Steganalysis with CNN Using Multi-channels Filtered Residuals

Abstract

In the current study of steganalysis, Convolutional Neural Network (CNN) have attracted many scholars’ attention. Recently, some effective CNN architectures have been proposed with better results than traditional Rich Models with Ensemble Classifiers. Inspired by the idea that Rich Models use various types of sub-models to enlarge different characteristics between cover and stego features, a scheme based on multi-channels filtered residuals is proposed for digital image steganalysis in this paper. This paper mainly focus on the stage of image processing, 3 high-pass filtered image residuals are fed to a deep CNN architecture to make full use of the great nonlinear curve fitting capability. As known, deep learning is powerful in pattern recognition, most previous networks only use single type of filtered residuals in steganalysis, varied high-pass filtered residuals can offer stronger features for CNN in this paper. After filtering, the residuals are superposed into a multi-channels residual map before training, this measure can involve a joint optimization of CNN’s parameters. But single residual map has no such effect. The experiment results prove that it’s an efficient way to provide a better detection performance, achieving an accuracy of 82.02% on Cropped-BOSSBase-1.01 dataset.