Mobility-Assisted Federated Learning for Vehicular Edge Computing.

This paper addresses the unique challenges of machine learning in smart public transportation systems, where data processing and model training are decentralized across smart buses. Traditional federated learning (FL) algorithms, mostly synchronous, struggle in this environment due to asynchronous communication patterns and limited Roadside Units (RSUs) coverage. To overcome these challenges, we introduce MARLIN (Mobility Assisted fedeRated LearnINg), an innovative asynchronous FL approach that utilizes the predictable movement of buses and the potential for Vehicle-to-Vehicle (V2V) communication. MARLIN employs buses as relays for server communication, enhancing interaction frequency and expediting FL convergence. Our experiments on the FMNIST dataset demonstrate that MARLIN significantly outperforms the existing asynchronous FL method [1], offering a viable solution for efficient data processing in intelligent public transportation systems.