Date of Award

12-11-2025

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (Ph.D.)

Department

Electrical & Computer Engineering

First Advisor

Kamrul Hasan

Abstract

Healthcare AI requires large datasets for clinical-grade performance, yet privacy regulations prevent direct inter-institutional data sharing. While traditional federated learning (FL) enables collaborative training without centralizing data, it suffers from four critical limitations: accuracy degradation, gradient leakage vulnerabilities, high communication costs, and limited interpretability. Accordingly, the overarching research goal is to design a privacy-first FL framework for e-health that preserves clinical-grade accuracy on obfuscated data while substantially reducing information leakage, minimizing communication overhead, and enabling interpretable decision support without sharing weights, gradients, or raw pixels. This thesis systematically addresses these challenges through a three-method framework that progressively enhances privacy, efficiency, and clinical utility. The first contribution develops ensemble-based FL that replaces vulnerable single-model clients with robust multi-architecture aggregation. This approach improves diagnostic accuracy while reducing membership-inference attack success from ∼90% to ∼75–80%, though computational overhead and residual privacy leakage persist. Building on this foundation, the second contribution integrates learnable encryption with Vision Transformers (ViTs), where medical data undergo block-wise permutation and pixel shuffling before training. Unlike CNNs that fail on scrambled data, ViTs’ global self-attention mechanism preserves discriminative features within encrypted images, substantially improving both accuracy and privacy protection. The final contribution eliminates gradient sharing entirely through lightweight [CLS] embedding aggregation, where clients transmit only compact 768-dimensional embeddings from locally trained ViTs, reducing communication overhead by 20×. These contributions collectively deliver a practical framework for privacy, accurate, and interpretable medical AI collaboration.

Download

Included in

Engineering Commons

Share

COinS