Awesome-LLM-Fingerprinting

This repository provides a collection of papers about LLM fingerprinting and is based on the paper entitled "SoK: Large Language Model Copyright Auditing via Fingerprinting".

Definition of LLM Fingerprinting

LLM fingerprinting is a passive approach that does not require modifying the model. Instead, it non-intrusively extracts a set of inherent yet distinctive characteristics that collectively serve as the model's "fingerprint". The core idea is that any model derived from the source model $M_o$ will preserve a statistically significant portion of this fingerprint.

In this paper and repository, we follow a classic narrow definition of LLM watermarking and fingerprinting. The primary difference between watermarking and fingerprinting is whether the method requires modifying the model's parameters and LLM fingerprinting is a non-intrusive method. Some existing papers embed "fingerprints" into LLMs by modifying their parameters. In this paper and repository, we classify these methods as LLM watermarking.

The primary advantage of LLM fingerprinting lies in its non-intrusive nature. Since it does not alter the model, it introduces no performance degradation. The computational overhead is typically confined, which is often less intensive than the fine-tuning required for watermarking. Consequently, fingerprinting offers a more flexible and widely applicable paradigm for copyright auditing, especially for models that are already in the public domain.

LLM Fingerprinting Benchmark

In this SoK, we also provide a comprehensive benchmark named LeaFBench for evaluating LLM fingerprinting methods.

Citation

If you find this repository useful, please cite our paper:

@article{shao2025sok,
    title={SoK: Large Language Model Copyright Auditing via Fingerprinting},
    author={Shao, Shuo and Li, Yiming and He, Yu and Yao, Hongwei and Yang, Wenyuan and Tao, Dacheng and Qin, Zhan},
    journal={arXiv preprint arXiv:2508.19843},
    year={2025}
}

List of Papers