Date of Award
2024
Publication Type
Thesis
Degree Name
M.Sc.
Department
Computer Science
Keywords
distributed machine learning; privacy-preserving machine learning; secret sharing; vertical federated learning
Supervisor
Dima Alhadidi
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 International License.
Abstract
Vertical federated learning (VFL) enables multiple parties, each owning different features of the same sample space, to collaboratively train a machine learning model while retaining their data locally. Each party keeps its data local in this setting, exchanging only intermediate computed results. However, these intermediate results are susceptible to various inference attacks, potentially compromising privacy. Recent solutions for VFL leverage cryptography, trusted execution environments, or differential privacy to protect privacy. However, these approaches often introduce significant computational and communication overheads, limiting the participation of low-resource devices or negatively impacting model accuracy. This thesis introduces FedMod, a novel privacy-preserving VFL approach that leverages multi-server secret sharing to achieve security and efficiency. FedMod addresses the limitations of existing methods, offering a robust solution that maintains accuracy comparable to centralized settings while significantly outperforming state-of-the-art privacy-preserving VFL approaches in computation time and communication cost. Our experimental results highlight FedMod's consistency across varying numbers of parties and middle servers, demonstrating its scalability and reliability. Moreover, FedMod excels in training time, delivering substantially faster results than all other high-performance approaches, making it particularly well-suited for real-time applications. Even when compared to methods using advanced compression techniques, FedMod maintains competitive communication cost while offering superior computational efficiency.
Recommended Citation
Mojallal, Kasra, "FedMod: Vertical Federated Learning Using Multi-Server Secret Sharing" (2024). Electronic Theses and Dissertations. 9622.
https://scholar.uwindsor.ca/etd/9622