[2E6-GS-10n] AI application

6:15 PM - 6:30 PM JST(9:15 AM - 9:30 AM UTC)

[2E6-GS-10n-04] Robust Watermark Embedding Method Based on Diffusion Generation Process

〇Taishi Nishiyama¹, Kazunori Kamiya² (1. NTT Social Informatics Laboratories, 2. NTT Security Holdings/NTT Social Informatics Laboratories)

Keywords:

watermark,diffusion model

Diffusion models can generate high-quality and diverse images from text prompts. However, they raise legal and ethical risks, including copyright infringement, misinformation, fraud, and impersonation. We embed an invisible watermark into intermediate states of the reverse diffusion process, rather than adding it to the final generated image. The reverse diffusion timestep for injection is a critical design parameter, since it determines the balance between watermark strength and generation quality. We propose a method that optimizes the trade-off among watermark robustness, generation quality, and imperceptibility, and automatically determines the injection timestep using an evaluation function. Furthermore, following prior work, we adopt a watermark concealment process to guide the model, enabling robust embedding while preserving generation quality. As a result, the embedded watermark remains robust to post-processing (e.g., cropping, filtering, and upscaling) with minimal visual distortion.