Stylometry for real-world expert coders: a zero-shot approach.

Area of Science:

• Computer Science
• Software Engineering
• Digital Forensics

Background:

• Code stylometry identifies software authors using stylistic features.
• Existing methods often use limited, artificial datasets and struggle with unknown authors.
• Applications include plagiarism detection, code audits, and assignment verification.

Purpose of the Study:

• To challenge existing code stylometry assumptions by using real-world open-source code.
• To develop a method for recognizing authors not present in the training data (out-distribution authors).
• To create and utilize a novel, large-scale dataset for code stylometry.

Main Methods:

• Assembled a new dataset of 114,400 code snippets from 104 authors.
• Developed and trained a K-nearest neighbors (k-NN) classifier on this dataset.
• Evaluated performance on both in-distribution and out-distribution authors.

Main Results:

• Achieved 69% accuracy for in-distribution authors, exceeding the state-of-the-art by over 20%.
• Maintained high performance with out-distribution authors, achieving 71% accuracy.
• Demonstrated the effectiveness of the k-NN classifier on real-world code data.

Conclusions:

• The developed approach successfully identifies authors from real-world open-source projects.
• The method is robust in recognizing authors outside the initial training set.
• This work significantly advances code stylometry for practical applications.