FlagThis

Abstract

This paper introduces GoodDiffusion, a proactive copyright protection mechanism for diffusion bridge models designed to prevent unauthorized use. Unlike passive watermarking or quality-degradation defenses, GoodDiffusion implements model-level control that provides high-quality generation for authorized queries with valid signatures while refusing to generate for unauthorized inputs. To counter white-box attacks that can recover static signatures, the authors propose a Learnable Signature Network (LSN) to generate sample-specific signatures, ensuring robust protection against surrogate signature recovery.