From Construction to Injection: Edit-Based Fingerprints for Large Language Models
Researchers have developed a new framework for creating robust fingerprints for large language models (LLMs) to prevent unauthorized use. The proposed method, called Code-mixing Fingerprints (CF), uses low-perplexity code-mixing under high-complexity constraints to balance imperceptibility and detectability. Additionally, a technique called Multi-Candidate Editing (MCEdit) creates redundant trigger-target mappings that degrade gracefully when models are modified, ensuring persistent ownership verification. AI
IMPACT Enhances LLM security by providing more robust methods to track ownership and prevent misuse.