Theses and Dissertations

ORCID

0009-0003-8878-2659

Issuing Body

Mississippi State University

Advisor

Chen, Zhiqian

Committee Member

Chen, Jingdao

Committee Member

Mittal, Sudip

Committee Member

Swan II, John

Date of Degree

8-12-2026

Original embargo terms

Immediate Worldwide Access

Document Type

Dissertation - Open Access

Major

Computer Science (Artificial Intelligence & Robotics)

Degree Name

Doctor of Philosophy (Ph.D.)

College

James Worth Bagley College of Engineering

Department

Department of Computer Science and Engineering

Abstract

The rapid expansion of data-driven methodologies in healthcare and finance is increasingly constrained by ethical, regulatory, and security requirements that limit centralized data sharing. Sensitive information—including electroencephalogram (EEG) recordings and inter-institutional financial transaction logs—remains siloed, reducing opportunities for collaborative model development and limiting generalizability across heterogeneous environments. In healthcare, epileptic seizure prediction is challenged by inter-patient variability, scarce and non-IID EEG data distributions, and the need for real-time inference on resource-constrained wearable devices. In finance, anti-money laundering (AML) detection is complicated by decentralized and dynamically evolving transaction networks, temporal-structural dependencies, limited labeled data, and information-sharing restrictions. This dissertation proposes a unified privacy-preserving Federated Learning (FL) framework to address these challenges across both domains. For seizure prediction, a federated transformer based architecture is developed that incorporates client-level personalization through a hypernetwork mechanism, enabling adaptive model specialization while maintaining global knowledge aggregation. To support deployment on wearable devices, a knowledge distillation strategy produces lightweight, patient-specific models that preserve predictive performance while improving computational efficiency. For financial crime detection, a federated multi-view temporal Graph Convolutional Network (TGCN) is introduced to collaboratively model structural and temporal transaction patterns across institutions. Asemi-supervised pseudo-label propagation mechanism mitigates label scarcity, while federated aggregation enables cross-institutional learning without exposing raw transactional data. Experimental evaluations under heterogeneous and non-IID conditions demonstrate improved robustness, scalability, and predictive performance relative to centralized and baseline federated approaches. The proposed seizure prediction framework enhances preictal state detection across diverse EEG profiles, while the federated multi-view TGCN strengthens AML detection in dynamic financial ecosystems under limited supervision. Overall, this work advances scalable, privacy-aware collaborative intelligence and contributes to the development of trustworthy and deployable artificial intelligence systems capable of secure cross-institutional learning in privacy-sensitive domains.

Download

Share

COinS