No Cover Image

E-Thesis 32 views 11 downloads

From Algorithm to Application: Enhancing Federated Learning with Adaptive Aggregation and Gradient Protection / YI HU

Swansea University Author: YI HU

  • 2025_Hu_Y.final.71338.pdf

    PDF | E-Thesis – open access

    Copyright: the author, Yi Hu, 2025. Distributed under the terms of a Creative Commons Attribution 4.0 License (CC BY 4.0)

    Download (6.5MB)

DOI (Published version): 10.23889/SUThesis.71338

Abstract

Federated Learning (FL) has emerged as a promising paradigm for decentralised machine learning, allowing multiple clients to collaboratively train models without sharing their private data. This distributed approach is particularly relevant in domains with stringent data privacy requirements, such a...

Full description

Published: Swansea University 2025
Institution: Swansea University
Degree level: Doctoral
Degree name: Ph.D
Supervisor: Xie, X
URI: https://cronfa.swan.ac.uk/Record/cronfa71338
Abstract: Federated Learning (FL) has emerged as a promising paradigm for decentralised machine learning, allowing multiple clients to collaboratively train models without sharing their private data. This distributed approach is particularly relevant in domains with stringent data privacy requirements, such as finance, healthcare, and edge computing. However, despite its advantages, FL faces three critical challenges: (1) efficient model aggregation, (2) protection against privacy leakage, and (3) real-world applicability in complex domains. This thesis addresses these challenges by proposing novel strategies to optimise aggregation mechanisms, enhance its privacy protection capabilities, and explore its application in financial modelling. First, we introduce Element-Wise Weights Aggregation for FL (EWWA-FL), a novel optimisation technique that improves global model convergence by adopting element-wise adaptive weighting. Traditional FL aggregation methods, such as FedAvg, assign a single proportion to each local model without considering the varying importance of individual model parameters. In contrast, EWWA-FL assigns unique proportions to each element within local model weights, ensuring more precise updates that account for dataset heterogeneity among clients. Experimental results demonstrate that EWWA-FL significantly improves both convergence speed and final model accuracy, outperforming FedAvg, FedOpt, and FedCAMS across multiple benchmark datasets. By incorporating an element-wise approach, EWWA-FL provides a more adaptive and fine-grained aggregation strategy that enhances FL’s performance in both Independent and Identically Distributed (IID) and Non-IID settings. Second, we propose AdaDefence, a privacy-preserving defence mechanism against gradient leakage attacks in FL. While FL eliminates the need for raw data sharing, recent attacks have demonstrated that an adversary can reconstruct private training data from shared gradients, posing a severe privacy risk. To counteract this, AdaDefence introduces a gradient stand-in approach, wherein local clients replace actual gradients with modified gradients before sending them to the server. This method prevents attackers from reconstructing private data while maintaining model utility. AdaDefence effectively defends against state-of-the-art attacks such as Deep Leakage from Gradients (DLG), Generative Regression Neural Network (GRNN), and Inverting Gradient (IG) without significantly compromising model accuracy. Our extensive empirical analysis shows that AdaDefence provides strong privacy guarantees while ensuring minimal performance degradation, making it a practical and scalable solution for real-world FL deployments. Finally, we explore the real-world application of FL in financial modelling, particularly in Cross-Stock Trend Integration (CSTI) for enhancing stock price prediction. Traditional financial models suffer from data fragmentation, where different financial institutions and stock markets operate in silos, limiting predictive power. To overcome this, we develop a FL-based approach that enables multiple financial institutions to collaboratively train stock prediction models without exposing sensitive trading data. This approach leverages cross-stock trend integration, allowing predictive models to learn patterns from multiple stocks while preserving privacy.Our experimental results demonstrate that federated cross-stock learning improves predictive accuracy and model robustness, outperforming conventional single-stock prediction methods.By enabling secure, multi-institution collaboration, this work highlights the potential of FL in advancing financial modelling while ensuring regulatory compliance and data confidentiality. By addressing these fundamental aspects, optimisation and protection, this thesis makes significant contributions to the field of FL. The proposed methodologies collectively enhance FL’s efficiency, security, and real-world applicability. Through EWWA-FL, FL models achieve faster and more reliable convergence. By introducing AdaDefence, FL gains stronger privacy protections against gradient-based attacks. Finally, by demonstrating FL’s potential in cross-stock trend integration, this thesis showcases how FL can be deployed in privacy-sensitive financial applications. These contributions pave the way for more efficient, secure, and scalable FL systems, advancing their adoption in a wide range of domains, including healthcare, autonomous systems, and financial technology.
Keywords: Deep Learning, Federated Learning, Finance
College: Faculty of Science and Engineering

Similar Items