Source Code Vulnerability Detection Based On SAR-GIN

With the development of the Internet era, software technology is being used more and more commonly, and the detection of vulnerabilities in the corresponding software must be efficient and accurate. However, software vulnerabilities are diverse, and mining vulnerabilities through source code requires a high level of professional experience for developers. Previous vulnerability detection solutions either relied on expert-defined features or used recursive neural networks only for code sequences, making it difficult to extract complex features of vulnerabilities in the traditional code space. In recent years, with the booming development of artificial intelligence technology, some scholars have started to try to combine graph neural networks to extract graph structure information of source code for software vulnerability detection. In this paper, by introducing a method based on Graph Isomorphism Network (GIN) combined with a self-attention aggregation mechanism for vulnerability detection, we achieve superior detection results. By assigning different attention weights to each layer, the self-attention mechanism helps to solve the problem of too smooth node representation and poor comprehensive performance of graph-level output. In this paper, we propose Self Attention Readout Graph Isomorphism Network (SAR-GIN), which efficiently extracts and utilizes global information at all depths by combining GIN with a self-attention mechanism in generating global graph-level features. Meanwhile, the experimental results on CWE-400, CWE-190 datasets, combined with the discrete Fourier transform, show that the modeling approach in this paper achieves superior results compared to other models.