Source code authorship attribution

To attribute authorship means to identify the true author among many candidates for samples of work of unknown or contentious authorship. Authorship attribution is a prolific research area for natural language, but much less so for source code, with eight other research groups having published empirical results concerning the accuracy of their approaches to date. Authorship attribution of source code is the focus of this thesis.We first review, reimplement, and benchmark all existing published methods to establish a consistent set of accuracy scores. This is done using four newly constructed and significant source code collections comprising samples from academic sources, freelance sources, and multiple programming languages. The collections developed are the most comprehensive to date in the field.We then propose a novel information retrieval method for source code authorship attribution. In this method, source code features from the collection samples are tokenised, converted into n-grams, and indexed for stylistic comparison to query samples using the Okapi BM25 similarity measure. Authorship of the top ranked sample is used to classify authorship of each query, and the proportion of times that this is correct determines overall accuracy. The results show that this approach is more accurate than the best approach from the previous work for three of the four collections.The accuracy of the new method is then explored in the context of author style evolving over time, by experimenting with a collection of student programming assignments that spans three semesters with established relative timestamps. We find that it takes one full semester for individual coding styles to stabilise, which is essential knowledge for ongoing authorship attribution studies and quality control in general.We conclude the research by extending both the new information retrieval method and previous methods to provide a complete set of benchmarks for advancing the field. In the final evaluation, we show that the n-gram approaches are leading the field, with accuracy scores for some collections around 90% for a one-in-ten classification problem.