Cyber Operations Gyms to Train Autonomous Cyber Defense Agents for NATO
Alexander Velazquez, Johannes F. Loevenich, Tobias Hürten, Konrad Wrona, Paulo H. L. Rettore, Vasil Boshnakov, Frederica Free-Nelson, Tracy Braun, Roberto Rigolin F. Lopes · 2025
This paper discusses challenges and potential solutions for training Autonomous Cyber Defense (ACyD) agents for coalition scenarios, particularly at the tactical edge. We propose the use of an Autonomous Cyber Operations (ACyO) gym to train specialized agents for coalition operations, tailored to the unique characteristics of tactical edge environments. Key characteristics of the tactical edge include distributed networking, intermittent connectivity, resource constraints, contact with hostile (cyber) forces, connections over untrusted network infrastructure, and interaction with an unpredictable physical environment. Our research also considers how these agents can be designed and deployed in a scalable manner while managing limited resources. We discuss why this work is relevant to NATO and the importance of federated learning to enable effective decentralized learning without compromising operational secrecy or data integrity. ACyD agents can enhance the resilience and security of constrained, contested, and highly dynamic military network environments. The results of this research will provide actionable insights and advanced methodologies for developing robust autonomous capabilities to defend against sophisticated cyber threats within critical military networks.