Vulnerability detection with feature fusion and learnable edge-type embedding graph neural network

Abstract

Deep learning methods are widely employed in vulnerability detection, and graph neural networks have shown effectiveness in learning source code representation. However, current methods overlook non-relevant noise information in the code property graph and lack specific graph neural networks designed for code property graph. To address these issues, this paper introduces Leev, an automated vulnerability detection method. We developed a graph neural network tailored to the code property graph, assigning iterative vectors to diverse edge types and integrating them into the message passing between nodes to enable the model to extract hidden vulnerability information. In addition, virtual nodes are incorporated into the graph for feature fusion, mitigating the impact of irrelevant features on vulnerability information within the code. Specifically, for the FFMPeg+Qemu, Reveal, and Fan et al. datasets, the F1 metrics exhibited improvements of 7.02%, 21.69%, and 27.74% over the best baseline, correspondingly.