Research On Image Blind Steganalysis Methods

In recent years, terrorist activities has become increasingly frequent, the image steganography is often used by terrorists to transmit information and instruction. It is proved that the steganography technology has been applied to achieve illegal intent, which has caused great harm to the state and social security. It deeply stimulated the research and application of image steganalysis which is one of the secret confrontation technologies. Image universal steganalysis, which is an important branch of image steganalysis, is becoming the research emphasis of image steganalysis. In this thesis, we research on the key problems on image universal steganalysis, and achieved creative improvements in methodology, which involve each layer of analysis process, including image layer, feature layer, classifier layer and decision layer, the contributions of our work are as follows:Firstly, the mechanism of image universal steganalysis was analyzed. To solve the problems that principle framework of former image steganalysis is too general to describe the mechanism of current steganalysis, the paper made optimization on the principle framework, combined with the fact of the application, and provide a relative concrete principle, theory and theoretical model which include feature extraction, feature pretreatment, classification, and evaluation. This part of the work is the most important components in the framework and can provide theoretical support for the research on image universal steganalysis.Secondly, to solve the problem that the content of the image will affect the accuracy of steganalysis, we proposed a steganalysis method based on sample pretreatment, designed a semi-blind and blind analysis algorithms based on clustering, and analyzed the impact of the sample content to the accuracy of steganalysis by experiments. Experiment results shows that our methods can reduce the image content impact on detection accuracy and enhance the accuracy of steganalysis in the view of image layer optimization.Thirdly, to solve the problem that the uncertain informations will affect the accuracy of steganalysis, we proposed a steganalysis method based on uncertainty reasoning, take steganalysis as a kind of uncertain problem, constructed a steganalysis algorithm framework based on evidence reasoning, and proposed a semi-blind and blind steganalysis algorithm based on uncertainty reasoning. Experiment results shows that our method reduced the impact of uncertainty on the steganalysis accuracy and enhanced the steganalysis accuracy in the view of decision layer optimization.Fourthly, to solve the problem that the single domain and single feature can not reflect the complete characteristic of the new complex steganographics, we proposed a steganalysis method based on multi-angle feature, and proposed a semi-blind steganalysis algorithm based on symmetric cross entropy and a blind steganalysis algorithm. Experiment results shows that our methods can ensure the enrichment of feature sets at a low dimension and enhanced the accuracy of steganalysis in the view of feature layer optimization.Finally, To solve the problem that enhancement of the steganalysis generalization ability in single nonlinear classification system was limited, we proposed a full blind steganalysis method based on ensemble linear classifier, which can improve the generalization ability and enhance the accuracy of steganalysis in the view of classifier layer optimization.