Accepted Manuscript

Research on Image Steganography Analysis Based on Deep Learning

Ying Zou, Ge Zhang, Leian Liu

PII: S1047-3203(19)30091-4
DOI: https://doi.org/10.1016/j.jvcir.2019.02.034
Reference: YJVCI 2495

To appear in: J. Vis. Commun. Image R.

Received Date: 27 August 2018

Revised Date: 26 February 2019
Accepted Date: 26 February 2019

Please cite this article as: Y. Zou, G. Zhang, L. Liu, Research on Image Steganography Analysis Based on Deep
Learning, J. Vis. Commun. Image R. (2019), doi: https://doi.org/10.1016/j.jvcir.2019.02.034

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
 Research on Image Steganography Analysis Based on Deep Learning
Ying Zou1,a , Ge Zhang2,3,b*and Leian Liu1,c
1
College of Information Science and Technology, Zhongkai University of Agriculture and
Engineering, Guangzhou, China
2
School of Computer and Information Engineering, Henan University, Kaifeng 475004,China
3
Henan Key Laboratory of Big Data Analysis and Processing, Henan University, Kaifeng
475004, China
a
[email protected], [email protected], [email protected]
*Corresponding author:Ge Zhang([email protected])

Abstract: Although steganalysis has developed rapidly in recent years, it still faces many difficulties and
challenges. Based on the theory of in-depth learning method and image-based general steganalysis, this paper
makes a deep study of the hot and difficult problem of steganalysis feature expression, and tries to establish a
new steganalysis paradigm from the idea of feature learning. The main contributions of this paper are as
follows: 1. An innovative steganalysis paradigm based on in-depth learning is proposed. Based on the
representative deep learning method CNN, the model is designed and adjusted according to the characteristics
of steganalysis, which makes the proposed model more effective in capturing the statistical characteristics such
as neighborhood correlation. 2. A steganalysis feature learning method based on global information constraints
is proposed. Based on the previous research of steganalysis method based on CNN, this work focuses on the
importance of global information in steganalysis feature expression. 3. A feature learning method for low
embedding rate steganalysis is proposed. 4. A general steganalysis method for multi-class steganography is
proposed. The ultimate goal of general steganalysis is to construct steganalysis detectors without
distinguishing specific types of steganalysis algorithms.
Key words: Steganalysis, Steganography, Feature Learning, Deep Learning, Convolutional Neural Network,
Transfer Learning, Multitask Learning

Chapter 1 Introduction
With the rapid development of information technology and the rapid popularization of the Internet,
digital media has become an important carrier for military, commercial and other organizations as well as
individuals to obtain and transmit information. But at the same time, because the digital communication in the
Internet is vulnerable to the threat of eavesdropping, malicious interference and other activities, people pay
more attention to the security issues such as privacy protection and data integrity in the process of information
transmission than ever before. The traditional solution uses encryption technology to convert the information
to ciphertext for transmission. But its shortcoming is that the encrypted ciphertext is usually disordered. It is
easy for an attacker to notice the existence of secret communication, which stimulates the attacker's enthusiasm
for decoding. At the same time, it may also lead to information being interfered or intercepted, resulting in the
failure of information transmission. In the above context, a new concept of communication security has been
gradually accepted and recognized: communication security means not only that the content of information to
be transmitted is secure, but also that the existence of the act of transmitting secret information is unknown.
Therefore, steganography, which is characterized by "camouflage" in the transmission of information, has
attracted more and more attention.
The basic principle of steganography is to hide the secret information which needs to be transmitted into
the redundant information by using the insensitive redundant information of human perception system existing
in common carriers, and to realize the transmission of secret information by means of carrier transmission.
Because the process of information hiding into the carrier usually does not change the normal perception
characteristics of the carrier, it is difficult for potential attackers to detect the existence of secret information,
thus ensuring the information security covert transmission. At the same time, it can also combine encryption,
scrambling, coding and other technologies, making it difficult to extract hidden information even if it is
detected by third parties, thus further ensuring the security of information transmission. In order to explain
steganography more vividly, we describe the "prisoner problem" as an example. Alice and Bob are prisoners in
different cells in the same prison. They are under the care of Warden Eve. Alice and Bob are going to discuss
plans for a joint jailbreak. Specific plans need to be negotiated through the exchange of information. However,
according to prison regulations, their communications need to be checked by the warden, so they can not
communicate in plaintext. In this case, Bob and Mice need to take more covert communication measures. One
consideration is to use encryption technology, that is, to hide the content of information, but because the
encrypted information is a mess of code, it is easy to arouse Eve's suspicion. Thus, in this case, a safer
consideration is to hide information in everyday objects, and make the hidden objects look normal, which can
reduce Warden Eve's vigilance and ensure the smooth transmission of information.

