Vulnerability Estimation of DNN Model Parameters with Few Fault Injections

摘要

The reliability of deep neural networks (DNN) againsthardware errors is essential as DNNs are increasingly employed in safetycriticalapplications such as automatic driving. Transient errors in memory,such as radiation-induced soft error, may propagate through the inferencecomputation, resulting in unexpected output, which can adversely triggercatastrophic system failures. As a first step to tackle this problem, this paperproposes constructing a vulnerability model (VM) with a small numberof fault injections to identify vulnerable model parameters in DNN. We reducethe number of bit locations for fault injection significantly and developa flow to incrementally collect the training data, i.e., the fault injection results,for VM accuracy improvement. We enumerate key features (KF) thatcharacterize the vulnerability of the parameters and use KF and the collectedtraining data to construct VM. Experimental results show that VMcan estimate vulnerabilities of all DNN model parameters only with 1/3490computations compared with traditional fault injection-based vulnerabilityestimation.