Reassessing Code Authorship Attribution in the Era of Language Models

TL;DR

This paper evaluates the effectiveness of large language models for code authorship attribution, revealing their strengths and limitations in identifying coding styles across diverse datasets.

Contribution

It provides the first comprehensive empirical analysis of transformer-based language models for code authorship attribution on multiple datasets.

Findings

Large language models show promising results in CAA tasks.

Model interpretability reveals how LMs understand coding styles.

Insights suggest future directions for improving CAA accuracy.

Abstract

The study of Code Stylometry, and in particular Code Authorship Attribution (CAA), aims to analyze coding styles to identify the authors of code samples. CAA is crucial in cybersecurity and software forensics for addressing, detecting plagiarism, and supporting criminal prosecutions. However, CAA is a complex and error prone task, due to the need for recognizing nuanced relationships between coding patterns. This challenge is compounded in large software systems with numerous authors due to the subtle variability of patterns that signify the coding style of one author among many. Given the challenges related to this task, researchers have proposed and studied automated approaches that rely upon classical Machine Learning and Deep Learning techniques. However, such techniques have historically relied upon hand-crafted features, and due to the often intricate interaction of different features (e.g., formatting, etc.), have key limitations in properly characterizing authorship, and are sensitive to adversarial code perturbations. Recently, transformer-based Language Models (LMs) have shown remarkable efficacy across a range of software engineering tasks, and in the authorship attribution on natural language in the NLP domain. However, their effectiveness in CAA is not well understood. As such, we conduct the first extensive empirical study applying two larger state-of-the-art code LMs, and five smaller code LMs to the task of CAA to 6 diverse datasets that encompass 12k code snippets written by 463 developers. Furthermore, we perform an in-depth analysis of our studied models' performance on CAA using established machine learning interpretability techniques. The results of our analysis illustrate important findings that illuminate the behavior of LMs in understanding stylometric code patterns during the task of CAA, and point towards important directions for future work.