Effectiveness of anomaly detection methods in network traffic under class imbalance using the detection reliability index
Radiotekhnika · 30.09.2026 · (226) · 99-106
DOI: 10.30837/rt.2026.3.226.08 ↗Kurz gefasst
Fachartikel in der Zeitschrift Radiotekhnika, online erschienen am 30.09.2026. Verfasst von H.V. Pekurovskyi und 2 weiteren Autor:innen.
Abstract (Originalsprache)
The paper addresses the problem of network traffic anomaly detection under class imbalance, which remains a critical challenge for modern intrusion detection systems due to the limited availability of labeled attack data and the inherent scarcity of anomalous traffic compared to normal traffic. Since supervised learning approaches require representative samples of malicious traffic for every attack type, which is practically infeasible given the constant emergence of new attacks, the study employs one-class learning methods trained exclusively on normal traffic. The authors perform a comparative analysis of three such methods — a multilayer-perceptron-based autoencoder, an isolation forest, and a one-class support vector machine (OCSVM) — on a subset of the widely used CICIDS2017 dataset, which combines records of normal traffic and traffic from a LOIC HTTP DDoS attack. The merged dataset contains 755,663 records with a class imbalance ratio of IR = 4.90, and the feature space consists of 76 numerical network-flow characteristics extracted with the CICFlowMeter tool. The key scientific contribution of the work is a new evaluation metric, the Detection Reliability Index (DRI), which, unlike conventional F1-score and ROC-AUC metrics, explicitly accounts for the operational burden of false positive alerts (FPR) by introducing a multiplicative penalty in the denominator of a weighted F-like score. The metric parameters were the recall weight beta = 0.8 and the penalty coefficient alpha = 0.5, which are justified by the asymmetry between the cost of a missed attack and the cost of a false alarm for security administrators. Experimental results show that the autoencoder achieves the highest precision (0.7245), ROC-AUC (0.9154), and proposed DRI value (0.6342) owing to its minimal false positive rate (FPR = 0.049). The isolation forest attains the highest recall (0.637) but generates substantially more false alarms (FPR = 0.169), which considerably reduces its practical applicability and results in the lowest DRI among the compared methods (0.5496). The one-class SVM, trained on a limited subsample due to the computational cost of the method on large datasets, occupies an intermediate position (DRI = 0.5835). The findings confirm that maximizing recall without accounting for the false alarm rate is a suboptimal strategy for operational intrusion detection systems, and that the proposed DRI metric reveals practical advantages of methods that remain hidden when using traditional classification quality measures. The work is of practical relevance for cybersecurity practitioners selecting anomaly detection methods for deployment in real network environments with limited administrator capacity for processing security alerts, and it calls for further validation on other attack types.
Zitieren
APA
Pekurovskyi, H., Kudrenko, S., & Kotukh, Y. V. (2026). Effectiveness of anomaly detection methods in network traffic under class imbalance using the detection reliability index. Radiotekhnika, 99-106. https://doi.org/10.30837/rt.2026.3.226.08
Vancouver
Pekurovskyi H, Kudrenko S, Kotukh YV. Effectiveness of anomaly detection methods in network traffic under class imbalance using the detection reliability index. Radiotekhnika. 2026:99-106. doi:10.30837/rt.2026.3.226.08
BibTeX
@article{Pekurovskyi2026,
author = {Pekurovskyi, H.V. and Kudrenko, S.O. and Kotukh, Y. V.},
title = {Effectiveness of anomaly detection methods in network traffic under class imbalance using the detection reliability index},
journal = {Radiotekhnika},
year = {2026},
number = {226},
pages = {99--106},
doi = {10.30837/rt.2026.3.226.08}
}
EUVIMED