*Result*: GTVD: a multi-level aggregation vulnerability detection method based on full-dependency program graph.

*In modern software development life cycles, proactive vulnerability discovery and remediation play crucial roles in ensuring application security. However, current deep learning-based vulnerability detection methods frequently face limitations due to overly simplistic feature extraction procedures and inadequate handling of long-range dependency relationships. In this paper, we present GTVD, a graph-based vulnerability detection framework for C/C++ source code, which addresses these challenges through three key innovations. First, we introduce the Full Dependency Program Graph (FDPG), a novel intermediate representation that comprehensively encodes both syntactic structures and semantic relationships within source code. This advancement overcomes the feature representation constraints inherent in conventional code attribute graphs. Our architecture employs a hierarchical Graph Neural Network to systematically extract structural patterns from the FDPG representation, ensuring a robust analysis of the program's inherent structures. At the core of our feature extraction mechanism lies the Multi-Level Message Aggregation (MLMA) strategy. This innovative approach enables progressive integration of information across multiple neighborhood orders, effectively capturing both local and global program dependencies. To mitigate feature degradation in long-range dependencies, we develop an Adaptive Weighted Aggregation (WAG) mechanism that dynamically adjusts feature contributions during graph-level representation learning. Comprehensive evaluations on three large-scale public datasets demonstrate GTVD's superior performance. Our method achieves an average improvement of 7.76% across four evaluation metrics compared to the baseline, thereby confirming our method's enhanced capability to identify complex vulnerability patterns. [ABSTRACT FROM AUTHOR]*