Path-sensitive code embedding via contrastive learning for software vulnerability detection

Abstract

Machine learning and its promising branch deep learning have shown success in a wide range of application domains. Recently, much effort has been expended on applying deep learning techniques (e.g., graph neural networks) to static vulnerability detection as an alternative to conventional bug detection methods. To obtain the structural information of code, current learning approaches typically abstract a program in the form of graphs (e.g., data-flow graphs, abstract syntax trees), and then train an underlying classification model based on the (sub)graphs of safe and vulnerable code fragments for vulnerability prediction. However, these models are still insufficient for precise bug detection, because the objective of these models is to produce classification results rather than comprehending the semantics of vulnerabilities, e.g., pinpoint bug triggering paths, which are essential for static bug detection. This paper presents ContraFlow, a selective yet precise contrastive value-flow embedding approach to statically detect software vulnerabilities. The novelty of ContraFlow lies in selecting and preserving feasible value-flow (aka program dependence) paths through a pretrained path embedding model using self-supervised contrastive learning, thus significantly reducing the amount of labeled data required for training expensive downstream models for path-based vulnerability detection. We evaluated ContraFlow using 288 real-world projects by comparing eight recent learning-based approaches. ContraFlow outperforms these eight baselines by up to 334.1%, 317.9%, 58.3% for informedness, markedness and F1 Score, and achieves up to 450.0%, 192.3%, 450.0% improvement for mean statement recall, mean statement precision and mean IoU respectively in terms of locating buggy statements.