Scholarlink

<p>Metamorphic testing (MT) provides an effective oracle-free strategy for evaluating deep neural networks under semantics-preserving perturbations. However, existing MT prioritization methods typically rely on static global heuristics and do not adequately capture class-dependent variation in model behavior or the complementary roles of predictive uncertainty and interpretability drift. This limitation is especially important in safety-critical vision applications. We propose a class-adaptive uncertainty–interpretability metamorphic test case prioritization framework (CUI-MTP) for vision-based deep learning systems. Rather than ranking individual samples, the framework prioritizes executable metamorphic test cases and formulates prioritization as a class-conditioned optimization problem that jointly models probabilistic instability and saliency-level behavioral change. Predictive uncertainty and interpretability drift are integrated through multi-objective Bayesian optimization with expected hypervolume improvement. Experiments on CIFAR-10, Fashion-MNIST, and ISIC2019 with ResNet-18, ResNet-50, and ConvNeXt-Base show that the proposed method consistently outperforms representative oracle-free baselines across datasets, architectures, and Top-<i>N</i> budgets, with statistically significant improvements (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(p&lt;0.01\)</EquationSource> </InlineEquation>). Qualitative Grad-CAM case studies further show that the framework captures not only obvious prediction failures but also high-risk cases with substantial saliency drift. Overall, the results demonstrate a practical and scalable strategy for robust, safety-aware validation of vision-based deep learning systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Class-adaptive oracle-free metamorphic test case prioritization framework for vision-based deep neural networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月16日

查看全文
  • Junhan Li,
  • Radziah Mohamad,
  • Johanna Ahmad

摘要

Metamorphic testing (MT) provides an effective oracle-free strategy for evaluating deep neural networks under semantics-preserving perturbations. However, existing MT prioritization methods typically rely on static global heuristics and do not adequately capture class-dependent variation in model behavior or the complementary roles of predictive uncertainty and interpretability drift. This limitation is especially important in safety-critical vision applications. We propose a class-adaptive uncertainty–interpretability metamorphic test case prioritization framework (CUI-MTP) for vision-based deep learning systems. Rather than ranking individual samples, the framework prioritizes executable metamorphic test cases and formulates prioritization as a class-conditioned optimization problem that jointly models probabilistic instability and saliency-level behavioral change. Predictive uncertainty and interpretability drift are integrated through multi-objective Bayesian optimization with expected hypervolume improvement. Experiments on CIFAR-10, Fashion-MNIST, and ISIC2019 with ResNet-18, ResNet-50, and ConvNeXt-Base show that the proposed method consistently outperforms representative oracle-free baselines across datasets, architectures, and Top-N budgets, with statistically significant improvements ( \(p<0.01\) ). Qualitative Grad-CAM case studies further show that the framework captures not only obvious prediction failures but also high-risk cases with substantial saliency drift. Overall, the results demonstrate a practical and scalable strategy for robust, safety-aware validation of vision-based deep learning systems.