Research On Image Steganalysis Technology Based On Spatial Rich Model

Image steganography is an important covert communication technology,which uses small changes in pixel values or DCT coefficients to hide secret information.However,steganography has advantages and disadvantages.While improving communication security for national security,military intelligence,and business secrets,it also facilitates illegal activities such as stealing confidential information and destroying information integrity.Steganalysis,as a confrontation technique of steganography,mainly analyzes the relevant features extracted by the carrier to determine whether there is hidden information.In recent years,many steganalysis features have achieved good performance.The research progress of steganalysis is conducive to detecting the security of steganography;at the same time,it helps prevent the leakage of confidential information and curb the spread of harmful information such as viruses,which has important theoretical and practical significance.Spatial image steganalysis is mainly divided into steganalysis methods based on traditional manual features and deep learning.At present,the mainstream and most advanced traditional steganalysis features are based on rich model features.In the steganalysis framework based on deep learning,filters with rich model features are also used to initialize the preprocessing layer.Therefore,in order to improve the detection performance of spatial image steganalysis algorithms,this research studies the traditional steganalysis algorithms and deep learning steganalysis methods based on rich models.The main contents include:(1)Steganalysis of spatial image combined with hybrid kernel feature mapping.This work is mainly aimed at the feature dimensionality of the rich model reaching tens of thousands of dimensions.When using the public data set BOSSbasev1.01,the existing methods cannot directly perform kernel transformation and approximate mapping on the features of the rich model.This thesis proposes an idea of segmenting and then mapping the features of the highdimensional rich model,and then using the ensemble classifier to classify the mapped features,and constructing a new hybrid kernel function structure to improve the feature approximation mapping algorithm,thereby improving the performance of steganalysis.(2)Rich model feature selection algorithm based on embedding probability and intra-class distance.Because the rich model feature extraction scheme counts a large number of features that are invalid to steganalysis,this seriously affects the classification accuracy and classification cost.However,the traditional Fisher criterion feature selection algorithm ignores the correlation between features.Therefore,it is proposed to simplify the features according to different embedding probabilities first,and then use intra-class distance and Person correlation coefficient to measure the attribute separability,select more effective features according to the measurement value,and reduce the feature dimension.(3)Image steganalysis based on separable convolution and local source residual learning.Although the steganalysis of the rich model and the selected channel version has achieved a high detection accuracy,the rich model has a high feature dimension,calculation is difficult,and the algorithm's robustness is poor.In the image steganalysis method based on deep learning,the improvement of its detection performance is usually achieved through structural extension and heuristic techniques.Therefore,this part combines rich model and deep learning,introduces separable convolution,local source residual and global covariance pool to improve the CNN network framework and improve detection performance.In summary,this research proposes a feature mapping algorithm and a feature selection method based on the features of the rich model,and combines the features of the rich model,proposes a steganalysis network based on separable convolution and local source residual learning.Finally,a large number of experiments verify the effectiveness of the proposed method.