A Rapid Evaluation Technology for SEU in Convolutional Neural Network Circuits

Abstract

This paper proposes a rapid Single Event Upset (SEU) evaluation platform that can perform fault injection at the algorithm level in Convolutional Neural Network (CNN), which is designed by software and hardware co-design. This proposed platform analyzes the layer structure and input parameters of CNN, generates the corresponding fault list and then performs fault injection at the algorithm-level in software side and mapped to the registers of hardware acceleration. Finally, the operation results of CNN after fault injection is analyzed to evaluate the robustness of CNN against SEU, and identify the sensitive areas of SEU. In this paper, SEU fault injection experiments are carried out on YOLOv2 neural network. Experimental results show that fault injection and sensitivity evaluation based on this method can analyze the overall robustness of CNN against SEU quickly and effectively. From the analysis of experimental results, the algorithmic nodes which are sensitive to SEU are located, and the efficiency of this rapid evaluation technology is also verified.