A Vulnerability Detection Method for Intermediate Code Based on a Relational Dependency Graph

With the advancement of information technology, the rapid increase in software systems has led to a growing demand for research on vulnerability detection. In recent years, the fast-paced development of deep learning has introduced new approaches and directions for this field. Existing source code vulnerability detection methods often fail to explicitly preserve the semantic information related to vulnerabilities in the source code, leading to difficulties in extracting vulnerable code features and high false-positive rates in detection. To address these issues, this study proposes a vulnerability detection method for intermediate code based on a relational dependency graph (RDG). After converting the source code into an intermediate representation, it is mapped to an RDG, which accurately expresses the data and control flow dependencies within the program. This approach retains critical node features and refines the types of control flow and data flow edges. By leveraging this method, it becomes possible to trace program data and control dependencies, understand data flow directions, and analyze interactions between different parts of the program, enabling the precise identification and interpretation of potential security vulnerabilities. Drawing on concepts from code property graphs (CPG), this method enhances the understanding and capture of complex relationships between nodes in graph-structured data. By performing vulnerability detection on intermediate code, this approach avoids false positives and negatives caused by language or platform-specific differences, thereby enabling deeper identification and interpretation of complex vulnerability patterns in source code. Experimental results on the SARD dataset demonstrate that the proposed method achieves a high precision rate of 98.7% for detecting command injection vulnerabilities (CWE78), significantly outperforming traditional methods.