However, steganography is also a true double-edged sword, which provides people with reliable and
secure means of Internet communication, at the same time, it may also provide convenience for organizations
and individuals with malicious intentions or improper purposes [11-14]. In fact, in recent years, there have
been reports about the use of steganography in espionage, terrorist attacks, crimes and other activities. In 2001,
some mainstream media in the United States, such as CNN and US Today, reported the news of secret
communications between Al Qaeda members using steganography. Reported that bin Laden gang will attack
the target map, action instructions and other information hidden in pornography, sports chat and other websites.
According to Die Zeit, an al-Qaida suspect was arrested in Berlin in May 2011 and police found him carrying a
memory card. Later, after being cracked by experts in charge of computer criminal investigation in the German
Federal Criminal Police Bureau, it was found that on the surface, only one pornographic video named
"KickAss" appeared on the card, but in fact, 141 text documents were hidden in the video, including a large
number of Al-Qaida action reports, future action plans and so on. In June 2010, the Federal Bureau of
Investigation (FBI) successfully arrested 10 Russian agents in New Jersey, which caused a great diplomatic
shock between the United States and Russia.
The FBI said that by eavesdropping on the conversations between the Murphy couple, who were all
suspects, they found that Murphy had sent his wife Cynthia to South America to hand over something
"invisible" to someone. According to the FBI analysis, these so-called "invisible things" are likely to be
recorded by digital steganography. They also claimed to have found devices for digital steganography in the
homes of three suspected Russian agents, and inferred that Russian agents used steganography to communicate
information with relevant Russian intelligence agencies. According to a Reuters report in July 2015, a report
released by security manufacturer FireEye shows that Russian hackers have successfully invaded the U.S.
defense system and captured several computers of the U.S. Department of Defense using data from Twitter
that appears to be ordinary photos. These seemingly ordinary Twitter images hide information and instructions
to activate malicious programs that have been implanted into the target computer. With the help of the
camouflage of Twitter pictures, the transmission of commands controlling malicious programs easily avoids
the detection of most detection systems. In April 2014, mainstream media in China, such as CCTV and Xinhua
News Agency, reported that the mobile Trojan Horse "immortal Trojan Horse II" infected millions of mobile
phones, and carried out fishing attacks, remote control of users'mobile phone photos, stealing users' online
silver and other activities that seriously endangered users'personal privacy and property security. It is reported
that, unlike the normal Trojan horse direct command mode, the Trojan horse cunningly disguises malicious
code instructions as a common picture through steganography, in order to avoid the "pursuit" of mobile phone
security software.
It can be seen from above that illegal or malicious use of digital steganography has brought serious harm
to national information security, business and personal privacy and property security [15-19]. In this case, how
to effectively supervise the use of steganography in real life, prevent or block the malicious or illegal use of
steganography in real time has become an urgent need of military and security departments in various
countries. Because of this, Steganalysis, as a countermeasure technology of steganography, came into being,
and has attracted the attention of governments and scientific research institutions. Steganalysis is a technique
to determine whether there is additional information hidden in the carrier or not, even to estimate the amount
of information embedded in the carrier, and to obtain the content of the hidden information by analyzing the
statistical characteristics of the carrier. It plays the role of Warden Eve in the model of "prisoner problem"
mentioned above. The research of steganalysis technology is of great significance in preventing the leakage of
confidential information, combating terrorism and criminal activities, and maintaining Internet security.
Based on the theory of in-depth learning method and image-based general steganalysis as the research
object, this paper focuses on the hot and difficult problem of feature expression in steganalysis, and tries to
establish a new framework of steganalysis from the new idea of feature learning. The specific research
contents of this paper mainly include the following aspects:
1. A digital image steganalysis framework based on depth learning is proposed.
Aiming at the problems in steganalysis, especially in feature expression. Considering feature learning, a
new steganalysis paradigm based on in-depth learning is proposed. Based on the representative deep learning
method CNN, the proposed model is designed and adjusted according to the characteristics of steganalysis,
which makes the proposed model more effective in capturing the statistical characteristics such as
neighborhood correlation Related to steganalysis, and automatically expresses the features of steganalysis
effectively through learning. Different from traditional methods based on artificial design features, this method
integrates feature extraction module and classification module into a trainable network model framework, and
automatically learns features and realizes classification in the form of data-driven, thus greatly reducing the
need for human experience and time. On this basis, the detection performance is further improved by model
fusion and other methods.
2. A steganalysis feature learning method based on global statistical information constraints is proposed.
This work focuses on the importance of global information in steganalysis feature expression, and
introduces how to use global statistical information to make the model learn better feature expression under the
framework of feature-based learning steganalysis. Referring to the idea of transfer learning, this paper
proposes a CNN model based on global statistical information constraints. By calculating auxiliary features to
obtain additional global statistical information, and then introducing global statistical information into CNN
model in the form of auxiliary tasks in migration learning, regularization constraints are applied to training
CNN model, so that it can learn better feature expression.
3. A feature learning method for low embedding rate steganalysis is proposed.
In the field of steganalysis, the detection of low embedding rate encrypted images is a key concern. The
difficulty lies in the small amount of embedding information, which makes the change of image statistical
characteristics relatively small, so it is more difficult to detect. To solve this problem, based on the idea of
Transfer Learning, this paper proposes to enhance the feature learning on low embedding rate dense image
datasets by migrating the prior information of features learned from high embedding rate dense image datasets
of CNN, so as to improve the detection performance of CNN model for low embedding rate dense image.
4. A general steganalysis method for multi-class steganography is proposed.
From the current research status of general steganalysis algorithms, the "universality" of existing
methods only reflects that different steganalysis algorithms can use the same algorithm steps to construct
detectors. However, in the process of constructing the detector, we need to know the specific types of edge
information of the steganography algorithm to be detected, and use these edge information to generate the
dense image for training. There are many possible steganography algorithms in real application scenarios.
Steganalysts often find it difficult to know which steganography users use. Therefore, it is urgent and
significant to study steganalysis detectors for different steganalysis algorithms (known or even unknown). This
chapter will explore this issue. On the basis of previous research on Steganalysis Based on in-depth learning
and combined with the idea of multi-task learning, this paper tries to propose a steganalysis method for
steganographic images generated by different steganographic algorithms.

