Multi-Feature Fusion Methods For Universal Steganalysis In Images

With the rapid development of computer science and technology, steganography technology has been widely used in many fields. However, due to its ease of use, low cost and other reasons, therefore it’s easy to find that steganography is often used in bad faith inevitably. Steganalysis does exactly the opposite. It can play a role to prevent malicious purposes by detect whether carriers contain any secret information. Universal steganalysis is a pattern recognition process. The traditional universal steganalysis does the detection by use of single feature and the detection result is often less than ideal. Today, many scholars introduce the idea of feature fusion to fuse multiple classical features in order to achieve better performance. However, there are still some problems in fusion detection for the current universal steganalysis, such as how to choose the basic features for fusion; how to take full advantage of the basic features of information to obtain better fusion detection result is still unsolved and so is how to make universal steganalysis able to handle mass data. In this paper, we focus on JPEG format images and consider how to do multi-feature fusion in universal steganalysis based on feature-level fusion and decision-level fusion. The main work of this thesis includes:1. Consider how to select a superior features combination to do feature fusion in order to achieve better steganalysis effect, this paper draws on the idea of Huffman tree and use correlation coefficient of canonical correlation analysis as its heuristic factor, proposed a method which in tree structure to select a superior features combination in multi-feature fusion blind detection in the image steganography and apply this method to the multi-feature fusion universal steganalysis. Experiments indicate that our method can get a features combination which holds superior fusion performance by fuse the basic features of the features combination serially in limited computing complexity condition when we deal with steganography blind detection for JPEG images.2. In the condition of given many features, this paper propose a multi-feature fusion for universal steganalysis based on the improved version of AdaBoostSVM. Compared to the original AdaBoostSVM, our method makes full use of the weak classifiers which are generated during the intermediate process, and improve the rules to update sample weights. Experiments indicate that the method can get better detection effect when it deals with multiple steganography and various embedding rates compared to three kinds of single features, serial fusion method and original AdaBoostSVM method. At the same time, we figure out the impact of the different parameters have on the final detection effect through the experiments.3. In order to deal with mass data, we propose a parallelizable multi-feature fusion method based on AdaBoostSVM, and show how to apply the method in MapReduce framework. The method divide the original training set, make subset executed in parallel on multiple processors, and assemble all the result to get the final classify results. Contrast experiments, as well as the parallel performance tests in multi-machine Hadoop cluster, show that the method can obtain a good parallel performance and won’t lower the detection rate significantly, which makes the method can handle mass data.Finally, a comprehensive comparison was made among three proposed multi-feature fusion in universal steganalysis method and explain the applications scenarios of each method.