Fairness Can Save Lives: A MAB Based Client Selection Strategy for Federated Learning Towards IoV Assisted ITS

Federated learning (FL) can not only improve training efficiencies for various intelligent recognition tasks (IRTs) in intelligent transportation system (ITS) with lower training latency, but also enhance data privacy during neural network training. Meanwhile, Internet of vehicles (IoV), supporting data transmission in ITS, is characterized by dynamic wireless environments and dense access demands. Thus, client selection is essential for FL arrangement to mitigate network unreliability resulted from the scarcity of spectrum resources, and improves the training performances of FL models. Thus, most scheduling strategies mainly aim to minimize the total latency for training, ignoring the consideration of client selection fairness. However, in ITS, if fairness is not taken into account, the trained global model would be out of balance with inaccurate recognition results, leading to unsatisfactory services quality, and even life-threatening accidents. Based on this fact, we design a multi-armed-bandit (MAB)-based client selection strategy to improve the training efficiency and reduce the latency for FL-based IRTs within ITS. Particularly, considering a high requirement of security for ITS related wireless services, we further propose a novel fairness measurement index in the reward function of MAB strategy to balance the delay and fairness performances of scheduling. Finally, we compare the proposed strategy with a gossip based and a delay-oriented client selection strategy. By comparing the performances with the previous strategies, the simulation results verified that the proposed scheduling strategy is the best in terms of scheduling fairness, with the highest accuracy and lower latency. That is to say, the proposed scheduling strategy can significantly improve the security of ITS with fairer, accurate and fast FL technologies.