FuguReport

GoodDiffusion: Proactive Copyright Protection for Diffusion Bridge Models via Learnable Sample-specific Signatures

Authors Shixi Qin, Zhiyong Yang, Shilong Bao, Zitai Wang, Qianqian Xu, Qingming Huang
Affiliations Beijing Academy of Artificial Intelligence / Chinese Academy of Sciences / University of the Chinese Academy of Sciences
Categories Method / Copyright Protection / Proactive control with authentication queries, Method / Model Security / Conditional signature assignment, Application / Diffusion Model Usage Control / Controlled generation rejection
License CC BY 4.0

Abstract Overview

This paper proposes GoodDiffusion, a proactive copyright protection method for diffusion bridge models that embeds authorization into the generation process itself. Instead of relying on post-hoc watermarking or merely degrading outputs, the protected model generates high-quality results only for inputs carrying valid signatures and otherwise maps unauthorized requests to a predefined warning image. The authors first analyze a naive fixed-signature design and show it is vulnerable in white-box settings because a surrogate signature can be recovered by gradient-based optimization. To address this, they introduce a Learnable Signature Network that produces sample-specific, input-conditioned signatures, aiming to prevent a universal bypass signature from transferring across inputs.

Novelty

The main novelty is reframing backdoor-style mechanisms as a constructive, model-level access-control tool for copyright protection in diffusion bridge models. A second distinctive contribution is the use of learnable sample-specific signatures, rather than a single static trigger, to counter white-box signature recovery and non-transferably bind authorization to each input.

Results

Across CelebA and ImageNet image-to-image tasks including super-resolution, inpainting, and deblurring, GoodDiffusion blocks unauthorized use by producing warning outputs while preserving usable generation quality for authorized inputs. On CelebA, the reported abuse rate is reduced from 100% for unprotected models to near zero, with less than 0.06% in the summarized main cases, while error rates for authorized users remain low. Security experiments further show that recovered surrogate signatures can break the static-signature version, whereas the sample-specific design keeps surrogate generations poor and abuse rates at or near zero.

Key Points

  1. GoodDiffusion internalizes authorization into diffusion bridge generation, allowing signed inputs to generate target images and unsigned inputs to be redirected to a warning image.
  2. The paper theoretically and empirically shows that naive static signatures are fragile in white-box settings because adversaries can recover transferable surrogate signatures.
  3. The Learnable Signature Network generates input-conditioned signatures that improve security by preventing a universal recovered signature from generalizing across inputs.

References

This page was created using generative AI such as GPT-5, Claude Opus 4, Gemini 3, Gemini 3.1 Flash Image, and their higher-end successor versions. No guarantee can be made regarding its contents.