FedBroadcast Exploit Broadcast Channel for Fast Convergence in Wireless Federated Learning

With the fast development of modern networking technologies, the transmission rate and reliability of wireless networks have been greatly improved. Meanwhile, the fast-developing Internet of Things (IoT) devices provide continually increasing computation capability. Federated learning (FL) was proposed to leverage communication and computation resources to perform machine learning (ML) tasks on IoT devices. In wireless FL systems, devices train ML models with local data sets, and a base station (BS) aggregates these trained models so that data privacy is guaranteed by the isolation of data sets. However, in the existing works, the same global model is transmitted to devices individually over the wireless channel multiple times, while the updated local models are received only by the BS, which ignores the wireless broadcast channel, incurs large communication overhead, and slows down the overall training process. In this article, we propose the FedBroadcast protocol to efficiently exploit the shared wireless broadcast channel for the two-way model transmission in FL. In the downloading step, we let BS broadcast the global model once for all scheduled devices, and design a dynamic programming-based algorithm to schedule devices optimally. In the uploading step, we also leverage the wireless broadcast channel and select some devices to receive all the updated local models without waiting for model downloading in the next round. Finally, to solve the potential block-cyclic sampling problem brought by device scheduling, we adopt pluralistic averaging modification, which improves the convergence performance under extreme data distributions. Extensive experiments demonstrate that FedBroadcast outperforms the existing wireless FL under different system settings.