Professional Documents
Culture Documents
IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 8, NO. 1, JANUARY 2013
I. INTRODUCTION
least significant bits) planes. In the past several years, the quantitative steganalysis for embedding in MLSB has attracted many
attentions, and some relevant quantitative steganalysis methods
have been proposed for the following two categories of typical
MLSB steganography (for description, is used to denote the
number of bit planes containing the embedded message.):
1) Randomly select samples and replace their LSBs with the
message fragments with size of . This category is called
as TMLSB steganography (T is the shortening of typ.
ical), such as T2LSB steganography for
2) Randomly select bits from the LSB planes and replace
them with the message bits. This category is called as
IMLSB steganography (I signifies the independence
of the effects on different bit planes), such as I2LSB
steganography for
.
Currently, the quantitative steganalysis methods for the above
two categories of typical MLSB steganography can be mainly
classified into two categories: 1) the weighted stego image steganalysis; and 2) the structural steganalysis.
The weighted stego image steganalysis was first proposed by
Fridrich et al. for LSB replacement [3]. In 2005, Yu et al. [6] extended it into the quantitative steganalysis for TMLSB steganography. Subsequently, Yu et al. [7] further improved the extended
result in [6]. The weighted stego image steganalysis is simple,
and outperforms the others for the high embedding ratio, especially for the embedding ratio closing to 1. However, the weighted
stego image steganalysis usually estimates the low embedding
ratio with larger error than the structural steganalysis.
The structural steganalysis was also first proposed for LSB replacement [1], [2]. In 2007, Ker and Luo et al. [8], [9] introduced
it into the quantitative steganalysis for 2LSB replacement. And
in 2008, the structural steganalysis method for LSB replacement
in [2] was extended to the quantitative steganalysis for TMLSB
and IMLSB steganography by Yang et al. [10]. The structural
steganalysis degrades rapidly when the embedding ratio is large,
especially close to 1. But the structural steganalysis usually estimates the low embedding ratio with less error.
Recently, Pevn et al. [11] proposed a general method for
constructing quantitative steganalyzers from features used in
blind steganalysis, and constructed the quantitative steganalyzer with SVR (Support Vector Regression) for steganography
in spatial domain which also works for MLSB steganography.
However, the constructed quantitative steganalyzer still offer
lower accuracy for bit plane replacement steganography.
Because the stego image with a low embedding ratio is usually detected more difficultly, the steganalysis of the stego image
with a low embedding ratio has been one of the important and
YANG et al.: PIXEL GROUP TRACE MODEL-BASED QUANTITATIVE STEGANALYSIS FOR MLSB STEGANOGRAPHY
difficult issues of steganalysis. Although the structural steganalysis can estimate the low embedding ratio with less error than
others, the existing structural steganalysis methods for MLSB
steganography only utilize the correlation between two adjacent samples. For LSB replacement, the structural steganalysis
methods utilizing the correlation among more than two adjacent
pixels can usually estimate the low embedding ratio with higher
precision [4], [5], [12]. Intuitively, one should also be able to
improve the estimation accuracy for the low embedding ratio of
MLSB steganography by utilizing the correlation among more
adjacent pixels.
Therefore, this paper mainly discusses how to utilize the
correlation among more than two pixels to estimate the embedding ratio of image MLSB steganography. It is constructed
as follows. In Section II, the pixel group trace model is built.
And some statistical characteristics of images are described
in Section III. Then, the quantitative steganalysis methods for
TMLSB and IMLSB steganography are given in Section IV. In
Section V, the statistical characteristics of images are analyzed
by experiments, and the quantitative steganalysis methods are
improved. In Section VI, a series of experimental results are
given. The paper closes in Section VII with a conclusion. In
order not to obscure the main ideas, some necessary but lengthy
details of derivation are put in two Appendices.
II. PIXEL GROUP TRACE MODEL FOR ANALYZING MSLB
STEGANOGRAPHY
Throughout this paper, an image is represented by a succession of pixels
, where the
index denotes the location of a pixel in the image, denotes
the number of pixels in the image, and denotes the number
of bits used to store a pixel. Let
and
denote the high
bits and the least significant bits of the
value , respectively. A -tuple
is named as a pixel group composed of adjacent pixels
. After MLSB steganography, a pixels
LSBs are possibly replaced by secret messages, but its high
bits are impossibly altered. Thus, a trace set of pixel
groups is defined based on the difference between the high
bits of
and
as follows:
(1)
where
. The trace sets are closed under MLSB steganography. In order to trace the influence of MLSB steganography
to the structures of pixel groups values, each trace set
is
partitioned into
trace subsets based on the LSBs of the
pixels in the pixel groups as follows:
(2)
where
.
