AST-GAT-BiLSTM: A structure-aware deep learning framework for smart contract vulnerability detection

Smart contracts (SCs) execute application logic on blockchains, and defects in their code can compromise decentralized applications and the assets they manage. Vulnerability detection requires representations that capture both contextual patterns in source-token sequences and the syntactic relationships between code elements. We present AST-GAT-BiLSTM, a deep learning framework that combines bidirectional long short-term memory (Bi-LSTM) representations with graph attention network (GAT) representations of abstract syntax trees (ASTs). The semantic branch models token context, while the graph branch learns from syntactic structure; their representations are combined through weighted fusion for vulnerability prediction. The evaluation uses data associated with CBGRU and Tianchi and considers reentrancy and timestamp dependence. Comparisons are reported separately for these two vulnerability types. An AST alone does not establish control-flow or data-flow relations, and classification metrics do not demonstrate computational efficiency. The conclusions are therefore limited to the evaluated tasks and data rather than unseen vulnerability classes or large-scale deployment.