OFDMA-F²L: Federated Learning With Flexible Aggregation Over an OFDMA Air Interface

Federated learning (FL) can suffer from communication bottlenecks when deployed in mobile networks, limiting participating clients and deterring FL convergence. In this context, the impact of practical air interfaces with discrete modulation schemes on FL has not previously been studied in depth. This paper proposes a new paradigm of flexible aggregation-based FL ( $\text{F}^{2}\text{L}$ ) over an orthogonal frequency division multiple-access (OFDMA) air interface, termed as “OFDMA- $\text{F}^{2}\text{L}$ ”, allowing selected clients to train local models for various numbers of iterations before uploading the models in each aggregation round. We optimize the selections of clients, subchannels and modulation scheme, adapting to channel conditions and computing power. Specifically, we derive an upper bound on the optimality gap of OFDMA- $\text{F}^{2}\text{L}$ capturing the impact of these selections, and show that the upper bound is minimized by maximizing the weighted sum rate of the clients per aggregation round. A Lagrange-dual based method is developed to solve this challenging mixed integer program of weighted sum rate maximization, revealing that a “winner-takes-all” policy provides the almost surely optimal client, subchannel, and modulation selections. Experiments on multilayer perceptrons and convolutional neural networks show that OFDMA- $\text{F}^{2}\text{L}$ with optimal selections can significantly improve the training convergence and accuracy, e.g., by about 18% and 5%, compared to potential alternatives.

I. Introduction

As a framework for distributed online computing and model training, federated learning (FL) has shown significant potential for applications, e.g., Internet-of-Things, autonomous driving, remote medical care, etc. [1]. FL enables individual mobile clients to train a global model collectively without releasing their data [2]. In particular, each client trains its local model independently relying on its local dataset and sends the gradient of the local model to a server. The server aggregates the gradients and broadcasts the aggregated global parameter to assist the clients in their local training.