Failure Characteristics of FPGA-Based Convolutional Neural Networks under RF Interference

Abstract

FPGAs have been widely used in multiple parallel computing of convolutional neural networks and realizing the parallel acceleration of multiplication, benefitting from its tremendous scalability and flexibility. So far, most of the researches is based on the design and optimization of FPGA-based convolutional neural network accelerators, which lacks researches on the electromagnetic immunity of convolutional neural network accelerators. Therefore, this article mainly studies the failure characteristics of FPGA-based convolutional neural networks accelerators under RF (Radio Frequency) interference. This paper implements two different convolutional neural networks LetNet-5 and YOLO based on a self-made FPGA test board. RF interference is injected into the core pins of the FPGA through a method which is similar to DPI (Direct Power Injection). Next, we observe the failure phenomenon and study the failure characteristics. It is found that under electromagnetic interference, an unusual phenomenon occurred. The research in this paper can provide a reference for the electromagnetic compatibility design of neural network accelerators.