Image Steganalysis Based on Multi-Scale and Attention Mechanism
Regarding the weak capability of current steganalysis algorithms to characterize features of complex texture regions in images,a steganography analysis model based on multi-scale feature fusion and attention mechanism was proposed.Firstly,the SRM(spatial rich model)filters are used to preprocess the input image to extract the noise component residuals so as to reduce the influence of the content of the image itself.Then,the multi-scale parallel net-work is used to extract the signal to enhance the learning of subtle features.Subsequently,the attention mechanism is introduced to carry out adaptive weighting of the features,emphasizing the role of important channel features in clas-sification,while simultaneously reducing the influence of non-important channel features on the classification.Final-ly,a covariance pooling method is proposed to model the correlation between features after deep neural network learning,and Newton iteration method is selected to solve the square root matrix to make network training more effi-cient.Experimental results demonstrate that the accuracy of the proposed model reaches 88.6％under the condition of a 0.5 bit per pixel embedding rate of the WOW(wavelet obtained weights)algorithm,which proves the effective-ness of the proposed method.
NETL
NSTL
万方数据
