From Construction to Injection: Edit-Based Fingerprints for Large Language Models
The framework proposes an end-to-end system to create reliable model fingerprints for LLMs. It tackles the trade-off between making fingerprints hard to accidentally activate and ensuring they remain effective even when the model is changed. The solution involves constructing code-mixing fingerprints and injecting them using a multi-candidate editing technique, achieving robust ownership verification without harming model performance.