Research And Implementation Of JPEG Steganalysis Based On Fractal Residual Deep Learning Framework

The mainstream format of image transmission in the Internet social platform is JPEG.People send JPEG images on the social platform every day to share their daily life.Hundreds of millions of JPEG image data are transmitted on the Internet every day.Among them,some criminals use steganography to hide reactionary,violent and terrorist information in JPEG.images.As an anti-steganography technique,steganalysis technology is gradually regarded as an important research topic in the field of image content security.In recent years,although the deep learning method has achieved excellent results in the field of steganography detection,there are many problems,such as lack of optimization of depth and width control of the model,poor generalization performance and so on.Based on the deep learning framework of fractal and residual,this paper studies JPEG steganalysis technology with advanced detection performance and generalization from the two directions of balancing network depth and width and feature learning.Based on the above problems,this paper has done the following research:1.Research on the lack of optimization of the depth and width control of the steganalysis model.Existing steganalysis models either deepen the depth of the convolutional neural network or widen the width of the network by parallelizing multiple convolutional neural network subnets to improve the performance of steganalysis.However,these models are either too wide or too deep,making training time-consuming.Therefore,combined with the depth fractal network,a fractal residual network FRNet with depth width control is designed in this paper.In FRNet,a fractal residual module high-pass filtering strategy is proposed as a network preprocessing module to extract JPEG steganographic signals;and a deep feature extraction unit is designed with reference to the downsampling module of ResNet.Compared with the downsampling module designed by the existing steganalysis model,the depth feature extraction unit introduces two convolutional layers with a size of 1 × 1 convolution kernel to combine the feature maps between different channels to further enhance the steganographic signal.The simulation results show that the steganalysis performance of FRNet on the J-UNI WARD steganography dataset with a quality factor of 95 and a payload of 0.3bpnzAC is 11.52%higher than that of J-XuNet and 10.12%higher than that of WangNet;The steganalysis performance on the J-UNIWARD steganography dataset with a quality factor of 95 and a payload of 0.3bpnzAC is 2.54%higher than that of SRNet.2.Research on poor generalization performance of steganalysis models.Existing steganalysis models,such as the FRNet proposed in this paper,have greatly reduced detection performance when performing steganalysis on unknown image sources.Therefore,in order to improve the practicability of the model,this paper proposes a triplet steganalysis network TripletStegNet with reference to the Siamese neural network.This is the first time that the idea of the triplet network has been successfully transferred to the field of JPEG steganalysis.Different from FRNet’s method of directly identifying steganographic signals for classification,TripletStegNet performs classification by learning meaningful distance differences between three sets of input images;TripletStegNet optimizes the Siamese neural network.It adds a feature extraction branch to the Siamese neural network,which allows the network to learn additional meaningful distances,thereby improving the generalization performance of the model.The simulation results confirm that the steganalysis performance of TripletStegN et on the J-UNIWARD steganography dataset with a quality factor of 75 and a payload of 0.1 bpnzAC is 25.25%higher than SRNet,24.39%higher than FRNet,and 6.2%higher than ASSAF;In the mismatch experiment,the performance degradation of TripletStegNet steganalysis is the smallest,only 5.2-6.1%.3.Design and,implementation of an image steganalysis system.The existing JPEG steganalysis system mainly detects specific steganographic algorithms,which cannot meet the needs of accurate detection of unknown image sources on the Internet,making it less practical.In order to prevent the images shared by users on social platforms from being abused by criminals to hide harmful information,this paper designs an image steganalysis system for mobile and WEB terminals by invoking the FRNet and TripletStegNet steganalysis model proposed above.The application supports multiple functions such as uploading images,performing steganalysis on images,issuing inspection reports,and saving historical records,which can conveniently ensure social security.