Machine Learning for Multi-Class Intrusion Detection in IoT and IIoT Edge Environments
Journal of Eye-Innovation in Security Analysis · 12.08.2026 · 1 (20261) · 45
DOI: 10.68990/jiesa.202614 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Eye-Innovation in Security Analysis, online erschienen am 12.08.2026. Verfasst von Irfan Uddin und 2 weiteren Autor:innen. Die Arbeit nennt 6 Literaturangaben.
Abstract (Originalsprache)
With the growing number of Internet of Things (IoT) and Industrial Internet of Things (IIoT) deployments, the attack surface has expanded on resource-constrained edge nodes, where intrusion-detection models must detect a wide variety of threats while controlling false alarms. In this paper, we present and evaluate a leakage-aware machine-learning framework using the cybersecurity dataset Edge-IIoTset. The empirical analysis is performed on 157,800 labeled records from the ML-ready subset, which contains normal traffic and 14 attack types. The pipeline removes timestamps, host addresses, raw payload/content strings, textual request fields, target labels, and constant variables to mitigate direct identifier and content leakage. It retains 34 nonconstant numeric protocol features. Class-weighted logistic regression, random forest, and extremely randomized trees are compared in 80/20 hold-out experiments stratified for binary attack detection and 15-class traffic classification. Random forest achieved 99.54% binary accuracy, 0.9911 a macro-F1, 0.9822 Matthews correlation coefficient (MCC), 0.9997 ROC-AUC. It achieves 91.76% accuracy and, 0.8731 macro-F1, and 0.9104 MCC for multiclass classification, beating the linear baseline and extra-trees model. The hold-out set perfectly separates DDoS-UDP, DDoS-ICMP, and DDoS-TCP, whereas Fingerprinting and web-oriented attacks are still more difficult. The rare MITM class exhibits full recall but low precision and shows a class-imbalance trade-off. Transport-layer variables, especially TCP acknowledgment, destination-port, sequence, and length fields, dominate the impurity-based importance. The results support tree ensembles for this benchmark, but also show why device-separated, time-aware, and cross-dataset validation is needed before operational deployment.
Zitieren
APA
Uddin, I., AlDawla, N., & Alserry, A. A. (2026). Machine Learning for Multi-Class Intrusion Detection in IoT and IIoT Edge Environments. Journal of Eye-Innovation in Security Analysis, 1(20261), 45. https://doi.org/10.68990/jiesa.202614
Vancouver
Uddin I, AlDawla N, Alserry AA. Machine Learning for Multi-Class Intrusion Detection in IoT and IIoT Edge Environments. Journal of Eye-Innovation in Security Analysis. 2026;1(20261):45. doi:10.68990/jiesa.202614
BibTeX
@article{Uddin2026,
author = {Uddin, Irfan and AlDawla, Najran and Alserry, Amel Ali},
title = {Machine Learning for Multi-Class Intrusion Detection in IoT and IIoT Edge Environments},
journal = {Journal of Eye-Innovation in Security Analysis},
year = {2026},
volume = {1},
number = {20261},
pages = {45},
doi = {10.68990/jiesa.202614}
}
EUVIMED