Evaluating Trojan Attack Vulnerabilities in Autonomous Landing Systems for Urban Air Mobility
Description
This study examines vulnerabilities in the autonomous landing systems of Urban Air Mobility (UAM) vehicles, with a particular focus on Trojan attacks targeting Convolutional Neural Networks (CNNs) used for navigation. Trojan attacks introduce covert triggers into CNN models, leading to misclassifications under specific conditions while maintaining normal functionality in typical scenarios. Our research specifically investigates the susceptibility of Autonomous Aerial Vehicles (AAVs) employing the widely adopted DroNet model for real-time obstacle avoidance. To assess these vulnerabilities, we curated a custom dataset consisting of over 5,000 images of landing pads, categorized under both normal and Trojan-triggered conditions. Our analysis revealed a significant drop in model accuracy— from 96.4% under clean data conditions to 73.3% when Trojan triggers were present. This substantial decline highlights the severe impact of Trojan attacks, particularly during critical landing operations. These findings emphasize the urgent need for robust defense mechanisms to mitigate Trojan-induced threats within UAM systems, which play a crucial role in navigation, obstacle avoidance, and secure communication with ground control. As UAM technologies become increasingly integral to transportation and emergency response frameworks, ensuring their cybersecurity is imperative to protecting both lives and infrastructure.