Leakage-Controlled Attack Classification with Cyber Kill Chain Stage Mapping for IoT Devices in SOHO Environments
Electronics · 20.11.2026 · 15 (22) · 5421
DOI: 10.3390/electronics15225421 ↗Kurz gefasst
Fachartikel in der Zeitschrift Electronics, online erschienen am 20.11.2026. Verfasst von Benjamin Aziz und 2 weiteren Autor:innen. Die Arbeit nennt 45 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Machine-learning (ML) intrusion detection studies on the CICIoT2023 IoT dataset routinely report near-perfect accuracy. We show that, for the widely used class-balanced six-class release, such figures are an artefact of the evaluation protocol. A data audit finds that 39.0% of the 1.8 million records are exact duplicates introduced by oversampling: under the conventional random 80/20 split, 48.6% of test records have an identical copy in the training set. We also find that the inter-arrival-time feature behaves as a capture-time stamp, achieving 87.4% accuracy on its own. Under the conventional protocol, Random Forest reaches 99.86% accuracy. With duplicates removed, the shortcut feature excluded, capture-time-blocked five-fold cross-validation and nested hyperparameter tuning, the best of seven models (XGBoost) reaches a macro-F1 of 0.733 (95% CI ± 0.098), and the three tree ensembles are statistically indistinguishable. At a realistic 1% attack prevalence, alert precision falls to 8.3%. On an independently processed, duplicate-free release of the full CICIoT2023 capture, holding out entire capture files lowers macro-F1 from 0.803 under a random split to 0.720, and the confusions between DDoS and DoS and among benign, reconnaissance and web-attack traffic recur. Each attack category is mapped to a single Cyber Kill Chain (CKC) stage, so the stage labels are derived deterministically from the attack categories and no new stage classifier is proposed. Instead, we evaluate what reporting results at stage level adds beyond an attack label: merging the DDoS and DoS categories into a single Delivery stage removes a confusion that the flow features of both releases cannot resolve; a previously unseen Delivery subtype is still assigned to the Delivery stage; and cost-sensitive stage decisions reduce missed attacks. Stage labels are assigned at category level, so the analysis does not infer an attack’s progression through the kill chain over time. SHAP analysis is repeated on the leakage-controlled model. We release the audit and the evaluation code so that results on this dataset can be reported reproducibly.
Zitieren
APA
Aziz, B., Makiyah, E. H., & Oragwu, U. (2026). Leakage-Controlled Attack Classification with Cyber Kill Chain Stage Mapping for IoT Devices in SOHO Environments. Electronics, 15(22), 5421. https://doi.org/10.3390/electronics15225421
Vancouver
Aziz B, Makiyah EH, Oragwu U. Leakage-Controlled Attack Classification with Cyber Kill Chain Stage Mapping for IoT Devices in SOHO Environments. Electronics. 2026;15(22):5421. doi:10.3390/electronics15225421
BibTeX
@article{Aziz2026,
author = {Aziz, Benjamin and Makiyah, Estabraq H. and Oragwu, Ukamaka},
title = {Leakage-Controlled Attack Classification with Cyber Kill Chain Stage Mapping for IoT Devices in SOHO Environments},
journal = {Electronics},
year = {2026},
volume = {15},
number = {22},
pages = {5421},
doi = {10.3390/electronics15225421}
}
EUVIMED