Towards Robust and Secure Autonomous Cyber Defense Agents in Coalition Networks
Johannes F. Loevenich, Tobias Hürten, Florian Spelter, Erik Adler, Johannes Braun, Linnet Moxon, Yann Gourlet, Thomas Lefeuvre, Roberto Rigolin F. Lopes · 2024
This paper outlines a methodology for developing and verifying robust and secure autonomous cyber defense (ACD) agents using hybrid artificial intelligence (AI) models to protect coalition networks. The motivation is to gather quantitative evidence throughout the entire lifecycle of the agent to support its robustness and security. The methodology includes a hierarchical AI agent architecture that emphasizes real-time threat detection, dynamic adaptation, and secure data handling. We propose a hybrid threat and attack model that combines different methods to identify potential attacker goals and capabilities, and generate and catalog potential threats for mitigation. In addition, we outline a W-shaped development process for formal verification of AI models and discuss the challenges for security and robustness throughout the life cycle. Thus, this work defines a starting point for addressing the challenges of training AI models in a complex and evolving cyber environment, and reports on ongoing research in the area of autonomous cyber defense systems for military applications.