Boosting Image Steganalysis Under Universal Deep Learning Architecture Incorporating Ensemble Classification Strategy
Ante Su, Xianfeng Zhao · IEEE Signal Processing Letters · 2019
Image steganalysis based on convolutional neural networks (CNNs) has achieved remarkable performance. However, all existing CNN-based steganalysis methods form an ensemble by merging the outputs of independently trained networks with the same architecture. In this letter, we propose a universal CNN architecture incorporating ensemble classification strategy. Any CNN-based steganalysis with the proposed architecture needs to train only one model to form an ensemble and can boost detection accuracy in both spatial and JPEG steganography. In particular, we propose a new method to construct subspaces for training well-designed base learners. In addition, a novel voting fusion structure automatically optimized with the training process is proposed. Experimental results on the public dataset demonstrate that the proposed architecture can further improve the performance of CNN-based steganalysis. Source code is available via GitHub (https://github.com/Ante-Su/CAECS).