HGVul: A Code Vulnerability Detection Method Based on Heterogeneous Source-Level Intermediate Representation

Vulnerability detection on source code can prevent the risk of cyber-attacks as early possible. However, lacking fine-grained analysis has rendered existing solutions still suffering from low performance; besides, explosive growth open-source projects dramatically increased complexity and diversity code. This paper presents HGVul, a vulnerability method based heterogeneous intermediate representation The key proposed is handling source-level (SIR) without expert knowledge. It first extracts graph SIR with multiple syntactic-semantic information. Then, HGVul splits into different subgraphs according to various semantic relations, which are used obtain information conveyed by types edges. Next, neural network attention operations deployed each subgraph learn representation, captures subtle effects node neighbors their representation. Finally, learned feature representations utilized perform detection. Experiments conducted datasets. F1 reaches 96.1% sample-balanced Big-Vul-VP dataset 88.3% unbalanced Big-Vul dataset. Further experiments actual project datasets prove better performance HGVul.