Privacy-Preserving Active Learning for Autonomous Urban Air Mobility Routing

The article explores the challenges of developing a privacy-preserving active learning system for autonomous urban air mobility (UAM) routing, which requires real-time decision-making, strict privacy requirements, and continuous learning from sensitive operational data.

Details

The author's research into federated learning systems for smart cities led them to explore the intersection of privacy and autonomous mobility, specifically in the context of urban air mobility (UAM) routing. Traditional centralized learning approaches were found to be incompatible with UAM routing due to the sensitive nature of flight path data, the dynamic nature of airspace policies, and the need for continuous learning and real-time decision-making. The author experimented with combining active learning, where the system intelligently selects the most valuable data points for learning, with privacy-preserving techniques like differential privacy and federated learning. The proposed system architecture includes an edge layer for local routing decisions, a fog layer for regional policy enforcement, and a cloud layer for global model training with privacy preservation. By quantifying prediction uncertainty using Bayesian deep learning, the system can identify flight scenarios where the optimal route is ambiguous due to conflicting constraints or novel situations, and flag these for human review or additional learning.