Multimodal hybrid recurrent framework with selective subpattern activation for smart contract vulnerability detection

Detecting vulnerabilities in smart contracts is a critical challenge for blockchain security, as flaws such as reentrancy, timestamp dependence, and infinite loops have resulted in severe financial losses in decentralized systems. Accurate and interpretable detection of these vulnerabilities remains challenging due to the complex semantics of smart contract code. In this study, we propose a multimodal hybrid recurrent framework for smart contract vulnerability detection that integrates sequential and structural code representations. The framework introduces a Selective Subpattern Activation (SSA) mechanism, which highlights vulnerability-indicative code subpatterns during the pattern extraction phase and provides interpretable insights into model predictions. Pattern-based features enhanced by SSA are processed using a Bidirectional Gated Recurrent Unit (BiGRU), while structural features derived from control and data flow representations are modeled using a Bidirectional Long Short-Term Memory (BiLSTM) network. The proposed approach is evaluated on a publicly available Ethereum smart contract dataset using five independent experimental runs, with results reported as averages. The results show that the framework achieves an accuracy of 92.16% and an F1 score of 88.83% for reentrancy vulnerability detection, achieving higher performance compared to baseline deep learning and graph-based models. Ablation experiments are performed to demonstrate the contribution of the SSA mechanism to both detection performance and interpretability.