A Dual-Path Network for Source-Robust Spatial Image Steganalysis

As the existing literatures show, deep learning based steganalysis has achieved much better detection performance compared to traditional hand-crafted feature based methods. However, most methods based on deep learning are quite vulnerable to keep great performance under the image-source mismatch scenario, which is the inconsistency of the source of the training set and testing set. To improve this drawback, we specially design the architecture of the proposed network. Moreover, we first introduce the bilinear pooling into steganalysis to reduce the data source dependence of the model for further improving the performance. Experimental results show that the proposed work can outperform other networks in the image-source mismatch scenario.