Adversarial Attacks and Resilience in Edge AI for Autonomous UAV and V2X Systems
Abstract
The rapid integration of artificial intelligence (AI) into edge computing has enabled real-time, low-latency decision-making in safety-critical autonomous systems, including unmanned aerial vehicles (UAVs), V2X-connected vehicles, and roadside units. However, edge AI models are vulnerable due to limited computational resources, physical exposure, and decentralized architectures, operating under strict latency, energy, and memory constraints. This chapter makes three contributions. First, it develops a three-dimensional adversarial attack taxonomy based on attack phase, targeted subsystem, and realization mode. Second, it presents a comparative UAV-V2X vulnerability analysis using reported attack success rates, including 97% to 35% UAV navigation degradation, 84.8% physical stop-sign misclassification, and 80% black-box LiDAR spoofing. Third, it proposes the Principled Resilience Framework (PRF), a hardware-budget-aware three-layer defense architecture whose performance estimates require future experimental validation.