Dynamic Hardware Defense for High-Centrality Nodes in Graph Convolutional Networks

With the growing reliance on graph data processing in safety-critical applications, ensuring the reliability of graph convolutional network (GCN) hardware systems has become paramount, especially in radiation-prone environments where single-event upsets (SEUs) pose significant risks. This article presents a comprehensive design framework for a fault-tolerant GCN system, addressing the unique challenges of SEU susceptibility through a novel fault-aware centrality measure and a dynamic hardware defense (DHD) strategy. Our approach begins with the development of a fault-aware centrality measure to precisely model the distribution of critical nodes within graph data. Leveraging this measure, we design a DHD strategy that integrates partial circuit reinforcement and high-centrality node marking, which dynamically route the dataflow of the most influential nodes in the graph to reinforced circuit units. The proposed system architecture is rigorously validated through extensive experiments, demonstrating significant improvements in fault tolerance. Specifically, the DHD strategy achieves improvements in hardening efficiency of $2.53\times $ , $2.70\times $ , and $2.85\times $ , respectively, compared with the traditional full triple modular redundancy (FTMR) strategy. Neutron radiation testing further validates the robustness of the system, showing effective mitigation of fault propagation under extreme conditions. By maintaining minimal hardware overhead and offering dynamic, cost-effective protection, this design framework provides a reliable solution for deploying GCNs in safety-critical applications.