A Software Code Defect Detection Method Based on the Fusion of Transformer and Graph Neural Networks

Abstract

Software defect detection plays a key role in improving system reliability, security, and maintainability. This study proposes a hybrid Transformer–GNN framework to jointly model code semantics and program structure. Source files are normalized, segmented at the function level, and converted into token sequences, while multi-relation graphs are generated to describe syntax, control flow, and data dependencies. The Transformer branch captures long-range contextual information from code tokens, whereas the GNN branch learns structural interactions among statements, execution paths, and variables. Their representations are adaptively integrated through a gated fusion module for defect classification. Experiments show that the proposed method outperforms conventional classifiers, sequence-oriented neural networks, and standalone Transformer or GNN models. Ablation results also verify the importance of semantic encoding, structural learning, data-flow information, and gated fusion. Overall, combining lexical context with graph-based program relations provides a more accurate, robust, and interpretable solution for software defect identification.