You are on page 1of 11

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO.

4, AUGUST 2012

1181

New Channel Selection Rule for JPEG Steganography


Fangjun Huang, Member, IEEE, Jiwu Huang, Senior Member, IEEE, and Yun-Qing Shi, Fellow, IEEE

AbstractIn this paper, we present a new channel selection


rule for joint photographic experts group (JPEG) steganography,
which can be utilized to find the discrete cosine transform (DCT)
coefficients that may introduce minimal detectable distortion for
data hiding. Three factors are considered in our proposed channel
selection rule, i.e., the perturbation error (PE), the quantization
step (QS), and the magnitude of quantized DCT coefficient to be
modified (MQ). Experimental results demonstrate that higher
security performance can be obtained in JPEG steganography via
our new channel selection rule.
Index TermsChannel selection, JPEG, magnitude, perturbation error (PE), quantization step (QS).

I. INTRODUCTION

TEGANOGRAPHY is a secret communication approach


that can transmit information without arousing suspicion
of the existence of the secret communication. The carrier of
steganography can be various kinds of digital media such as
image, audio, and video, etc. Due to the common use of JPEG
images, JPEG steganography has attracted much attention by
the researchers in this area.
There are two most important approaches that have been utilized to improve the security performance of JPEG steganography. The first one is matrix embedding [1], [2]. Through it,
less alternation needs to be made to the cover image while embedding the same amount of information bits. The second one is
channel selection, i.e., trying to find those coefficients that may
introduce minimal detectable distortion for data hiding [3].
In [3], Fridrich et al. presented the perturbed quantization
(PQ) steganography with wet paper codes, in which the secret
message bits are embedded while the cover object is processed
with an information-reducing operation. They pointed out that
if the senders could confine the embedding changes to those
channels of the processed cover object whose values are the
most uncertain, higher security performance will be obtained.
Manuscript received November 09, 2011; revised February 13, 2012; accepted April 16, 2012. Date of publication May 07, 2012; date of current version July 09, 2012. This work was supported in part by 973 Program of China
(2011CB302204), in part by the National Natural Science Foundation of China
(61173147, U1135001), in part by the Fundamental Research Funds for the Central Universities, and in part by the State Key Laboratory of Information Security (Institute of Software, Chinese Academy of Sciences). The associate editor
coordinating the review of this manuscript and approving it for publication was
Dr. Mauro Barni.
F. Huang and J. Huang are with the School of Information Science and Technology, Sun Yat-Sen University, GD 510006, China (e-mail: huangfj@mail.
sysu.edu.cn; isshjw@mail.sysu.edu.cn).
Y.-Q. Shi is with the Department of Electrical and Computer Engineering,
New Jersey Institute of Technology, Newark, NJ 07102 USA (e-mail: shi@njit.
edu).
Color versions of some of the figures in this paper are available online at
http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TIFS.2012.2198213

Note that the application of PQ channel selection rule has covered a wide range of information-reducing operations such as
lossy compression, downsampling and A/D conversion, etc. As
an example, Fridrich et al. [3] have implemented the PQ channel
selection rule in JPEG steganography, which is based on the assumption that the raw, uncompressed image is available to the
sender. According to the PQ channel selection rule, the secret
message bits are embedded into those changeable coefficients
whose rounding errors are close to 0.5. Since the unrounded coefficients (i.e., the DCT coefficients that have been divided by
the quantization steps and not yet rounded) can be rounded either up or down without a noticeable difference, embedding the
secret message bits into these coefficients will be more difficult to be perceived. For example, suppose that the unrounded
( is a small posDCT coefficient is with the value of
itive value), this coefficient will be rounded to 3 in standard
, which
JPEG compression, and the rounding error is
is close to 0.5. If the secret bit 0 is embedded into this coefficient, the quantized DCT (i.e., the DCT coefficients having
been divided by the quantization steps and rounded) coefficient
will be changed to 2 if only the least significant bit (LSB) of
the coefficient is modified. The error introduced by the embedding is
. As seen, the embedding error introduced by data
hiding is close to the rounding error in standard JPEG compression. It will not introduce a noticeable distortion to the carrier
JPEG image and hence can improve the security performance
of the steganographic schemes. For simplicity, the difference
between the rounding error and embedding error is called perturbation error (PE) in this paper. As seen, the philosophy behind PQ channel selection rule is to embed the secret message
bits into those changeable coefficients whose rounding errors
are close to 0.5, i.e., the perturbation errors are close to zero.
However, since in standard JPEG compression, the number of
generated DCT coefficients whose rounding errors are close to
0.5 is rather small, the application of PQ channel selection rule
will be restricted by the few changeable DCT coefficients. In
[3], Fridrich et al. have exemplified how to apply the PQ channel
selection rule in JPEG steganographic schemes with wet paper
codes. They take a singly compressed JPEG file as the cover
image and embed the secret message bits when recompressing
the JPEG image for the second time with a quality factor lower
than the previous one. While the quantization steps in a singly
compressed image and the doubly compressed image satisfy a
certain linear relationship, a lot of changeable DCT coefficients,
i.e., the coefficients with rounding error of 0.5, will be generated. Hence, the PQ channel selection rule can be easily applied
in the double JPEG compressing process [3].
In [4], Fridrich et al. presented two other updated versions
of PQ: texture-adaptive PQ (PQt) and energy-adaptive PQ
(PQe). Through considering the local block characteristics
such as texture complexity and energy capacity, better security

