Reinforcement Learning of Defensive Strategies Against Attacks in Partially Observable Local Area Networks
Advances in Knowledge-Based Systems, Data Science, and Cybersecurity · 2025
DOI: 10.54364/cybersecurityjournal.2025.1117 ↗Kurz gefasst
Fachartikel in der Zeitschrift Advances in Knowledge-Based Systems, Data Science, and Cybersecurity, 2025. Verfasst von Takudzwa V. Banda und 1 weiteren Autor:innen.
Abstract (Originalsprache)
Enterprise Local Area Networks (LANs) face increasing threats from self-propagating worm ransomware that disrupts operations and encrypts data. Traditional intrusion detection systems rely on static rules that fail against adaptive malware. Reinforcement Learning (RL) offers a promising approach for developing autonomous cybersecurity agents capable of learning optimal defensive strategies through interaction with their environment. However, existing studies often rely on simplified network models, single unreliable detectors, and overlook unsafe defensive actions that can cause data loss and network downtime. This study develops an adaptive RL-based defense framework in Microsoft’s CyberBattleSim, formalized as a Partially Observable Markov Decision Process (POMDP). The framework integrates three key mechanisms: (1) a multi-detection system combining network- and nodelevel detectors, (2) probabilistic shielding to reduce unsafe actions and preserve data and network availability, and (3) a trust-based belief update mechanism that weights detector outputs. Experimental results demonstrate that multi detection system improves both training stability and defensive performance compared to single detection system. Probabilistic shielding minimizes data and network availability losses, both of which are key indicators of LAN resilience, while enabling the agent to learn an optimal policy. The trust system further enhances tactical precision by ensuring that belief updates accurately reflect detector reliability and confidence.
Zitieren
APA
Banda, T. V., & Rens, G. B. (2025). Reinforcement Learning of Defensive Strategies Against Attacks in Partially Observable Local Area Networks. Advances in Knowledge-Based Systems, Data Science, and Cybersecurity. https://doi.org/10.54364/cybersecurityjournal.2025.1117
Vancouver
Banda TV, Rens GB. Reinforcement Learning of Defensive Strategies Against Attacks in Partially Observable Local Area Networks. Advances in Knowledge-Based Systems, Data Science, and Cybersecurity. 2025. doi:10.54364/cybersecurityjournal.2025.1117
BibTeX
@article{Banda2025,
author = {Banda, Takudzwa V. and Rens, Gavin B.},
title = {Reinforcement Learning of Defensive Strategies Against Attacks in Partially Observable Local Area Networks},
journal = {Advances in Knowledge-Based Systems, Data Science, and Cybersecurity},
year = {2025},
doi = {10.54364/cybersecurityjournal.2025.1117}
}
EUVIMED