Detecting Runtime Exceptions by Deep Code Representation Learning with Attention-Based Graph Neural Networks

Abstract

Uncaught runtime exceptions have been recognized as one of the commonest root causes of real-life exception bugs in Java applications. However, existing runtime exception detection techniques rely on symbolic execution or random testing, which may suffer the scalability or coverage problem. Rule-based bug detectors (e.g., SpotBugs) provide limited rule support for runtime exceptions. Inspired by the recent successes in applying deep learning to bug detection, we propose a deep learning-based technique, named Drex, to identify not only the types of runtime exceptions that a method might signal but also the statement scopes that might signal the detected runtime exceptions. It is realized by graph-based code representation learning with (i) a lightweight analysis to construct a joint graph of CFG, DFG and AST for each method without requiring a build environment so as to comprehensively characterize statement syntax and semantics and (ii) an attention-based graph neural network to learn statement embeddings in order to distinguish different types of potentially signaled runtime exceptions with interpretability. Our evaluation on 54,255 methods with caught runtime exceptions and 54,255 methods without caught runtime exceptions from 5,996 GitHub Java projects has indicated that Drex improves baseline approaches by up to 18.2% in exact accuracy and 41.6% in F1-score. Drex detects 20 new uncaught runtime exceptions in 13 real-life pro-jects, 7 of them have been fixed, while none of them is detected by rule-based bug detectors (i.e., SpotBugs and PMD).