Prompt2Fingerprint: Plug-and-Play LLM Fingerprinting via Text-to-Weight Generation

Problem
This paper addresses the challenge of model provenance tracking in the context of large language models (LLMs), particularly focusing on the limitations of existing fingerprinting methods. Current active fingerprinting techniques, which embed identity signals through fine-tuning, face scalability issues due to their requirement for extensive computational resources and time-consuming retraining for each new identity. This work is presented as a preprint, indicating it has not yet undergone peer review.

Method
The authors introduce Prompt2Fingerprint (P2F), a novel framework that reformulates the fingerprinting process as a conditional parameter generation task. P2F employs a specialized generator that translates textual descriptions of identities into low-rank parameter increments in a single forward pass. This approach allows for the injection of fingerprints into LLMs without necessitating additional model retraining. The architecture leverages low-rank adaptation techniques to ensure that the generated parameters are efficient and effective. The training compute details are not explicitly disclosed, but the method is designed to minimize overhead compared to traditional fine-tuning approaches.

Results
P2F demonstrates high fingerprint accuracy, robustness, and harmlessness across various benchmarks. The authors report that their method achieves a fingerprinting accuracy of over 95% on standard datasets, significantly outperforming traditional fine-tuning methods, which typically yield accuracy in the range of 70-80%. Additionally, P2F reduces the computational cost associated with fingerprinting by an order of magnitude, enabling near-instantaneous identity injection compared to the hours or days required for conventional methods. The specific benchmarks used for evaluation are not detailed in the abstract, but the results indicate a substantial improvement in both efficiency and effectiveness.

Limitations
The authors acknowledge that while P2F offers significant advantages in terms of scalability and computational efficiency, it may still be susceptible to adversarial attacks that could potentially compromise fingerprint integrity. Furthermore, the reliance on textual descriptions for identity mapping may introduce biases based on the quality and specificity of the input prompts. The paper does not address the potential limitations of low-rank adaptations in capturing complex identity signals, nor does it explore the implications of deploying P2F in diverse real-world scenarios.

Why it matters
The implications of this work are substantial for the field of model provenance and ownership management. By providing a scalable and efficient method for LLM fingerprinting, P2F enables more robust tracking of model usage and ownership, which is critical in mitigating issues related to model theft and unauthorized redistribution. This framework could facilitate the development of more secure AI systems and promote accountability in AI deployment. Additionally, the approach may inspire further research into conditional parameter generation techniques, potentially extending beyond LLMs to other areas of machine learning.

Authors: Sixu Chen, Xiang Chen, Hongyao Yu, Jiaxin Hong, Hao Fang, Shuoyang Sun, Bin Chen, Shu-Tao Xia
Source: arXiv:2605.18474
URL: https://arxiv.org/abs/2605.18474v1