Entropy and Mobility-Based Model Assignment for Multi-Model Vehicular Federated Learning

Machine Learning (ML) is extensively employed for key functions of Connected and Autonomous Vehicles (CAVs), where many models are executed simultaneously within a vehicle to provide diverse applications, from perception to planning and control. One of the most appealing ML approaches for CAVs is Federated Learning (FL) due to its privacy-preserving nature and distributed learning capabilities. However, current FL approaches mostly focus on single-model training and are unsuitable for parallel training of multiple models. Multi-model FL involves training multiple ML models to perform different tasks, often simultaneously, to meet the demands of different applications within the same context. In this way, this work introduces MELRO, an FL model assignment algorithm based on link duration, training latency, and data entropy from CAVs. MELRO balances computing resources and addresses high vehicle mobility while considering the heterogeneity of data and availability of resources in CAVs. The assignment algorithm takes advantage of data transmitted periodically by CAVs, such as beacons, to calculate link duration and training latency, define the model assignment matrix for CAVs, and consider data entropy. Finally, MELRO increases accuracy for FL applications by at least 11.76% while reducing training latency by 25% and maintaining computational resource usage.