Electronic Theses and Dissertations

Author

Sagar Pathak

Date

2024

Document Type

Thesis

Degree Name

Master of Science

Department

Computer Science

Committee Chair

Dipankar Dasgupta

Committee Member

Mohd. Hasan M Ali

Committee Member

Myounggyu M Won

Abstract

Data exhibits a natural distribution, and effective machine learning models rely on substantial datasets. However, much data remains inaccessible due to privacy and security risks in traditional centralized settings, which require data collection on a central server. Federated Learning (FL) addresses this by moving computation to the data source instead of the server. Despite this, FL faces challenges from data and model poisoning attacks due to its distributed nature. Verifying the authenticity of clients in FL is difficult. Proposed solutions include statistical analysis of client updates, hardware-based isolation, Differential Privacy (DP), and Homomorphic Encryption (HE). However, these solutions have limitations and significant trade-offs, such as the privacy-utility trade-off. This research proposes a novel approach to fortify the FL environment using Zero-Trust (ZT) inspired continuous verification of client updates for model poisoning attacks and filter ensembles for data poisoning attacks. Our experiment demonstrates improved results against these attacks.

Comments

Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to ProQuest.

Notes

Open Access

Share

COinS