Fair and equitable client selection in federated learning for heterogeneous IoV networks

<p dir="ltr">Over the past decade or so, rapid advancements in the domains of Internet of Things (IoT), edge computing, and artificial intelligence have facilitated the interconnection of billions of devices, sensors, and objects. Consequently, a large volume of data is being generated and real-time valuable insights from the same via learning-based mechanisms are highly indispensable so as to make informed decisions. Traditional machine-learning approaches mandate the data to be transmitted to a centralized server for processing purposes. However, this requires considerable computational resources and hence raises significant privacy concerns pertinent to the clients’ data. To tackle the said challenge, Federated Learning (FL) was introduced by Google as a decentralized machine learning-based mechanism, wherein a global model is collaboratively trained via the aggregation of the local parameters’ updates from the participating clients, i.e., instead of sending the raw data from the clients for the FL training purposes. Accordingly, FL, as of today, has gained considerable attention of researchers from both academia and industry, and has been extensively employed across a diverse range of applications. Client selection, nevertheless, remains a critical issue in FL since not all the clients could be employed for the training process owing to a number of system constraints. In particular, in case of heterogeneous and resource-constrained environments, a careful FL client selection process is indispensable to avoid the risk of sub-optimal model performance. Whilst a number of FL client selection frameworks have been proposed over the years, they do not take into consideration issues pertinent to fair and equitable client selection within the context of heterogeneous Internet of Vehicle (IoV) networks.</p>