Universial Steganalysis Based On Images

Steganography is the art of secret communication. It encompasses methods of transmitting secret messages through innocuous cover carriers in such a manner that the very existence of the embedded messages is undetectable. As the counter part, steganalysis is one of the most important parts of detecting hidden information, and it focuses on multimedia such as image, audio and video. By observing the multimedia, steganalysis tries to detect the hidden messages without any knowledge of the steganography algorithms and secret messages. Nowadays, a lot of steganalysis algorithms have been proposed. According to the application's targets, those algorithms can be classified into two classes: specific steganalysis and universal steganalysis. Most of them are specific algorithms and usually they have strong pertinence and thus a high rate of detection, but considering the diversity of the current steganography tools, those algorithms are much restricted in the application. As for universal steganalysis, they can be trained to detect hidden messages embedded by almost any methods, so they have a wide range of application, but currently the correct detection rate is not as good as specific steganalysis.In this thesis, we first study the existing specific steganalysis, and make improvement on one specific steganalysis algorithm---RS algorithm. Based on that, we further analysis the universal steganalysis, through the methods of multivariate analysis (ANOVA), we identify 4 specific IQMs that are most consistent and accurate vis-à-vis the effects of steganography out of the 26 IQMs, then using BP neural networks to carry out the machine learning, and finally applied it to classify the stego images. Experiment results indicate that this method has a wide range of application. Finally, we combine this with the specific steganalysis, which improves the rate of detection greatly.