FPEdit: Robust LLM Fingerprinting through Localized Parameter Editing
A new framework called FPEdit has been developed by researchers to fingerprint large language models (LLMs) in order to safeguard against unauthorized distribution and commercial misuse. Existing techniques often involve a compromise: intrinsic methods necessitate complete access to parameters, while backdoor strategies utilize statistically unusual triggers that can be easily identified. FPEdit utilizes knowledge editing to embed semantically coherent natural language fingerprints through sparse, targeted adjustments to model weights. It incorporates Promote-Suppress Value Vector Optimization, which boosts the likelihood of target tokens while minimizing the presence of competing tokens, thereby ensuring effective fingerprinting without compromising model performance. The full paper can be found on arXiv with the identifier 2508.02092.
Key facts
- FPEdit is a framework for fingerprinting LLMs.
- It uses knowledge editing to inject natural language fingerprints.
- Fingerprints are injected through sparse, targeted weight modifications.
- Promote-Suppress Value Vector Optimization enhances target tokens and suppresses competitors.
- Current fingerprinting methods have a trade-off between parameter access and trigger detectability.
- The paper is published on arXiv with ID 2508.02092.
- LLMs are valuable intellectual assets vulnerable to unauthorized redistribution.
- FPEdit aims to protect against fine-tuning and black-box deployment theft.
Entities
Institutions
- arXiv