Study On JPEG Image Steganalysis

Steganograhy is to embed unperceivable data in Medias of images, audio or video. Steganalysis is the art and science of detecting the existence of such suspicious information by various means. Some kinds of Medias, such as JPEG and GIF images, are very likely to be used for steganography, since there are used popularly and convenient to spread and circulate. Techniques of attacking such media are very important; this paper mainly aims at to detect JPEG images.The study of steganalysis mainly consists feature extraction and classifier selection and design. In recent, Shi proposed a Markov based feature extraction approach, and it detection rate outperformed the existing steganalyzers of JPEG.This paper's contribution mainly concentrates on the following two parts:First, we improved that Markov based feature extraction approach. We tried to improve that approach in the following four ways. We experimented JPEG 1-D and 2-D array, higher order difference of JPEG DCT coefficients, extracting features from intra-block and inter-block and separating the DCT coefficients to several groups. The experiments demonstrated that using intra-block JPEG 2-D array and second order difference can significantly improve detecting rate.Second, we have proven the equivalence between a new implementation of kernel canonical correlation analysis and traditional implementation with less condition, and use it as the classifier of steganalysis, the effect is better than support vector machines (SVM). The traditional kernel canonical correlation analysis works on the kernel spectrum-based feature space. In recent, Huang proposed that it can implementation kernel canonical correlation analysis on the kernel generated Hilbert space under certain condition, and gave a proof based on the statistical theory. This paper gives another method to prove it, and this method requires less limiting condition. SVM have been widely used in steganalysis, for its convenience and high performance. We have extracted 18 kinds of feature based on Markov process, and classify them with SVM. For the best feature, we also use the classifier based on KCCA to reclassify it. The experimental work shows that the KCCA based classier outperform SVM, especially when the embedding rate is low.