1556-6013/$31.00 2012 IEEE

1182

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 4, AUGUST 2012

Fig. 1. Block diagram of JPEG compression process.

performance can be obtained via forming two corresponding


new channel selection rules. However, as that in PQ, since
the number of changeable DCT coefficients in standard JPEG
compression is rather limited, these two new channel selection rules still have some limitations when applied to JPEG
steganography directly.
In [5], Kim et al. developed the PQ channel selection rule.
Different from the original PQ channel selection rule [3],
[4], all the nonzero quantized alternating current (ac) DCT
coefficients can be regarded as changeable coefficients. Those
changeable coefficients are divided into blocks, and in each
block one or more coefficients which may introduce minimal
PE are selected for data hiding. For simplicity, we call this
developed PQ channel selection rule as D-PQ. In [5], Kim
et al. provided a simple and practical scheme to apply the
D-PQ channel selection rule in JPEG steganography, which is
based on the modified binary Hamming codes and allows more
than one embedding change in each coefficient block. This
new embedding algorithm is called modified matrix encoding
(MME). According to the number of allowable changing
bits in each coefficient block, the MME schemes are called
MME2, MME3, etc. Recent research has demonstrated that the
D-PQ channel selection rule can also be applied in other JPEG
steganographic schemes [6][8] in which the Bose, Chaudhuri
and Hocquenghem (BCH) codes or Reed-Solomon codes [9]
are utilized and hence improve the security performance.
In this paper, we present a new channel selection rule for
JPEG steganography. Different from that in PQ and D-PQ, not
only the PE, but also the quantization step (QS) and the magnitude of the quantized DCT coefficient to be modified (MQ),
are considered. These three factors are nonlinearly combined
together to form a new channel selection rule to find the coefficients which may introduce minimal detectable distortion for
data hiding. Via applying our new channel selection rule with
the same embedding strategy as that in MME, the experimental
results demonstrate that higher security performance can be obtained in JPEG steganography.
The rest of this paper is organized as follows. In Section II,
our new channel selection rule is introduced. Experimental results are given in Section III, and the conclusion is drawn in
Section IV.
II. PROPOSED CHANNEL SELECTION RULE
APPLICATIONS

AND ITS

A. Introduction of JPEG Compression


The key steps of the JPEG compression process [10] are
shown in Fig. 1, which illustrates the special case of grayscale
image compression. Note that the color image can be approximately regarded as multiple grayscale images with some minor
changes [10]. As seen, by applying the two-dimensional DCT

Fig. 2. Standard JPEG quantization table corresponding to QF

to the nonoverlapping and consecutive 8 8 blocks of the


original image data, the DCT coefficients will be obtained.
The obtained coefficients are then fed to the quantizer. In
the quantizer, the DCT coefficients are first divided by the
quantization steps and then rounded to obtain the quantized
DCT coefficients. At last, the quantized DCT coefficients are
entropy encoded to have the compressed image data.
For simplicity, the coefficients that have been divided by
quantization steps and not yet rounded are called unrounded
DCT coefficients, and those that have been divided by the
quantization steps and rounded are called quantized DCT
coefficients, respectively.
B. Proposed Channel Selection Rule
Different from that in PQ and D-PQ, not only the PE (i.e., perturbation error), but also the other two new factors are considered in our channel selection rule to find the coefficients which
may introduce minimal detectable distortion for data hiding.
The first one is QS (i.e., the quantization step). As we know,
in JPEG compression, the quantization steps associated with
different frequencies are different. Fig. 2 illustrates the standard JPEG quantization table corresponding to the quality factor
(QF) of 80. As seen, in general the quantization steps associated
with low frequencies are smaller than those with the high frequencies according to the zig-zag order. Because the inverse discrete cosine transform (IDCT) is a linear transformation, even if
the two unrounded DCT coefficients are with the same rounding
error, the coefficient associated with the smaller QS may result
in less information loss in the obtained JPEG image. In a similar
way, we can deduce that if the two unrounded DCT coefficients
to be modified are with the same PE, the coefficient associated
with the smaller QS may result in less distortion in the obtained
stego image.
The second one is MQ (i.e., the magnitude of the quantized
DCT coefficient to be modified). As we know, the JPEG
steganography can be viewed as superimposing a weak signal
(i.e., the secret message bits) onto a strong background signal
(i.e., the quantized DCT coefficients). In general, the detection

