VulSAT: Source Code Vulnerability Detection Scheme Based on SAT Structure

The number of software vulnerabilities have increased rapidly, and their forms have shown the characteristics of complexity and diversity, which has brought severe challenges to software systems. Deep learning can automatically learn the features of code and is widely used in source code vulnerability detection. Some studies treat codes as text sequences, ignoring the structural relationship between lines of code. Other studies extract the code as a graph structure, but ignore the order relationship of code lines and do not make full use of the dependencies between code lines. This paper proposes a C/C++ source code vulnerability detection scheme based on SAT (Structure-Aware Transformer) structure. It first standardizes the source code, then extracts the PDG (program dependency graph) representation of source code, next convert each node into a vector representation using sentence embedding, which retains the dependencies between nodes. Finally input it into the SAT architecture model for training, and judge whether the input program contains vulnerabilities according to the output results. We compared VulSAT with a variety of vulnerability detection methods (i.e., Devign, SySeVR, VulCNN, Checkmarx, VulDeePecker), and selected C/C++ functions in the SARD vulnerability dataset as experimental objects. The performance of accuracy rate and false positive rate has been improved to a certain extent.