Image Steganalysis Algorithm Based On Image Block Regression Learning And Deep Residual Network

Image steganography is to embed the secret information into the image on the premise of meeting the imperceptibility.However,while realizing the secret communication,it is also likely to be used by criminals,thus endangering national security.Image steganalysis is the opposite of image steganography.Its main purpose is to detect whether the image transmitted in the channel has secret information to prevent crimes.Therefore,the research on image steganalysis is of great significance.In recent years,CNN has occupied an advantage in image steganalysis with its superior performance,among which the algorithm of selecting channel information as auxiliary information has achieved a good momentum of development.However,this method also has some shortcomings: the extraction of channel information is time-consuming;the sample imbalance of learning channel information with pixel as branch task;the false detection rate is still high and so on.Therefore,under the guidance of deep learning theory and on the basis of previous research on spatial image steganalysis,after studying the existing spatial adaptive steganalysis algorithm and coding algorithm,the innovative work of this thesis is summarized as follows:(1)An image steganalysis algorithm based on multi-task learning and image block regression learning is proposed.The backbone network is used for image steganalysis,and the branch network is used for image block regression learning.After the research of steganography,it is found that the more complex the texture region,the greater the probability of being embedded,and the more the number of pixel change.Therefore,this thesis proposes an algorithm that takes image block regression learning as branch task,in which the label in each image block is the number of changed pixels in each image block.The larger the label value is,the complex of texture region is.If it is 0,the image block is located in the smooth region.By learning these information,the branch network can assist the backbone network to recognize the texture area and get better results.Experimental results show that,compared with other image steganalysis based on the learning of selection-channel-aware algorithms,the proposed method reduces the false detection rate by about up to 1% and and the lowest is about 0.2% respectively.In addition,by comparing the cover image with the stego image,it can be found that the changed position is fixed in different embedding rates and different steganography.For this reason,a method based on transfer learning is also proposed,that is,first train the network with high embedding rate(0.5bpp),when train the network with low embedding rate(0.1bpp-0.4bpp),preload the network with high embedding rate over 0.1bpp,and then fine tune it.The experimental results show that the maximal reduction of false detection rate is about 1% and it shortens nearly half of the training time in the low embedding rate.(2)Based on image block regression as auxiliary task,auxiliary a new deep residual network for image steganalysis is proposed.Combined the concept of ”Deep is Better” and related knowledge of steganalys,and a variety of residual modules are added to the backbone network.Finally,a deep residual network composed of nine residual blocks is designed,in which the noise information is weak,and pooling layer is essentially a low-pass filter,which will weaken beneficial feature.For this reason,the main line of each residual block does not use pooling layer,but chooses to set the convolution kernel stride to reduce the dimension.The experimental results show that the proposed method not only solves the problem that the false detection rate is still lower than that of the method of directly adding the selection-channel-aware information at low embedding rate,but also reduces the false detection rate by 2% at high embedding rate.