NEURAL-NETWORK-BASED METHOD FOR EARLY CYBERATTACK WARNING WITH INTERPRETABLE RISK AND RELIABILITY ASSESSMENT
Infocommunication technologies and electronic engineering · 2026-11 · 6 (2) · 95-107
DOI: 10.23939/ictee2026.02.095 ↗Kurz gefasst
Fachartikel in der Zeitschrift Infocommunication technologies and electronic engineering, 2026-10. Verfasst von Andriy Sydun. Die Arbeit nennt 15 Literaturangaben.
Abstract (Originalsprache)
Early warning of multistage cyberattacks requires identifying threat precursors rather than merely classifying events after malicious activity has occurred. This study develops an adaptive correlation–temporal cyberattack prediction method that integrates heterogeneous security telemetry, cyber threat intelligence, relationships between attack stages, and behavioral anomalies to estimate the probability and timing of a transition to a critical attack phase. The method assigns adaptive weights to events based on source reliability, recency, recording quality, and alert duplication; correlates related observations; and evaluates the temporal coherence of attack-stage transitions using MITRE ATT&CK. External indicators of compromise are matched to asset configurations and vulnerabilities, while deviations from verified normal behavior provide evidence of potentially unknown threats. A machine-learning model estimates interval-specific attack probabilities, which are calibrated and aggregated across the prediction horizon. Ensemble-based uncertainty assessment, telemetry-availability masks, and adaptive alert thresholds support conservative risk estimates and control false-warning burden. Warning priority reflects attack likelihood, asset criticality, and the time available for preventive measures. An evaluation protocol is specified using the open IoT-23 dataset, comprising 23 scenarios and approximately 325.3 million network flows. It separates scenarios across training, validation, and testing, applies a 60-minute observation window and a 30-minute forecast horizon, and includes comparisons with Random Forest, XGBoost, and Transformer models, ablation analysis, and robustness checks for telemetry loss and statistical drift. The reported numerical comparison indicates an F1-score of 0.917, PR-AUC of 0.954, and 2.8 false warnings per 100 asset-days for adaptive correlation–temporal cyberattack prediction method, the median warning lead time is 19.6 minutes. The findings suggest that combining temporal event coherence with contextual threat evidence may improve warning usefulness. However, the comparative figures are not experimentally verified and require independent validation on datasets with reliable attack-stage timestamps and time-stamped cyber threat intelligence.
Zitieren
APA
Sydun, A. (2026). NEURAL-NETWORK-BASED METHOD FOR EARLY CYBERATTACK WARNING WITH INTERPRETABLE RISK AND RELIABILITY ASSESSMENT. Infocommunication technologies and electronic engineering, 6(2), 95-107. https://doi.org/10.23939/ictee2026.02.095
Vancouver
Sydun A. NEURAL-NETWORK-BASED METHOD FOR EARLY CYBERATTACK WARNING WITH INTERPRETABLE RISK AND RELIABILITY ASSESSMENT. Infocommunication technologies and electronic engineering. 2026;6(2):95-107. doi:10.23939/ictee2026.02.095
BibTeX
@article{Sydun2026,
author = {Sydun, Andriy},
title = {NEURAL-NETWORK-BASED METHOD FOR EARLY CYBERATTACK WARNING WITH INTERPRETABLE RISK AND RELIABILITY ASSESSMENT},
journal = {Infocommunication technologies and electronic engineering},
year = {2026},
volume = {6},
number = {2},
pages = {95--107},
doi = {10.23939/ictee2026.02.095}
}
EUVIMED