Learning Code Anomalies for Bug Detection with Graph Neural Networks

This paper proposes Adl, a neural network-based approach that learns to detect bugs as anomalies. Compared with the classical anomaly detection approach, Adl eliminates the need for designing rule templates. Furthermore, Adl can generalize to bug types unseen during training. Adl also outperforms existing learning-based bug detectors in sample efficiency (i.e., requiring less training data) and self-explainability (i.e., explaining its predictions without external interpretability methods). An extensive evaluation of Adl shows that 1) Adl achieves significantly higher accuracy than Gated Graph Neural Network (GGNN) and Graph Interval Neural Network (GINN), two notable deep bug detection models, on two existing bug datasets (i.e., Bugs.jar and BugSwarm); 2) on Defects4J, Adl outperforms Infer, a prominent static analysis tool, and FICS, the latest anomaly detection method; and 3) on JBench, Adl surpasses RacerD—a state-of-the-art static race detector—without prior training on concurrent programs.