When embedding messages into an image by MLSB
steganography, if the bits in the embedding positions are different from the message bits, they should be modified, viz.
the bits 1 (and 0) should be flipped to be 0 (and 1); otherwise,
217
218
IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 8, NO. 1, JANUARY 2013
(9)
(12)
However, when applying the nonnegative mask to the pixel
group with noise level , if some pixels LSBs have been
flipped by MLSB steganography, they would be reflipped to the
natural versions. Because the pixel group noise in the natural
image is usually lower than it in the stego image, with the increase of the embedding ratio, the pixel group noise level will
increase with lower probability, and decrease with higher probability. When applying the nonpositive mask
to the pixel
group with noise level , if some pixels LSBs have been
flipped by MLSB steganography, they would be shifted to the
values farther from the natural versions, viz. the noise would
be enhanced. Then with the increase of the embedding ratio,
the pixel group noise level will increase with higher probability, and decrease with lower probability. Therefore, with the
increase of the embedding ratio of MLSB steganography, the
statistic
will decrease, and the statistic
will increase, viz. when
,
(13)
(10)
are the integers in
and not all
where
of them are 0.
For the pixel groups whose noise levels are , based on the
difference between the noise levels of them before and after
applying the nonnegative mask to them, they can be divided
into regular pixel group set
, singular pixel group set
and unchanged pixel group set
as follows:
1) If
, then
;
2) If
, then
;
3) If
, then
.
Similarly, they can also be divided into regular pixel group set
, singular pixel group set
and unchanged pixel
group set
with the nonpositive mask
.
(14)
where
(15)
(16)
In [1], Fridrich et al. have explained the reasons of (13) and
(14) for LSB steganography. In Section V, the rationality of
(11)(14) will be further explained by experiments.
For the pixel groups in the trace subset
, from the definition of the noise level function in (6) and Proposition 1, one can
obtain the following relationships among
and the noise
levels of the pixel groups before or after applying the masks
or
to them, viz (17), shown at the bottom of the page.
(17)
YANG et al.: PIXEL GROUP TRACE MODEL-BASED QUANTITATIVE STEGANALYSIS FOR MLSB STEGANOGRAPHY
(18)
where
219
(19)
is used to judge whether applying the mask
will increase, decrease or maintain the noise levels of the pixel
groups in
which originally equal to . Applying the formulas (18) to (15) will yield the following function about the
cardinalities of the trace subsets:
(20)
where
.
Similarly, the function (16) can also be rewritten as follows:
where
denotes the -fold Kronecker
tensor product of the transition probability submatrix , and
..
.
..
.
..
..
.
(24)
(21)
For cover images, the statistical characteristics (11) and (12)
denote that the values of (20) and (21) are approximately the
constant 0. For stego images, the statistical characteristics (13)
and (14) denote that with the increase of the embedding ratio,
(20) and (21) decreases and increases strictly, respectively.
Therefore, it is possible to apply the functions (20) and (21) to
the quantitative steganalysis for MLSB steganography.
where
(26)
and
(27)
The right item in (27) equals to the dot product between the
-th row vector of the matrix
and the
vector
. Thus, the formula (23) can be derived.
One can subtract the first row from the 2nd, 3rd, , and -th
rows in the submatrix , and add the obtained 2nd, 3rd, , and
220
IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 8, NO. 1, JANUARY 2013
where
(34)
..
.
..
.
..
(28)
..
.
(35)
and
equivwhich owns an eigenvalue
alent eigenvalues
. It implies that
is only singular when
, viz.
. Thus, from the property of
the Kronecker tensor product, when
, the inverse of the
matrix
is
L
L
(36)
L
..
.
..
.
..
..
.
(29)
For TMLSB steganography, one can directly count the cardinality of each trace subset in the stego image, and estimate
the statistics (20) and (21) of the cover image, viz.
and
, by replacing the item
in (5) with
. Then, the following two estimation equations of
the embedding ratio can be derived based on the statistical characteristics of cover images given by (11) and (12):
(30)
(31)
where
denotes the
-th
row vector of the matrix
. Then (30) and (31) can
be translated into the following general equations respectively
(The details are given in Appendix A):
L
is a lattice in the -dimensional integer
space.
