TrFP fingerprints LLMs from how representations change with depth
Original title: From Representations to Computational Trajectories: Intrinsic Fingerprinting for Large Language Model Provenance Verification
Fingerprints stay under 0.3 MiB and remain stable after fine-tuning, pruning, quantization, merging, and permutation.
The paper presents TrFP, an intrinsic fingerprint for large language model provenance that records how representations evolve across network depth rather than comparing parameters or static representations. Compact, permutation-invariant trajectory descriptors are stored as the fingerprint. Across multiple LLM families, the authors report reliable separation of related and unrelated models and robustness to fine-tuning, pruning, quantization, model merging, parameter permutation, and adaptive attacks, with storage under 0.3 MiB.
Unconfirmed: The robustness and discrimination claims are the authors’ experimental results, not an independent audit.
- The method does not modify the model or embed ownership marks.
- Fingerprints use permutation-invariant descriptors of representation trajectories.
- Storage is reported as less than 0.3 MiB.
- Tested transformations include merging, quantization, pruning, and adaptive attacks.
From the original article
1 min read at arxiv.orgLarge language models (LLMs) are valuable digital assets, yet their weights can be easily modified after unauthorized acquisition, making model provenance difficult to verify. Model fingerprinting provides a promising solution for model ownership protection without modifying the model or embedding ownership information.
By Haiyong Tang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.