Robust Watermarking via Dual Guidance
Yuhang Zhang, Yuanman Li, Li Dong, Xia Li · 2024
In recent years, digital watermarking methods based on deep learning have been extensively studied due to their crucial role in protecting copyright information and forensic tracing. Existing robust watermarking methods have attained notable imperceptibility and robustness through end-to-end training, as well as by employing techniques such as adversarial training and simulated attacks. However, these methods often neglect the watermark encoding process during the embedding stage. Typically, they use duplicated or upsampled watermarks for embedding, which disregards the characteristics of the cover image and limits the performance of the watermarking method. To address this limitation, we propose a robust watermarking via dual guidance. During the embedding stage, we utilize both global semantic features and local features of the cover image to guide the watermark encoding and embedding processes. The global semantic feature guides the watermark encoding, enabling an adaptive encoding process based on the cover image, thereby enhancing the imperceptibility and robustness of the watermarking method. Local features are used to generate embedding masks that adjust the strength of the watermark embedding. In the extraction stage, we introduce a multi-scale attention-based decoder to extract the watermark from the global features of the embedded image. Experimental results demonstrate that our method outperforms existing watermarking models.