Learning selection channels for image steganalysis in spatial domain

Advanced steganography modifies the complex regions of digital media to embed secret messages, while steganalysis aims to detect whether the digital media contains secret messages or not. It is well recognized that the content adaptivity which adopted in steganography should also be considered for steganalysis to improve detection accuracy, and thus the embedding locations are weighted (so called selection channel) to build steganalytic detectors. However, the existing selection channels incorporated into steganalysis are all manually designed and keep constant even in the whole training stage of deep learning based steganalysis. Therefore, the handcrafted and fixed selection channels leave much room for improvement in steganalysis. In this paper, we propose to learn the selection channels in an end-to-end manner. Our steganalytic scheme has two parts: selection channel network and steganalysis network. These two networks are trained together. The selection channel network learns and outs the selection channels for the steganalysis network, and the steganalysis network integrated with the learned selection channels predicts the final steganalysis results. Our experiments under various conditions show that the learned selection channels considerably improve the detection accuracy of steganalytic schemes against content-adaptive steganography, and also exhibit high universality and robustness in real-world environments. (C) 2020 Elsevier B.V. All rights reserved.