Early Soft Error Reliability Assessment of Convolutional Neural Networks Executing on Resource-constrained IoT Edge Devices

Machine learning (ML) algorithms offer solutions across diverse domains, including resource-constrained devices. Despite advances in performance optimization and reduced-precision implementations, ML susceptibility to soft errors from radiation remains unclear. This work uses virtual platforms (VPs) to conduct a comprehensive soft error reliability assessment, at early design phases, of ML algorithms for Arm processors. The test strategy integrates confidence metrics, extensive fault injection (FI) campaigns, application profiling, and fault classifications. The main goals comprise (i) analyzing the consistency of SOFIA, a JIT-based framework, against event-driven simulators (mean mismatch ±4%, worst-case ±8%) and (ii) investigating correlations between FI results, NN-optimized kernels, and reduced-precision CNNs for IoT devices (up to 60% critical faults), aiming to promote software-based mitigation techniques (up to 98% critical faults mitigated). Results covering over 14.8 million FIs highlight SOFIA’s consistency, offering insights into balancing performance and reliability in multithreaded IoT edge platforms.