Source code author identification with unsupervised feature learning

Abstract

Automatic identification of source code authors has many applications in different fields such as source code plagiarism detection, and law suit prosecution. This paper presents a new source code author identification system based on an unsupervised feature learning technique. As a method of extracting features from high dimensional data, unsupervised feature learning has obtained a great success in many fields such as character recognition and image classification. However, according to our knowledge it has not been applied for source code author identification systems. Therefore, we investigated an unsupervised feature learning technique called sparse auto-encoder as a method of extracting features from source code files. Our system was evaluated with several datasets and results have shown that performance is very close to the state of art techniques in the source code identification field.

Introduction

Automatic identification of source code authors is an important activity in many fields in computing. Among those fields, source code plagiarism detection is one of the main fields which can be easily benefited by the source code author identification. According to Zobel (2004), “students may plagiarize by copying code from friends, the Web or so called “private tutors”. Therefore, detection and prevention of source code plagiarism is essential for programming assignments.

Chao et al. (2006) have mentioned that “A quality plagiarism detector has a strong impact to law suit prosecution”. Furthermore, Lange and Mancoridis (2007) have pointed out that, source code author identification is useful in criminal justice and corporate litigation as well.

In this research we investigated a novel technique in machine learning called unsupervised feature learning (Raina et al., 2007) in order to improve the performance of source code author identification systems. Recently unsupervised feature learning algorithms have obtained a huge attention in the machine learning community (Raina et al., 2007) and have been successfully used for many applications. However, according to our knowledge it has not been tried for source code author identification. Therefore, we investigated one such learning technique called sparse auto-encoder (Coates et al., 2010).

This paper is organized as follows. In Section 2, we present previous works related to source code author identification. The architecture of our system is described in Section 3. In Section 4, we describe training, and testing of the system. Finally, we conclude this paper by presenting our conclusions and discussing further improvements.