Enhancing neural network robustness: Laser fault injection resistance in 55-nm SRAM for space applications

Abstract The integration of artificial intelligence (AI) with satellite technology is ushering in a new era of space exploration, with small satellites playing a pivotal role in advancing this field. However, the deployment of machine learning (ML) models in space faces distinct challenges, such as single event upsets (SEUs), which are triggered by space radiation and can corrupt the outputs of neural networks. To defend against this threat, we investigate laser-based fault injection techniques on 55-nm SRAM cells, aiming to explore the impact of SEUs on neural network performance. In this paper, we propose a novel solution in the form of Bin-DNCNN, a binary neural network (BNN)-based model that significantly enhances robustness to radiation-induced faults. We conduct experiments to evaluate the denoising effectiveness of different neural network architectures, comparing their resilience to weight errors before and after fault injections. Our experimental results demonstrate that binary neural networks (BNNs) exhibit superior robustness to weight errors compared to traditional deep neural networks (DNNs), making them a promising candidate for spaceborne AI applications.