Effect Analysis of Low-Level Hardware Faults on Neural Networks using Emulated Inference

Artificial Neural Networks (ANN) are increasingly deployed in various applications and devices using hardware accelerators. However, faults in the processing hardware can affect the output of the ANN and, thus, the reliability of the application using it. Analyzing the effect of hardware faults on the application at design time is essential but non-trivial. We introduce a framework to emulate ANN inference on hardware resource descriptions under hardware faults. Hardware architecture, scheduling, and fault models are fully adaptable. An in-depth controlled experiment shows how hardware faults jeopardize any robustness guar-antees. Benchmark experiments on state-of-the-art ANN demonstrate the scalability of our framework.