Smart Contract Bytecode Vulnerability Detection with Neighborhood Constrained Cross Attention

Abstract

Smart contract vulnerability detection has gained increasing attention due to growing financial losses from hacker attacks. Existing deep learning methods either rely on a single feature type or lack effective interaction among heterogeneous features, limiting vulnerability representation. To address this, we propose neighborhood constrained cross attention. It uses the control-flow graph’s k-hop neighborhood as a structural prior to restrict bidirectional interactions between graph features and sequence features to local regions likely associated with the same execution logic, thereby reducing noise from global attention. Self-attention is further applied within each branch to model long-range dependencies. Experiments show that NCCA-Det achieves accuracies of 94.87%, 92.62%, and 92.94% on three common vulnerability types, significantly outperforming comparative methods, and thus offers a complementary solution for bytecode-level smart contract vulnerability detection.