Real-Time Detection of White and Black Box Adversarial Attacks on Deep Reinforcement Learning UAV Guidance

This paper explores the vulnerability of Deep Reinforcement Learning (DRL) models used in Uncrewed Aerial Vehicles (UAVs) to adversarial attacks. It evaluates popular white and black box attacks, including Fast Gradient Sign Method (FGSM), Carlini & Wagner (C&W), Projected Gradi-ent Descent (PGD), Zeroth-Order-Optimisation (ZOO), Zeroth-Order-Optimisation-Adam (ZOO-Adam), and Genetic Attack (GA). The study assesses attack success and real-time attack capabilities, crucial for UAV operations. It highlights the need for robust defenses and proposes an adversarial detector using an Auto-encoder and compares it with an LSTM detector. The paper emphasizes the importance of addressing adversarial threats and proposes mitigation methods.