Deep Learning on Spatial Rich Model for Steganalysis

Recent studies have indicated that deep learning for steganalysis may be a tendency in future. The paper novelly proposes a faeture-based deep learning classifier for steganalysis. Analysis shows SRM features are suitable to be the input of deep learning. On this basis, a modified convolutional neural network (CNN) is designed for detection. In the initial layers, taking the thought of ensemble classifier for reference, we extract L subspaces of the entire SRM feature space, and process each subspace respectively. In the deeper layers, two different structures are designed. One is complex in structure and hard to train, but achieve better detection accuracy; the other is simple in structure and easy to train, but the achieved detection accuracy is a little worse. Experiments show that the proposed method achieves comparable performance on BOSSbase compared to GNCNN and ensemble classifier with SRM features.