A Somewhat Robust Image Watermark against Diffusion-based Editing Models

• DOI: 10.48550/arxiv.2311.13713
• Publication Date: 2023-01-01
• Authors: Mingtian Tan; Tianhao Wang; Somesh Jha
• Keywords: Image editing; Code (set theory); High fidelity

Recently, diffusion models (DMs) have become the state-of-the-art method for image synthesis. Editing based on DMs, known their high fidelity and precision, inadvertently introduced new challenges related to copyright infringement malicious editing. Our work is first formalize address this issue. After assessing attempting enhance traditional watermarking techniques, we recognize limitations in emerging context. In response, develop a novel technique, RIW (Robust Invisible Watermarking), embed invisible watermarks leveraging adversarial example techniques. technique ensures extraction accuracy of $96\%$ watermark after editing, compared $0\%$ offered by conventional methods. We provide access our code at https://github.com/BennyTMT/RIW.