusion method can estimate the embedding
ratio with higher accuracy than the individual method.
Index Terms— steganalysis, fusion, embedding ratio, local
variance, support vector regression
I. INTRODUCTION
Steganography is the art of hiding the very pres-
ence of communication by embedding secret message
into innocuous looking covers, such as digital images
[1]. Contrarily, one of the main goals of steganalysis
is to detect the stego object generated by steganogra-
phy. Steganography and steganalysis have been the key
technologies of multimedia information security [2], [3].
Technically, steganography is considered broken when
the mere presence of secret message can be established
[1]. However, in order to extract the secret message, the
investigators need more details of stego object, such as
the length of secret message and the modification ratio of
samples [4], [5]. Steganalysis that can estimate the length
of secret message or the modification ratio of samples is
called as quantitative steganalysis [6]. The estimation of
the secret message’s length or the modification ratio can
not only be used to distinguish the stego objects, but also
help to the estimation of stego positions and the search
of stego key [4], [5], [7].
Nowadays, there have been many quantitative ste-
ganalysis methods for different steganography methods.
This paper is based on “Error Correction of Sample Pair Analysis
Based on Support Vector Regression,” by C. Yang, F. Liu, and X. Luo,
which appeared in the Proceedings of the 3rd International Conference
on Multimedia Information Networking and Security (MINES), Shang-
hai, China, Nov. 2011.
c
2011 IEEE.
This work was supported in part by the National Natural Science
Foundation of China (Grant Nos. 60970141, 60902102, 61170032 and
61272489), the Fund of Innovation Scientists and Technicians Out-
standing Talents of Henan Province (Grant No. 094200510008), and
the Doctoral Dissertation Innovation Fund of Zhengzhou Information
Science and Technology Institute (Grant No. BSLWCX201002).
For example, for the popular least significant bit (LS-
B) replacement, researchers have proposed many corre-
sponding quantitative steganalysis methods, such as RS
(regular and singular groups) method [8], DIH (difference
image histogram) method [9], SPA (sample pair analysis)
method [10], WS (weighted stego image) method [11]
and some improved variant of them. For multiple least
significant bit planes replacement, LSB matching, ±K,
stochastic modulation, JSteg, F5, OutGuess and so on,
some relevant quantitative steganalysis methods also have
been designed. And some researcher presented to de-
sign quantitative steganalyzers from the features in blind
steganalysis [12], [13]. Additionally, some researches
on error analysis of quantitative steganalysis have been
published. In 2005, Rainer B
¨
ohme proposed multiple
regression models as a method for quantitative evaluation
of the accuracy in quantitative steganalysis with respect
to various moderating factors [14]. In 2006, on the basis
of the results in [14], Rainer B
¨
ohme and Andrew D. Ker
presented a rationale for a two-factor model for sources of
error in quantitative steanalysis, and analyzed the effects
of some factors on the two error components [15]. In
2007, Andrew D. Ker derived the error distribution of
the least squares steganalysis for cover images [16]. In
the past 2011, the authors of this paper used the support
vector regression to learn the prediction function of the
estimation error of the SPA method [17].
The above works drive the further researches on quan-
titative steganalysis. But, we all know that for different
images, different quantitative steganalysis methods will
obtain different results. We know that fusion of multiple
results or features would generate more accurate results
[18], [19]. Therefore, we try to fuse the existing quan-
titative steganalysis methods to estimate the embedding
ratio more accurately. In this paper, we consider the main
factors influencing the estimation errors of the structural
steganalysis and weighted stego image steganalysis, and
use the support vector regression to fuse these two typical
quantitative steganalysis. The experimental results verify
the validity of the proposed fusion method.
JOURNAL OF SOFTWARE, VOL. 8, NO. 3, MARCH 2013
doi:10.4304/jsw.8.3.731-736