On Principles of Efficiency for Federated Learning

Recent technological advancements have spurred an unprecedented surge in data generation across a wide array of sources, including mobile phones, Internet of Things (IoT) devices, autonomous vehicles, tablets, and other edge computing technologies. In this landscape, Federated Learning (FL) has emerged as a crucial machine learning paradigm that enables collaborative model training across various institutions or devices, crucially preserving the privacy of raw and sensitive data. This approach is particularly vital in today’s context, where the handling of sensitive information, such as personal medical records and individual activities, is increasingly subject to stringent data protection regulations like the GDPR. These regulations highlight the challenges of data transfer across borders, making centralized training models less feasible. FL offers a significant advantage by enabling nearly seven billion IoT devices and three billion smartphones worldwide to contribute to research and development securely and efficiently. The paradigm’s ability to leverage extensive data and computing power while ensuring privacy makes it an indispensable solution, especially in fields like healthcare, where there is a pressing need for privacy- preserving large-scale studies. The importance of federated learning is underscored by its potential to harness this global network of devices, offering a scalable, secure, and efficient method for data utilization in the face of growing privacy concerns and regulatory challenges. While FL presents compelling advantages, it is imperative to confront and research the inherent challenges, particularly from optimization and efficiency standpoints. This thesis addresses the central challenge of improving the efficiency of FL systems by introducing principled solutions that reduce carbon emissions, communication costs, and computational overhead—without compromising model performance. It is structured around four key principles for enabling efficient federated learning: minimizing communi- cation emissions, exploiting training sparsity, aligning representations through anchors, and integrating secure layers for dropout tolerance. The first part of the thesis presents a rigorous methodology for quantifying the energy consumption and carbon footprint of FL, revealing that, depending on configuration, FL can be significantly more energy-intensive than centralized training—underscoring the need for energy-aware FL design. Building on this insight, the thesis proposes ZeroFL, a novel sparsification framework that reduces both computation and communication costs in on-device FL while improving accuracy under high sparsity levels. Next, it introduces FedAnchor, a federated optimization method that aligns client latent representations using server-based anchor data and a contrastive loss, enabling faster convergence and greater robustness in the semi-supervised learning scenarios. Finally, the thesis proposes vFedSec, the first efficient, dropout-tolerant, and privacy-preserving protocol for vertical FL, offering significant speedups over homomorphic encryption-based methods while ensuring secure training in the presence of unreliable client connectivity. Together, these contributions advance the state of the art in federated learning by systematically addressing its most pressing efficiency bottlenecks. The thesis offers not only concrete algorithms and empirical validations but also a generalizable framework of principles that can guide future research and deployment. By bridging optimization, energy awareness, and privacy preservation, this work lays the foundation for scalable, secure, and sustainable federated learning in real-world settings.

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2025

Author dc:creator

• Qiu, Xinchi

Advisor dc:contributor.advisor

• Lane, Nicholas

• Federated Learning
• Machine Learning

Licence

• https://www.repository.cam.ac.uk/bitstreams/e7847770-fb65-4783-9982-a9b3e3f91a8c/download
• http://purl.org/NET/rdflicense/allrightsreserved

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.121047

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/388899