Image Steganalysis via Multi-Column Convolutional Neural Network
Ke Qi, Liu Dong Ming, Daxing Zhang · 2018
Deep learning that jointly studies and extracts features is very promising for steganalysis. In this article, we design a simple but effective Multi-column Convolutional Neural Network (MCNN) based on steganalysis architecture for images. The proposed MCNN architecture allows the input image to be of arbitrary size or resolution. In particular, by utilizing filters with receptive fields of different sizes, the features learned by each column CNN are adaptive to variations in payloads. Comprehensive experiments on standard dataset show that MCNN model can detect the state of arts steganographic algorithms with a high accuracy. It also outperforms several recently proposed CNN-based steganalyzers in conditions of the same embedding key stego and cover-source mismatch scenarios.