IMPROVING THE TRAINING PROCEDURE OF THE ISOLATION FOREST MODEL IN A HYBRID INTRUSION DETECTION SYSTEM FOR SDN-ENABLED IOT NETWORKS
Infocommunication technologies and electronic engineering · 2026-11 · 6 (2) · 85-94
DOI: 10.23939/ictee2026.02.085 ↗Kurz gefasst
Fachartikel in der Zeitschrift Infocommunication technologies and electronic engineering, 2026-10. Verfasst von Taras Havryliv und 1 weiteren Autor:innen. Die Arbeit nennt 20 Literaturangaben.
Abstract (Originalsprache)
This paper improves the training procedure of the Isolation Forest model in a hybrid intrusion detection system for software-defined IoT networks. Unlike existing procedures, it forms the training set from a separate buffer of data samples that the other two detectors of the system did not flag and includes a time-bounded lock of training with a fallback to the previous procedure, which makes it possible to keep the vote of the model during a sustained attack and to raise the recall of its detection. The final decision in the system is taken by a majority vote of three detectors: an entropy detector, a heuristic detector and the Isolation Forest model, which is the only one of them trained during operation. It is shown that periodic retraining of the model on the contents of a short sliding window during a sustained attack makes its vote uninformative: the model starts to treat attack samples as normal and, after the attack, falsely flags normal traffic. It is also established that the quality of a model trained on a small set depends strongly on the random composition of that set, and that even a minor share of attack samples in it noticeably weakens the model. The training procedures were compared for four attack classes, for an attack of reduced intensity, after a legitimate load increase and for a start of the module during an attack. According to simulation on 120 stream realisations of the system's traffic generator, the recall of a distributed attack that is recognised by only one of the other two detectors rose from 0.33 to 0.89. The price is a temporary increase in the false-alarm rate when the legitimate load grows sharply, by more than four times, and the loss of this gain if such an attack lasts longer than the lock. Compared with training the model once, an advantage in mean recall is not proven, but the improved procedure reduces the dependence of model quality on the random composition of the training set and keeps the model able to adapt to changes in normal traffic. Without additional settings the new version of the detection module reproduces the behaviour of the previous one. The results can be used to configure the detection module in the controller of a software-defined network; the time before the lock is released is a parameter that the operator chooses depending on what matters more: fewer false alarms or resilience to sustained attacks.
Zitieren
APA
Havryliv, T., & Shpur, O. (2026). IMPROVING THE TRAINING PROCEDURE OF THE ISOLATION FOREST MODEL IN A HYBRID INTRUSION DETECTION SYSTEM FOR SDN-ENABLED IOT NETWORKS. Infocommunication technologies and electronic engineering, 6(2), 85-94. https://doi.org/10.23939/ictee2026.02.085
Vancouver
Havryliv T, Shpur O. IMPROVING THE TRAINING PROCEDURE OF THE ISOLATION FOREST MODEL IN A HYBRID INTRUSION DETECTION SYSTEM FOR SDN-ENABLED IOT NETWORKS. Infocommunication technologies and electronic engineering. 2026;6(2):85-94. doi:10.23939/ictee2026.02.085
BibTeX
@article{Havryliv2026,
author = {Havryliv, Taras and Shpur, Olha},
title = {IMPROVING THE TRAINING PROCEDURE OF THE ISOLATION FOREST MODEL IN A HYBRID INTRUSION DETECTION SYSTEM FOR SDN-ENABLED IOT NETWORKS},
journal = {Infocommunication technologies and electronic engineering},
year = {2026},
volume = {6},
number = {2},
pages = {85--94},
doi = {10.23939/ictee2026.02.085}
}
EUVIMED