Research On Image Steganalysis Techniques Based On Fusion

The technology of image steganalysis plays an important role in the field of informationsecurity and becomes very important. In recent years, a number of detection techniques forimage steganography have achieved fruitful research results, which exhibit excellent detectionaccuracy under the laboratory environments. However, the existing steganalysis algorithms aredifficult to obtain high detection accuracy when applied to the Internet conditions, in which boththe embedding algorithm and the secret message length are unknown, and the image sourceshave a variety of image content and texture. Therefore, it is of great significance to makeresearches on the steganalysis of these conditions.Combined with the information fusion techniques, the thesis focuses on the study ofsteganalysis of these conditions separately. The main contributions of this thesis are summarizedas follows:1. Information fusion techniques and its application in image steganalysis are discussed.Firstly, the fundamental conception of information fusion techniques is presented. Then, severaltypical decision level fusion methods are discussed in detail. Finally, the state-of-the-art of imagesteganalysis based on fusion is outlined, and the frame of the image steganalysis techniquesbased on fusion is proposed.2. Most existing steganalysis techniques assume that the embedding rate of test image isknown. However, in practice, the embedding rate of test image is unknown. To address thisproblem, an image steganalysis scheme based on Dempster-Shafer evidence theory is proposed.Firstly, various classified results are acquired by using the multi-rate classifiers established in thetraining phase. Secondly, these classified results are converted to evidences, and then, theweighted coefficients are assigned to these classified results based on the missed detection ratesand the false alarm rates of different classifiers. Finally, the decision is obtained byDempster-Shafer evidence theory based on weighted coefficients. Experimental results show thatour method improves detection accuracy under the rate-unknown. Moreover, the proposedmethod could achive the request of practice application through adjusting the accuracy and falsealarm.3. The existing detection algorithms are difficult to obtain high detection accuracy when theembedding algorithm of the stego-images is unknown. To deal with this condition, we attempt topropose an image steganalysis scheme based on boosting fusion. We can first obtain variousclassifying results by establishing multi-embedding algorithm training models, and then acquirethe performance of these classifies according to the boosting algorithm. In the end, the final decision-making result is obtained by combinational rule based on weighted coefficients. Thedetection work is presented to attack the current spatial domain and JPEG steganographicmethods. Extensive experimental results show that our proposed method can effectively detectthe unknown steganography methods. The performance of detection method using boostingfusion is better than the method without boosting fusion.4. Most of the current blind detection techniques do not consider the influence of imagecontents on the steganalysis performance. In this thesis, an image steganalysis based on imagecontent and fuzzy integral fusion is proposed. In the training phase of this proposed method, theinput images are first divided into several classes according to image complexity, each class istrained separately and the fuzzy measure for each class is calculated. In the testing phase, the testimage is sent to the well-trained classifiers to acquire the local decision-making values, and then,these values are fused by fuzzy integral to obtain the final distinguishing results. Experimentalresults on several sets of images demonstrate that the proposed steganalyzer significantlyenhance the detection accuracy of prior art.Finally, the research work in this thesis is summarized and the further research directions offusion steganalysis are discussed.