QoS-Aware Resource Bargaining for Federated Learning Over Edge Networks in Industrial IoT

This work addresses the problem of optimal resource allocation for federated learning (FL) over wireless edge-networks to maximize the quality of Industrial Internet of Things (IIoT)-based services in terms of delay and revenue. The resource-constrained nature of IIoT devices necessitates the efficient utilization of their computation and communication resources for performing FL. Additionally, the selection of an optimal set of IIoT devices by a service-provider is necessary to perform FL with a required QoS. Furthermore, to ensure the participation of the desired IIoT devices, it is necessary for the service provider to properly incentivize these devices. Although few works aim to address these issues, the possibility of partial contribution of CPU power and dataset by IIoT devices and their corresponding incentivization is overlooked by researchers. Hence, in this work, we propose FedBarg, a game-theoretic scheme, in which we model the aforementioned problem using a Nash bargaining game. We show that the corresponding optimization problem is non-convex. Thereafter, we propose a joint coalition formation game- and utility theory-based scheme to find the optimal solutions. Simulations show that FedBarg improves the incurred delay by 16.7% and revenue of the users and service provider by 14.7% and 11.1%, respectively, compared to existing benchmarks.