Abstract
Version control systems (VCSs) have almost become the de facto standard for the management of open-source projects and the development of their source code. In VCSs, source code which can potentially be vulnerable is introduced to a system through what are so called commits. Vulnerable commits force the system into an insecure state. The farreaching impact of vulnerabilities attests to the importance of identifying and understanding the characteristics of prior vulnerable changes (or commits), in order to detect future similar ones. The concept of change classification was previously studied in the literature of bug detection to identify commits with defects. In this paper, we borrow the notion of change classification from the literature of defect detection to further investigate its applicability to vulnerability detection problem using semi-supervised learning. In addition, we also experiment with new vulnerability predictors, and compare the predictive power of our proposed features with vulnerability prediction techniques based on text mining. The experimental results show that our semi-supervised approach holds promise in improving change classification effectiveness by leveraging unlabeled data.