A Software Vulnerability Detection Model Combined with Graph Simplification

Code vulnerability detection is an important issue in software security, however, when analyzing large programs, the diversity and complexity of the syntax and semantic of real-world vulnerability code makes it difficult to efficiently extract vulnerability features. In addition, real-world exploit code contains a large number of redundant features unrelated to the vulnerability, which will further impact detection efficiency. Traditional methods are limited by predefined vulnerability patterns, while existing deep learning-based methods usually retain all code information to extract features. To address this challenge, we define a new code representation called Code Simplified Graph (CSG) and then propose a new vulnerability detection method, GSVD. This method uses CSG as the code representation and CodeBERT as the semantic extractor, and finally extracts the vulnerability features by combining a CNN model with an attention mechanism. Our approach represent the code graphically, eliminating as much redundant information irrelevant to vulnerabilities as possible while preserving structural information. The experimental results show that GSVD has a good performance in both accuracy and speed of detection comparing with the previous methods.