Deep Neural Networks (DNNs) face challenges during deployment due to covariate shift, i.e., data distribution shifts between development and deployment contexts. Fine-tuning adapts pre-trained models to new contexts requiring smaller labeled sets. However, testing fine-tuned models under constrained labeling budgets remains a critical challenge. This paper introduces MetaSel, a new approach tailored for DNN models that have been fine-tuned to address covariate shift, to select tests from unlabeled inputs. MetaSel assumes that fine-tuned and pre-trained models share related data distributions and exhibit similar behaviors for many inputs. However, their behaviors diverge within the input subspace where fine-tuning alters decision boundaries, making those inputs more prone to misclassification. Unlike general approaches that rely solely on the DNN model and its input set, MetaSel leverages information from both the fine-tuned and pre-trained models and their behavioral differences to estimate misclassification probability for unlabeled test inputs, enabling more effective test selection. Our extensive empirical evaluation, comparing MetaSel against 11 state-of-the-art approaches and involving 68 fine-tuned models across weak, medium, and strong distribution shifts, demonstrates that MetaSel consistently delivers significant improvements in Test Relative Coverage (TRC) over existing baselines, particularly under highly constrained labeling budgets. MetaSel shows average TRC improvements of 28.46% to 56.18% over the most frequent second-best baselines while maintaining a high TRC median and low variability. Our results confirm MetaSel’s practicality, robustness, and cost-effectiveness for test selection in the context of fine-tuned models.

Successful deployment of Deep Neural Networks (DNNs), particularly in safety-critical systems, requires their validation with an adequate test set to ensure a sufficient degree of confidence in test outcomes. Although well-established test adequacy assessment techniques from traditional software, such as mutation analysis and coverage criteria, have been adapted to DNNs in recent years, we still need to investigate their application within a comprehensive methodology for accurately predicting the fault detection ability of test sets and thus assessing their adequacy. In this paper, we propose and evaluateTEASMA, a comprehensive and practical methodology designed to accurately assess the adequacy of test sets for DNNs. In practice,TEASMAallows engineers to decide whether they can trust high-accuracy test results and thus validate the DNN before its deployment. Based on a DNN model's training set,TEASMAprovides a procedure to build accurate DNN-specific prediction models of the Fault Detection Rate (FDR) of a test set using an existing adequacy metric, thus enabling its assessment. We evaluatedTEASMAwith four state-of-the-art test adequacy metrics: Distance-based Surprise Coverage (DSC), Likelihood-based Surprise Coverage (LSC), Input Distribution Coverage (IDC), and Mutation Score (MS). We calculated MS based on mutation operators that directly modify the trained DNN model (i.e., post-training operators) due to their significant computational advantage compared to the operators that modify the DNN's training set or program (i.e., pre-training operators). Our extensive empirical evaluation, conducted across multiple DNN models and input sets, including large input sets such as ImageNet, reveals a strong linear correlation between the predicted and actual FDR values derived from MS, DSC, and IDC, with minimum$R^{2}$values of 0.94 for MS and 0.90 for DSC and IDC. Furthermore, a low average Root Mean Square Error (RMSE) of 9% between actual and predicted FDR values across all subjects, when relying on regression analysis and MS, demonstrates the latter's superior accuracy when compared to DSC and IDC, with RMSE values of 0.17 and 0.18, respectively. Overall, these results suggest thatTEASMAprovides a reliable basis for confidently deciding whether to trust test results for DNN models.