HUANG et al.: NEW CHANNEL SELECTION RULE FOR JPEG STEGANOGRAPHY

1183

threshold of perception for an embedded signal is proportional


to the magnitude of the background signal. Thus, embedding
the secret bits into those coefficients with relatively large
magnitude will be more difficult to be detected. In our channel
selection rule, the magnitude of the quantized DCT coefficient
is thus considered. Note that MQ is relevant to QS, since
in general the quantized DCT coefficient with relative large
magnitude belongs to the relatively low frequency (which may
have small QS) according to the zig-zag order.
Suppose the raw, uncompressed image is available to the
sender. Let
denote the elements in
quantization matrix, and
denote unrounded and quantized DCT coefficients of mode
in the th block, respectively, where
represents the total number of 8 8 blocks in JPEG image. The
relation between
and
is as follows:

JPEG steganography, the quantized DCT coefficients with minimal detectable distortion, which are computed according to (3),
will be selected for data hiding.
As seen from (3) in our new channel selection rule, the impacts of the presented two new factors QS and MQ are controlled by
and , respectively. Different values of
and
may result in different anti-detection capability of our new
channel selection rule. Since the statistics of images and the
extracted features of todays universal JPEG steganalyzers are
complex, moreover the three factors, i.e., PE, QS and MQ are
not independent and they would interact with each other, it is not
easy to establish a direct link between the anti-detection capability and the control parameters
and
theoretically. However, in the next section, our various experiments will be given
to demonstrate that via introducing these two new factors, i.e.,
QS and MQ separately into our proposed channel selection rule,
the detection capability of todays universal JPEG steganalyzers
can be decreased significantly. Moreover, both of these two control parameters, and , can be selected in a wide range. Our
experimental results will also demonstrate that with the combination of QS and MQ, the detection capability of todays universal JPEG steganalyzers can be decreased even further.

round

(1)

where round
rounds the element to its nearest integer. Note
that in (1),
represents the quantized DCT coefficient that is
obtained in JPEG compression without secret message embedding. When some secret bit is embedded into the quantized DCT
coefficient
, the coefficient after data hiding is denoted by
. The aforementioned perturbation error can be represented
as
(2)
where
is a function that returns the absolute value of the corresponding element . In (2),
and
represent the above-mentioned embedding error and rounding error,
respectively. Note that if the secret message bit is embedded
through changing the LSB of the quantized DCT coefficient,
will be equal to
(while the embedded message bit is
the same as the LSB of
) or
(while the embedded
message bit is different from the LSB of
). According to (2),
when the fractional part of the unrounded DCT coefficients
is as close as 0.5, minimal perturbation error will possibly be
obtained. Especially, when
is equal to
, no perturbation
error will be introduced and
is equal to zero.
In our new channel selection rule, the PE will be considered
with QS and MQ simultaneously. According to the previous
analysis, in order to reduce the detectable distortion that may
be introduced by data hiding, we need to embed the secret message bit into those coefficients which are associated with small
values of PE and QS, whereas with large value of MQ. Thus,
the detectable distortion in our channel selection rule is defined
as
(3)
In (3),
and
are two parameters that are used
to control the impacts caused by
and
. Since in most
of the JPEG steganographic schemes, secret message bits are
only embedded into those nonzero quantized DCT coefficients,
the parameter is generally selected as 0. Otherwise, the parameter can be selected as a relatively small constant, e.g.,
the value 1. When our new channel selection rule is applied in

C. Application of Our Channel Selection Rule


MME2 and MME3 utilize the modified binary Hamming
codes to embed secret message bits into the DCT coefficient
block with the length of
. Via a new matrix embedding methodology, MME2 and MME3 provide a simple and
practical approach to apply the D-PQ channel selection rule in
their embedding algorithm.
Here, we present a brief example to explain how to apply
our proposed channel selection rule with MME2 embedding
strategy. Without loss of generality, we assume that one
of the secret message blocks with
bits is represented by
, and the corresponding nonzero quantized
ac DCT coefficient block for message embedding is represented by
. Note that the corresponding
unrounded DCT coefficient block
is on
hand since the raw, uncompressed image is supposed to be
available. As in the original MME2 scheme, for any coefficient
in block
that needs to be changed, it will be modified as
follows:
if
if
if
if

