Deep residual network for halftone image steganalysis with stego-signal diffusion

Abstract

More and more convolution neural network (CNN) models are used for image steganalysis, which show superior performances than traditional steganalytic methods. However, no researches on halftone image steganalysis by CNN have yet been carried. In this paper, a novel residual CNN model with stego-signal diffusion for halftone image steganalysis is proposed and achieves state-of-the-art detection accuracy. Considering inverse halftoning can reconstruct the gray-scale image from the halftone image, inverse halftoning is used to preprocess the halftone image, which can diffuse the stego-signal to neighboring pixels. As a result, the difference between the cover and stego image is magnified on the texture. Then, the residual block is utilized to construct the CNN model, since it could preserve the stego-signal better than plain network, and the magnified difference allows the network to better identify cover and stego images. A series of experiments are conducted on a large-scale dataset. The detection accuracy is improved by the magnified difference, and our proposed model outperforms the previous methods.