Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Code Fingerprints: Disentangled Attribution of LLM-Generated Code

Jiaxun Guo, Ziyuan Yang, Mengyu Sun, Hui Wang, Jingfeng Lu, Yi Zhang

TL;DR

Experimental results demonstrate that DCAN achieves reliable attribution performance across diverse settings, highlighting the feasibility of model-level provenance analysis in software engineering contexts.

Abstract

The rapid adoption of Large Language Models (LLMs) has transformed modern software development by enabling automated code generation at scale. While these systems improve productivity, they introduce new challenges for software governance, accountability, and compliance. Existing research primarily focuses on distinguishing machine-generated code from human-written code; however, many practical scenarios--such as vulnerability triage, incident investigation, and licensing audits--require identifying which LLM produced a given code snippet. In this paper, we study the problem of model-level code attribution, which aims to determine the source LLM responsible for generated code. Although attribution is challenging, differences in training data, architectures, alignment strategies, and decoding mechanisms introduce model-dependent stylistic and structural variations that serve as generative fingerprints. Leveraging this observation, we propose the Disentangled Code Attribution Network (DCAN), which separates Source-Agnostic semantic information from Source-Specific stylistic representations. Through a contrastive learning objective, DCAN isolates discriminative model-dependent signals while preserving task semantics, enabling multi-class attribution across models and programming languages. To support systematic evaluation, we construct the first large-scale benchmark dataset comprising code generated by four widely used LLMs (DeepSeek, Claude, Qwen, and ChatGPT) across four programming languages (Python, Java, C, and Go). Experimental results demonstrate that DCAN achieves reliable attribution performance across diverse settings, highlighting the feasibility of model-level provenance analysis in software engineering contexts. The dataset and implementation are publicly available at https://github.com/mtt500/DCAN.

Code Fingerprints: Disentangled Attribution of LLM-Generated Code

TL;DR

Experimental results demonstrate that DCAN achieves reliable attribution performance across diverse settings, highlighting the feasibility of model-level provenance analysis in software engineering contexts.

Abstract

The rapid adoption of Large Language Models (LLMs) has transformed modern software development by enabling automated code generation at scale. While these systems improve productivity, they introduce new challenges for software governance, accountability, and compliance. Existing research primarily focuses on distinguishing machine-generated code from human-written code; however, many practical scenarios--such as vulnerability triage, incident investigation, and licensing audits--require identifying which LLM produced a given code snippet. In this paper, we study the problem of model-level code attribution, which aims to determine the source LLM responsible for generated code. Although attribution is challenging, differences in training data, architectures, alignment strategies, and decoding mechanisms introduce model-dependent stylistic and structural variations that serve as generative fingerprints. Leveraging this observation, we propose the Disentangled Code Attribution Network (DCAN), which separates Source-Agnostic semantic information from Source-Specific stylistic representations. Through a contrastive learning objective, DCAN isolates discriminative model-dependent signals while preserving task semantics, enabling multi-class attribution across models and programming languages. To support systematic evaluation, we construct the first large-scale benchmark dataset comprising code generated by four widely used LLMs (DeepSeek, Claude, Qwen, and ChatGPT) across four programming languages (Python, Java, C, and Go). Experimental results demonstrate that DCAN achieves reliable attribution performance across diverse settings, highlighting the feasibility of model-level provenance analysis in software engineering contexts. The dataset and implementation are publicly available at https://github.com/mtt500/DCAN.
Paper Structure (35 sections, 9 equations, 10 figures, 4 tables)

This paper contains 35 sections, 9 equations, 10 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Problem Formulation
  5. Overview
  6. Feature Extraction
  7. Feature Disentanglement
  8. Optimization and Learning Objectives
  9. Source Classification Loss
  10. Representation Consistency Loss
  11. Total Objective
  12. Dataset Construction
  13. Dataset Overview
  14. Construction Pipeline
  15. Prompt Design
  16. ...and 20 more sections

Figures (10)

  • Figure 1: Overview of the DCAN framework.
  • Figure 2: Dataset construction pipeline
  • Figure 3: Dataset Diversity and Complexity Analysis. (a) Difficulty Landscape: Visualizes the distribution of task difficulties across major algorithmic domains. (b) Fine-grained Topic Distribution: Provides a breakdown of sample counts across specific algorithmic tags. (c) Category Composition: Illustrates the ratio of Easy, Medium, and Hard tasks within each category.
  • Figure 4: Syntactic Distributional Differences Analysis. The distinct medians and interquartile ranges reveal consistent generative personas across different LLMs.
  • Figure 5: Stylometric Analysis of Generated Comments. (a) illustrates the varying degrees of helpfulness (density) across models, while (b) reveals distinct formatting preferences in annotation placement.
  • ...and 5 more figures