Research And Implementation Of Image Steganalysis Based On Deep Learning

Image steganography was a technique which ensured communication security by concealing information within images.Image steganalysis aimed at analyzing steganography techniques to distinguish between cover images and stego images.Most deep learning-based image steganalysis techniques deepen the network by stacking convolutional layers.However,as the resolution of feature maps decreasing during downsampling,a large amount of key detailed information was disappeared,which led to insufficient capture ability of weak steganographic signals.In addition,existing methods mostly used cross-entropy loss as feedback information for backpropagation,but this loss function did not explicitly optimize the features to reduce the similarity between the cover image and stego image,resulting in low discriminative power of the generated features.These problems limited the detection accuracy of image steganalysis.This thesis conducted research around these issues and proposed solutions based on existing image steganalysis techniques.Firstly,to address the issue of inadequate extraction of weak steganographic signals,this thesis proposed a multi-resolution steganalysis network(MRS-Net).MRS-Net designed a multi-resolution module that used parallel convolutions to extract low-dimensional detail information by maintaining a high-resolution branch in the backbone network,and extracted high-dimensional semantic information through parallel lowresolution branches.The multi-scale feature fusion helped shallow information to be better transmitted to deep layers and effectively supplemented the detail information.In addition,to reduce redundant features while extracting diversified features,MRS-Net used attention modules to analyze the sensitivity of different channels and positional information in the image,making the model better adapted to the distribution pattern of steganographic signals and focusing on regions with rich steganographic signals.The experimental results of multiple benchmark experiments on the BOSSBase 1.01 dataset showed that the accuracy of MRS-Net improved by 9.9%and 3.3%compared to YeNet and SRNet,respectively,demonstrating its superior feature extraction capability.Secondly,to enhance the discriminative ability of the model,this thesis proposed a loss optimization method based on metric learning,which designed two penalty terms targeting on steganalysis.The penalty terms measured the similarity between cover images and stego images,then used an optimal distance metric to learn their differences,thus improving the discriminative power of the features.The optimized loss function was guided by labeled information from both the classification task and the penalty terms.To balance the weights of the two tasks,a penalty coefficient was designed to achieve optimal performance of the overall loss.The optimized loss function was incorporated into the training of MRS-Net,utilizing the parallel connections of MRS-Net to penalize multi-scale features simultaneously and compare the differences between features more comprehensively.Under the same experimental environment,MRS-Net achieved the highest accuracy improvement of 2.6%by incorporating the optimized loss function,indicating that the proposed loss optimization method can effectively enhance the discriminative power of the features.