Chapter 2 Relevant Research Summary

Neil E. Johnson of George Mason University in the late 1990s in the United States first began the study
of steganalysis. Later, Dartmouth College, Massachusetts Institute of Technology, New York State University,
Purdue University, New Jersey Institute of Technology, Wet Stone Corporation, IBM Corporation, Microsoft
Corporation and other institutions have carried out research in this direction, and most of them have received
strong support from the United States Department of Defense, the National Security Agency and other
government departments. In addition, the United Kingdom, Germany, France, Russia, Japan, Finland and other
countries have also actively invested in the research and published important results. At present, Jessica
Fridrich, Andrew D. Ker, Yun Q. Shi, Haw Farid, Niels Provos, Andreas Westfeld and other well-known
research experts are in the field. In China, the Institute of Automation, the Institute of Information Engineering,
the Chinese Academy of Sciences, the Beijing Institute of Electronic Technology Application, Tsinghua
University, Sun Yat-sen University, China University of Science and Technology, Shenzhen University,
Beijing University of Posts and Telecommunications, Dalian University of Technology, Tianjin University,
Shanghai University, Tongji University, PLA University of Information Engineering, Hunan University and
other universities and research institutes are all located. The research has been carried out and supported by
national "863", "973", "National Natural Science Foundation" and other projects or funds. International well-
known journals and conferences in the field include: IEEE Transactions on Informarion Forensics and Security
(TIFS), Information Hiding and Multimedia Security (IH&MMSec), Media Watermarking, Security, and
Forensics (MWSF), International Workshop on Information Forensics and Security (WIFS), International
Workshop on Digital Watermarking (IWDW), etc. As of October 2016, the National Conference on
Information Hiding and Multimedia Information Security has successfully held 13 sessions.
Although steganalysis research has made continuous progress in recent years, the so-called "one foot tao,
one foot devil tall", steganography research is also developing rapidly, and has brought new challenges to
steganalysis. Advanced steganography can make better use of the variability and complexity of natural image
content to conceal the existence of weak steganographic signals. At the same time, the existence of large
amounts of image data on the Internet helps to make the carrier images with hidden information relatively safe
to be mingled among them. These factors make steganalysis like "looking for a needle in a haystack". The
increasing difficulty of steganalysis also makes the traditional heuristic-based Steganalysis Based on artificial
design features face a more severe test. Therefore, there is an urgent need to develop some new methods for
steganalysis. In recent years, with the rise of in-depth learning methods, feature learning has gradually been the
focus of machine learning research. In-depth learning uses deep neural networks to automatically learn
effective feature representation in the form of data-driven, and quickly overturns the research framework based
on artificial features in many machine learning related application fields, becoming a new research paradigm.
The idea of in-depth learning is expected to bring new research ideas for steganalysis tasks. As a
confrontational technology, the research of steganalysis needs to pay attention to the development of
steganalysis, so as to achieve "targeted", so this chapter first introduces the research status of steganalysis, and
then introduces the development of steganalysis. At the same time, in order to better understand the in-depth
learning methods used in this paper.

Chapter 3 Steganalysis Based on Convolutional Neural Network

3.1 introduction
At present, the most concerned problem in steganalysis is feature expression. Good feature expression
plays a key role in the final detection performance. In traditional steganalysis methods, this part is done by
human. However, it is a very laborious and heuristic method to select features manually. Whether effective
features can be selected depends largely on human experience, and the adjustment of feature parameters takes
a lot of time and energy. Up to now, although great progress has been made in feature extraction of
steganalysis, many effective features have emerged. Because of the difficulty of steganalysis itself and the
continuous development and progress of steganography, the design of steganalysis features has brought serious
challenges. Especially in recent years, the content adaptive steganography algorithm can maintain more
complex statistical characteristics such as neighborhood correlation, which makes steganalysts need to
consider more complex image statistical characteristics for effective detection, and the difficulty of artificial
design features is also increasing, which puts forward higher requirements for Designers'experience and
knowledge.
In view of the problems faced by steganalysis, unlike traditional methods, this work pioneered the
introduction of in-depth learning into steganalysis [20-23]. Considering the problem of feature expression in
steganalysis from the perspective of feature learning, a new idea to solve the problem of steganalysis was
proposed. In this chapter, a steganalysis framework based on in-depth learning is proposed. Based on the
representative deep learning model, Convolutional Neural Network (CNN), combined with the characteristics
of steganalysis task, the network structure is further designed and selected, which makes the deep learning
model more suitable for learning the feature expression related to steganalysis. Particularly noteworthy is that
this chapter proposes a new framework for steganalysis, which proves the validity of feature-based learning
based on in-depth learning model in steganalysis through experimental analysis. It has guiding and referential
significance for the next development direction of steganalysis.
3.2 Steganalysis feature learning framework based on CNN.
The feature of in-depth learning is that the complex representation is fitted by multi-layer non-linear
transformation structural units, and the effective feature expression is learned automatically through training.
For specific application areas, the key to the success of in-depth learning is whether the structure design of in-
depth model is suitable for capturing task-related statistical characteristics. At present, the most concerned
statistical feature in steganalysis research is neighborhood dependencies. In fact, with the continuous
development of steganography, the new steganography algorithm can maintain some simple statistical
characteristics of images, but it is difficult to completely maintain complex statistical characteristics such as
neighborhood correlation. At present, the most effective artificial features in the field of image steganalysis are
based on the modeling of neighborhood correlation to distinguish the changes brought about by steganographic
embedding operations. Due to the lack of accurate image statistical models, and the current advanced content
adaptive steganography usually embeds information into image regions with complex texture and noise, the
neighborhood correlation of these regions is more complex, and it is difficult to effectively model. In
traditional methods, feature design patterns based on human experience are facing great challenges. In this
study, by using the successful experience of traditional steganalysis methods for reference, the important
information of neighborhood correlation is also taken into account in the selection and structure design of deep
learning model, so that the designed deep learning model can automatically mine the neighborhood correlation
of steganalysis and learn effective feature expression. Therefore, a steganalysis model based on CNN is
proposed. CNN, as a representative in-depth learning method, is a kind of neural network with special network
structure. It contains local connection, weight sharing, downsampling and other structural design ideas. It can
directly learn more effective feature expression from data. From the perspective of steganalysis, the
convolution structure in CNN is conducive to capturing the high-order statistical information of the correlation
between adjacent pixels, which is very important for steganalysis. At the same time, weight sharing, pooling
and other structures also greatly reduce the training parameters, making the network model can handle larger
size and number of pictures. Therefore, CNN is selected as the basic model in the framework of deep learning
in this paper. Furthermore, on the basis of the traditional CNN structure, this work takes more into account the
characteristics related to steganalysis tasks. By designing and selecting appropriate model structure units and
parameters, the proposed model can effectively learn the expression of steganalysis features.

