Digital Image Steganalysis Based on Visual Attention and Deep Reinforcement Learning

Recently, the adaptive steganography methods have been developed to embed secret information with the minimal distortion of images. As the opposite art, steganalysis methods, especially some convolutional neural network-based steganalysis methods, have been proposed to detect whether an image is embedded with secret information or not. The state-of-the-art steganography methods hide secret information in different regions of an image with different probabilities. However, most of the current steganalysis methods extract the steganalysis features from different regions without discrimination, which reduces the performance of the current deep-learning-based steganalysis methods when attacking the adaptive steganography methods. In this paper, we propose a new self-seeking steganalysis method based on visual attention and deep reinforcement learning to detect the JPEG-based adaptive steganography. First, a region is selected from the image by a visual attention method, and a continuous decision is then made to generate a summary region by reinforcement learning. Thereby, the deep learning model is guided to focus on these regions that are favorable to steganalysis and ignore those regions that are unfavorable. Finally, the quality of training set and the detection ability of steganalysis are improved by replacing the mis-classified training images with their corresponding summary regions. The experiments show that our method obtains the competitive detection accuracy, compared with the other state-of-the-art advanced detection methods.