RWP: a robust watermarking plugin for attribution and protection in stable diffusion models.

Abstract

Diffusion models have achieved remarkable success in content generation, driving the rapid development of various customized models. However, this progress also presents significant challenges in provenance tracking, including the misuse of black-box models and the dissemination of deceptive generated content. Watermarking has emerged as a potential solution to address these issues. However, existing methods face limitations in complex AI-driven commercial scenarios, including inaccurate user tracing and ineffective prevention of model misuse. To overcome these challenges, we propose RWP, a robust watermarking plugin specifically designed for AI commercial applications. Our approach embeds traceable watermarking information directly into the diffusion model's generation process by incorporating a reversible encryption module that produces unique transformation sequences. This design enables the tracing of a large number of users in commercial diffusion models. Unlike previous methods that require fine-tuning or retraining the diffusion model architecture, our method introduces no modifications to the diffusion model itself and can be integrated as a plugin. This significantly improves the deployment efficiency of watermarking systems. Experimental results show that our method improves tracking accuracy by 20-50% under both image-to-image translation and inpainting attacks compared with AquaLoRA, while maintaining high visual fidelity, thereby demonstrating its practical robustness and effectiveness. Moreover, the proposed framework expands the traceable user capacity from the existing 48-64 bits to more than 400 bits, enabling substantially finer-grained user identification. These results validate the practicality of RWP in large-scale commercial scenarios and provide reliable technical support for copyright protection and accountability in AI-generated content.