FCADD: Robust Watermarking Resisting JPEG Compression with Frequency Channel Attention and Distortion De-gradient

Recently, the majority of deep learning-based robust watermarking techniques exhibit proficiency in concealing secret information, thereby providing effective protection for the copyright information of images. However, during the digital transmission process, images typically undergo JPEG compression, which significantly compromises the integrity of the watermarks. We found that most current methodologies are inadequately equipped to mitigate the adverse impacts of JPEG compression. In this work, we proposed a robust watermarking method against JPEG compression by applying Frequency Channel Attention and Distortion De-gradient mechanism (FCADD). Specifically, we introduce a only forward propagation mechanism in the attack simulation layer that bypasses the difference caused by image processing, such as compression, dropout, during transmission, allowing a portion of the loss function’s gradient to be passed back to the encoder. Additionally, for the sake of model stability and generalizability, we assign a certain probability to both the JPEG simulation function and a lossless function in different training batches to participate in the training process. Furthermore, we introduce Frequency Channel Attention Blocks (FCA Block) in the feature extraction process, leveraging the information on the image frequency domain to achieve better visual effects. Extensive experiments have demonstrated that our approach outperforms the current state-of-the-art (SOTA) models in JPEG compression across any quality factor, and also exhibits superior visual effects and robustness against other distortions such as Gaussian filtering and Dropout.