Chapter 4 Steganalysis Feature Learning Based on Global Statistical Information Constraints

4.1 introduction
In the previous chapter, a CNN-based steganalysis framework is proposed, which can automatically
acquire effective steganalysis feature expression by feature learning. In this chapter, further research is carried
out on the basis of previous research work. This work focuses on the importance of global information in
steganalysis feature expression, and introduces how to use global statistical information to enable the model to
learn better feature expression under the framework of feature-based learning steganalysis. Global statistical
information in steganalysis generally refers to the statistical information based on the whole image, which
plays an important role in steganalysis. Generally, steganography can be seen as adding noise to an image. At
the same time, the noise is very weak relative to the image content itself, and can be easily masked by the
complex image content itself. From the long-term research experience of steganalysis, steganalysts need to
consider two aspects of information at the same time in order to better capture the slight changes brought by
the modification of steganalysis operations in some local areas of the image. The first aspect is the local
information, that is, how to describe the relationship between the elements in the local area of the image. It
often involves image characteristics such as complex correlation between pixels. The second aspect is global
information, that is, considering the relationship between local elements of the image, further global image
statistics is needed to make the image changes caused by steganography more significant in the image feature
space. Both of them play an extremely important role in steganalysis.
Through the analysis of the steganalysis model based on CNN proposed in the previous chapter, this
paper considers that the convolution structure in the model has advantages in capturing the correlation between
the pixels in the local area of the image, but in the use of global information in the process of feature extraction,
CNN usually fuses the information of the local area layer by layer through the scaling of the convolution layer
or the pooling operation, which exists in itself. Restrictions. To solve this problem, this chapter proposes a
method based on model regularization constraints. The goal is to use model regularization technology to
introduce global information and achieve more effective steganalysis feature learning.
4.2 Relevant Work
Researchers have long noticed the importance of global image statistics for steganalysis. In traditional
methods based on artificial features, how to use global information has always been a problem to be
considered in feature design. In early steganalysis feature extraction, image global statistical histogram and
statistical moments based on histogram are usually used as features. These methods make use of the global
statistical characteristics of images to distinguish the differences between images before and after
steganography.
In the steganalysis method based on CNN proposed in the previous chapter, the convolution structure of
CNN is suitable for describing the relationship between the pixels in the local area of an image, and through
hierarchical convolution, it can automatically learn more complex image local relationship patterns. In terms of
global information utilization, it is different from traditional steganalysis methods. In the CNN model, the local
area information of the image is fused by pooling. Through multi-layer pooling operation, the high-level
features of CNN fuse the local information in a larger range of images from the low-level layer by layer.
Compared with the traditional global statistical methods such as histogram and co-occurrence matrix, pooling
is a region-based statistical method. It is different from the fine-grained statistics of calculating multiple bin
(groups) from the whole graph in histogram and co-occurrence matrix. In pooling, a response value is
calculated from the corresponding region, and it is easy to lose some useful global information for steganalysis.
This chapter tries to study how to make better use of global information and promote more effective feature
learning based on CNN model.

4.3 Steganalysis Feature Learning Model Based on Global Information Constraints

In order to make better use of global statistical information in steganalysis feature learning and use the
idea of transfer learning for reference, a CNN model based on global statistical information constraints is
proposed in this paper. By calculating auxiliary features to obtain additional global statistical information, the
information is introduced into the CNN model in the form of auxiliary tasks in transfer learning, and the
training of CNN model is regularized to make it learn better feature expression.
The method proposed in this paper enlightens the successful application of self-transfer learning in some
other research fields based on in-depth learning. Transfer learning is a machine learning method that helps new
task learning by transferring knowledge in a task or data domain. The key to its success lies in the fact that the
first two tasks are related and have some commonalities. At the same time, one task can provide information
useful to another task and not easily captured by itself. In the field of target recognition, Oquab et al. proposed
the task of transferring the mid-level image feature expression learned from supervisory training on large-scale
data sets to target recognition on small-scale data sets. The method is to use some deep network layer
parameters after training on large-scale data sets when training on small-scale data sets. Ji et al. used similar
ideas in the research of video-based human behavior recognition. In this method, the behavior information of
long continuous frames is introduced into the 3D-CNN model through auxiliary motion features. There are
other similar tasks that utilize auxiliary tasks or features. Unlike these works, this work focuses on steganalysis.
The goal of this work is to introduce global statistical information into CNN model to improve the
performance of model detection by referring to the idea of transfer learning.
 Chapter 5 Feature Learning for Low Embedding Rate Steganalysis
5.1 introduction
In the previous two chapters, the most concerned feature expression problem in steganalysis is studied
from the perspective of feature learning. A steganalysis model based on CB,_r1N is proposed, and the
expressive ability of the model is improved by means of model structure design and prior information
introduction. On the basis of previous research work on steganalysis feature learning based on CNN, this
chapter focuses on steganalysis feature learning for low embedding rate steganalysis image detection.
The difficulty of steganalysis is that the steganographic noise signal introduced into images by
steganographic operations is usually very weak, and the difference between images before and after
steganographic operations is very small, and this difference can easily be masked by the difference between
different image contents. When the amount of information embedded in the image decreases gradually, the
difficulty of steganalysis will be further increased. In fact, Steganalysis for the detection of low embedding rate
encrypted images has always been a challenging problem. The difficulty is that, compared with high
embedding rate embedding, low embedding rate embedding requires fewer image elements to be modified and
fewer changes to the statistical characteristics that may be used by steganalysts. Especially in recent years, the
content adaptive steganography proposed in this paper can hide steganographic noise signals into complex
texture regions of images which are difficult to model statistically when embedding information. Therefore,
under the condition of low embedding rate, adaptive steganography can give full play to its advantages and is
more difficult to detect.
Based on the research work of steganalysis feature learning in the previous chapters, this paper finds that
the training difficulty of the proposed steganalysis model based on CNN is especially obvious when faced with
low embedding rate steganalysis task, and even can not converge. It is worth emphasizing that the standard of
low embedding rate is not accurately defined in steganalysis. The main reason is that the difficulty of
steganalysis needs to consider both the steganographic algorithm to be detected and the embedding rate, and it
is difficult to define the absolute low embedding rate according to the detection difficulty. The low embedding
rate mentioned in this paper is a relative concept and relative to the same steganography algorithm. Although
the detection difficulty may vary greatly under the condition of uniform embedding rate of different algorithms,
for the same algorithm, the detection difficulty will gradually increase with the decrease of embedding rate
until it is difficult to distinguish between normal image and dense image. Experiments show that the training of
CNN model can not converge when the embedding rate of the classified image in the training set is less than
0.1 BPP when training and detecting steganography with the CNN model proposed in Chapter 3. In fact,
although deep learning networks such as CNN have good feature learning ability, there are a lot of training
parameters in these deep learning network models, which are easy to fall into local optimum in the training
process, so there are some training difficulties. This situation is particularly evident in the training of small
data sets. For steganalysis tasks, the difference between the encrypted image and the normal image is smaller
under the condition of low embedding rate. At the same time, the content of different images is greatly
different, which makes CNN more easily misled by the difference of image content, and it is difficult to
distinguish steganographic noise signals. Aiming at this problem, based on the idea of transfer learning, this
work proposes to enhance the feature learning on low embedding rate dense image data set by transferring the
prior information of the features learned from high embedding rate dense image data set of CNN, so as to
improve the detection performance of CNN model for low embedding rate dense image. This method has a
good effect in the face of the most advanced content adaptive steganography.

