Grant: Discovery Grants Program - Individual

The main goal of this proposal is to establish a long-lasting research program in the cybersecurity of intelligent autonomous multi-agent systems. Recent developments of intelligent autonomous vehicles and multi-agent systems necessitate the study of their cybersecurity in terms of detection, identification, and mitigation against various cyberattacks to protect and ensure the safety of their users. Autonomous vehicles are equipped with many sensors, actuators, and network connections, making them extremely vulnerable to cyberattacks.  Existing cybersecurity solutions have been developed for cyber-physical systems focusing on power systems, embedded devices, IoT, and individual robotic systems. These results have limited usefulness in autonomous multi-agent systems where continuous interactions between agents require a cooperative approach in cybersecurity. The proposed research aims to develop a new cyberattack detection, identification, and mitigation framework for intelligent autonomous multi-agent systems using state-of-the-art machine learning techniques. We propose to apply deep reinforcement learning, deep neural networks, and resilient encrypted control techniques to develop new monitoring, identification, and control methods for such systems. The algorithms will have real-time learning capabilities, will be able to detect and identify cyberattacks, and will provide resilient control for autonomous multi-agent systems. Since multi-agents are inherently distributed systems, the proposed detection, identification, and control algorithms will be distributed in nature while aiming to achieve the global objective of the multi-agent network. The system intelligence will be used to distinguish between simple faults and sophisticated cyberattacks. The novelty of this research program is fourfold: (1) Development of a hybrid adversary detection system with an embedded learning capability in the autonomous systems deep neural network that monitors agents' sensors, actuators, and communication channels. (2) A new method that reduces the malicious agent's adverse effects on a multi-agent reinforcement learning system, including actor-critic architecture. The method will achieve the overall goal by increasing the cumulative reward of all individual agents and reducing the malicious agents' harm to the multi-agent system. (3) A novel encrypted control method that provides security of the processed data and the control loop for individual agents, where the information is encrypted in both communication channels and the control computation. This new control paradigm will encrypt not only interaction between agents but will provide secure actuator inputs based on encrypted sensor data and encrypted states of the controller. (4) A new heterogeneous multi-robot test bed for experimentation with autonomous multi-agent systems at various levels, including static and mobile sensors, ground-based autonomous agents, and micro-aerial vehicles.