Improving Cost Learning for JPEG Steganography by Exploiting JPEG Domain Knowledge

Although significant progress has been achieved recently in automatic learning of steganographic cost, the existing methods designed for spatial images cannot be directly applied to JPEG which are more common media daily life. The difficulties migration mainly caused by characteristics $8\times 8$ DCT mode structure. To address issue, this paper we extend an cost scheme JPEG, where proposed called JEC-RL (JPEG Embedding Cost with Reinforcement Learning) is explicitly tailor It works embedding action sampling mechanism under reinforcement learning, a policy network learns optimal policies via maximizing rewards provided environment network. Following domain-transition design paradigm, composed three modules, i.e., pixel-level texture complexity evaluation module, feature extraction and mode-wise rearrangement module. These modules operate serial, gradually extracting useful features from decompressed image converting them into elements, while considering including inter-block intra-block correlations simultaneously. gradient-oriented way provide stable reward values using wide architecture equipped fixed preprocessing layer basis filters. Extensive experiments ablation studies demonstrate that method can achieve good security performance against both advanced feature-based modern CNN-based steganalyzers.