Chapter 7 Enhanced image watermarking through cross-attention and noise-invariant domain learning
Agnibh Dasgupta, Xin Rong Zhong · 2025
Image watermarking deals with embedding and extracting watermarks in a cover image, focusing on improving generalization and robustness against various manipulations. Recent deep learning methods enhance these aspects mainly using convolutional techniques for embedding and including targeted augmentations during training. Vision transformers, known for their self-attention mechanism, efficiently handle longrange dependencies and global context, showing promise for image watermarking. However, their potential for watermark embedding, especially using cross-attention, has not been fully explored. This study introduces a robust image watermarking approach utilizing cross-attention and invariant domain learning, presenting two significant advancements. First, we propose a watermark embedding method based on a multihead cross-attention mechanism. This mechanism enables an effective exchange of information between the cover image and the watermark, pinpointing the best spots for embedding based on semantic relevance. Second, we suggest learning a domain representation that is invariant to both semantic shifts and noise, potentially broadening the scope for improving image watermarking techniques.