PRTS: Test Sample Selection Based on Category Probability Repair for DNN Testing

Selecting appropriate test samples is a critical and frequently utilized step in optimizing deep neural network (DNN) testing. However, DNN models often exhibit high confidence in incorrect predictions, limiting the effectiveness of traditional selection methods that rely on model predictions to detect overconfident faults. To address this issue, this work introduces probability repair–based test sample selection (PRTS), a novel approach that repairs DNN probability distributions to improve test sample selection. We first categorize test samples into trustworthy and untrustworthy subsets based on uncertainty scores. Unlike conventional methods that rely solely on model‐predicted class probabilities, we design feature similarity–based class probabilities (FSCP) to quantify uncertainty scores by measuring the discrepancy between these two probability distributions. Additionally, a repair process is incorporated to mitigate overconfidence and underconfidence in FSCP, explicitly targeting untrustworthy data instances to reduce the influence of erroneous neighbors and enhance overall efficiency. Extensive experiments were conducted across diverse datasets and DNN architectures, comparing PRTS with 12 state‐of‐the‐art baseline methods. The results demonstrate that PRTS generally outperforms existing approaches in terms of the number and diversity of revealed faults and improvements in DNN model accuracy, thereby serving as a robust and efficient solution to enhance the reliability of DNN testing.