Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FNF: Functional Network Fingerprint for Large Language Models

Yiheng Liu, Junhao Ning, Sichen Xia, Haiyang Sun, Yang Yang, Hanyang Chi, Xiaohui Gao, Ning Qiang, Bao Ge, Junwei Han, Xintao Hu

TL;DR

FNF proposes a training-free fingerprinting method for LLM provenance based on functional network activity extracted from Transformer activations using CanICA. By thresholding spatial maps and computing Spearman correlations across time courses of functional networks from multiple inputs, it builds a cross-model fingerprint that remains stable under fine-tuning, data-variant training, pruning, architectural changes, permutation, and weight repackaging. The approach demonstrates strong discriminative power between models sharing an origin and unrelated models, outperforming the baseline REEF in cross-architecture settings and offering a practical tool for IP protection and model auditing. Its interpretability stems from ICA-decomposed networks, providing a functional-network perspective on model lineage and robustness to common obfuscation techniques.

Abstract

The development of large language models (LLMs) is costly and has significant commercial value. Consequently, preventing unauthorized appropriation of open-source LLMs and protecting developers' intellectual property rights have become critical challenges. In this work, we propose the Functional Network Fingerprint (FNF), a training-free, sample-efficient method for detecting whether a suspect LLM is derived from a victim model, based on the consistency between their functional network activity. We demonstrate that models that share a common origin, even with differences in scale or architecture, exhibit highly consistent patterns of neuronal activity within their functional networks across diverse input samples. In contrast, models trained independently on distinct data or with different objectives fail to preserve such activity alignment. Unlike conventional approaches, our method requires only a few samples for verification, preserves model utility, and remains robust to common model modifications (such as fine-tuning, pruning, and parameter permutation), as well as to comparisons across diverse architectures and dimensionalities. FNF thus provides model owners and third parties with a simple, non-invasive, and effective tool for protecting LLM intellectual property. The code is available at https://github.com/WhatAboutMyStar/LLM_ACTIVATION.

FNF: Functional Network Fingerprint for Large Language Models

TL;DR

FNF proposes a training-free fingerprinting method for LLM provenance based on functional network activity extracted from Transformer activations using CanICA. By thresholding spatial maps and computing Spearman correlations across time courses of functional networks from multiple inputs, it builds a cross-model fingerprint that remains stable under fine-tuning, data-variant training, pruning, architectural changes, permutation, and weight repackaging. The approach demonstrates strong discriminative power between models sharing an origin and unrelated models, outperforming the baseline REEF in cross-architecture settings and offering a practical tool for IP protection and model auditing. Its interpretability stems from ICA-decomposed networks, providing a functional-network perspective on model lineage and robustness to common obfuscation techniques.

Abstract

The development of large language models (LLMs) is costly and has significant commercial value. Consequently, preventing unauthorized appropriation of open-source LLMs and protecting developers' intellectual property rights have become critical challenges. In this work, we propose the Functional Network Fingerprint (FNF), a training-free, sample-efficient method for detecting whether a suspect LLM is derived from a victim model, based on the consistency between their functional network activity. We demonstrate that models that share a common origin, even with differences in scale or architecture, exhibit highly consistent patterns of neuronal activity within their functional networks across diverse input samples. In contrast, models trained independently on distinct data or with different objectives fail to preserve such activity alignment. Unlike conventional approaches, our method requires only a few samples for verification, preserves model utility, and remains robust to common model modifications (such as fine-tuning, pruning, and parameter permutation), as well as to comparisons across diverse architectures and dimensionalities. FNF thus provides model owners and third parties with a simple, non-invasive, and effective tool for protecting LLM intellectual property. The code is available at https://github.com/WhatAboutMyStar/LLM_ACTIVATION.
Paper Structure (21 sections, 5 equations, 4 figures, 13 tables, 1 algorithm)

This paper contains 21 sections, 5 equations, 4 figures, 13 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Preliminaries
  4. Method
  5. Functional Activation Consistency
  6. Extract Functional Networks in LLMs
  7. Functional Network Fingerprint
  8. Results
  9. Experiments Setting
  10. Evaluation
  11. Baseline Method
  12. Fine-tuning
  13. Training-data Variants
  14. Model Pruning
  15. Model Merging
  16. ...and 6 more sections

Figures (4)

  • Figure 1: The total framework of Functional Network Fingerprint.
  • Figure 2: Functional network activity of LLaMA2-7B-hf and LLaMA2-7B-chat-hf across different samples.
  • Figure 3: Functional network activity of LLaMA2-7B-chat-hf and fine-tuned variants trained on different amounts of data.
  • Figure 4: Examples of functional network activity from Amber and OpenLLaMA-7B.