Robust Onboard DNNs under Single Event Upsets

Le résumé fourni par la source

The deployment of Deep Neural Networks (DNNs) for onboard satellite applications, such as raw data processing and Earth observation, is rapidly expanding. However, Single Event Upsets (SEUs) caused by high-energy particles induce bit flips in memory-stored model parameters, compromising system reliability. Hardware redundancy and radiation-hardened components incur substantial cost and resource overhead. To address this, we develop a simulation tool to evaluate DNN robustness under SEUs and conduct extensive measurements across various bit-flip scenarios. Our results reveal that DNNs are particularly vulnerable to multi-bit upsets. We propose a lightweight, software-based method that combines 10-bit weight normalization with BCH encoding, correcting up to 3-bit errors with only 21.44 MB overhead. Experimental results on MNIST and CIFAR-10 using multiple models demonstrate that our method reduces vulnerable SEU events by over 100× compared with the baselines, with less than 3% accuracy loss, making it highly suitable for resource-constrained satellite platforms.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.