Protecting Visible Watermarks Against Diffusion-Based Inpainting via Localized, Robust, and Reversible Latent...

The rapid spread of generative content challenges copyright protection and content identification. Visible watermarks, as a conventional mechanism for ownership attribution, are increasingly required in generated content. However, their localized and fragile nature makes them easily removable, especially via diffusion-based inpainting, which can seamlessly erase watermarks and reconstruct background regions. Although prior studies have explored adversarial perturbations to hinder such inpainting, they suffer from several limitations: 1) trade-off between effectiveness and imperceptibility; 2) vulnerability to diffusion-based purification attacks; and 3) irreversible perturbation injection, which prevents subsequent image modification. To address these issues and enhance visible watermark protection, we propose a localized perturbation method in the latent space, alleviating the trade-off between perturbation effectiveness and imperceptibility in pixel space. To implement perturbation injection, we apply reversible transformations to the latent vectors and leverage natural images as keys, making the perturbations more resistant to diffusion-based purification attacks without adversarial training. Moreover, authorized users can achieve near-lossless recovery by removing the perturbation from the latent space with the correct key. We further explore its potential application to generative video watermark protection. Experiments show that our method achieves strong protection effectiveness and imperceptibility on natural images as well as generative content, making it more suitable for practical applications. Code is available at https://github.com/charles335cs/MarkShield.