An end-to-end automatic methodology to accelerate the accuracy evaluation of deep neural networks under hardware transient faults

• Introduces A-Mean, a unified and end-to-end automatic methodology for rapid evaluation of hardware transient faults on DNNs, achieving up to 922.80× speedup over TensorFI+. • Estimates both general classification accuracy and application-specific safety-critical misclassification (SCM) using silent data corruption rates of basic operations and a two-level mean calculation mechanism. • Employs a max-policy to handle non-sequential structures and a worst-case scheme to compute enlarged SCM and halved accuracy, enabling conservative reliability assessment. • Provides an easy-to-use automatic tool, publicly available, for prompt fault evaluation across diverse DNN models and datasets with minimal accuracy loss (0.77% on average).