Analysis of 5.2 Content Adaptive Steganography

In order to better elaborate the research methods proposed in this chapter, the characteristics of content
adaptive steganography algorithm are introduced and analyzed before the research methods are specifically
introduced. Content adaptive steganography is one of the most secure steganography algorithms. The core idea
of this kind of algorithm is to make use of the complex statistical characteristics of texture regions in images,
which make it difficult to model effectively. For a given embedding rate, the information is embedded in
texture regions first, so that the influence of steganography modification can be covered up by the complex
statistical characteristics of texture.
Fig. 5.1 shows the embedding position of information after embedding a carrier image in a representative
content adaptive algorithm database according to different embedding rates. The carrier image contains the
rough surface of the cliff, the vegetation layer on the cliff, the edge of the building and other areas with
complex texture, as well as the smooth areas of the sky and the building wall. When embedding information
with WOW, it can be found that when embedding rate is low, the information is basically embedded in the
region with the most complex texture orientation. With the increase of embedding rate, the location of
embedding information gradually expands from texture-rich regions to relatively smooth regions. For example,
when the embedding rate is 0.6 bpp, part of the information is embedded in the smoother wall area between the
building and the cliff. The map reflects the priority of embedded region selection in content adaptive algorithm,
that is, according to the texture complexity from high to low. This kind of steganographic algorithm has more
advantages in low embedding rate, and it is more difficult to detect than the early steganographic algorithm
which randomly chooses the embedding area.
Another feature of this kind of algorithm is that, in the process of increasing embedding rate, the location
of information embedding usually diffuses gradually along the extension of the region with the richest texture.
That is to say, the embedding area corresponding to high embedding rate generally includes the embedding
area corresponding to low embedding rate, and the embedding area with relatively low embedding rate can be
regarded as the extension of texture area used in low embedding rate. From the perspective of steganalysis,
since high-embedding usually first uses texture regions similar to low-embedding, the statistical characteristics
of images affected by high-embedding are similar to that of low-embedding. At the same time, higher
embedding modifies more image elements, leaving more "traces" that can be used for steganalysis.
5.3 Method
Through the research and analysis of the characteristics of content adaptive steganography in the
previous section, we find out the reasons why it is difficult to detect content adaptive steganography in low
embedding rate. At the same time, we find that there are some commonalities in the influence of different
embedding rates on the statistical characteristics of images. This paper considers that there is a priori
information that can be used by low embedding rate steganalysis in high embedding rate steganographic
images which are relatively easy to detect. Based on the idea of transfer learning, this paper proposes to
transfer auxiliary information from high embedding rate image to CNN model training for detecting low
embedding rate image. This method is different from the traditional steganalysis, which treats different
embedded images independently. In traditional steganalysis, when steganographic images with different
embedding rates are detected by the same steganographic algorithm, independent training classifiers are
usually trained according to different embedding rates. The proposed method hopes to introduce information
from high embedding rate images to help low embedding rate dense image detection. In the previous chapter,
the transfer learning method is also used to introduce the priori information of auxiliary tasks into the CNN
model. It uses auxiliary features and achieves the constraints of prior information by means of mixed training
of auxiliary tasks and target tasks. This transfer learning method of mixed training of auxiliary task and target
task has similar application in some computer vision tasks. Unlike the methods in the previous chapter, the
goal of this chapter is to improve the detection performance of low embedding rate steganography, and the
way of transferring information is similar to pre-training without using additional auxiliary features. Pre-
training ideas similar to those in this paper are used in some computer vision work such as target recognition.
The proposed framework consists of two modules, which correspond to target task and source task respectively.
The target task is low embedding rate image detection, and the source task is also called auxiliary task, which
is high embedding rate image detection. Firstly, the source task is trained, then the trained features are
transferred to the target task, and the target task is trained to improve the training performance of the target
task.

