A Multi-Tier Federated Learning Architecture for Scalable and Efficient Model Training in Autonomous Vehicles

Abstract

Since autonomous vehicles have been adopted so quickly, there are now more diverse and rich data streaming in real time from vehicle sensors, asking for machine learning models that focus on privacy and efficiency. A new solution called Federated Learning (FL) can learn from large amounts of distributed data without sharing those data or putting them at risk. Yet, the classic approach to FL, where a big server handles and guides learning for many clients, has challenges with keeping up with growth, established communication requirements and rate of convergence in challenging vehicle conditions. Based on our findings, we introduce the MTFL architecture for autonomous vehicle ecosystems which helps improve model training, allow larger scales and save communication resources. Our architecture involves three tiers: (i) client vehicles which conduct local model training; (ii) edge nodes in the region which take models from nearby vehicles and mix them; and (iii) a powerful cloud server which combines models from many regional edge nodes and sends the final result to local vehicles. We explain the complete mathematical setup of the MTFL framework, expressing the loss function, communication cost model and guaranteeing convergence. In experiments with synthetic AV data across 50 vehicle clients and 5 edge clusters, our MTFL model was found to be better than centralized FL in accuracy, how fast it trained and communication bandwidth. With this method, training efficiency improves by a significant 25% and communication use by 40%, without adversely affecting results. The design for MTFL architecture gives a scalable and realistic approach to using federated learning in the future of intelligent transportation systems.