Research On Digital Image Steganalysis In Network Environment

In recent years,the network security situation has become more and more serious.As a new tool for multi-media security,digital image steganography plays an increasingly important role in the process of information transmission.As a reverse analysis technology,steganalysis has made significant progress after years of research.These detection algorithms can be incredibly sensitive and accurate in the laboratory environment.But when applied to the actual network environment,the image characteristics are complex and the detector has no information about the embedding algorithm and payload.This condition inevitably results in the so-called model mismatch problem and the detection performance will degrade,often dramatically.The emergence of large image data provides us with a practical solution.In the current network environment,it is easy to obtain a large number of images which come from different sources or have different sizes and processing history.It can solve the traditional mismatch problem by retrieving cover images form the data set with the same prior probabilistic distribution.This paper studies the unsupervised universal steganalysis methods combined with image retrieval and outlier detection.In addition,the deep learning technique has been applied to the field of steganalysis in recent years as a useful tool to solve complex problems.Based on the powerful data processing,feature extraction,this paper studies a supervised steganalysis algorithm by deep learning networks.The main contributions of this thesis are summarized as follows:1.The basic concepts and research status of information hiding,digital image steganography and steganalysis are discussed.The existing steganalysis methods are summarized in the perspective of machine learning.The problems of steganalysis when applied to the network environment are analyzed and the solution are given respectively.2.In order to solve the cover source mismatch problem,an unsupervised universal steganalysis method based on image source retrieval and outlier detection is proposed.Firstly,the imaging principle of digital camera and the inherent statistical characteristics is briefly introduced.Then,the camera source identification algorithm and the traditional low-dimensional outlier detection algorithm are introduced respectively.The detection framework is provided.Finally,the performance of the proposed algorithm is compared with the traditional method based on one-class classifier.The experiment results show that the algorithm proposed in this paper behaves better than the one-class classifier based method,which not only eliminates the influence of cover source mismatch,but also can detect multiple steganographic algorithms.3.In order to solve the problem about steganalysis on original,smoothing and sharping heterogeneous images,an unsupervised universal steganalysis method based on image source retrieval and outlier detection is proposed.Firstly,two commonly used image filtering algorithms are briefly introduced and the characteristics of image filtering are discussed.We use the outlier detection methods suitable for high dimensional space to solve the so-called “curse of dimension” problem.The experiment results show that the performance of the proposed method is better than traditional outlier detection method.However,due to the high dimension of rich model feature,the extraction process and the outlier detection can be really time consuming,so the detection efficiency can be relative low.4.In order to solve the problem of high dimension of rich model features,a supervised steganalysis algorithm based on deep learning network is proposed in this thesis.Firstly,the basic structure of convolutional neural network and the rich model features for JPEG image are introduced.Then we design a deep learning network suitable for rich model features and the network structure and parameter setting are given.Accounting for the powerful data processing and GPU parallel computing ability,it is possible to solve the problem of high dimension and increasing the number of training samples can improve the steganalysis performance to a certain extent.The proposed method performs better than ensemble classifier and it is proved that the deep learning can be applied to the field of steganalysis.In the end,the research work of this thesis is summarized and the further research direction about steganalysis in the real world scenario is discussed.