Image Steganalysis Of Deep Convolutional Network Based On Network Structure Enhancement

With the continuous advancement of Internet technology and digital technology,information security in the communication process has attracted more attention than ever before.Steganalysis,as a means against steganography,aims to determine whether a given object contains secret information,which is of great research value.Image steganalysis methods based on artificial features are time-consuming and labor-intensive,and there are many difficulties.Now,how to improve the accuracy of steganalysis has been a major research hotspot in the field of information hiding.Thanks to the development of deep learning,steganalysis has achieved great progress and made significant breakthroughs.Taking digital image as the research object,this thesis studies the feature representation and learning of image steganalysis of deep convolutional network based on network structure enhancement.The main work is as follows:(1)The thesis proposes an image steganalysis method based on two-way convolutional neural network.The new network combines two sub-networks to improve the generalization and representation ability of the network.Then,the design of convolution kernel of preprocessing layer is closely related to effective feature representation.In this model,highpass filters and random values are combined to initialize the weights of preprocessing layer in two sub-networks,aiming at extracting comprehensive features and achieving global optimization in a unified network.In order to highlight the global and accuracy of the features,average pooling and maximum pooling are used to the two sub-networks respectively.The experimental results demonstrate that the model designed in this thesis achieves a considerable improvement compared with the existing advanced CNN models,and has obvious advantages over the traditional steganalysis methods based on artificial features.(2)This thesis designs an image steganalysis model based on enhanced residual network.To overcome the difficulties of gradient vanishing and degradation in deep networks,the model takes residual network as the basic structure.In the preprocessing module,the network uses high-pass filters for initialization and truncation linear unit as the activation function,which can obtain residual features,eliminate extreme data and control the feature data within a certain range.To effectively enhance the network structure.This thesis adds a convolutional block attention module,which is a hybrid domain attention mechanism combining channel and space.It helps the model to capture significant features more effectively and suppress unnecessary features.Through experiments,we can get the model proposed in this thesis achieves good performance.