Soft error detection based on graph neural networks

In the space radiation environment, soft errors induced by high-energy particles can lead to Silent Data Corruption (SDC). To mitigate the impact of SDC, this paper proposes a soft error detection method based on Graph Neural Networks (GNNs). First, the Instruction Dependency Graph (IDG) of a program is extracted to represent instruction features and their dependencies. Second, the IDG is leveraged to train a GNN model to predict the SDC Vulnerability of instructions. Finally, a detection method based on SDC-vulnerable paths is proposed to identify critical soft error propagation paths within the IDG and implement redundancy hardening for instructions along these paths. Experimental results demonstrate that, compared to existing approaches, the proposed SDC vulnerability prediction model significantly reduces prediction error without requiring extensive Fault Injection (FI) experiments. Under identical performance overhead, the proposed soft error detection method substantially enhances the detection rate.