Research And Implementation Of Joint Steganalysis Model Based On Deep Learning

In recent years,network security has received great attention,among which steganography and steganalysis techniques have been constantly improving in mutual confrontation.Steganography is responsible for embedding secret information,and steganalysis is responsible for detecting the existence of secret information.With the development of adaptive steganography,steganalysis research tasks become more challenging.The rapid development of deep learning has brought new research ideas to steganalysis.The thesis focuses on the problem of image steganalysis in deep learning,mainly including the following 3 parts:(1)Research on the multi-image domain joint steganalysis of grayscale images.Existing gray-scale image steganalysis work only focuses on spatial or transform domain features in isolation,but in fact the interference of steganography will be reflected in both spatial and transform domain.Based on this,a cross-domain detection mechanism is proposed,which simulates the spatial rich model(SRM)and the transform domain discrete cosine transform residual feature extraction method to jointly extract the steganographic signals in the spatial and transform domains.At the same time,nonlinear detection mechanism is proposed to simulate the MinMax method of spatial SRM in order to adapt to the nonlinear embedding state of steganographic signals.In order to enhance the sensitivity of the model to the steganographic signals,a transfer learning method is proposed to initialize the model network with low embedding rate image set.Finally,the gray image joint steganalysis model Wang-Net is designed and implemented.The simulation results show that Wang-Net performed well in WOW steganography with an embedding rate of 0.2,S-UNIWARD steganography with an embedding rate of 0.2,WOW steganography with an embedding rate of 0.4,and S-UNIWARD steganography with an embedding rate of 0.4,whose accuracy rates of under-operative steganography detection were 81.2%,77.7%,92.0%,and 88.8%,which were 4.6%,5.8%,3.8%,and 4.1%higher than that of Zhu-Net.(2)Research the color channel joint steganalysis of color images.In view of the current color image steganalysis work that does not consider the steganographic traces between color channels,a cross-channel detection mechanism is proposed.We design the vector feature extraction strategy(VFES),multi plane feature extraction method 1(MpFES1)and multi plane feature extraction method 2(MpFES2).In VFES,the color image pixel vector of red,green and blue is used as the basic unit to jointly extract the steganographic residuals in color channels and between color channels.In MpFES1,the plane along X and Y axes is used as the basic unit to jointly extract the steganographic residuals in spatial multi planes.In MpFES2,the spatial plane of xoy,xoz and yoz is used as the basic unit to jointly extract the steganographic residuals within and between color channels of color images.The simulation results show that the average steganography detection accuracy of MPFES2 model was 92.6%,which was about 4.0%and 1.0%higher than that of the conventional feature extraction strategy(CFES)and the channel feature extraction strategy(ScFES),respectively.Although the average steganography detection performance of VFES model is slightly worse than ScFES,it is better than CFES,and we can research the preservation method of spatial structure of pixel color component residuals in the future.Although the average steganography detection performance of MpFES1 model is poor,the detection accuracy of its new vector steganography algorithm is more than 4.5%higher than that of channel by channel steganography algorithm,and we can study the weakening method of interference cause by the same order residual sum operation in the future.In summary,DRN-S has the best steganalysis ability among them,which is used as the joint steganalysis model of color image.(3)Design and implement a joint steganalysis platform,which not only combines the functions of gray image and color image steganalysis,but also jointly extracts the spatial and transform domain steganographic signals in the gray image steganalysis system,and jointly extracts the steganographic signals in and between color channels.The simulation results show that this joint steganalysis platform could perform the detection functions of gray-scale WOW steganography,gray-scale S-UNIWARD steganography,color image channel-by-channel steganography,color image vector steganography,which had certain application value.