Research On Image Steganalysis Model Based On Deep Learning

The security issues in the data era are becoming increasingly prominent,and steganography and steganalysis are two important research directions of information security.Steganography refers to hide secret information in carriers such as digital media,which is conducive to safe transmission of information and covert communication.However,because of the visual nature of digital images and the redundancy of coding,they are often used by attackers to hide and steal confidential information or transmit terrorist information,which makes steganography become a means of attack and endangers national security and social stability.Steganalysis plays an important role in preventing malicious steganography.Traditional steganalysis faces difficulties in feature extraction and large amount of calculation.Deep learning can use network learning for feature extraction,which has become a new research method in the field of steganalysis.In this thesis,image steganalysis based on deep learning is studied,and two steganalysis methods are proposed.The main work is as follows:(1)To extract steganographic signals well and enhance the feature extraction ability of the network,an improved deep residual steganalysis model combining attention mechanism is proposed.The model’s pre-processing layer uses SRM to obtain noise residue and proposes an activation strategy that can widen the relative values of residual feature maps to distinguish features better without changing the overall data distribution range.The feature extraction part is composed of depth separable convolutional layers and residual network blocks that can enhance the feature extraction ability.In addition,a channel attention mechanism is introduced in the residual block to enhance the model’s learning ability of steganographic signals and further improve the detection accuracy.Under the S-UNIWARD steganographic algorithm with an embedding rate of 0.2bpp,the detection accuracy reaches 79.3%,which is 5.7% higher than the GBRAS-Net model.(2)To address the problem of incomplete residual feature extraction by the filter in the preprocessing stage,a residual feature channel recombination method is proposed to enhance the pre-processing stage’s ability to extract key features.First,the SRM filter is used to obtain the residual feature,and recombing the incomplete residual feature channels to enhance the feature.Then,the residual feature map adjusts the weights adaptively during the training process in order to relocate weights in recombined channels and input them into the feature extraction module,which can enhance the network’s learning of key features.Finally,a neural network model is constructed for verification,and the residual feature recombination module is also applied to the existing model.Experiments show that the network with the introduced residual feature recombination has better detection accuracy.This research fully utilized the knowledge of steganography and the existing foundation of steganalysis models to achieve the goal of constructing a deep learning-based image steganalysis model.Through experiments verified that the proposed model has better detection accuracy,which has certain theoretical significance and practical application value.