부산대학교 도서관

With ever-growing data amount generated on spacecrafts such as satellites, it becomes necessary to process part of those data in space with machine learning techniques before transmitting them to the ground. The key challenge for such in-space intelligence is its robustness due to the harsh environment those spacecrafts operate in, especially the single-event upset (SEU) that can cause the in-memory model weight to be flipped between 0 and 1. This work first builds a simulation platform that can efficiently and accurately inspect the robustness of Deep Neural Networks (DNNs) against SEUs. Atop the platform, we perform the first measurement study to demystify the DNN robustness against SEUs under both single-bit and multi-bit error settings. The results highlight how fragile DNNs could be in space, especially when the exponent bits within its weights are flipped. To this end, we propose an effective, model-transparent, and low-overhead approach to enhance the DNN robustness against SEUs. Its key idea is to offline scale up the weight of each layer to amortize the impacts from the flipped exponent bits, and then scale down the output with a scaling factor during execution. Extensive experiments show that our approach can reduce the DNN vulnerability by up to $50,000 \times$, and thus make in-space intelligence robust enough even for critical tasks.