This research introduces a scalable fingerprinting framework for large language models that enables owners to verify model ownership while maintaining robustness against detection evasion.

• Prioritizes scalability to support numerous fingerprints within a single model
• Defends against fingerprint leakage and coalition attacks from users
• Reduces false discovery rates while maintaining fingerprint effectiveness
• Provides concrete implementation strategies for model protection

For security teams, this research offers practical approaches to establish model provenance and protect intellectual property in an increasingly complex AI deployment landscape.

Scalable Fingerprinting of Large Language Models