Reproducibility Data for Explainable Machine Learning Evaluation of Cyberattack Detection in Smart Manufacturing IIoT Systems
Description
This dataset contains reproducibility artifacts for a machine-learning intrusion-detection study using the public WUSTL-IIoT-2021 industrial IoT testbed. The deposited files include frozen temporal-block split indices and summaries, grouped cross-validation results for logistic regression, random forest, XGBoost and LightGBM, out-of-fold LightGBM probabilities, calibration and validation-selected operating thresholds, primary held-out test metrics, block-bootstrap uncertainty results, attack-family recall, and reviewer-stage robustness analyses. The latter include a strictly later attack-free benign evaluation, leave-one-attack-family-out testing, dominant-feature ablations, a group-structured permutation control, detailed partition audits, model fit-time variability, and block-bootstrap global/local SHAP stability analyses. The raw WUSTL-IIoT-2021 dataset is third-party source data and is not redistributed. The primary near-perfect benchmark result should be interpreted as block-held-out within-campaign discrimination rather than future-time, endpoint-disjoint, cross-site, or unseen-attack generalization.