Chapter 6 General Steganalysis for Multi-class Steganography

6.1 introduction
The ultimate goal of general steganalysis is to construct a steganalysis detector without distinguishing
specific types of steganalysis algorithms, which can simultaneously detect different types of known or even
unknown steganalysis. However, from the current research status of generic steganalysis algorithms, the
"universality" of existing methods only reflects that different steganalysis algorithms can use the same
algorithm steps to construct detectors. However, in the process of constructing the detector, we need to know
the specific types of edge information of the steganography algorithm to be detected, and use these edge
information to generate the dense image for training. In fact, the detection performance of steganalysis
algorithms given in most previous studies is the result of "one-to-one" scenarios, and then the steganalysis
image embedded with the same steganalysis algorithm in the training set is detected with the detector of
Training I. At present, the general steganalysis research mainly focuses on this "one-to-one" framework for the
following reasons. Firstly, the image modification caused by steganography is usually very weak, which is
easily concealed by the difference between the content of the image itself. This also makes it difficult to
establish an effective image statistical model in steganalysis. Researchers hope to use more edge information
to build a more accurate image statistical model. Secondly, the influence of different steganography
embedding mechanisms on the statistical characteristics of images is also very different. Similar to the problem
caused by image heterogeneity in image classification, when constructing detectors in steganalysis, a similar
"heterogeneity" problem needs to be considered if multiple steganographic algorithms are considered at the
same time. In fact, due to the deficiencies of feature generalization performance in traditional feature-based
steganalysis methods, there are great challenges to solve these problems. However, from the practical point of
view, the "one-to-one" steganalysis model is not enough. There are many possible steganography algorithms in
real application scenarios. Steganalysts often find it difficult to know which steganography users use.
Therefore, it is urgent and significant to study steganalysis detectors for different steganalysis algorithms
(known or even unknown). In this chapter, we will explore this issue. On the basis of previous research on
Steganalysis Based on in-depth learning and combined with the idea of multi-task learning, a steganalysis
method for steganographic images generated by different steganographic algorithms is proposed.
6.2 Relevant Work
 At present, in the field of image steganalysis, there are relatively few studies on the construction of
detectors that can simultaneously detect different steganographic algorithms, especially for the early JPEG
domain steganographic algorithms. From these works, we can summarize three kinds of solutions to problems.
In the first method, the training data set is composed of cover image and stego image, which are generated by
various steganography algorithms. Then the feature is extracted from the training data set and a two-class
classifier is trained to distinguish the normal image from the stego image. In Pevny's and Ffidrich's work, four
different JPEG domain steganography algorithms are selected for hybrid training. After extracting DCT
domain features, two classes of SVM classifiers are trained. So the trained classifier can distinguish the normal
image from the image generated by different steganographic algorithms, even for the steganographic algorithm
which is not used in training, it has certain detection effect. This method is a more intuitive mixed training
mode. In the second method, the training data set is constructed in the same way as in the first method, and the
steganalysis features are extracted. The difference is that this method regards steganalysis as a multi-class task,
and trains multi-class classifiers to distinguish between normal images and specific embedding algorithm types.
In the method proposed by Pevny and Fridrich [142,143], seven JPEG domain steganography algorithms and
dense images are generated according to different embedding rates. The training set is composed of normal
images, and representative DCT domain features are extracted for training. In multi-classification tasks, the
multi-classifiers constructed by them are used to distinguish seven categories including normal image classes
and six types of steganography algorithms. In the training of multi-classifier, they adopt the method of
"max.wins". That is to say, they train 21 second-class SVMs for seven classes of images, and count the results
of 21 second-class classifications. The classes that get the most votes in seven classes are the final decision
types of multi-classifications. The main objective of this kind of method is to distinguish steganographic
algorithm types concretely. The disadvantage of this method is that it needs to train several binary classifiers to
distinguish two sets of binary classifiers. When the number of steganographic types increases, the
computational cost of this method is high. The third method is to use one. class Classification. In this method,
the training data set contains only the normal image, but does not use the encrypted image. Then a class of
classifier is trained with the features extracted from the normal image, and the boundary of the normal image
class in the feature space is found by training.

SVM detects steganography in JPEG domain. In addition to OC. SVM, Pevny and Fridrich tried One
Class Neighbor Machine (OC-NM). Compared with the other two methods, this type of method can be trained
independently of steganographic algorithm, but the problem is that the decision boundary of one class of
classification is more difficult to determine than that of two and multi-class classification, so the detection
effect of these methods is not ideal. In addition, the existing content adaptive steganography and other
advanced methods can maintain the statistical characteristics of complex natural images, which will bring
greater challenges to this type of methods.
By analyzing the advantages and disadvantages of various methods, the goal of this chapter is to solve
the problem of binary classification. Even if the detectors that have to be constructed can successfully
distinguish between cryptographic carriers and normal carriers in the face of images embedded in different
steganography algorithms. Among the three methods mentioned above, the performance of multi-
steganographic detector has an important relationship with feature expression, but these methods all depend on
the features extracted manually. In the first and second methods, the steganalysis features corresponding to
different steganographic algorithms are extracted separately, and the related characteristics between different
algorithms are not utilized in the process. With the help of steganalysis research based on in-depth learning,
this work incorporates feature expression, an important module, into the learning process. At the same time,
drawing on the idea of multi-task learning, this paper hopes to use the complementary information between
different steganography algorithms to improve the detection performance and generalization ability of
detectors in the process of feature learning and classification.
6.3 Steganalysis Framework Based on Multitask CNN
This section introduces the proposed steganalysis framework based on Multitask CNN. Firstly, the
training set of multi-steganographic algorithm is constructed, and then the CNN model is trained by multi-task
learning to optimize the binary classification task of normal image/secret image and the classification task of
multi-steganographic algorithm. Steganographic classification task is an auxiliary task. Its purpose is to help
the main task, i.e. image/dense image binary classification task training. The final trained detector can detect
many steganographic algorithms at the same time. How to choose the structure of CNN model and how to
carry out multi-task learning in the whole framework is related to the final detection performance of the
method.

