718 C. Yang et al.
out in the other-LSB planes ([16] has given the reason for two
least significant bits (2LSB) embedding). Some relevant works
have reported that secret message can not only be embedded into
LSB plane, but also be embedded into multiple least significant
bit (MLSB) planes in [17, 18].
For the replacement of MLSB planes, there have been
some steganalysis methods proposed for two distinct MLSB
steganography paradigms—TMLSB steganography (‘T’means
this embedding paradigm is typical.) and IMLSB steganography
(‘I’ means the messages are embedded into each bit plane
independently.), which embed equal ratio of message bits into
MLSB planes. And these steganalysis methods for MLSB
steganography can mainly be classified as two categories:
structural steganalysis [16, 19–21] and WS steganalysis [22–
24]. However, the embedders likely embed message into
different bit planes with different ratios independently, and
this category of MLSB steganography is called as ID-MLSB
steganography (where ‘ID’ denotes that the message lengths
may be embedded into different bit planes independently
with different ratios). Furthermore, some other steganography
methods can be classified into the ID-MLSB steganography,
such as the adaptive steganography in [18]. Therefore, the
quantitative steganalysis of ID-MLSB steganography should be
very important to the forensics steganalysis. For this category
of MLSB steganography, the steganalysis methods above all
fail to estimate the embedding ratio in each bit plane. And so
far, only the literature [25] presented a method to estimate the
embedding ratio in each bit plane based on the SPA model
for MLSB replacement steganography (MSPA model) in [21].
The obtained method was called IDMSPA method, and would
be called ID3SPA method when being used to estimate the
embedding ratios in the three LSB planes. In [26], the IDMSPA
method also has been adapted for the quantitative steganalysis
of a category of adaptive steganography that embeds message
bits into different bit planes with different ratios based on the
block’s noise level of MLSB planes.
Although the methods in [25, 26] can estimate the embedding
ratio in each bit plane of MLSB planes, they all own the inherent
defect—their performances degrade rapidly with the increase of
embedding ratio in any bit plane, and when the embedding ratio
in any stego bit plane is middle or large, they will fail to estimate
the embedding ratios accurately. Contrarily, the WS steganalysis
usually performs more stably and owns good performance for
the case of a high embedding ratio [21]. Therefore, we try
to estimate the embedding ratios of ID-MLSB steganography
based on WS more stably and accurately than the existing
methods. The main results of this paper are as follows.
(i) A new WS with multiple weight parameters is defined
for ID-MLSB steganography.
(ii) A theorem is proved that when the squared Euclidean
distance between the WS and the cover image is
minimal, the obtained weight parameters are equal to
the embedding ratios in the MLSB planes.
(iii) Based on this theorem and an estimated cover
image, a quantitative steganalysis method is proposed
for estimating the ratio of message hidden into
each bit plane. Experimental results show that
the new steganalysis method performs more stably
with the change of embedding ratio than typical
structural steganalysis, and outperforms the structural
steganalysis on estimation accuracy when the
embedding ratio in any bit plane is middling or large.
The remainder of this paper is organized as follows. Section 2
introduces the related works, including the WS steganalysis for
LSB steganography and the ID-MLSB steganography briefly.
Section 3 describes the proposed weighted stego steganalysis
for ID-MLSB steganography. Finally, a series of experimental
results are given in Section 4. The paper closes in Section 5 with
the conclusions.
2. RELATED WORKS
2.1. WS steganalysis for LSB steganography
The WS steganalysis was proposed by Fridrich and Goljan
[12], and occupies an unusual position in the steganalysis of
LSB replacement. Unlike the structural detectors, this method
does not use the pixel group analysis on which almost every
other reasonably accurate detector relies, but has fairly good
accuracy; moreover, it retains its estimation accuracy when
embedding changes are not distributed evenly over the cover
[15]. In the following, the basic principle of WS steganalysis
for LSB replacement will be introduced briefly.
In [12], let X ={x
i
}
n
i=1
be a set of integers in the range
[0, 255] representing a grayscale cover image whose size is
n = M × N. The value of x
i
after flipping its LSB will be
denoted as ¯x
i
, viz. ¯x
i
= x
i
+ 1 − 2(x
i
mod 2). Let S ={s
i
}
n
i=1
denote the stego-image after embedding pn (0 ≤ p ≤ 1) bits
using LSB replacement in pn pixels randomly selected from
the cover image X. Fridrich and Goljan defined the WS with
weight parameter q as follows:
S
(q)
={s
(q)
i
},
s
(q)
i
= s
i
+
q
2
(¯s
i
− s
i
),
(1)
where 0 ≤ q ≤ 1, i = 1, 2,...,n. Then, Fridrich and Goljan
proved that the weight parameter q and the embedding ratio p
satisfy the following relationship:
p = arg min
q
1
n
n
i=1
(s
(q)
i
− x
i
)
2
, (2)
which shows that S
(p)
is the closest WS to the cover image in the
least square sense among all WSs S
(q)
for 0 ≤ q ≤ 1. Based on
this, Fridrich and Goljan formulated the procedure of estimating
the unknown embedding ratio p from the stego-image as a
The Computer Journal, Vol. 55 No. 6, 2012
at National University of Defense Technology on December 17, 2012http://comjnl.oxfordjournals.org/Downloaded from
评论0
最新资源