VulDeCRG: A Vulnerability Detection System based on Code Representation Graph

With the advancement of computer technology, software systems have become more complex, leading to an increase in software vulnerabilities. Deep learning-based methods have gained popularity for vulnerability detection due to their ability to extract complex features and generalize well. However, the use of deep learning requires converting code into vector representations. Existing methods often convert code into graph or sequence representations to extract information. Graph representations capture structural information, while sequence representations preserve programming logic. We developed VulDeCRG, a function-level vulnerability detection system. VulDeCRG utilizes Code Representation Graph (CRG), which combines graph and sequence representation techniques. It employs a graph neural network with a node selection network to extract vulnerability features. We implemented VulDeCRG using CRG and evaluated it on two datasets, Big-Vul and FFMPeg+Qemu. Extensive experiments showed that VulDeCRG slightly outperforms state-of-the-art vulnerability detection methods, achieving a 6.61% increase in precision and a 3.97% improvement in F1 score.