Research On Universal Steganalysis Of Spatial Domain Images Based On Rich Model

As a mainstream information hiding method,image steganography has been favored by many researchers.In order to better ensure the security of secret information transmission,a large number of steganographic algorithms have been proposed to embed secret information in images.Especially with the introduction of adaptive steganographic algorithms,the secret information embedded by steganography becomes more difficult to be detected,greatly improving the security of steganographic images.However,some illegal messages are also embedded into images using steganography by criminals for secret dissemination and criminal activities,thus evading the monitoring of public security departments,posing a threat to the safety of the public and the country.To address the adverse consequences brought about by the abuse of image steganography,image steganalysis technology has been proposed by scholars to distinguish whether images transmitted over the network contain secret information.Image steganalysis technology mainly includes spatial domain and transform domain aspects,corresponding to the detection of message embedding in spatial domain and transform domain steganography.This paper mainly studies steganalysis technology in the spatial domain.Based on a rich model,traditional spatial domain image universal steganalysis methods and deep learning-based spatial domain image universal steganalysis methods are studied separately to improve the performance of steganalysis algorithms.This paper mainly includes the following aspects:(1)A spatial domain image steganalysis method combining fusion features and feature mapping is proposed.This method is based on the traditional steganalysis method: firstly,fusion features are used to replace existing single steganalysis features to obtain richer steganalysis features to enhance the ability to capture disturbances caused by steganography on the original image;then,the extracted features are subjected to PCA for dimensionality reduction processing,and the dimensionality-reduced features are used for feature mapping combined with non-linear mixed kernel.Finally,FLD integrated classifier is used for classification.The experimental results prove that,compared with the existing algorithms,this method has a greater improvement in the detection rate when the steganography embedding rate is low,and effectively improves the detection rate of the steganographic image.(2)A spatial domain image steganalysis method based on local source residual learning and spatial pyramid pooling is proposed.This method is based on deep learning-based steganalysis method: firstly,in the preprocessing layer,SRM filtering kernel is used to initialize the convolutional kernel to filter the input image to obtain richer residuals;then,separable convolution modules and local source residual learning modules are added to the network to obtain more effective steganalysis features and improve the performance of steganalysis network.Finally,spatial pyramid pooling is used to replace the traditional global average pooling to model local features and effectively reduce computation.Experimental results demonstrate that the proposed steganalysis network effectively improves the detection rate of steganographic images.