Vulnerability Estimation of DNN Model Parameters with Few Fault Injections

The reliability of deep neural networks (DNN) against hardware errors is essential as DNNs are increasingly employed in safety-critical applications such automatic driving. Transient memory, radiation-induced soft error, may propagate through the inference computation, resulting unexpected output, which can adversely trigger catastrophic system failures. As a first step to tackle this problem, paper proposes constructing vulnerability model (VM) with small number fault injections identify vulnerable parameters DNN. We reduce bit locations for injection significantly and develop flow incrementally collect training data, i.e., results, VM accuracy improvement. enumerate key features (KF) that characterize use KF collected data construct VM. Experimental results show estimate vulnerabilities all DNN only 1/3490 computations compared traditional injection-based estimation.