Chapter 7 Summary
With the rapid development and popularization of computer and Internet technology, the problem of
information security on the network has become increasingly prominent. Steganography, as a representative
covert communication technology, can be used illegally by terrorist organizations, spies and criminal gangs
while ensuring communication security, thus endangering national security, public and personal privacy
security and other issues. Therefore, the research of steganalysis technology for detecting and blocking
steganography is of great significance.
Although steganalysis has developed rapidly in recent years, it still faces many difficulties and challenges.
Based on the theory of in-depth learning method and image-based general steganalysis, this paper makes a
deep study of the hot and difficult problem of steganalysis feature expression, and tries to establish a new
steganalysis paradigm from the idea of feature learning. The main contributions of this paper are as follows:
1. An innovative steganalysis paradigm based on in-depth learning is proposed. In the field of
steganalysis, feature expression is the focus and difficulty. Traditional methods depend on artificial design
features. With the challenge of steganography, artificial design features become more and more difficult. This
paper introduces in-depth learning into steganalysis from the perspective of feature learning. Based on the
representative deep learning method CNN, the model is designed and adjusted according to the characteristics
of steganalysis, which makes the proposed model more effective in capturing the statistical characteristics such
as neighborhood correlation Related to steganalysis, and automatically learns the effective expression of
steganalysis features, thus greatly reducing the need for human experience and time. In order to better
understand the working principle of CNN for steganalysis, this paper visualizes the features learned, and finds
that the high-pass-like filter core is more conducive to capture the patterns related to steganographic noise. At
the same time, in order to improve the generalization ability and detection performance of the model, this
paper considers the method of model fusion, trains different CNN models separately, and fuses the output of
the model. The experimental results show the effectiveness of the proposed method. The detection
performance of the proposed method based on CNN is close to that of the traditional feature-based method.
Finally, this paper analyses the performance of CNN using Ensemble Classifier to classify the features learned,
and gets better detection performance by combining CNN features and artificial features.
2. A steganalysis feature learning method based on global information constraints is proposed. Based on
the previous research of steganalysis method based on CNN, this work focuses on the importance of global
information in steganalysis feature expression. By summing up the experience of traditional steganalysis
methods, we have noticed that global statistical information plays an important role in steganalysis feature
expression. In CNN model, the local region information of image is fused by pooling to make use of the
information in a larger range, which is a region-based statistics. There are some limitations in using global
statistical information. In order to make better use of global statistical information in steganalysis feature
learning, this paper proposes to obtain additional global statistical information by calculating auxiliary features,
then introduce global information into CNN model in the form of auxiliary tasks, and regularize the training of
CNN model, so as to promote it to learn better feature expression. In the experiment, representative adaptive
steganography WOW and S-UNIWARD are detected on BOSSbase library to verify the effectiveness of the
proposed method. The results show that the detection performance of the proposed method based on global
constraints is much better than that of the previous CNN-based method.
3. A feature learning method for low embedding rate steganalysis is proposed. Steganalysis for low
embedding rate image detection has always been a challenging problem. Generally, the difference between the
images before and after steganography is very small, and this difference is easily concealed by the difference
between different image content, so it is difficult to detect. As the embedding rate decreases, the number of
image elements that need to be modified in low embedding rate embedding is less, and the change of statistical
characteristics that may be used by steganalysts is less, so it becomes more difficult to detect. Based on the
research work of steganalysis feature learning in the previous chapters, experiments show that the proposed
steganalysis model based on CNN is especially difficult to train and even unable to converge in the face of low
embedding rate steganalysis task. To solve this problem, this work is based on the idea of transfer learning.
First, the network is pre-trained by high embedding rate data sets, and then the features learned are transferred
to low embedding rate detection model for further fine-tuning training. Through this method, the prior
information of high embedding rate image is transferred to low embedding rate image detection, which
improves the detection performance of CNN model for low embedding rate image. In the experiment, the
effectiveness of the proposed method is verified by detecting WOW and UNIWARD steganography
algorithms. The experimental results on BOSSbase database show that the method based on migration learning
significantly improves the detection ability of steganalysis method based on CNN at low embedding rate.
4. A general steganalysis method for multi-class steganography is proposed. The ultimate goal of general
steganalysis is to construct steganalysis detectors without distinguishing specific types of steganalysis
algorithms. However, in the current steganalysis research, most of the methods can not be truly universal. In
the process of constructing the detector, we need to know the specific type of edge information of the
steganography algorithm to be detected, and use these edge information to generate the dense image for
training. From the practical application point of view, the current method can not meet the requirements.
Therefore, it is urgent and significant to study steganalysis detectors for different steganalysis algorithms (even
unknown ones). In this work, the problem is explored and studied. On the basis of previous research work on
Steganalysis Based on in-depth learning, combined with the idea of multi-task learning, this paper attempts to
propose a more general method for multi-steganalysis detection. This method constructs two related
steganalysis tasks in CNN training, that is, the binary classification task to distinguish the normal image from
the carrier image and the multi-classification task to distinguish the type of steganalysis algorithm. The goal of
this work is to hope that "fine-grained" multi-classification tasks can introduce more abundant shared
information into the model, and promote the performance of the two-category directors. In the experiment,
four representative adaptive steganography algorithms are selected to generate a hybrid training set, and multi-
task learning is used to train the known and unknown algorithms in the training set. The results show that the
proposed method based on Multi-task has better performance than the direct hybrid training method.
Declarations
Ethical Approval and Consent to participate: Approved.
Consent for publication: Approved.
Availability of supporting data: We can provide the data.

Competing interests
These no potential competing interests in our paper. And all authors have seen the manuscript
and approved to submit to your journal. We confirm that the content of the manuscript has not
been published or submitted for publication elsewhere.
Funding
This work was supported by the Science and Technology Planning Project of Guangdong
Province under Grant (2017A070709012) , the quality resource sharing course project-
“Computer Network” (Official document by Department of education of Guangdong province
([2015] no. 133)), the provincial-level characteristic specialty-“Network Engineering” and the
provincial teaching team-“Teaching team of basic core course of computer major” (Official
document by Department of education of Guangdong province ([2017]214)), 2018 higher
education teaching reform project of Guangdong province-“Reform and practice of the training
mode of network engineering talents based on the cooperation of school-school and school-
enterprise” (Official document by Department of education of Guangdong province ([2018] no.
180)), and the university-level high-quality resource sharing course project-“Network Security
Technology” and “Wireless Sensor Network and RFID Technology”. And NSFC (No. 61802114,
61802113), Scientific Research Foundation of the Higher Education Institutions of Henan
Province (18A520021, 18A120001), Project supported by the Jiangsu Key Laboratory of Image
and Video Understanding for Social Safety (Nanjing University of Science and Technology
No.30916014107)
Conflict of interest
There is no conflict of interest.

Author’s contributions
All authors take part in the discussion of the work described in this paper. The author Ying Zou
wrote the first version of the paper, and did part experiments of the paper, Ge Zhang initiated the
project . Leian Liu revised the paper in different version of the paper.

Acknowledgements
The authors thank the editor and anonymous reviewers for their helpful comments and valuable
suggestions.

