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

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Model fingerprinting embeds identifiers into generative model outputs to protect intellectual property rights (IPR).
• Existing fingerprinting methods for text-to-image (T2I) models are vulnerable to collusion attacks, where multiple attackers combine models to remove fingerprints.

Purpose of the Study:

• To develop a robust fingerprinting method for T2I models with anti-collusion capabilities.
• To address the vulnerability of current fingerprinting techniques against coordinated attacks.

Main Methods:

• Proposed a novel method encoding fingerprints into the coefficients of a personalized normalization module (PNM) within T2I models.
• Introduced an anti-collusion mechanism using lossless function-invariant parameter transformations to degrade colluded model quality.
• Implemented a worst-case optimization strategy for enhanced robustness against model-level attacks.

Main Results:

• Achieved high fidelity and robustness in T2I generation and editing tasks, with fingerprint extraction accuracy exceeding 99.5%.
• Demonstrated proactive robustness against collusion attacks by significantly increasing the Fréchet Inception Distance (FID) of colluded models.
• Enabled efficient creation of multiple fingerprinted model copies via PNM reparameterization without retraining.

Conclusions:

• The proposed fingerprinting method offers a robust solution against collusion attacks for T2I models.
• The technique effectively protects intellectual property while maintaining high image generation quality.
• This work represents a significant advancement in securing generative AI models against unauthorized use and redistribution.