Communication Efficient Federated Learning for 6G Wireless Networks

Abstract

The advancement of 5G communication technology has stimulated the development of Edge AI, and the transition to 6G will further accelerate the growth in the volume of data generated at edge devices. Despite the advancement, many systems still rely on traditional centralized learning, in which raw data collected at edge devices are uploaded to a central server. The rapid growth in the volume of data generated at edge devices and significant privacy concerns over collecting sensitive raw data make traditional centralized learning impractical. Federated learning (FL) has emerged as a key distributed learning technology to train on decentralized data without compromising privacy. However, FL introduces a significant communication bottleneck due to the iterative transmission of high-dimensional model updates for large-scale deep learning models, especially under limited communication resources.

To address the communication bottleneck, model sparsification and over-the-air computation (AirComp) have been studied. Particularly, global sparsification schemes compatible with AirComp have a limitation in that they lack an effective mechanism to compensate for the locally accumulated errors. In this dissertation, we propose a communication-efficient federated learning algorithm, coined random FLARE (R-FLARE), to overcome the limitation. In R-FLARE, all devices and the server share a common set of randomly selected indices to maintain the benefits of AirComp, while a novel selective error compensation mechanism is introduced to effectively utilize the accumulated error of each device.

We theoretically analyze R-FLARE using the $l_{2}$-norm based error compensation under a non-convex loss function, demonstrating the convergence rate of $O(1/\sqrt{T})$ over $T$ communication rounds, which is identical to those of the state-of-the-art algorithms. Furthermore, extensive simulations confirm that the proposed R-FLARE algorithm achieves superior communication efficiency and faster convergence speed compared to conventional and state-of-the-art algorithms under both ideal and practical wireless channel environments.

In conclusion, the proposed R-FLARE algorithm is expected to be a practical solution for enabling fast and communication-efficient distributed learning in future 6G Edge AI environments by overcoming severe communication resource constraints.