Research On Lightweight Image Steganalysis Based On Convolutional Neural Network

Steganography and steganalysis are currently important research hotspots in the field of information hiding.The misuse of steganography causes a lot of security risks,such as:commercial criminals use steganography to accomplish covert communication to achieve information theft.Image steganography is a technique to embed secret information by modifying the complex area of digital images to achieve the purpose of stealth communication,but the special characteristics of steganography lead to its possible use in illegal fields,so the study of steganalysis has important research significance and practical value for maintaining information security.Steganalysis is classified into two types according to its technical fundamentals: traditional steganalysis methods and deep steganalysis methods.Traditional steganalysis models due to their need for a significant amount of illuminating knowledge and design filtering kernels,while deep learning-based steganalysis approaches use the tremendously powerful characterization learning capability of the network to extract statistical features of images and use a huge number of learnable parameters to agent manual design.Therefore,to solve the deployment dimension problems of existing deep steganalysis methods such as high reduction and slow convergence,this paper aims to study various lightweight deep learning networks and a priori knowledge decision theory to construct a lightweight steganalysis model for images based on deep learning,which further reduces the parameters in the network and the network fitting time while ensuring the model detection accuracy.The research of this paper is as follows.(1)In order to achieve fast fitting of steganalysis and reduce convergence time,a lightweight steganalysis model(CNet)based on embedding probability map is proposed for the first time to achieve faster convergence speed than other models with fewer parameters.Firstly,a semantic segmentation model based on pixel level is used and applied to extract the embedding probability map of the carrier image;then,a pseudo-twin network mechanism is used to perform parallel feature extraction on the dual-stream input to achieve the reduction of the number of parameters and input the obtained feature map into the SE-Block attention mechanism;finally,the embedding probability map corresponding to the image under the experimental demonstration helps the network to acquire prior knowledge and achieve convergence faster.Finally,it is experimentally demonstrated that the embedding probability map corresponding to the images helps the network to acquire prior knowledge faster and achieve convergence with 50% lower number of parameters than the same level model.(2)In order to solve the problem of redundant parameters of existing steganalysis,an image lightweight steganalysis model(Light Net)based on multi-order statistical discrepancy and combined with lightweight network structure is proposed to significantly reduce model parameters while ensuring detection accuracy on par with other models.First,the model is improved by studying multi-level residual convolution blocks to realize the combination of different size perceptual fields and improve the steganographic noise extraction ability of the model;Then,the multi-order pooling scheme is investigated,and the feature maps after multiscale residual block operations are put into the first-order(average pooling)and second-order(covariance pooling)statistical layers,and are placed into the final classifying layer to auxiliary the model for classification.In terms of experimental results,the steganalysis suggested in this paper can maintain the identical detection accuracy with only 1?2% of the number of parameters of other models.