.
References
[1]Johnson, Neil F., and Sushil Jajodia. "Exploring steganography: Seeing the unseen." Computer 31.2
(1998).
[2]Fridrich, Jessica, Miroslav Goljan, and Rui Du. "Detecting LSB steganography in color, and gray-scale
images." IEEE multimedia 8.4 (2001): 22-28.
[3]Cheddad, Abbas, et al. "Digital image steganography: Survey and analysis of current methods." Signal
processing 90.3 (2010): 727-752.
[4] Zhang L, Song M, Liu Z, Liu X, Bu J, Chen C (2013) Probabilistic graphlet cut: exploiting spatial
structure cue for weakly supervised image segmentation. In: Proceedings of the IEEE conference on
computer vision and pattern recognition. IEEE, Portland, pp. 1908–1915
[5]Cachin, Christian. "An information-theoretic model for steganography." International Workshop on
Information Hiding. Springer, Berlin, Heidelberg, 1998.
[6]Marvel, Lisa M., Charles G. Boncelet, and Charles T. Retter. "Spread spectrum image
steganography." IEEE Transactions on image processing 8.8 (1999): 1075-1083.
[7]Artz, Donovan. "Digital steganography: hiding data within data." IEEE Internet computing 5.3 (2001):
75-80.
[8] Junwei Han, Dingwen Zhang, Gong Cheng, Lei Guo, Jinchang Ren. Object detection in optical remote
sensing images based on weakly supervised learning and high-level feature learning. IEEE Transactions
on Geoscience and Remote Sensing, 53(6): 3325-3337, 2015.
[9]Luo, Weiqi, Fangjun Huang, and Jiwu Huang. "Edge adaptive image steganography based on LSB
matching revisited." IEEE Transactions on information forensics and security 5.2 (2010): 201-214.
[10]Zhang L, Hong R, Gao Y, Ji R, Dai Q, Li X (2016) Image Categorization by Learning a Propagated
Graphlet Path. IEEE T-NNLS 27(3):674–685.
[11]Lin, Chang-Chou, and Wen-Hsiang Tsai. "Secret image sharing with steganography and
authentication." Journal of Systems and software 73.3 (2004): 405-414.
[12]Hussain, Mehdi, and Mureed Hussain. "A survey of image steganography techniques." (2013).
[13] Junwei Han, Xiang Ji, Xintao Hu, Dajiang Zhu, Kaiming Li, Xi Jiang, Guangbin Cui, Lei Guo,
Tianming Liu. Representing and retrieving video shots in human-centric brain imaging space. IEEE
Transactions on Image Processing, 22(7): 2723-2736, 2013.
[14] Tuo Zhang, Lei Guo, Kaiming Li, Changfeng Jing, Yan Yin, Dajiang Zhu, Guangbin Cui, Lingjiang Li,
Tianming Liu. Predicting functional cortical ROIs via DTI-derived fiber shape models. Cerebral cortex,
22(4): 854-864, 2012.
[15] Zhang, L., Gao, Y., Xia, Y., Dai, Q., & Li, X. (2015). A fine-grained image categorization system by
cellet-encoded spatial pyramid modeling. IEEE transactions on industrial electronics, 62(1), 564-571.
[16] Dingwen Zhang, Deyu Meng, Junwei Han. Co-saliency detection via a self-paced multiple-instance
learning framework. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(5): 865-878,
2017.
[17] Zhang, L., Gao, Y., Xia, Y., Lu, K., Shen, J., & Ji, R. (2014). Representative discovery of structure
cues for weakly-supervised image segmentation. IEEE Transactions on Multimedia, 16(2), 470-479.
[18] Dingwen Zhang, Junwei Han, Chao Li, Jingdong Wang, Xuelong Li. Detection of Co-salient Objects
by Looking Deep and Wide. International Journal of Computer Vision, 120(2): 215-232, 2016.
[19] Gong Cheng, Peicheng Zhou, Junwei Han. Learning rotation-invariant convolutional neural networks
for object detection in VHR optical remote sensing images. IEEE Transactions on Geoscience and
Remote Sensing, 54(12): 7405-7415, 2016.
[20] Dumitrescu, Sorina, Xiaolin Wu, and Zhe Wang. "Detection of LSB steganography via sample pair
analysis." International Workshop on Information Hiding. Springer, Berlin, Heidelberg, 2002.
[21] Junwei Han, Dingwen Zhang, Xintao Hu, Lei Guo, Jinchang Ren, Feng Wu Background prior-based
salient object detection via deep reconstruction residual. IEEE Transactions on Circuits and Systems for
Video Technology, 25(8): 1309-1321, 2015.
[22] Junwei Han, King Ngi Ngan, Mingjing Li, Hong-Jiang Zhang. Unsupervised extraction of visual
attention objects in color images. IEEE Transactions on Circuits and Systems for Video Technology,
16(1): 141-145, 2006.
[23] Zhang, L., Xia, Y., Ji, R., & Li, X. (2015). Spatial-aware object-level saliency prediction by learning
graphlet hierarchies. IEEE Transactions on Industrial Electronics, 62(2), 1301-1308.

Author details
Zhongkai University of Agriculture and Engineering, No. 501, Zhongkai Rd, Haizhu District,
Guangzhou , School of Computer and Information Engineering, Henan University, Kaifeng
475004,China

Ying Zou was born in Henan, china in 1979. She is currently a lecturer in
Zhongkai University of Agriculture and Engineering, China. She is a member of
China Computer Federation. Her research interests include: network security
technology, IoT technology and machine learning. In recent five years, as the first
author has published 1 paper, chaired three software copyrights.
E-mail: [email protected]
 Ge Zhang received the B.S. and M.S. degrees in computer science from the
Zhongnan University of Economics and Law, Wuhan, China, in 2005 and
2007, respectively, and received the Ph.D. degree from Huazhong University
of Science and Technology, Wuhan, China,in 2015. Now he is an assistant
professor in in School of Computer and Information Engineering, Henan
University, Kaifeng, China. His research interests include image processing,
streaming media, network modeling, and measurement.
E-mail：[email protected]

Leian Liu is an associate professor in computer science and technology, Zhongkai University of
Agriculture and Engineering, China. He received the Ph.D. in Circuits and Systems from South
China University of Technology. He is a member of Guangdong Institute of Electronics and China
Computer Federation. Currently, he is interested in the research areas such as IoT technology,
network security, and embedded technology, etc.
E-mail:[email protected]