Toward Efficient and Reliable Wireless Federated Learning: System Design and Theoretical Analysis

Author: ORCID icon
Wei, Xizixiang, Electrical Engineering - School of Engineering and Applied Science, University of Virginia
Shen, Cong, EN-Elec & Comp Engr Dept, University of Virginia

In recent years, the surge in data generation by wireless edge devices has propelled the integration of artificial intelligence (AI) across various domains, including computer vision and natural language processing. Traditionally, ML models are trained in a centralized fashion, where edge devices transmit their data to a central server. However, this centralized approach incurs substantial communication resource costs and poses significant privacy risks. Federated Learning (FL) emerges as an innovative solution, addressing these concerns by decentralizing the learning process. The quest for communication efficiency is central to the optimization of FL, as it represents a critical bottleneck that impacts its scalability and effectiveness. This dissertation concentrates on enhancing the reliability and efficiency of both uplink and downlink wireless transmissions within the FL framework.

This dissertation introduces a pioneering convergence analysis for FL under the condition of simultaneous noisy communications in both directions, establishing the criteria essential for ensuring FL's convergence over noisy channels. Followed by the theoretical results, two novel designs are proposed for both uplink and downlink transmissions in multiple-input multiple-output (MIMO) and single-input single-output (SISO) systems. For MIMO systems, the proposed ``random orthogonalization" method, leveraging the massive MIMO characteristics of channel hardening and favorable propagation, facilitates natural over-the-air model aggregation without the need for channel state information at the transmitters (CSIT) in the uplink stage, and offers an efficient model broadcasting technique during the downlink stage. In the realm of SISO systems, our strategy also eliminates the necessity for CSIT through the use of orthogonal sequences, providing robust and flexible differential privacy (DP) guarantee at both the item and client levels.

Extensive numerical experiments conducted with real-world datasets affirm the effectiveness and efficiency of the proposed methods, highlighting their potential to significantly enhance the FL process by addressing its communication challenges.

PHD (Doctor of Philosophy)
Machine Learning, Federated Learning, Wireless Communication
Issued Date: