Strategic Input Selection for Deep Neural Networks Reliability Evaluation

Assessing the reliability of large Deep Neural Networks (DNNs) is becoming increasingly challenging as models and datasets grow, making exhaustive reliability assessment unfeasible through either radiation testing or fault simulation. In this work, we propose a robust input selection methodology to assess the reliability of large image classification DNNs. Our methodology is based on the confidence distribution of model predictions. We validate this approach through a cross-layer analysis, combining neutron beam experiments and software fault simulations (SWFS) on state-of-the-art large DNNs. Our results show that DNN confidence has a strong impact on the misclassification rates. Our input selection method exhibits up to 22× higher misclassification rates than the commonly used random input selection approach. These findings highlight the limitations of relying only on random selection when estimating the fault sensitivity of large DNNs and emphasize the need for prior input analysis.