Research On Blind Steganalysis Of JPEG Images In Real World

As one of the most important techniques of protecting information security, steganalysis is always a hot research area in the field of multimedia information security. After many years of research, fruitful achievements have been made in JPEG blind steganalysis techniques. However, current blind steganalysis methods may not work well in real world. This is because that the real-world steganalysis is more complicated, the statistical characteristics of images are more various, and the image amount is more excessive. All above reasons usually cause mismatch and large computation problem of current steganalyzer. As a result, the practical use of current blind steganalysis methods has been limited. Therefore, it is significant to make researches on the blind steganalysis techniques for JPEG images which can be applied into the real world.Based on the analysis of the advantages and disadvantages of current JPEG blind steganalysis methods and the distribution characteristics of blind steganalysis features, this thesis focuses on the study of JPEG blind steganalysis framework, rapid detection methods for large-scale steganalysis, and steganalysis methods for key guilty JPEG images. The main contributions of this thesis are summarized as follows:1.The basal concept of information hiding is introduced. The basic theory, research status, and the development prospect of steganography and steganalysis are summarized. Especially the current blind steganalysis techniques are classified from the perspective of machine learning and the advantages and disadvantages of each classification are summarized.2. For solving the problem that current steganalysis methods based on supervised learning always encounter embedding algorithm mismatch and cover source mismatch problems, a blind steganalysis framework for JPEG image combining with the semi-supervised learning and soft margin SVM is proposed. The proposed framework contains two detection phases, and utilizes the information in test set for improving detection performance. Reliable blind steganalysis for JPEG image is realized only using cover images for training. The experimental results show that the proposed framework can contribute to improving the detection accuracy of steganalysis detector based on one-class classifier, and have good robustness under different cover source mismatch conditions.3. For solving the problem that current blind steganalysis methods cannot work well in large-scale steganalysis, a rapid steganalysis framework for large-scale JPEG steganalysis problem based on outlier detection is proposed. First, dimension reduction is performed on the high-dimensional blind steganalysis features to reduce the computation complexity. Second, two methods are designed to measure the outlier degree according to different possible scenarios of the stego samples number in the test set. By this way, the cover source mismatch problem is circumvented. At last, a method to estimate the number of outliers among the test sample set is designed to identify the stego samples. The proposed method does not need any training process, thereby avoiding mismatch problem between the training and test samples and significantly reducing computation complexity. The experimental results show that the proposed method is superior to the conventional method in detection performance and speed in most cases.4. For solving the blind steganalysis issue of key guilty image, a JPEG steganalysis framework combining the retrieval of image-inherent statistical properties and unsupervised outlier detection is proposed. First, cover images with similar image-inherent statistical properties to the test image are searched from massive image database to establish an aided sample set. Outlier detection is then performed in a test set composed of the test image and its aided sample set to judge the type of the test image. The experimental results show that the proposed features can measure the image-inherent statistical properties of JPEG images well, and the whole detection paradigm demonstrates better performance than the steganalysis strategy using a mixed image set for training.Finally, the research work of this thesis is concluded and further research work is proposed.