(4)

and
is the coefficient after having been
where
modified. The message bits
are embedded into
the corresponding coefficient block
according
to the following three steps.
1) Conduct matrix embedding with
codes to find
the coefficient that may need to be modified in block .
Suppose the quantized ac DCT coefficient that needs to be
changed in block is
. According to (3),
we can compute the distortion
when the coefficient
is changed, as in (4). Note that in this step, if no coefficient
needs to be changed in this block via matrix embedding, all

1184

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 4, AUGUST 2012

Fig. 3. Block diagram of SVM training and testing.

the message bits


can be embedded into
the coefficient block without any modification, and the
following two steps will be neglected.
2) Then, the sender can find the pairs of numbers
in coefficient block
such that
,
where
is the function that converts
decimal value to binary vector, and
represents the
ExclusiveOr operation. Note that for any , there are
such pairs
in
coefficient block , which can be enumerated easily. For
each pair,
and
are computed
respectively according to (3) to find the distortion that
will be introduced. Suppose that among all these pairs
of coefficients, the coefficient pair
may have
the least distortion (please note that the corresponding
distortion is
).
3) Compare
with
. If
,
the matrix embedding is conducted with
codes, and only one coefficient in this block is modified
according to (4); otherwise, the matrix embedding is conducted with
codes, and the two coefficients
and
in this coefficient block are modified according
to (4).
In the same way, all the message bits in different blocks can
be embedded into their corresponding coefficient blocks.
Our new channel selection rule can also be applied with
MME3 embedding strategy in the same way except it allows
three embedding changes in each coefficient block to obtain
the minimal detectable distortion. For more information about
MME2 and MME3, please refer to [5].

our group in different scenarios with different cameras, 1543


images were downloaded from NRCS [11], and the remaining
1096 images are from CorelDraw image data set [12]. All the
5000 images are converted to the grayscale images and then
central-cropped into the size of 512 512, which are uncompressed images and will be used as the input of N-MME2 and
N-MME3. In the following, these uncompressed images will be
called input image and the JPEG compressed image without any
message embedding will be called cover image. The stego and
cover images are with the quality factor of 80.
To demonstrate the efficiency of our new channel selection
rule, four universal JPEG steganalyzers presented in [13][16]
are employed, denoted by ClbJFMP-274 [13], MP-486 [14],
ClbMP-324 [15] and POMM-98 [16], respectively, where the
numbers 274, 486, 324 and 98 denote the total number of
features utilized, Clb stands for calibration technique [17],
JF stands JPEG features [18], MP for Markov features [19],
and POMM represents the partially ordered Markov models
[16]. To the best of our knowledge, these four steganalyzers
are among the most effective universal JPEG steganalyzers in
detecting todays JPEG steganographic schemes.
The LibSVM toolbox [20] is employed in the classification.
The training and testing processes are exemplified in Fig. 3. As
seen, the training data is selected firstly, and then the kernel
of LibSVM is determined. In order to eliminate the effect of
randomness incurred by different kernels of LibSVM, two different kernels, i.e., radial basis function (RBF) kernel and linear
kernel, are utilized in our testing. When RBF kernel is selected,
the hyper-parameters
are optimized using five-fold crossvalidation over a fixed grid of values. The grid search is performed on the following multiplicative grid:

III. EXPERIMENTAL RESULTS AND ANALYSIS


In this section, experimental results and analysis are presented to demonstrate the improvements of our new channel
selection rule on JPEG steganography. The proposed channel
selection rule is applied with the same embedding strategy as
that in MME2 and MME3 using the modified binary Hamming
codes. The algorithms utilizing our new channel selection rule
are called new MME2 (N-MME2) and new MME3 (N-MME3),
respectively. According to (3), two new factors QS and MQ are
considered in our proposed channel selection rule. If the two
control parameters corresponding to QS and MQ are selected
as
, our channel selection rule will be the
same as D-PQ. Thus N-MME2 and N-MME3 will degrade into
MME2 and MME3, respectively.
The test image data set consists of 5000 uncompressed images. Among them, 2631 images were taken by members of

After the optimized parameter pair


is found, it will be
used on the training data to get the training model through the
svmtrain function of LibSVM toolbox. Note that if the linear
kernel is selected, the svmtrain function of LibSVM toolbox
will be utilized directly to find the training model. At last, the
training model is used to predict the testing data.
In our experiments, the secret message bits are randomly generated, and the embedding rates are represented in terms of bpnc
(bits per nonzero quantized DCT coefficients) values. Because
the stego images of N-MME2 and N-MME3 may be generated
with different parameters
and , and different embedding

