An Image Steganalysis Model Combined with the Directionality of Residuals

In recent years, many steganalysis methods using convolutional neural networks have been proposed. In the existing steganalysis networks, in order to enhance steganalysis noise and reduce the impact of image content, the high-pass filter is applied to extract residuals. However, the residual is usually directly input into a network for feature extraction, without considering further processing to enhance the statistical feature extraction of the subsequent network. Furthermore, the processing of convolutional layer in a network can be viewed as horizontal and vertical scanning maps, and the form of directions is simple. In this paper, to enrich directional features, the directionality of residuals is incorporated into the learning of network. Before feature extraction, residuals are rearranged in the direction of minor-diagonal. In addition, local binary pattern is applied to the residual map to obtain the correlation between each element in residual map and its multi-directional adjacent elements. Three spatial steganography algorithms, WOW, HUGO and S-UNIWARD, are selected in the simulation. The simulation results show that the incorporation of residual directionality into convolutional neural network can improve the steganalysis performance of the network.