CNN-based Reversible Data Hiding for JPEG Images

In the field of Joint photographic experts group (JPEG) reversible data hiding (RDH), due to the weak correlation between the adjacent alternating current (AC) coefficients in the JPEG image, the existing JPEG RDH methods cannot effectively find those extension coefficients with high embedding efficiency and prioritize them for carrying message bits. In this paper, a new convolutional neural network (CNN)-based JPEG RDH scheme is proposed. First, the Laplacian distribution model is applied to roughly pre-estimate the expansion probability of the AC coefficients. Then, the approximate pre-estimated expansion probability and the actual expansion probability of the AC coefficients are used to train the carefully designed CNN-based estimation model, and the embedding efficiency of each AC coefficient can be calculated through the output of the CNN model. In the embedding stage, a new adaptive embedding strategy called coefficient selection strategy is proposed, which is more efficient that those previously proposed selection strategies based on block selection and frequency selection. Finally, the AC coefficient with greater embedding efficiency will be preferentially used for data hiding. Extensive experimental results demonstrate the effectiveness of our proposed CNN-based method compared with the state-of-the-art JPEG RDH methods.