The existing deep learning-based methods for source code vulnerability detection often suffer from severe loss of syntax and semantics in target code， and neural network models allocating weights to the graph nodes （edges） in target code unreasonably. To address these issues， a method named VulATGCN for detecting source code vulnerabilities was proposed on the basis of Code Property Graph （CPG） and Adaptive Transformer-Graph Convolutional Network （AT-GCN）. In the method， CPG was used to represent source code， CodeBERT was combined for node vectorization， and graph centrality analysis was employed to extract deep structural features， thereby capturing the code’s syntax and semantic information in multi-dimensional way. After that， AT-GCN model was designed by integrating strengths of Transformer-based self-attention mechanism， which excels at capturing long-range dependencies， and Graph Convolutional Network （GCN）， which is proficient at capturing local features， thereby realizing fusion learning and precise extraction of features from regions with different importance. Experimental results on real vulnerability datasets Big-Vul and SARD show that the proposed method VulATGCN achieves an average F1 score of 82.9%， which is 10.4% to 132.9% higher than deep learning-based vulnerability detection methods such as VulSniper， VulMPFF， and MGVD， with an average increase of approximately 52.9%.
