Code authorship identification via deep graph CNNs
Chandler Holland, Navid Khoshavi, Luis G. Jaimes · 2022
Code authorship identification (CAI) is an emerging field useful to identify code plagiarism, settle copyright disputes, and it is in general, an essential tool for code forensics. However, scaling CAI has always been limited by domain knowledge and the burden of feature engineering. Recent works in CAI have demonstrated the utility of using Convolutional Neural Networks (CNN) to avoid this problem. Unfortunately, the use of CNNs for CAI imposes a rigid representation of programs, limiting expressions of dependency and semantic relationships. In this paper, we explore the use of Graph Deep CNN for CAI. Here, programs are represented by graphs which allow to express their complex relationships. We use a subset of Google Code Jam as a dataset and a methodology for transforming these programs into graphs. After extensive simulations, we show our methodology performs well in terms of scalability, accuracy, and loss.