A Deep Residual Multi-scale Convolutional Network for Spatial Steganalysis

Shiyang Zhang, Hong Zhang, Xianfeng Zhao, Haibo Yu

Abstract: Recent studies have indicated that Convolutional Neural Network (CNN), incorporated with certain domain knowledge, is capable of achieving competitive performances on discriminating trivial perturbation introduced by spatial steganographic schemes. In this paper, we propose a deep residual multi-scale convolutional network model, which outperforms several CNN-based steganalysis schemes and hand-crafted rich models. Compared to CNN-based steganalyzers proposed in recent studies, our model has a deeper network structure and it is integrated with a series of proven elements and complicated convolutional modules. With the intention of abstracting features from various dimensions, multi-scale convolutional modules are designed in three different ways. Besides, inspired by the idea of residual learning, shortcut components are adopted in the proposed model. Extensive experiments with BOSSbase v1.01 and LIRMMBase are carried out, which demonstrates that our network is able to detect multiple state-of-the-art spatial embedding schemes with different payloads.