Hierarchical Multi-Agent Reinforcement Learning for Autonomous Cyber Defense in Coalition Networks
Tobias Hürten, Johannes F. Loevenich, Florian Spelter, Erik Adler, Johannes Braun, Linnet Moxon, Yann Gourlet, Thomas Lefeuvre, Roberto Rigolin F. Lopes
Abstract
This paper presents a comprehensive methodology for autonomous cyber defense (ACD) through the deployment of distributed agents using multi-agent reinforcement learning (MARL). The primary objective is to maintain operational services distributed across three geographical locations interconnected by an untrustworthy network infrastructure (Internet). The ACD agents will be developed and tested in the Cyber Operations Research Gym (CybORG) environment, which simulates a wide range of network conditions and hosts both red (attacker) and blue (defender) agents. The proposed solution employs a centralized training approach with a decentralized execution strategy using Multi-Agent Proximal Policy Optimization (MAPPO). In addition, we discuss the use of an efficient communication protocol to digitize messages into an 8-bit format, using personalized message generation for the ACD agents, thereby enhancing their collaborative defensive capabilities. Furthermore, we outline a curriculum learning methodology and a hierarchical agent structure to optimize defense strategies and zone-specific tasks.