HUANG et al.: NEW CHANNEL SELECTION RULE FOR JPEG STEGANOGRAPHY

1185

Fig. 4. Detection accuracy rates of N-MME2 corresponding to different parameter combinations (with RBF kernel). (a) With increasing of
while
.
(2) With increasing of

rates (i.e., 0.05, 0.10, 0.15 and 0.20


), a lot of binary classifiers need to be conducted in our testing. For each binary classifier, we randomly select 1/2 stego images and the corresponding
1/2 cover images as the training data, and the remaining 1/2
stego and 1/2 cover images are used for testing. In our testing,
true positive means that the stego image is predicted as stego
image, and true negative means that the cover image is predicted
as cover image. Since the cover and stego images to be detected
have the same quantities in our experiments, the final detection
accuracy rate (AR) is computed as AR
TPR TNR ,
where TPR and TNR represent the true positive rate and true
negative rate, respectively.
The experiments in this section contain four parts. In the first
part, the efficiency of QS and MQ will be tested separately in detail with ClbJFMP-274. Note that in most cases, ClbJFMP-274
may have the best detection accuracy rates among the aforementioned four JPEG steganalyzers according to our experimental
results. In the second and third parts, the aforementioned four
JPEG steganalyzers will be utilized to test the anti-detection capability and universality of our new channel selection rule with
the combination of QS and MQ. Note that in these first three
parts, RBF kernel is employed for testing. In order to eliminate
the effect of randomness that may be incurred by different kernels, in the fourth part the linear kernel is also employed in our
testing for a demonstration.
A. Testing Contributions of QS and MQ Separately
In this section, we will test the contributions of the proposed
two new factors, i.e., QS and MQ, in our channel selection rule
separately. A series of experiments corresponding to different
values of
and
are conducted. Note that
and
are
two parameters that are utilized to control the strength of impact
caused by QS and MQ, respectively. For an illustration, the two
parameters are selected in the range of [0, 1]. The experimental

while

results about N-MME2 and N-MME3 are shown in Figs. 4 and


5, respectively, where the horizontal axes represent the values
of the parameters (i.e.,
and ) and the vertical axes represent the detection accuracy rates. In Figs. 4(a) and 5(a),
is
set to 0 and
is increased from 0 to 1 with the step of 0.1. In
Figs. 4(b) and 5(b),
is set to 0 and
is increased from 0
to 1 with the step of 0.1. The embedding rates are selected as
0.05, 0.10, 0.15 and 0.20
, respectively. As seen in Figs. 4
and 5, when embedding rate is 0.05
, the detection accuracy rates corresponding to N-MME2 and N-MME3 are near
random guessing (i.e., around 50%), no matter what values of
and are selected. However, with the increasing of embedding rate, the efficiency of utilizing QS or MQ is distinctive.
Our experiments also demonstrate that in our channel selection
rule the parameters
and
can be selected in a wide range.
That is, the two control parameters can be easily determined in
the range of (0, 1] to form a new channel selection rule.
Note that our proposed channel selection rule may still work
with the increasing of
or . According to our testing, when
is set to 0 and is selected in the range of (1, 1.5] (or is set
to 0 and
is selected in the range of (1, 3]), the anti-detection
capability of our proposed channel selection rule may still be
better than that of D-PQ. However, since the proposed two new
factors QS and MQ are not independent of PE, if they are set
with too strong impact, the impact of PE may be depressed (note
that in general, the magnitudes of QS and MQ are larger than
that of PE). Our experimental results demonstrate that when the
control parameter
or
is selected out of the range (0, 1],
the anti-detection of the obtained channel selection rule will be
less than that of some cases with selecting the two parameters in
the range of (0, 1]. For example, the channel selection rule with
has less anti-detection
capability than that with
in general. Thus, while applied with the proposed channel

1186

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 4, AUGUST 2012

Fig. 5. Detection accuracy rates of N-MME3 corresponding to different parameter combinations (with RBF kernel). (a) With increasing of
while
.
(2) With increasing of

Fig. 6. Detection accuracy rates of N-MME2 corresponding to different comand


(with RBF kernel).
binations of

selection rule, we recommend to select the two parameters


and
in the range of (0, 1], and good performance can easily
be obtained.
B. Testing Efficiency With Combination of QS and MQ
In this section, some experiments will be given to demonstrate that with the combination of QS and MQ, the anti-detection capability of our new channel selection rule can be improved even further. In our testing, the two parameters
and
are selected as
and
. The detection accuracy rates with
and
are also illustrated for a comparison. Figs. 6 and 7 illustrate the experimental results corresponding to N-MME2 and

while

Fig. 7. Detection accuracy rates of N-MME3 corresponding to different comand


(with RBF kernel).
binations of

N-MME3, respectively, where the horizontal axes represent the


values and the vertical axes represent the detection accuracy rates. As seen in Figs. 6 and 7, when the two parameters are selected as
, the security performance of N-MME2 and N-MME3 will be better than that with
and
.
When the two parameters are selected as
,
the security performance of N-MME2 and N-MME3 will be
higher than that with
and
too.
In Figs. 8 and 9, we also illustrate the receiver operating
characteristic (ROC) curves corresponding to N-MME2 and
N-MME3, respectively, where the horizontal axes represent

HUANG et al.: NEW CHANNEL SELECTION RULE FOR JPEG STEGANOGRAPHY

1187

Fig. 8. ROC curves of N-MME2 corresponding to different embedding rates (with RBF kernel). (a)
.

