TrFP: An Intrinsic Fingerprinting Provenance Method for LLMs Based on Computational Trajectories
Start here
The researchers propose TrFP, an intrinsic LLM fingerprinting method based on computational trajectories: a permutation-invariant trajectory descriptor captures the evolution of representations with network depth and constructs a fingerprint, enabling source verification without modifying the model. It reliably distinguishes related from unrelated models across multiple LLM families, remains robust under fine-tuning, pruning, quantization, model merging, parameter permutation, and adaptive attacks, requires less than 0.3 MiB of storage, and outperforms multiple SOTA methods in source discrimination.
Generated by AI from reports · Updated 13 days ago
Attention and developments
Attention, hourly
Now 24, 0% in 24 hours·Peak 24 (Nov 6 03:00)
Related research
This event's paper
- From Representations to Computational Trajectories: Intrinsic Fingerprinting for Large Language Model Provenance Verification
arXiv · 2026-10-26 · Deep reading
Progress timeline
- arXivFrom Representations to Computational Trajectories: An Intrinsic Fingerprinting Method TrFP for Large Language Model Provenance Verification
The researchers propose TrFP, an intrinsic fingerprinting method for large language models based on computational trajectories, which can verify model provenance without modifying the model. TrFP captures how representations evolve with network depth and uses permutation-invariant trajectory descriptors to build fingerprints, reliably distinguishing related from unrelated models across multiple LLM series. The fingerprints remain robust under fine-tuning, pruning, quantization, model merging, parameter permutation, and adaptive attacks, require less than 0.3 MiB of storage, and outperform multiple SOTA methods in provenance discrimination.