Research Of Image Steganalysis Based On Multi-domain Features

With the rapid development of computer science and network technology,Information security has drawn more and more attention. Steganography andsteganalysis are two important branches of information security, which restrict andpromote each other. Steganography will embed secret information in the cover files toform the stego files，in order to protect the cover files or hide the secret information.On the contrary，the main purpose of steganalysis is to detect whether to hide secretinformation or not. It is an important means to prevent illegal secure communication,but also is a measure to evaluate the information hiding algorithm security.This dissertation firstly introduces the basic theory and development of securecommunication and information hiding, a survey on the concept and the evaluatingindicators of steganography and steganalysis are presented subsequently, then wefocus on the recent development of steganalysis and point out the current problems.Based on the above analysis, this dissertation revolves around the image steganalysisbased on multi-domain features to conduct research; the main contents of thisdissertation are summarized as follows:An universal steganalysis method for JPEG images is proposed in thisdissertation. The proposed algorithm can achieve high detection correct ratio bycombining various statistical features, including DCT, spatial domain, and DWTfeatures. The combined statistical features can reveal the image difference betterbetween an original cover image and its stego-image. Experiments have beenconducted to show the effectiveness of the proposed method.Then, on the basis of the above study, in order to obtain higher detection ratiosand faster training speed for image steganalysis, a new steganalysis algorithm basedon Extreme Learning Machine (ELM) is also presented by combining with singlehidden layer feedforward neural networks (SLFN). The proposed algorithm firstextract some features in discrete cosine transform (DCT) and spatial domain from a JPEG image according to Pevny T’s algorithm, then the PCA is utilized to reduce theoriginal193-dimensional features and finally, ELM is used as classifying techniqueto construct a blindly steganalysis algorithm. The experimental results show thatcompared with SVM, ELM has faster learning speed and similar classificationaccuracy because it has smaller number of turning parameters and less number ofneurons. That is, ELM can be used in a blind steganalysis for all kinds of JPEGimages.