. (b)

. (c)

. (d)

Fig. 9. ROC curves of N-MME3 corresponding to different embedding rates (with RBF kernel). (a)
.

. (b)

. (c)

. (d)

the false positive rates and the vertical axes represent the true
positive rates. In each subfigure of Figs. 8 and 9, the embedding
rate is shown on the title. The legend shows the parameters
(i.e.,
and ) and the corresponding area under curve (AUC)
values for each curve. The ROC curves and the AUC values

demonstrate that: 1) the presented two new factors, i.e., QS and


MQ both have contributions to decrease the detection capability
of ClbJFMP-274 and 2) through combing these two factors, the
efficiency of our channel selection rule can be improved even
further.

1188

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 4, AUGUST 2012

Fig. 10. Detection accuracy rates of N-MME2 against three different JPEG steganalyzers (with RBF kernel). (a) MP-486. (b) ClbMp-324. (c) POMM-98.

C. Testing Results With Other JPEG Steganalyzers


To demonstrate the universality of our new channel selection
rule, other three JPEG steganalyzers, i.e., MP-486, ClbMP-324
and POMM-98, are utilized in our testing. Counting ClbJFMP-274, four universal JPEG steganalyzers are utilized. Two
of them are self-calibration based JPEG steganalyzers and the
other two do not rely on self-calibration technique. Various
features that can effectively distinguish the cover and stego
images of todays JPEG steganography are extracted, including
the Markov features [13][15], Partial Markov features [16],
DCT coefficient histogram features [13], and blocking effect
features [13], etc.
The new experimental results (conducted with MP-486,
ClbMP-324 and POMM-98) corresponding to N-MME2 and
N-MME3 are illustrated in Figs. 10 and 11, respectively, where
the horizontal axes represent the embedding rates (i.e.,
values) and vertical axes represent the detection accuracy rates.
The utilized JPEG steganalyzers are illustrated on the title
and the parameter values (i.e.,
and ) are depicted on the
top-left corner of each subfigure.
The same results can also be observed from Figs. 10 and 11.
That is: 1) both of these two factors, QS and MQ, have definite
contributions to decrease the detection capability of todays universal JPEG steganalyzers and 2) by combining the aforementioned two factors, the efficiency of our new channel selection
rule can be improved even further.
D. Testing Results With Linear Kernel
In order to eliminate the effect of randomness incurred by
different kernels of LibSVM, we also conduct the experiments
by LibSVM with linear kernel. Note that when RBF kernel is
selected, searching for optimized hyper-parameters
is a

very time-consuming task. We have not conducted the statistical


tests over multiple runs due to the time limitation before.
For each reported experimental result in this part, we have
conducted the experiments 20 times. In each run, the training
samples include randomly selected 1/2 stego images and the
corresponding 1/2 cover images, and the rest of the stego and
cover images are used for testing. The average detection accuracy
rates are shown in Figs. 12 and 13. The aforementioned four
JPEG steganalyzers, i.e., ClbJFMP-274, MP-486, ClbMP-324
and POMM-98, are utilized in our testing. The experimental
results corresponding to N-MME2 and N-MME3 are illustrated
in Figs. 12 and 13, respectively, where the horizontal axes
represent the embedding rates (i.e.,
values) and vertical
axes represent the detection accuracy rates. The utilized JPEG
steganalyzers are illustrated on the title and the parameter
values (
and
) are depicted on the top-left corner of
each subfigure.
As seen, no matter which linear kernel or RBF kernel is selected (note that some experiments have also been conducted
by us with polynomial kernel and the obtained results demonstrate that our proposed method is independent of the selected
classifiers), the detection capability of the aforementioned four
JPEG steganalyzers can be decreased significantly with using
our new channel selection rule. For example, while the control parameters are selected as
or
and the embedding rate is as high as 0.10
,
the detection accuracy rates obtained by the aforementioned
four steganalyzers are constantly less than 60%, which is near
a random guessing. Furthermore, with the increasing of embedding rate, the anti-detection capability of our proposed channel
selection rule will be more obvious. For example, when the embedding rate is increased to 0.15 or 0.20
, with using our
channel selection rule the detection accuracy rates obtained by

HUANG et al.: NEW CHANNEL SELECTION RULE FOR JPEG STEGANOGRAPHY

1189

Fig. 11. Detection accuracy rates of N-MME3 against three different JPEG steganalyzers (with RBF kernel). (a) MP-486. (b) ClbMp-324. (c) POMM-98.

