Image Steganalysis Via Augmented Utilization Of Selection-channel Information

Since digital image has become an important carrier of information diffusion in modern society,it is necessary to investigate information security on digital images.Image steganography can hide secret messages into images.Due to the complexity of the content within an image and the subtle modification caused by steganography,image steganography can achieve privacy communication without being noticed.In areas such as military,polity and finance,steganography is important due to its ability in ensuring security communication.However,if steganography is used illegally,a baleful influence will be exerted to personal privacy,society stabilization,or even national security.Thus,distinguishing whether an image carries secret information is of great importance.So far,most image steganographic algorithms are proposed under the framework of image content adaptivity,which aims to make it difficult for steganalytic detection by hiding information into noisy and textural areas.Nevertheless,on the opposite side of steganography,image steganalysis based on content adaptivity is just at the starting stage with little achievements.To improve the detection accuracy of steganalysis,this thesis is carried out based on the utilization of augmented selection-channel information.It contains two aspects:the traditional feature extraction method and the deep learning based neural network method.(1)Existing steganalytic methods use selection-channel information directly without taking the characteristics of selection-channel information into consideration.To cope with this problem,we mainly make better use of selection-channel information by two different ways:preprocessing and utilization method.By these two ways,pixels with large values in selection-channel information,which are more likely to be modified during the embedding process,are assigned higher weights in the final steganalytic feature.The detection accuracy of steganalytic algorithm can be effectively improved by making the designed feature capable to capture statistical difference resulting from steganography.Experimental results show that the performance of our proposed method is better than other steganalytic methods.(2)Since traditional steganalytic convolutional neural networks(CNN)may lose much usefulinformation,we propose a strategy to augment selection-channel information into CNN.By augmenting the larger values in selection-channel information,loss of useful information can be reduced.We prove the effectiveness of this strategy by experiments.(3)So far,selection-channel information is computed under the condition of available embedding payload.However,it is difficult to acquire such kind of information in real life.Hence we propose a strategy to evaluate selection-channel information with embedding payload unknown.Moreover,we incorporate this strategy with our proposed steganalytic feature and steganalytic convolutional neural network.According to our experimental results,higher detection accuracy can be achieved by our proposed strategy,compared with steganalytic algorithm without selection-channel information,which proves the effectiveness of our proposed strategy.In conclusion,we propose several strategies to make better use of selection-channel information for feature extraction method and convolutional neural network method respectively,under known and unknown embedding payload.The experimental results emphasize the importance of selection-channel information.