From Construction to Injection: Edit-Based Fingerprints for Large Language Models

The article presents an end-to-end injected fingerprinting framework for large language models (LLMs) aimed at improving ownership verification against unauthorized use. It introduces Code-mixing Fingerprints (CF) and Multi-Candidate Editing (MCEdit) to overcome imperceptibility trade-offs and maintain persistent behaviors despite model modifications. This framework is crucial for practitioners as it enhances the robustness of model ownership verification while minimizing impacts on model utility.