Multi-View Graph-Based Code Representation Learning for Vulnerability Detection

Vulnerability detection is essential for ensuring software reliability and security. Existing learning-based approaches often fail to fully capture fine-grained program semantics and heterogeneous structural dependencies. In particular, directly merging different code representations into a unified graph may introduce semantic ambiguity and weaken vulnerability-specific signals. To address this issue, we propose SMGCN, a semantic-aware multi-view graph framework for vulnerability detection. SMGCN independently models abstract syntax, control flow, and data flow as separate graph views, preserving their structural characteristics. A node importance recalibration mechanism is introduced to emphasize vulnerability-relevant nodes, and an attention-based fusion module adaptively integrates multi-view representations into a unified embedding for classification. Extensive experiments on four benchmark datasets demonstrate that SMGCN consistently outperforms representative learning-based and LLM-based baselines, achieving average improvements of 7.18% in accuracy and 11.0% in F1-score. Additional analyses further validate the effectiveness of multi-view modeling and node importance calibration. These results highlight the advantage of semantic-aware multi-view graph learning for robust vulnerability detection.