Replication Data for “Evidence Reuse in Metamorphic Testing of Vision-Based DNNs: A Multi-Granularity Characterization Across Models and Domains”
Description
This dataset provides the derived experimental data and analysis code supporting the study “Evidence Reuse in Metamorphic Testing of Vision-Based DNNs: A Multi-Granularity Characterization Across Models and Domains.” The study investigates the reusability of metamorphic-testing evidence at configuration, family, and concrete source–MR levels across vision-based deep neural networks and source domains. It evaluates two image-classification source domains (CIFAR-10 and an ISIC-derived dermoscopic source distribution), three frozen model architectures (ResNet-50, ConvNeXt-Base, and ViT-B/16), and ten deterministic MR configurations from four transformation families, yielding 1,567,170 source–MR executions. The repository contains deterministic source and MR manifests, source- and candidate-level experimental outcomes, protocol-audit records, model and checkpoint metadata, statistical-analysis outputs, figure-source data, supplementary-analysis results, and reproducibility code. The archived materials support analyses of MR effectiveness, transformation-intensity effects, configuration- and family-level transferability, directional failure-instance transfer, label-free response signals, and source-grouped JSD calibration under fixed manual-inspection budgets. The data support the study’s central evidence-reuse finding: configuration-level effectiveness can support initialization of a new MR suite, family-level evidence requires target-specific recalibration, and concrete source–MR failures require revalidation on the target model. The repository contains derived data and reproducibility materials only. The original CIFAR-10 and ISIC images and trained checkpoint binaries are not redistributed; source data should be obtained from their original providers under the applicable terms.
Files
Institutions
- University of Technology MalaysiaJohor, Johor Bahru