Image Steganalysis Based On Cross Layer Optimization Convolutional Neural Network

With the rapid development of Internet technology,Great changes have taken place in the way of information exchange,which brings convenience to people and also causes numerous information security issues.As an important direction of information hiding,steganography can realize covert communication without the detection of the third party by modifying redundant carriers such as text,images,and audio in the Internet,effectively guaranteeing communication security.The advancement of Internet technology has promoted the development of steganography,which has brought severe challenges to the research of steganalysis.As a countermeasure against steganography,steganalysis aims to detect whether the carrier contains secret information.How to design an efficient steganalysis detector is of great research value.As traditional steganalysis based on artificial features becomes more difficult with the advent of content adaptive technology.Therefore,in order to solve the above problems,this thesis takes digital image as the carrier,combines deep learning with steganalysis technology,and studies steganalysis based on cross layer optimized convolutional neural network.The main work includes the following two aspects:(1)A model of image steganalysis based on dense connected convolutional neural network is proposed.First of all,in view of the problems of deep convolutional networks such as gradient vanishing and gradient explosion,this thesis introduces the dense connected network into the steganalysis,and uses shortcut connections between the input and output of the convolutional layer to enhance feature transfer,and the reuse of features is realized.Secondly,in the network preprocessing layer,a set of predefined high-pass filters is used to obtain residual features for subsequent network learning,which greatly accelerates the convergence of the network.Then,to further improve the performance of the network,this thesis adds global covariance pooling to obtain the second-order statistics of higher-order features.Finally,through experimental results,it can be obtained that compared with the current advanced CNN steganalysis model,the detection performance of this model has been significantly improved.(2)An image steganalysis based on feedforward attention mechanism is proposed.First,in order to enable the network to achieve richer and more accurate feature expression,this thesis combines the feedforward attention mechanism with a custom CNN architecture to adaptively focus on the most informative areas in the image.Then,this thesis sets the feedforward attention mechanism and the CNN network into two parallel structures.The attention module feeds the threshold value obtained by extracting the CNN network features layer by layer to the CNN main network,so that the network can learn features of different levels,and enhance the classification capabilities of the network.Finally,the experimental results show that the model proposed in this thesis is significantly better than the current CNN steganalysis model.