Fingerprint Vector: Enabling Scalable and Efficient Model Fingerprint Transfer via Vector Addition

ACL ARR 2026 January Submission9521 Authors

Keywords: Large Language Model, Copyright Protection, LLM Fingerprinting, Backdoor Watermarking, Fingerprint Transfer

Abstract: Backdoor-based fingerprinting is a widely used technique for verifying the ownership of large language models, but it scales poorly when a single foundation model is fine-tuned into numerous downstream variants. Fingerprinting each model individually is costly, while inheritance-based approaches that embed fingerprints into the base model suffer from late-stage fingerprinting, instability under further training, and interference with downstream adaptation. We propose the \textbf{Fingerprint Vector}, a post-hoc ownership transfer mechanism that decouples fingerprint information from the base model. The fingerprint vector is obtained as the parameter-space difference between a fingerprinted model and its clean foundation, and can be directly applied to any structurally compatible downstream model without additional fine-tuning. Extensive experiments show that transferred fingerprints achieve effectiveness and harmlessness comparable to direct fine-tuning-based fingerprinting, while maintaining strong robustness under common post-deployment modifications. These results demonstrate that \textbf{Fingerprint Vector} enables efficient “fingerprint-once, transfer-many” ownership protection for large model families.

Paper Type: Long

Research Area: Safety and Alignment in LLMs

Research Area Keywords: security and privacy

Contribution Types: NLP engineering experiment, Theory

Languages Studied: English

Submission Number: 9521