Smart contract vulnerability detection method based on Bi-modal cross-attention mechanism

To address the problem that existing deep learning methods for smart contract vulnerability detection rely on single-modal feature extraction and insufficient contextual information capture, leading to relatively low detection accuracy, a smart contract vulnerability detection method based on the Bi-modal cross-attention mechanism was proposed. A specific attention mechanism was designed that simultaneously analyzed both contract source code and bytecode, achieving bidirectional mapping and complementary enhancement between high-level semantic features in source code and low-level execution flows in bytecode, thereby enriching feature representation. Residual connections were introduced to effectively preserve and transmit original feature information, mitigating the vanishing gradient problem in deep network training. Extensive testing on public datasets demonstrates that the proposed method improves detection accuracy by more than 2% compared to baselines. Ablation experiments confirm that cross-modal feature fusion and the design of the attention mechanism work in synergy with each other, significantly improving the detection performance. © 2025 Editorial Board of Journal on Communications. All rights reserved.