VULDA: Source Code Vulnerability Detection Via Local Dependency Context Aggregation on Vulnerability-Aware Code Mapping Graph.

Vulnerability detection is crucial in the field of software security. However, existing methods often suffer from interference caused by redundant information and insufficient cross-line semantic dependencies when handling large-scale and complex source code, which limits detection performance. To address these challenges, this paper proposes VULDA, a source code vulnerability detection method that integrates a Vulnerability-aware Code Mapping Graph (VCMG) with Local Dependency Context Aggregation (LDCA). VCMG significantly reduces redundancy in graph structures by aligning multi-granularity semantics to line-level nodes, thereby enhancing representational compactness. Additionally, it incorporates static heuristic rules and structural features to weight nodes, effectively improving the model’s sensitivity to key vulnerability-related code. Building upon this, the LDCA module aggregates both control-flow graph (CFG) and data-dependency graph (DDG) paths, achieving dual-context aggregation of logical and data semantics, which further enhances the model’s ability to express complex vulnerability patterns. Experimental results on multiple real-world datasets, including SARD, Reveal, and FFmpeg+Qemu, demonstrate that VULDA outperforms existing methods across various metrics, notably achieving a 23.09