Efficient, robust, and collusion-resistant fingerprinting for image diffusion models

The researchers propose a collusion-resistant fingerprinting method for text-to-image (T2I) diffusion models, intended to protect model intellectual property and prevent unauthorized redistribution. The paper points out a previously unexplored systematic weakness of existing generative model fingerprinting methods, namely the inability to resist collusion attacks in which multiple attackers merge models to remove or obfuscate fingerprints. The method encodes a bit-string fingerprint into the coefficients of a personalized normalization module (PNM) embedded in the model, can reliably recover the fingerprint from any generated image, and supports creating multiple fingerprinted copies without retraining by reparameterizing the PNM; the anti-collusion mechanism is based on a lossless function-invariant parameter transformation that substantially degrades the image generation quality of colluding models, making them largely unusable, and a worst-case optimization strategy is introduced to improve robustness to model-level attacks. Experiments show that the method achieves fingerprint extraction accuracy above 99.5% on multiple T2I generation and editing tasks, and the authors state that it is the first to achieve significant proactive robustness to collusion attacks, substantially increasing the FID of colluding models.