A multi-input deep learning model for C/C++ source code attribution

Authors

Richard J. Tindell II, James Madison UniversityFollow

Preferred Name

Jeff

Creative Commons License

This work is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 International License.

ORCID

https://orcid.org/0000-0002-5163-6784

Date of Graduation

5-8-2020

Semester of Graduation

Spring

Degree Name

Master of Science (MS)

Department

Department of Computer Science

Second Advisor

Nathan Sprague

Third Advisor

Brett Tjaden

Abstract

Code stylometry is applying analysis techniques to a collection of source code or binaries to determine variations in style. The variations extracted are often used to identify the author of the text or to differentiate one piece from another.

In this research, we were able to create a multi-input deep learning model that could accurately categorize and group code from multiple projects. The deep learning model took as input word-based tokenization for code comments, character-based tokenization for the source code text, and the metadata features described by A. Caliskan-Islam et al. Using these three inputs, we were able to achieve 90% validation accuracy with a loss value of 0.1203 using 12 projects consisting of 5,877 files. Finally, we analyzed the Bitcoin source code using our data model showing a high probability match to the OpenSSL project.