A cross-layer fault propagation analysis method for edge intelligence systems deployed with DNNs

Abstract To evaluate the impact of soft errors on convolutional neural networks (CNNs) deployed in edged computation systems, we propose a data-driven assessment strategy to characterize the propagation flow across hardware and software abstraction layers of the system in an interpretable way. Single-bit-flip injections in underlying hardware architecture are performed on virtual embedded system with a CNN-based image classifier deployed on it. We depict the local activation and global dependencies caused by soft errors across the system in form of a directed acyclic graph by using generative adversarial networks and Bayesian networks as data modeling methods. The cross-layer fault propagation paths and component sensitivities show that the deep neural networks like CNNs can effectively prevent the faults that may cause critical failures from propagating to the system output via the channel sparsity and regular pooling mechanism in the network pipelines.