Research On JPEG Image Mismatched Steganalysis

With the development of Internet communication and mutilmedia process technologies, steganography aiming to establish covert communication between trusting parties has been concernend by the public. Steganography refers to the technology of hiding secrect information in the other media covers and transmitting secrect information on the Internet without awareness of observers. While people get benefit from steganography, criminals use it to transmit secrect information threatening social security. Therefore, steganalysis, a technology to recognize stego samples from cover samples on the Internet is of great significance.Tranditional steganalysis relies on the assumption that steganalyst has obtained cover samples and stego samples from steganographier. In the practical steganalysis application we cannot always assure this assumption and the training dataset and testing dataset are usually mismatched. Currently, though many papers have pointed out the traditional steganalysis exhibits a lower detection accuracy in a mismatched condition, only a small amount of works attempted to mitigate the impact of mismatches. This paper studies the effect that each mismatched factor has on traditional steganalysis feature set in the eye of machine learning. For different steganalysis environments we propose a pool training mismatched steganalysis method based on local domain generalization and generalized transfer component analysis mismatched steganalysis method based on transfer learning.(1) Firstly, the traditional steganalysis is introduced in detail including several steganalysis feature sets and machine learning tools in steganalysis. Secondly, different mismatchend factors that have effect on traditional steganalysis are analyzed through a mismatched steganalysis framework. Finally, related mismatched steganalysis methods are introduced by category.(2) This paper proposes a pool training mismatched steganalysis method based on local domain generalization. This method introduces the concept of local domain of the test image and extracts the shared componet from traditional steganalysis feature set across the local domain by minizing the variance of local domain feature distribution, and maintaining a partial correlation between the training data and the label set at the same time. With the shared component, the mismatches can be reduced. Compared with previous pool training methods our method is more effective.(3) As the dependence of pool training mismatched steanalysis methods on diverse of training data, we propose generalized transfer component analysis for mismatched steganalysis based on transfer learning. This method adaptively correct a limited amount of training data features to reduce the mismatches between training data and testing data making it robust to different mismatched factors. Comparison with the previous methods our method can be effective when training set is small and not diverse enough in many mismatched conditions.