Research On Transform Domain Speech Steganalysis Based On Ensemble Learning

With the development of multimedia technology and computer network technology, steganography became a research focus of information security in internet age. Steganography has broad application prospect in copyright protection, covert communication and many other aspects. Steganography can be used to transfer some baleful messages by outlaws when it is used to protect our information. Thus it is necessary to filter or intercept that baleful information, which is the role of steganalysis. And steganalysis is also useful for steganography, which can help us to design more secure steganographic algorithms, they help each other.In view of the current problem of unsatisfactory effect on transform domain speeches steganalysis, new features are drawn to improve the effect. Firstly, the mel-cepstrum, combined histogram, moments of high-order statistics and co-occurrence matrix are analysed in this paper. Secondly, the new features are drawn from the speeches which are embedded secret messages in transform domain that contains DCT domain, DFT domain and DWT domain. Finally, the features are classified by SVM and the best detection accuracy is95%. The results show that mel-cepstrum, combined histogram and co-occurrence matrix are sensitive to speeches which are embedded secret messages in DCT domain and DWT domain; combined histogram is sensitive to speeches which are embedded secret messages in DFT domain.Because the individual feature carry limited information and single classifier’ classification effect is not ideal, directly the effect of steganalysis is not satisfied. To solve this problem, we did a lot of work. Firstly, based on the principle of information fusion, mel-cepstrum, combined histogram, moments of high-order statistics are fused in two each other. Secondly, the integration features are drawn from the speeches which are embedded secret messages in transform domain. Finally, the features are classified by ensemble learning algorithm and the best detection accuracy is99%which is better than SVM. The results show that the integration feature contains more information and integration feature is conductive to classify. Three kinds of integration features are sensitive to speeches which are embedded secret messages in DCT domain and DWT domain; combined histogram-mel-cepstrum and combined histogram-moments of high-order statistics are sensitive to speeches which are embedded secret messages in DFT domain.