The estimation equations (32) and (33) are both polynomial
equations of degree which could lead to up to roots for the
embedding ratio . The Proposition 2 describes how to select the
appropriate real root as the estimated embedding ratio under
the condition that the statistical characteristics given by (11) and
(12) are exactly satisfied, viz. two sides of the formulas (11) and
(12) are exactly equivalent.
Proposition 2: For the cover image with the statistical characteristics given by (11) and (12), when embedding random
messages into its LSB planes with the embedding ratio
by TMLSB steganography, should be the smallest nonnegative real root of (32) and (33).
Proof: From Theorem 1 and the derivation of (32) and (33),
it can be seen that the embedding ratio
must be one of the
real roots of (32) and (33). Then this proposition is proven by
reduction to absurdity.
1) If the embedding ratio
is not the smallest nonnegative
real root of the estimation equations (32) and (33), then the
estimation equations (32) and (33) would have another real
root
, which should satisfy (32) and (33). From
(5) and (29), one can obtain
,
where
would satisfy the statistical characteristics of
cover images given by (11) and (12).
2) Because
, the following equation
can be obtained:
(37)
The item
as
(32)
..
.
(33)
..
.
..
..
.
(38)
YANG et al.: PIXEL GROUP TRACE MODEL-BASED QUANTITATIVE STEGANALYSIS FOR MLSB STEGANOGRAPHY
221
where
where
.
Because
and
(43)
(39)
rewritten as
is approximately equivalent
which demonstrates that
to the vector of the cardinalities of the trace subsets in the
stego image with the embedding ratio
.
From the statistical characteristics of stego images given
by (13) and (14), the embedding ratio
would cause
and
, viz.
would not satisfy
the statistical characteristics of cover images given by (11) and
(12). This is in conflict with the result in 1) which is derived
from the precondition that
is not the smallest nonnegative
real root of (32) and (33).
Therefore, the embedding ratio should be the smallest nonnegative real root of (32) and (33).
(44)
Because the right item in (44) equals to the dot product between
the
-th row vector of the matrix
and the
vector
, the formula (41) can be derived.
Because
is not singular except for
, when
owns the following inverse matrix:
(45)
(46)
(41)
where
(47)
(48)
(49)
222
IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 8, NO. 1, JANUARY 2013
where
(50)
YANG et al.: PIXEL GROUP TRACE MODEL-BASED QUANTITATIVE STEGANALYSIS FOR MLSB STEGANOGRAPHY
223
(55)
Fig. 1. Regression results of the models (53) and (54) for different noise levels
: (a) and (c) the regression coefficients; (b) and (d) the stanand
dard estimation errors.
(56)
where
L
The regression models (53) and (54) are fitted to the data generated in the following experimental setup: 1) download 3000
images from http://photogallery.nrcs.usda.gov with originally
very high resolution in format tiff; 2) divide 3000 images
into three parts averagely, and convert these three parts of images to gray-scale bmp images with size of about 512 366,
768 549 and 1024 731, respectively (The tool used is Advanced Batch Converter 3.8.20 with the typical interpolation
filterbilinear.).
For the case of
and
, the mask
is set as (1,
2, 1) to regress the models (53) and (54). In Fig. 1, the solid
circles denote the regression results for the even noise levels,
and the hollow circles denote the regression results for the odd
noise levels. Fig. 1(a) and (c) show that the regression coefficients of the models (53) and (54) are all very close to 1. Especially, when the noise levels are less than 30, most of the regression coefficients are 1. Fig. 1(b) and (d) show that most of
the standard errors of the models (53) and (54) are less than
0.05. Especially, when the noise level is odd and larger than 5,
the standard error is significantly less than that of the neighbor
even noise level. Additionally, the confidences of the significant
tests for the models (53) and (54) are all 0.000 which is certainly
less than the typical confidence level 0.05. These demonstrate
that when
, the models (53) and (54) can be used
to well describe the relationship between
and
,
and the relationship between
and
, respectively.
For the case of
and
, the mask
is set as (2,
5, 2) to regress the models (53) and (54). One can obtain the
similar results to that of the case of
, though the models
with odd noise levels are not always fitted better than the models
with even noise levels. (For the reason of space, the detailed
experimental results are not shown.)
The consistency of the above regression results with (11) and
(12) nicely explains the rationality of the statistical characteristics given by (11) and (12).
