Reinforcement-Learning-Driven Integrated Detection and Mitigation of UAV GPS Spoofing Attacks
Jueming Hu, Mohammad Ammar, Bilal Hussain, Jae-Won Kim, Irfan Khan
Abstract
Unmanned Aerial Vehicles (UAVs) have demonstrated significant capabilities across various applications, including logistics, urban air mobility, surveillance, and defense. However, UAV operational effectiveness heavily depends on the Global Positioning System (GPS), which provides real-time navigation essential for mission success. UAV reliance on GPS introduces potential vulnerabilities to spoofing attacks, where adversaries transmit fictitious signals to disrupt navigation. This study proposes RLDM-UAV (Reinforcement Learning-driven integrated Detection and Mitigation of UAV GPS spoofing attacks), a novel reinforcement learning framework integrating detection and mitigation to ensure resilient UAV navigation. RLDM-UAV relies on GPS data and onboard camera inputs, eliminating the need for additional sensors. RLDM-UAV utilizes a Deep Q-Network (DQN)-based algorithm to enable dynamic GPS spoofing detection and adaptive switching between GPS-based and vision-based navigation policies based on the reliability of GPS signals. To enhance learning efficiency, we propose a novel experience replay mechanism that prioritizes incorrect detections. The average online computation time for the real-time detection and mitigation decision per time step is less than 23 ms on ARM-based CPUs. The performance of RLDM-UAV is evaluated against three types of GPS spoofing attacks: random attacks, replay attacks, and stealth attacks. The results demonstrate that RLDM-UAV achieves superior attack detection accuracies across all three types of GPS spoofing attacks and greater robustness under varying attack rates during testing compared to baseline methods.