Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Trained Without My Consent: Detecting Code Inclusion In Language Models Trained on Code

Vahid Majdinasab, Amin Nikanjam, Foutse Khomh

TL;DR

TraWiC is a model-agnostic and interpretable method based on membership inference for detecting code inclusion in an LLM’s training dataset, capable of detecting 83.87% of codes that were used to train an LLM.

Abstract

Code auditing ensures that the developed code adheres to standards, regulations, and copyright protection by verifying that it does not contain code from protected sources. The recent advent of Large Language Models (LLMs) as coding assistants in the software development process poses new challenges for code auditing. The dataset for training these models is mainly collected from publicly available sources. This raises the issue of intellectual property infringement as developers' codes are already included in the dataset. Therefore, auditing code developed using LLMs is challenging, as it is difficult to reliably assert if an LLM used during development has been trained on specific copyrighted codes, given that we do not have access to the training datasets of these models. Given the non-disclosure of the training datasets, traditional approaches such as code clone detection are insufficient for asserting copyright infringement. To address this challenge, we propose a new approach, TraWiC; a model-agnostic and interpretable method based on membership inference for detecting code inclusion in an LLM's training dataset. We extract syntactic and semantic identifiers unique to each program to train a classifier for detecting code inclusion. In our experiments, we observe that TraWiC is capable of detecting 83.87% of codes that were used to train an LLM. In comparison, the prevalent clone detection tool NiCad is only capable of detecting 47.64%. In addition to its remarkable performance, TraWiC has low resource overhead in contrast to pair-wise clone detection that is conducted during the auditing process of tools like CodeWhisperer reference tracker, across thousands of code snippets.

Trained Without My Consent: Detecting Code Inclusion In Language Models Trained on Code

TL;DR

TraWiC is a model-agnostic and interpretable method based on membership inference for detecting code inclusion in an LLM’s training dataset, capable of detecting 83.87% of codes that were used to train an LLM.

Abstract

Code auditing ensures that the developed code adheres to standards, regulations, and copyright protection by verifying that it does not contain code from protected sources. The recent advent of Large Language Models (LLMs) as coding assistants in the software development process poses new challenges for code auditing. The dataset for training these models is mainly collected from publicly available sources. This raises the issue of intellectual property infringement as developers' codes are already included in the dataset. Therefore, auditing code developed using LLMs is challenging, as it is difficult to reliably assert if an LLM used during development has been trained on specific copyrighted codes, given that we do not have access to the training datasets of these models. Given the non-disclosure of the training datasets, traditional approaches such as code clone detection are insufficient for asserting copyright infringement. To address this challenge, we propose a new approach, TraWiC; a model-agnostic and interpretable method based on membership inference for detecting code inclusion in an LLM's training dataset. We extract syntactic and semantic identifiers unique to each program to train a classifier for detecting code inclusion. In our experiments, we observe that TraWiC is capable of detecting 83.87% of codes that were used to train an LLM. In comparison, the prevalent clone detection tool NiCad is only capable of detecting 47.64%. In addition to its remarkable performance, TraWiC has low resource overhead in contrast to pair-wise clone detection that is conducted during the auditing process of tools like CodeWhisperer reference tracker, across thousands of code snippets.
Paper Structure (55 sections, 6 figures, 21 tables, 5 algorithms)

This paper contains 55 sections, 6 figures, 21 tables, 5 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. Large Language Models
  4. Code Clone Detection
  5. Membership Inference Attacks
  6. TRAWIC
  7. Definitions
  8. Motivating Example
  9. Next token prediction
  10. Name cloze prediction
  11. End-to-End Data Processing Example
  12. Methodology
  13. Pre-processing
  14. Inference
  15. Comparison
  16. ...and 40 more sections

Figures (6)

  • Figure 1: TraWiC's dataset inclusion detection pipeline
  • Figure 2: An example of how prefix and suffix are generated for NiCad and JPlag
  • Figure 3: Feature importance of the final dataset with different edit distance thresholds - SantaCoder
  • Figure 4: Correlation of different features in the generated dataset
  • Figure 5: Feature importance of Llama-2 and Mistral with different edit distance thresholds
  • ...and 1 more figures