L
.
The above formulas are related to the cardinalities of the trace
subsets in the cover images seriously. They are too complicated
to theoretically infer their monotony. Therefore, we count the
cardinalities of the trace subsets in the typical image lena.bmp
and above 3000 NRCS images, and apply them to (55) and (56)
to evaluate their monotony. Fig. 2(a) takes the typical image
lena.bmp as an example to show the relationship between the
embedding ratio and the values of
and
with
. It can be seen that with the increase of the embedding ratio,
and
would decrease and increase
strictly, respectively. This is consistent with the statistical characteristics given by (13) and (14). For the 3000 NRCS images,
whether
and
are consistent with the (13)
and (14) is judged based on the signs of the derivatives of (55)
and (56). Fig. 2(b) and (c) show that except that the noise level
is
and
decreases strictly and
increases strictly for most images when
and the embedding ratio is in [0, 1], though the ratio of strictly decreasing
and the ratio of strictly increasing
decrease
with the increase of the noise level. Additionally, for T3LSB
steganography, the similar results can also be obtained when
. These results demonstrate that most of the stego
images of TMLSB steganography satisfy the statistical characteristics given by (13) and (14).
For IMLSB steganography, From Theorem 2, both of
and
can be denoted as the following functions (The details are similar to Appendix B):
(57)
224
Fig. 2. When
and
images where
IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 8, NO. 1, JANUARY 2013
and the embedding ratio of T2LSB steganography is in [0, 1], (a) the relationship between the embedding ratio and the values of
for the image lena.bmp; (b) the ratio of the number of images where
decreases strictly; and (c) the ratio of the number of
increases strictly.
(58)
For I2LSB and I3LSB steganography, the experimental results
in the typical image lena.bmp and the 3000 NRCS images are
still consistent with the statistical characteristics given by (13)
and (14), and not supplied for the reason of space.
C. Improving Quantitative Steganalysis of TMLSB and IMLSB
Steganography
From the source of between-image error, the obvious idea to reduce it is to select the proper noise level which minimizes the absolute value of
or
and satisfies the statistical
characteristics given by (13) and (14). However, different images
usually own different proper noise levels. If one selects a certain
noise level , especially if the selected certain noise level is large,
the number of the pixel groups utilized will be too few to respond
to the embedded messages enough sensitively. In order to make
the most of the pixel groups and eliminate the outlier cases of (11)
and (12) as possible as we can, we select multiple different noise
levels which not only cause
or
very small,
but also well satisfy the statistical characteristics given by (13)
or (14), then combine the corresponding formulas (11) and (12)
to obtain the following statistical characteristic:
(59)
(and
) is the set of noise levels which not only cause
(and
) very small, but also well satisfy the
statistical characteristic given by (13) (and (14)).
Based on the statistical characteristic given by (59), the
following improved estimation equations can be obtained for
TMLSB and IMLSB steganography, respectively:
where
(60)
(61)
It can be seen that the accuracies of the improved estimation
equations (60) and (61) are seriously interrelated by the selected
noise level sets
and
. Ideally, one can apply various possible noise level sets to the (60) and (61), then select the proper
noise level sets
and
. Obviously, this would make the
selection of the proper
and
too complicated to be feasible. Therefore, the noise level sets
and
will be recommended by the following steps:
1) Select the initial noise level sets
and
based on the
regression coefficients. In our experiments, the noise levels
with which the models (53) (and (54)) own the regression
coefficients in [0.999, 1.001] are put into
(and
).
2) Eliminate the noise levels which do not satisfy the statistical characteristic (13) (and (14)) from
(and
). In
experiments, if the ratio of strictly decreasing
(and
strictly increasing
) is less than 95%, the corre(and
).
sponding value of would be eliminated from
3) Reorder the residual elements in
and
based on the
standard estimation errors of the regression models;
4) Select the first
noise levels obtained
by 3). If the selected noise levels belong to
before reordering, add them to the possible proper set
; oth. For
erwise, add them to the possible proper set
and
to (60) or
each value of num, apply
(61) to estimate the embedding ratios. When the interquartile range of the estimation errors is minimum, the corresponding noise level sets
and
are selected
as the recommended noise level sets
and
.