Fig. 12. Detection accuracy rates of N-MME2 against aforementioned four JPEG steganalyzers (with linear kernel). (a) ClbJFMP-274. (b) MP-486. (c) ClbMp324. (d) POMM-98.

the four JPEG steganalyzers may be decreased 15%30% or


even further.
IV. CONCLUSION
In this paper, we have presented a new channel selection rule
for JPEG steganography. Different from PQ (D-PQ), two new

factors, i.e., QS and MQ are considered besides PE. Various experimental results have demonstrated that the proposed two new
factors have significant contributions to decrease the detection
capability of todays universal JPEG steganalyzers. Moreover,
by combing these two new factors with PE simultaneously, the
efficiency of our new channel selection rule can be improved
even further. The proposed channel selection rule is easy to con-

1190

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 4, AUGUST 2012

Fig. 13. Detection accuracy rates of N-MME3 against aforementioned four JPEG steganalyzers (with linear kernel). (a) ClbJFMP-274. (b) MP-486. (c) ClbMp324. (d) POMM-98.

duct, which is almost as simple as PQ (D-PQ). However, the


improvement that can be obtained is distinctive.
REFERENCES
[1] R. Crandall, Some notes on steganography, [Online]. Available:
http://os.inf.tu-dresden.de/westfeld/crandall.pdf 1998, Steganography
Mailing List [Online]. Available:
[2] A. Westfeld, High capacity despite better steganalysis (F5-a steganographic algorithm), in Proc. Information Hiding, 4th Int. Workshop,
2001, vol. 2137, Lecture Notes in Computer Science, pp. 289302.
[3] J. Fridrich, M. Goljan, and D. Soukal, Perturbed quantization
steganography with wet paper codes, in Proc. ACM Workshop Multimedia and Security, Magdeburg, Germany, Sep. 2021, 2004, pp.
415.
[4] J. Fridrich, T. Pevn, and J. Kodovsk, Statistically undetectable
JPEG steganography: Dead ends, challenges, and opportunities, in
Proc. ACM Workshop on Multimedia and Security, Dallas, TX, Sep.
2021, 2007, pp. 314.
[5] Y. Kim, Z. Duric, and D. Richards, Modified matrix encoding
technique for minimal distortion steganography, in Proc. Information
Hiding, 8th Int. Workshop, 2007, vol. 4437, Lecture Notes in Computer
Science, pp. 314327.
[6] D. Schnfeld and A. Winkler, Reducing the complexity of syndrome coding for embedding, in Proc. Information Hiding, 9th Int.
Workshop, 2007, vol. 4567, Lecture Notes in Computer Science, pp.
145158.
[7] R. Zhang, V. Sachnev, and H. J. Kim, Fast BCH syndrome coding
for steganography, in Proc. Information Hiding, 11th Int. Workshop,
2009, vol. 5806, Lecture Notes in Computer Science, pp. 4858.
[8] C. Fontaine and F. Galand, How Reed-Solomon codes can improve
steganographic schemes, EURASIP J. Inform. Security, vol. 2009, pp.
110, 2009.
[9] T. K. Moon, Error Correction Coding, Mathematical Methods and Algorithms. Hoboken, NJ: Wiley, 2005.
[10] G. K. Wallace, The JPEG still picture compression standard, IEEE
Trans. Consumer Electron., vol. 38, no. 1, pp. 1834, Feb. 1992.
[11] NRCS Photo Gallery, [Online]. Available: http://photogallery.nrcs.
usda.gov, Available:

