Research On Universal Steganalysis Technology Based On Deep Learning

With the rapid development of Internet technology,multimedia information security has become an important research topic.Digital steganography is a kind of technique for secure communication by hiding secret information into digital objects.However,it can also be abused for illegal purposes,so the study of steganalysis is of great significance.In recent years,advanced steganography has brought more serious challenges to steganalysis.In this thesis,we focus on steganalysis,and tackle the problem of feature representation in the view of feature learning.The main work is as follows:A method based on highly modularized convolutional neural network is proposed to improve the detection accuracy of steganography in small embedding rate.Firstly,the fundamental network is built by repeating residual network units to extract the complex statistical properties of images.Then,extracting the channel information of residual image by adding the group convolution,it is very good to strengthen the signal characteristics from the hidden information.Finally,a large number of datasets are used to train the network.Experimental results show that the proposed method compared with the existing methods has better performance,and extract more effective image features.Meanwhile,using the residual network module as the template,the network model can be easily built to facilitate the adjustment and training.In order to improve the accuracy of detection,we propose a method based on the multiple activation module and residual structure.According to the speciality of the content adaptive steganography,we adopts multiple activation functions to capture more traces left by steganographic embedding and increase the diversity of features.Experiments show that this method is effective in detecting content adaptive steganography.