Layer-Sensitivity-Aware Hybrid Redundancy Hardening Design for Reliable CNN Accelerators

Convolutional neural networks (CNNs) are increasingly deployed in satellite remote sensing signal processing systems to support onboard image interpretation and low-latency intelligent processing. However, CNN accelerators implemented on SRAM-based field-programmable gate arrays (FPGAs) are highly vulnerable to radiation-induced soft errors, especially single-event upsets (SEUs), which can severely degrade inference reliability in space environments. To address this challenge, this work conducts a fine-grained sensitivity analysis of CNN accelerators under SEUs, revealing significant vulnerability variations across different convolution layers. Based on these observations, a layer-sensitivity-aware hybrid redundancy (LSAHR) hardening design is proposed. Experiments on VGG11 and ResNet18 using the UCM, NWPU45, and WHU19 datasets demonstrate that the proposed approach effectively improves network robustness against SEUs while achieving an efficient tradeoff between accuracy preservation and hardware overhead.