Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Provable Model Provenance Set for Large Language Models

Xiaoqi Qiu, Hao Zeng, Zhiyu Hou, Hongxin Wei

TL;DR

This work formalizes model provenance as a provable model-set identification problem and introduces Model Provenance Set (MPS), a sequential test-and-exclusion method that constructs a compact set of all true provenance models with a user-defined confidence $1 - \alpha$. It relies on a relative-distance test statistic and permutation-based $p$-values to detect provenance signals among a pool of candidates, enabling multi-source attribution with rigorous error control. The authors prove asymptotic coverage guarantees and efficiency, and demonstrate robust performance on a large Hugging Face benchmark, showing high coverage with small set sizes and improved reliability over empirical baselines. The approach is fingerprint-agnostic and suitable for attribution and auditing tasks, with practical implications for IP protection of LLMs.

Abstract

The growing prevalence of unauthorized model usage and misattribution has increased the need for reliable model provenance analysis. However, existing methods largely rely on heuristic fingerprint-matching rules that lack provable error control and often overlook the existence of multiple sources, leaving the reliability of their provenance claims unverified. In this work, we first formalize the model provenance problem with provable guarantees, requiring rigorous coverage of all true provenances at a prescribed confidence level. Then, we propose the Model Provenance Set (MPS), which employs a sequential test-and-exclusion procedure to adaptively construct a small set satisfying the guarantee. The key idea of MPS is to test the significance of provenance existence within a candidate pool, thereby establishing a provable asymptotic guarantee at a user-specific confidence level. Extensive experiments demonstrate that MPS effectively achieves target provenance coverage while strictly limiting the inclusion of unrelated models, and further reveal its potential for practical provenance analysis in attribution and auditing tasks.

Provable Model Provenance Set for Large Language Models

TL;DR

This work formalizes model provenance as a provable model-set identification problem and introduces Model Provenance Set (MPS), a sequential test-and-exclusion method that constructs a compact set of all true provenance models with a user-defined confidence . It relies on a relative-distance test statistic and permutation-based -values to detect provenance signals among a pool of candidates, enabling multi-source attribution with rigorous error control. The authors prove asymptotic coverage guarantees and efficiency, and demonstrate robust performance on a large Hugging Face benchmark, showing high coverage with small set sizes and improved reliability over empirical baselines. The approach is fingerprint-agnostic and suitable for attribution and auditing tasks, with practical implications for IP protection of LLMs.

Abstract

The growing prevalence of unauthorized model usage and misattribution has increased the need for reliable model provenance analysis. However, existing methods largely rely on heuristic fingerprint-matching rules that lack provable error control and often overlook the existence of multiple sources, leaving the reliability of their provenance claims unverified. In this work, we first formalize the model provenance problem with provable guarantees, requiring rigorous coverage of all true provenances at a prescribed confidence level. Then, we propose the Model Provenance Set (MPS), which employs a sequential test-and-exclusion procedure to adaptively construct a small set satisfying the guarantee. The key idea of MPS is to test the significance of provenance existence within a candidate pool, thereby establishing a provable asymptotic guarantee at a user-specific confidence level. Extensive experiments demonstrate that MPS effectively achieves target provenance coverage while strictly limiting the inclusion of unrelated models, and further reveal its potential for practical provenance analysis in attribution and auditing tasks.
Paper Structure (35 sections, 3 theorems, 24 equations, 6 figures, 2 tables, 1 algorithm)

This paper contains 35 sections, 3 theorems, 24 equations, 6 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Provenance relationship.
  4. Model similarity score.
  5. Model provenance set.
  6. Objective.
  7. Methodology
  8. The design of test statistic
  9. Hypothesis testing using $t$-statistics
  10. Sequential test-and-exclusion algorithm
  11. Experiments
  12. Conclusion
  13. Related Work
  14. Connection to Model Confidence Set
  15. Proofs
  16. ...and 20 more sections

Key Result

Theorem 3.1

For any model $f_i, f_j \in \mathcal{M}$, suppose that (a) prompts $\{x_t\}_{t=1}^N$ are sampled independently, (b) $\mathbb{E}[|d_{ij,t}|^{2+\delta}] < \infty$ for some $\delta > 0$, and (c) $\sigma^2_{ij} = \mathrm{Var}(d_{ij,t}) > 0$ whenever $\mu_i = \mu_j$. Then the permutation-based $p$-value

Figures (6)

  • Figure 1: Overview of the model provenance set framework. Given a target model $g$ and candidate models $\mathcal{M}$, our MPS implements a sequential test-and-exclusion procedure to build a provenance set $\hat{\mathcal{M}}$ satisfying $Pr(\mathcal{M}^*\subseteq \hat{\mathcal{M}}) \ge 1-\alpha$, where $\mathcal{M}^*$ is the true provenance set. Orange circles in $L$ indicate the distances between $g$ and true provenances, and $f^*$ is the most similar candidate model.
  • Figure 2: Examples of derivation chains with $\mathrm{TAM}\!=\!3$. Red nodes means the final child LLM versions after multiple fine-tuning.
  • Figure 3: Instruction template for prompt generation
  • Figure 4: Averaged coverage and set size of MPS procedure with different score functions with $\alpha=0.1$ (Top) and $\alpha=0.15$ (Bottom). $|\mathcal{M}|$ denotes the candidate set size; $\mathrm{TAM}$ is the number of true ancestor models.
  • Figure 5: Similarity score distributions for unrelated pairs (red) and lineage-related pairs at different derivation distances. 1-hop refers to parent–child models (blue), and 2-hop indicates grandparent–grandchild (green).
  • ...and 1 more figures

Theorems & Definitions (10)

  • Definition 2.1: Provenance relationship
  • Definition 2.2: True provenance set
  • Remark 2.3
  • Theorem 3.1: Valid $p$-value
  • Theorem 3.2: Coverage guarantee
  • Theorem 3.3: Asymptotic efficiency
  • Remark 3.4: Minimum detectable gap
  • proof
  • proof : Proof
  • proof