Frequency Modulation Aggregation for Federated Learning

TL;DR

This paper introduces a digital frequency broadband aggregation scheme for federated edge learning that reduces power consumption and improves performance over traditional analog methods by using MFSK and TBMA techniques.

Contribution

The paper proposes a novel digital aggregation method combining MFSK and TBMA, offering lower power consumption and better performance in federated learning communication.

Findings

Achieves no performance drop up to -10 dB SNR

Outperforms linear analog modulations in AWGN channels

Requires 14 dB less PAPR than analog schemes

Abstract

Federated edge learning (FEEL) is a framework for training models in a distributed fashion using edge devices and a server that coordinates the learning process. In FEEL, edge devices periodically transmit model parameters to the server, which aggregates them to generate a global model. To reduce the burden of transmitting high-dimensional data by many edge devices, a broadband analog transmission scheme has been proposed. The devices transmit the parameters simultaneously using a linear analog modulation, which are aggregated by the superposition nature of the wireless medium. However, linear analog modulations incur in an excessive power consumption for edge devices and are not suitable for current digital wireless systems. To overcome this issue, in this paper we propose a digital frequency broadband aggregation. The scheme integrates a Multiple Frequency Shift Keying (MFSK) at the transmitters and a type-based multiple access (TBMA) at the receiver. Using concurrent transmission, the server can recover the type (i.e., a histogram) of the transmitted parameters and compute any aggregation function to generate a shared global model. We provide an extensive analysis of the communication scheme in an additive white Gaussian noise (AWGN) channel and compare it with linear analog modulations. Our experimental results show that the proposed scheme achieves no drop in performance up to $-10$ dB and outperforms the analog counterparts, while requiring 14 dB less in peak-to-average power ratio (PAPR) than linear analog modulations.