Following above steps, the following recommended noise
level sets are obtained from the NRCS image set in Section V.A
for T2LSB, I2LSB, T3LSB and I3LSB steganography:
YANG et al.: PIXEL GROUP TRACE MODEL-BASED QUANTITATIVE STEGANALYSIS FOR MLSB STEGANOGRAPHY
225
Fig. 3. Performance comparison for T2LSB steganography: (a), (b), and (c) estimation errors in NRCS images; (d), (e), and (f) estimation errors in BOSS images.
Fig. 4. ROC curves for the embedding ratio 0.02 in (a) NRCS images and
(b) BOSS images.
Firstly, for T2LSB steganography, Fig. 3(a), (b), (d) and (e)
show that when the embedding ratios is smaller than 0.6, the
estimation error IQR of the T2Triple method is significantly
smaller than that of the T2SPA [10], 2Couples [8], 2-WS [7]
and SVR [11] methods (Fig. 3(b) and (e)), and the estimation
error median of the T2Triple method is approximately equivalent to that of the T2SPA method which is the best one among
the T2SPA, 2Couples, 2-WS and SVR methods (Fig. 3(a) and
(d)). Especially, when the embedding ratio is not larger than 0.3,
the estimation error IQR of the T2Triple method is smaller than
that of the T2SPA method by about 29% (0.004) for the NRCS
images (Fig. 3(b)), and by about 45% (0.007) for the BOSS images (Fig. 3(e)). The small biases and less IQRs make the mean
absolute errors of the T2Triple method also significantly smaller
than that of others (Fig. 3(c) and (f)). Moreover, the ROC curves
for the embedding ratio 0.02 show that the T2Triple method
can distinguish the stego images with low embedding ratio of
T2LSB steganography more reliably than others (Fig. 4).
226
IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 8, NO. 1, JANUARY 2013
Fig. 6. Performance comparison for T3LSB steganography [(a), (c), and (e)]
and I3LSB steganography [(b), (d), and (f)] in BOSS images: the scales of the
.
y-axis should be multiplied by
APPENDIX A
DERIVATION OF ESTIMATION EQUATIONS (32) AND (33)
Expand the first summation item in (30) by the components
of the vectors
and as follows:
VII. CONCLUSION
Since more than two adjacent pixels still own strong correlation, the quantitative steganalysis for MLSB steganography
was discussed based on a pixel group trace model presented in
this paper. And the following results have been obtained: 1) a
pixel group trace model was built, which considers the relationship among more than two adjacent pixels; 2) some statistical
(62)
YANG et al.: PIXEL GROUP TRACE MODEL-BASED QUANTITATIVE STEGANALYSIS FOR MLSB STEGANOGRAPHY
227
L
(71)
(63)
where L
. Because
can further be translated into
, (63)
L
Because
(64)
Similarly, the residual three summation items in (30) and (31)
can also be reduced. Then, the estimation equations (30) and
(31) would be translated into
(72)
Similarly, the residual three summation items in (46) and (47)
can also be reduced. Then the estimation equations (46) and (47)
would be translated into
(65)
(73)
(66)
From the binomial theorem, it can be obtained that
(67)
Applying (35) and (67) to (65) and (66) yields
(74)
From the binomial theorem, it can be obtained that
(68)
(69)
(75)
Applying (50) and (75) to (73) and (74) yields
Let
and apply it to (68) and (69). Then, the
general estimation equations (32) and (33) can be obtained.
(76)
(77)
APPENDIX B
DERIVATION OF ESTIMATION EQUATIONS (48) AND (49)
Expand the first summation item in (46) by the components
and as follows:
of the vectors
(70)
Let
and apply it to (76) and (77). Then, the general
estimation equations (48) and (49) can be obtained.
REFERENCES
[1] J. Fridrich, M. Goljan, and R. Du, Detecting LSB steganography in
color and gray-scale images, IEEE Multimedia, Special Issue on Security, vol. 8, no. 4, pp. 2228, Oct./Dec. 2001.
[2] S. Dumitrescu, X. Wu, and Z. Wang, Detection of LSB steganography
via sample pair analysis, IEEE Trans. Signal Process., vol. 51, no. 7,
pp. 19952007, Jul. 2003.
[3] J. Fridrich and M. Goljan, On estimation of secret message length
in LSB steganography in spatial domain, in Proc. SPIE, Security,
Steganography, and Watermarking of Multimedia Contents VI, E. J.
Delp, III and P. W. Wong, Eds., San Jose, CA, 2004, vol. 5306, pp.
2334.
228
IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 8, NO. 1, JANUARY 2013