The effect of weight errors on neural networks
Austin P. Arechiga, Alan J. Michaels · 2018
Neural networks have moved beyond theoretical equations and into real-world applications. In order for neural networks to be useful in real-world environments, their reliability in non-ideal computing conditions must be understood. Cyber and environmental threats can both affect the execution of neural network programs, but little is known about how well neural networks can withstand attack. This paper examines the robustness of Multilayer Perceptron (MLP) neural networks and Convolutional Neural Networks (CNNs) against weight errors. Specifically, Single Event Upset (SEU) style errors such as individual bit flips in memory are the focus. The classification accuracy of the networks is tested before and after bit flips to see if there's any correlation between neural network architecture and robustness to weight errors. The experimental results show that in general MLP networks are much more robust to weight errors than CNNs. For MLPs larger networks with more layers are more robust than smaller, shallower networks and for CNNs networks with larger kernels are more robust than networks with smaller kernels.