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 HuarXiv 2026 · arXiv

TL;DR

The Functional Network Fingerprint (FNF) is a novel, training-free method that protects the intellectual property of Large Language Models (LLMs) by analyzing their internal "brain-like" functional network activity. By measuring the dynamic co-activation patterns of neurons across just a few text samples, FNF can reliably identify if a suspect model was stolen or modified from a victim model, even if the thief altered the model's architecture, dimensions, or parameters.

What Problem Does This Paper Address?

Training state-of-the-art LLMs costs millions of dollars, making model weights highly valuable intellectual property. However, bad actors can easily steal open-source or proprietary weights, apply superficial modifications—such as fine-tuning, parameter permutation, or pruning—and claim the model is a completely original creation built from scratch.

Worse yet is a subtle plagiarism technique called "weight repackaging," where a developer alters the model's structural dimensions (e.g., adding layers or expanding MLP dimensions) while keeping the core inherited weights intact. Traditional defense mechanisms fail here: watermarking techniques degrade model performance, while existing representation-based fingerprints (like REEF) break entirely when comparing models with mismatched hidden dimensions or architectures. The authors set out to create a non-invasive fingerprinting method robust enough to track a model's lineage across different scales and structures.

Methods

Inspired by neuroscience techniques used to map functional connectivity in the human brain, the authors propose Functional Network Fingerprinting (FNF). The core idea is to treat neuron activations within an LLM block as fMRI-like signals.

The FNF pipeline consists of the following steps:

  1. Activation Extraction: Pass identical text sequences through both a victim model and a suspect model, extracting the raw neuron signals from the output of specific Transformer blocks.
  2. Network Disentanglement: Apply Canonical Independent Component Analysis (CanICA)—a standard neuroimaging tool—to group multi-sample signal matrices, reducing dimensionality via PCA whitening and isolating $K$ statistically independent spatial functional networks.
  3. Thresholding & Masking: Apply a fixed cutoff to isolate the most salient, core neurons within each functional network, creating a binary mask.
  4. Activity Time Courses: Compute the mean activation of these masked neurons over the token sequence, yielding a 1D dynamic time course for each network.
  5. Correlation Matrix: Calculate the Spearman rank correlation between all pairs of functional time courses across the two models. Spearman is chosen over Pearson because it evaluates qualitative trend consistency rather than absolute linear values, making it highly robust to monotonic transformations.
Figure 1. The total framework of Functional Network Fingerprint.

Datasets

The authors evaluated FNF using several standard NLP datasets to assess how different text distributions affect fingerprint reliability:

  • WikiText-2: The primary evaluation dataset used to extract and evaluate model similarity.
  • SQuAD & SST-2: Used in ablation studies to test the impact of dataset types and sequence lengths.
  • Random Token Dataset: A synthetic dataset of meaningless, randomly generated tokens used to evaluate the method's boundaries.

Key Insights and Ideas

  • "Neural Species" Lineage: Models originating from the same base weights share an underlying "neuroanatomical blueprint." Even if they are heavily fine-tuned or structurally expanded, their core functional networks co-activate in highly consistent patterns when exposed to the same stimuli.
  • Dynamics Over Weights: By focusing on the dynamic co-activation patterns of groups of neurons rather than static weight matrices, FNF is completely immune to parameter permutation (shuffling weights) and structural disguises.
  • Meaning Doesn't Matter, Length Does: Surprisingly, the functional networks align even when the input consists of completely random, meaningless tokens. The critical factor isn't semantic meaning, but sequence length; the models just need enough tokens to establish a statistically reliable activation trend.

Key Results

The authors compared FNF against Centered Kernel Alignment (CKA), a strong representation-based baseline:

  • Fine-Tuning & Pruning Robustness: FNF successfully identified the shared lineage between pre-trained LLaMA2-7B and its fine-tuned chat variant (FNF score of 0.9493). It also maintained strong correlation scores (>0.50) even after severe 50% parameter pruning using advanced methods like Wanda-sp, where CKA scores collapsed to near zero.
Figure 2. Functional network activity of LLaMA2-7B-hf and LLaMA2-7B-chat-hf across different samples.
  • Defeating Weight Repackaging: In cross-model comparisons against structural expansion (e.g., Mistral-7B vs. Solar-10.7B, which expands depth from 32 to 48 layers), FNF successfully detected their common origin (score: 0.5743), while CKA failed (score: 0.3226).
  • Low False Positives: When comparing truly unrelated models trained from scratch on different data (e.g., LLaMA2-7B vs. Amber or OpenLLaMA), FNF yielded very low scores (~0.20 to 0.30), proving it doesn't misidentify distinct models.

Limitations and Future Work

  • Sequence Length Dependency: FNF requires input sequences with a sufficient number of tokens (typically >100). On short-text datasets like SST-2 (averaging ~10 tokens per sample), the correlation estimation becomes significantly less reliable.
  • Hyperparameter Tuning: The number of target components ($K$) must be chosen empirically. While $K=64$ proved optimal in these experiments, very small or excessively large values can degrade fingerprint clarity.
  • Future Directions: The authors suggest that exploring how these functional networks map to specific safety boundaries or behavioral traits could open up new avenues for mechanistic interpretability and model auditing.

Summary updated

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.
BibTeX
@misc{liu2026fnf,
  author = {Yiheng Liu and Junhao Ning and Sichen Xia and Haiyang Sun and Yang Yang and Hanyang Chi and Xiaohui Gao and Ning Qiang and Bao Ge and Junwei Han and Xintao Hu},
  title = {{FNF: Functional Network Fingerprint for Large Language Models}},
  year = {2026},
  eprint = {2601.22692},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.22692},
}
Figure