CASCADED HYBRID ARCHITECTURE FOR NETWORK ANOMALY DETECTION BASED ON ISOLATION FOREST AND GAN AUTOENCODERS
Information and communication technologies, electronic engineering · 2026-06 · 6 (1) · 30-40
DOI: 10.23939/ictee2026.01.030 ↗Kurz gefasst
Fachartikel in der Zeitschrift Information and communication technologies, electronic engineering, 2026-04. Verfasst von A. Riy und 1 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 12 Literaturangaben.
Abstract (Originalsprache)
The research is dedicated to improving the security of telecommunication networks by developing a hybrid system for detecting anomalies in traffic. The problem of the work is due to the inefficiency of traditional signature-based Intrusion Detection Systems (IDS) against complex, time-distributed cyberattacks. A two-level architecture is proposed, which aggregates statistical and deep behavioral machine learning through an adaptive Decision Fusion module. The first level is represented by a statistical stream based on the Isolation Forest algorithm for rapid filtering of explicit spatial deviations. The second level is a deep behavioral stream based on the DGAE architecture, where a Multi-Layer Perceptron Autoencoder (MLP) acts as a feature extractor. To metrize the Uncertainty of predictions, a Generative Adversarial Network (WGAN-GP) and the stochastic Monte Carlo Dropout method are integrated. The consideration of the spatio-temporal context of connections is realized by the Sliding Window algorithm, which transforms isolated flow-features into time series. Experimental evaluation was performed on the CIC-IDS2017 dataset (2.83 million flows). Four strategies for ensemble independent signals were investigated: Score Fusion, Consensus Fusion, Stacking Fusion, and Cascade Fusion. It is proven that the Cascade Fusion method is optimal for stateless-environments (analysis of isolated flows). This strategy compensates for the high False Positive Rate of Reconstruction Error by sequential filtering through Mahalanobis Distance and Isolation Forest scores, ensuring the maximum F1-Score. During modeling with the Sliding Window, the "Anomaly Dilution" effect was detected and mathematically justified. It is established that the expansion of the input space from 78 to 234 features leads to a drop in the autoencoder's sensitivity: anomalous patterns are smoothed out by the successful reconstruction of legitimate neighboring vectors. In contrast, it is proven that the aggregation of time series fundamentally strengthens distance metrics. In the contextual multidimensional space, the pure Mahalanobis distance becomes self-sufficient and surpasses all existing ensemble algorithms in overall accuracy. A methodology for IDS deployment is formed: cascade ensembles are effective for memoryless packet inspection systems, while the latent distance metric is optimal for the preventive protection of telecommunication networks based on contextual flow analyzers.
Zitieren
APA
Riy, A., & Zablotskyi, S. (2026). CASCADED HYBRID ARCHITECTURE FOR NETWORK ANOMALY DETECTION BASED ON ISOLATION FOREST AND GAN AUTOENCODERS. Information and communication technologies, electronic engineering, 6(1), 30-40. https://doi.org/10.23939/ictee2026.01.030
Vancouver
Riy A, Zablotskyi S. CASCADED HYBRID ARCHITECTURE FOR NETWORK ANOMALY DETECTION BASED ON ISOLATION FOREST AND GAN AUTOENCODERS. Information and communication technologies, electronic engineering. 2026;6(1):30-40. doi:10.23939/ictee2026.01.030
BibTeX
@article{Riy2026,
author = {Riy, A. and Zablotskyi, S.},
title = {CASCADED HYBRID ARCHITECTURE FOR NETWORK ANOMALY DETECTION BASED ON ISOLATION FOREST AND GAN AUTOENCODERS},
journal = {Information and communication technologies, electronic engineering},
year = {2026},
volume = {6},
number = {1},
pages = {30--40},
doi = {10.23939/ictee2026.01.030}
}
EUVIMED