iSeal: cryptographic fingerprinting for reliable LLM ownership verification

The paper proposes iSeal, which it claims is the first fingerprinting method that can still reliably verify ownership when a model thief has end-to-end control of a suspected LLM. The method injects unique features into the model and external modules, combined with an error-correction mechanism and a similarity verification strategy, to resist verification-time attacks such as collusion fingerprint forgetting and response tampering. Against more than 10 attacks on 12 LLMs, iSeal achieved a 100% fingerprint success rate, while baseline methods failed under fingerprint forgetting and response tampering.