AI-Assisted Reversible Data Hiding Using Reinforcement Learning

Recently, artificial intelligence (AI) algorithm has been extensively utilized as an optimization tool in digital watermarking. However, existing works seldom consider the reversibility of the embedding framework, thereby neglecting the protection of sensitive carriers. In this paper, we propose an AI-assisted reversible data hiding (RDH) method based on reinforcement learning. In our method, the Q-Learning algorithm is introduced to address the optimization problem of RDH, i.e., adaptive two-dimensional (2D) mapping generation, and it is designed to simulate pixel modifications in 2D space, employing a Markov decision process formulation within the reinforcement learning paradigm. The new reward function is given to evaluate the effectiveness of 2D mapping based on the estimated embedding capacity and the distortion-capacity ratio. To ensure reversibility, a mapping adjustment strategy is implemented to update the environmental states. Experimental results show that the proposed method outperforms conventional 2D RDH and demonstrates competitive performance compared to other state-of-the-art RDH methods.