Action robust reinforcement learning for air mobility deconfliction against conflict induced spoofing

Increased dynamic drone usage has increased complexity in aerial navigation and often demands distributed local deconfliction. Due to the high velocities and few landmarks, robust deconfliction relies on precise positioning and synchronization. However, intentional spoofing attacks aimed at inducing navigation conflicts threaten the reliability of conventional techniques. Here, we address these concerns by establishing a baseline on the impact of novel conflict-inducing spoofing attacks on existing geometric navigation methods. Based on the impact of the attacks on the navigation, reinforcement learning (RL) strategy is used to counter the effects of spoofing attacks. In order to counter the effect of spoofing in randomized dynamic airspace conditions, a zero-sum action-robust (ZSAR) RL based on mixed Nash equilibrium objective is used. The proposed methodology yields an improved number of conflict-free paths while reducing average conflicts compared to existing state of the art RL strategies, thus making it suitable for deploying autonomous aircrafts.