Leader-Assisted Client Selection for Federated Learning in IoT via the Cooperation of Nearby Devices

Federated learning is a form of distributed learning in which each participating node is handled as a client and is responsible for the training of a model using only its local data. The participation of inappropriate clients, on the other hand, may have a substantial impact on reliability, which would entail a significant consumption of resources. Due to this difficulty, the aspect of client selection is challenging. Most of the proposed approaches are based on selecting clients from their resource’s information by gathering them in a centralized manner, which may have many drawbacks. In this paper, we propose a decentralized client selection. The gathering of resource information is accomplished through the collaboration of neighboring nodes, coordinated by the leader. The leader is elected using a leader election algorithm. Based on the gathered information, the leader then trains a lightweight deep learning model to select clients throughout the IoT context. The proposed approach has been developed and validated by experimenting with different complex scenarios.