Abstract
Abstract
DNNs, similar to traditional software systems, may exhibit defects that can lead to serious consequences, especially in safety-critical scenarios. As a result, the ability to detect such defects reliably has become increasingly important, where the quality of the test dataset plays a central role. In this work, we introduce a test data selection method which combines a self-attention mechanism with K-means clustering (i.e., SAKCL) in a coordinated manner. The self-attention component assigns different levels of importance to feature dimensions, helping highlight informative patterns within samples while reducing the effect of redundant information. Based on these refined features, K-means clustering is applied to organize the data and capture structural relationships across samples. Through this process, the selected test cases achieve a better balance between fault detection capability and diversity. Experiments conducted on four benchmark datasets (MNIST, CIFAR-10, Fashion-MNIST, and SVHN) and multiple DNN architectures (LeNet-1/5, ResNet-20, and VGG-16) show that SAKCL consistently performs better than existing methods. On average, it increases the proportion of error-revealing test cases (
FDR
) by more than 12%, improves diversity-related metrics (
KMNC
) by over 6.5%, and leads to a retraining accuracy improvement (
ΔAcc
) of + 3.319% compared with baseline approaches. Statistical analysis further supports the reliability of these improvements (
p
Cliff’s δ
> 0.8). Overall, the proposed method provides a practical and scalable solution for selecting effective test data in DNN quality assurance.