Comparing Machine Learning and Deep Learning Models With Attention Mechanisms for Multiple Source Code Authorship Attribution

Authorship attribution (AA) in source code is crucial for plagiarism detection, forensic analysis, and software maintenance, especially when author information is missing or ambiguous. This field is not as well explored as the written communication AA and there are new models to be tested on different sized dataset. This study compares traditional machine learning models, such as Random Forests trained on TF-IDF tokenized representations, with deep learning models, including BERT and an Attention-based Neural Network (AttentionNN), for the task of multiple authorship attribution. We evaluate different representations of source code, including raw tokenized text and Abstract Syntax Tree (AST) structures, to determine their effectiveness in distinguishing authorship. The models are tested on various sized but relatively small datasets compared to the existing experiments, the largest consisting of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3,756 \mathrm{C}$</tex> language functions from Google Code Jam spanning 2009-2018, assessing their accuracy, precision, recall, and training efficiency across varying author set sizes. Our findings highlight the strengths and limitations of each approach, showing that while deep learning models, particularly BERT trained on raw source code, achieve the highest accuracy, Random Forests offer competitive results with significantly lower computational cost. The study also explores the impact of function complexity on model performance, revealing that AttentionNN benefits from structural complexity while struggling with short functions. The results emphasize the trade-offs between accuracy, computational efficiency, and data representation, providing insights into the optimal model selection for source code authorship attribution.