LCF: A Unified Fingerprinting Framework for LLMs
A new research paper introduces LCF (Construction-Driven Injection), a unified framework for fingerprinting large language models (LLMs) to protect against unauthorized redistribution and commercial misuse. Existing methods separate fingerprint construction from injection, leading to issues like accidental activation of natural-language fingerprints and easy filtering of garbled fingerprints via perplexity-based detection. LCF jointly optimizes both stages, making injection aware of the trigger's linguistic structure for targeted optimization. The framework constructs code-mixing fingerprints using edit-based techniques, combining natural language and code to improve robustness and stealth. The paper is available on arXiv under ID 2607.25633.
Key facts
- LLMs are costly intellectual assets exposed to unauthorized redistribution and commercial misuse.
- Injected fingerprints are trigger-target pairs embedded in model behavior for ownership verification.
- Existing fingerprinting frameworks decouple construction from injection.
- Natural-language fingerprints are prone to accidental activation.
- Garbled fingerprints are easily filtered by perplexity-based detection.
- LCF jointly optimizes fingerprint construction and injection.
- LCF uses code-mixing fingerprints with edit-based techniques.
- Paper available on arXiv: 2607.25633.
Entities
Institutions
- arXiv