Study On The Methods To Alleviate The Negative Influence Of Cover Source Mismatch In Digital Image Steganalysis

In the laboratory environment,the digital image steganalysis has obtained relatively high detection accuracy.However,in the real-word,the application of digital image steganalysis is far behind digital image steganography.The application of digital image steganalysis to real word has to overcome many difficulties.Among them,Cover Source Mismatch(CSM)is one of the major difficulties.CSM occurs when a detection classifier for steganalysis is trained on one cover source and tested on another source.At present,the most effective way to alleviate CSM is to use mixed and large source training set,however,increasing the number of the training set also increases the training time and resource consumption.In this dissertation,we consider new approachs to study more effective ways to mitigate the effect of CSM from the following two perspectives:(1)Study on the interaction between image texture complexity and cover source mismatch in steganalysis,and then alleviate the effect of cover source mismatch through texture complexity.(2)Cover source mismatch is alleviated by controlling the effect of other noises except the steganography noise on the feature distribution of steganalysis.According to the above two problems,the main contents of this dissertation are as follows:A texture complexity measurement algorithm,along with different image sources,is used to develop a mathematical model based on two-way analysis of variance,aiming to study the interaction between image texture complexity and cover source mismatch.The study results show that there is a significant interaction between texture complexity and cover source mismatch.The study results also imply that we can alleviate the negative effects of CSM from the pespective of texture complexity.As an example,we propose a method based on texture complexity to mitigate the effect of CSM based on texture complexity measurement,experimental results show that this method can alleviate the effect of CSM on steganalysis to a certain extent.The distribution features of noises except the steganography noise in the image are extracted.The process of extracting the feature of steganalysis is inevitably affected by other noises or even the texture complexity,which,to a certain extent,causes to the CSM problem.In this dissertation,we extract the convolution residual of different sizes and weights to get the 8000 dimensional RND(Redundant Noise Distribution)feature composed of 60 sub models,which are used to describe the distribution feature of other noises besides the steganographic noise in the image.Based on the RND feature and JS divergence,we propose a similar image matching algorithm.The training set matched by the matching algorithm is the image set similar to the test set in other noise distribution,in which the same cover source image is the most similar.Experiments of the steganalysis of adaptive steganography and non-adaptive steganography under various mismatch scenarios proved that CSM can be effectively alleviated by training set matched.The training set matched by the matching algorithm can effectively alleviate the CSM problem.At the same time,from the statistical result of the matched images,we found that when the image library contains the homologous image of the test set,the matching algorithm will first match the homologous image in the image library,this also in a certain degree prove the effectiveness of the proposed matching algorithm.