Research On Image Steganalysis Based On Deep Learning

Image steganography and steganalysis,as research hotspots in the field of information hiding,have received an increasing number of attention from the academic community.Nowadays,the main research results of steganography are based on content adaptive steganography,the main idea of which is to embed the information in the texture area of the cover image to achieve the purpose that it is difficult to be detected.As the anti-detection technology of steganography,steganalysis are mutually reinforced compared with steganography.At present,the research hotspots of steganalysis based on deep learning are mainly focused on passive steganalysis.This article focuses on the passive steganalysis problem of steganography and the problem of steganographic pixel location.At present,in the research of image steganalysis based on deep learning,most of them can not avoid the loss of steganographic signal as the network is deepened,as well as need traditional steganalysis methods to add prior knowledge of the selected channel.This paper proposes an image passive steganalysis network SERGO?NET which combines residual learning and channel modeling.In the pre-processing layer of the network,combining SRM filter with specific weight and Bayar filter which learns weights automatically to extract the image residual noise image.In the feature extraction layer,a SER?BN residual module is proposed.This module is used to avoid the loss problem of steganographic information,and the fusion attention mechanism can further expand the weight of the effective channel,so that the accuracy of detection is promote.In the experiment,based on the BOSSbase v1.01 data source,the content adaptive steganography S-UNIWARD is used to set the embedding rate from 0.1bpp to 0.4bpp to obtain images with different payloads.Compared with the passive steganalysis network Yenet,using the network SERGO?NET in this paper,the detection accuracy is improved by 2% to 3%.Since passive steganalysis can only judge whether there contain embedded information in the image,it cannot identify the location of the embedded information,in order to further enhance the practicability of image steganalysis.This paper expands the research goal of image steganalysis as steganography location of adaptive steganography and non-adaptive steganography LSB matching,and proposes an end-to-end image steganography location network PSL?NET.An image is input at the input end,and the steganography location image of this image is obtained at the output end.In the preprocessing layer,the SRM filter is used to extract the residual noise image.In the deep residual layer,deep residual learning is used to enhance the expression ability of steganographic features.In the pixel prediction layer,using the mask image which marked the actual position of the pixel to learn supervisedly and strengthen the perception of network for local steganographic pixels,treating smooth or textured pixels without discrimination,predict the probability of each pixel of the image whether has been steganographic.Experiments on the BOSSbase v1.01 data source show that the network PSL?NET described in this article can detect the steganographic pixels of the steganographic images embedded by content adaptive steganography S-UNIWARD at different payloads,and also can detect the steganographic pixels of steganographic image embedded by non-content adaptive steganography LSB matching.