Incorporating Clustering Modification Directions into Reinforcement Learning Based Cost Learning Framework

Content-adaptive image steganography embedding cost learning frameworks based on deep learning can generate a more exquisite embedding probability map within a short time, and such methods have reached remarkable security performance compared to conventional hand-craft based methods and received increasing attention in recent years. However, existing Reinforcement Learning (RL)-based schemes are typically based on single-step state machine, making it difficult for further improvement. This paper ex-tends the existing RL-based framework into two steps to enhance the simulated stego images from policy network to improve the performance, that is, during the training pro-cess, similar to the conventional methods, a module will be added after the policy network, the current embedding direction is adjusted according to the sign of modification directions of the neighborhood. The experimental results show that the proposed module not only improve the performance during the training process, but also enhance the actual security performance compared with single-step based frameworks when countering mul-tiple steganalyzers.