[12] CorelDraw Image CD, [Online]. Available: http://www.corel.com,


Available:
[13] T. Pevn and J. Fridrich, Merging Markov and DCT features for
multi-class JPEG steganalysis, in Proc. SPIE Electronic Imaging,
Security, Steganography, and Watermarking of Multimedia Contents
IX, 2007, vol. 6505, pp. 650503.1650503.13.
[14] C. Chen and Y. Q. Shi, JPEG image steganalysis utilizing both intrablock and interblock correlations, in IEEE Int. Symp. Circuits Systems,
Seattle, WA, May 1821, 2008.
[15] F. Huang and J. Huang, Calibration based JPEG steganalysis, Sci. in
China Series F: Inform. Sci., vol. 52, no. 2, pp. 260268, 2009.
[16] J. Davidson and J. Jalan, Steganalysis using partially ordered Markov
models, in Proc. Information Hiding, 12th Int. Workshop, Calgary,
Alberta, Canada, Jun. 2830, 2010.
[17] J. Fridrich, M. Goljan, and D. Hogea, Steganalysis of JPEG images:
Breaking the F5 algorithm, in Proc. Information Hiding, 5th Int.
Workshop, 2002, vol. 2578, Lecture Notes in Computer Science, pp.
310323.
[18] J. Fridrich, Feature-based steganalysis for JPEG images and its implications for future design of steganographic schemes, in Proc. Information Hiding, 6th Int. Workshop, 2004, vol. 3200, Lecture Notes in
Computer Science, pp. 6781.
[19] Y. Q. Shi, C. Chen, and W. Chen, A Markov process based approach
to effective attacking JPEG steganography, in Proc. Information
Hiding, 8th Int. Workshop, LNCS 4437, 2007, pp. 249264.
[20] C.-C. Chang and C.-J. Lin, LIBSVM: A Library for Support Vector Machines 2001 [Online]. Available: http://www.csie.ntu. edu.tw/~cjlin/
libsvm, Available:
Fangjun Huang (M07) received the B.S. degree
from Nanjing University of Science and Technology,
China, in 1995, and the M.S. and Ph.D. degrees from
Huazhong University of Science and Technology,
China, in 2002 and 2005, respectively.
From June of 2009 to June of 2010, he was
a Postdoctoral Researcher in the Department of
Electrical and Computer Engineering, New Jersey
Institute of Technology. Now, he is with the School
of Information Science and Technology, Sun Yat-Sen
University, China. His research interests include
digital forensics and multimedia security.

HUANG et al.: NEW CHANNEL SELECTION RULE FOR JPEG STEGANOGRAPHY

1191

Jiwu Huang (M98SM00) received the B.S. degree from Xidian University, China, in 1982, the M.S.
degree from Tsinghua University, China, in 1987, and
the Ph.D. degree from the Institute of Automation,
Chinese Academy of Science, in 1998.
He is currently a Professor with the School of Information Science and Technology, Sun Yat-sen University, Guangzhou, China. His current research interests include multimedia forensics and security.
Dr. Huang received the China National Funds for
Distinguished Young Scientists from NSF of China in
2003. He received the First-Grade Award in science research from the National
Ministry of Education, China, in 2004. He and his students won the best student paper award in the 8th International Workshop on Digital Watermarking
(Pusan, Korea) in 2008. He serves as a member of IEEE CASS Multimedia
Systems and Applications Technical Committee, and a member of IEEE SPS
Information Forensics and Security Technical Committee. He is an Associate
Editor for IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY,
LNCS Transactions on Data Hiding and Multimedia Security (Springer), and
EURASIP Journal on Information Security (Hindawi).

Yun-Qing Shi (M88SM92F05) received the


B.S. and M.S. degrees from Shanghai Jiao Tong
University, Shanghai, China, and the M.S. and Ph.D.
degrees from University of Pittsburgh, Pittsburgh,
PA.
He joined the Department of Electrical and
Computer Engineering, New Jersey Institute of
Technology, Newark, in 1987, and is now a Professor there. His research interests include digital
data hiding, forensics and information assurance, visual signal processing and communications, motion
analysis, theory of multidimensional systems and signal processing. He is an
author/coauthor of more than 250 papers, one book and five book chapters, and
an editor of ten books. He holds 20 U.S. patents.
Dr. Shi obtained the Innovators Award 2010 by New Jersey Inventors Hall
of Fame for Innovations in Digital Forensics and Security. His U.S. patent 7
457 341 entitled System and Method for Robust Reversible Data Hiding and
Data Recovery in the Spatial Domain won 2010 Thomas Alva Edison Patent
Award by Research and Development Council of New Jersey. He is the founding
Editor-in-Chief of LNCS Transactions on Data Hiding and Multimedia Security (Springer), an Associate Editor of Journal on Multidimensional Systems
and Signal Processing (Springer) and three other journals. He served as an
Associate Editor of IEEE TRANSACTIONS ON SIGNAL PROCESSING and IEEE
TRANSACTIONS ON CIRCUITS AND SYSTEMS (II), an Associate Editor of International Journal of Image and Graphics, and a Guest Editor of special issues
for a few journals, the technical program chair of IEEE ICME07, co-technical
chair of IWDW06, 07, 09, 10, 11, IEEE MMSP05, co-general chair of IEEE
MMSP02, a Distinguished Lecturer of IEEE CASS (200203). He is a member
of a few IEEE technical committees.

You might also like