Noise Non-Differentiable in Deep Learning End-to-End Image Watermarking Models
Xinting Wu, Ying Huang, Hu Guan, Baoning Niu, Fangpeng Lan · 2023
In recent years, deep learning-based watermarking algorithms have been widely studied due to the powerful feature extraction capability of deep learning. The deep learning watermarking framework consists of three parts: encoder, noise layer, and decoder. In order to better coordinate the simultaneous update of encoder-decoder parameters, end-to-end training is often used, and this training method requires that the noise layer must be composed of differentiable attacks, otherwise the update of model parameters cannot be achieved through the back propagation effect. In this paper, we analyze the causes of noise non-differentiable and summarize the existing literature methods to solve the non-differentiable problem, and analyze the differences between different methods by comparing the experimental findings. An iterative training approach is proposed to improve the robustness effect of the model in the face of real attacks by collaborative codec training and learning of real noise distribution, with an average bit prediction accuracy (BPA) of 97.56%.