Using Classification Techniques to Determine Source Code Authorship

The ability to test authorship of source code is a useful technique. We present techniques for using statistical machine learning to accomplish this task. We translate source code into abstract syntax trees and then split up the trees into functions. The tree for each function is considered a document, with a given author. This collection is fed to an SVM package using a kernel that operates on tree structured data. The classifier is trained with source code from two authors, and is then able to prediction which of the two authors wrote a new function. We achieve between 67% and 88% classification accuracy over the set of programs we examined. This demonstrates that our techniques are useful either alone, or in concert with other methods.