Efficient, Robust, and Anti-Collusion Fingerprinting of Image Diffusion Models

Summary

This paper presents a fingerprinting method for text-to-image diffusion models that encodes bit-string fingerprints into the coefficients of a personalized normalization module (PNM), so they can be recovered from any generated image. It adds an anti-collusion mechanism based on lossless function-invariant parameter transformations, which the authors report degrades the generation quality of colluded models, and reports fingerprint extraction accuracy above 99.5% across several T2I generation and editing tasks.