Theses and Dissertations

ORCID

0000-0002-5264-1382

Advisor

Perkins, Andy

Committee Member

Chen, Zhiqian

Committee Member

Pillai, Nisha

Committee Member

Nanduri, Bindu

Date of Degree

8-12-2026

Original embargo terms

Immediate Worldwide Access

Document Type

Dissertation - Open Access

Major

Computer Science

Degree Name

Doctor of Philosophy (Ph.D.)

College

James Worth Bagley College of Engineering

Department

Department of Computer Science and Engineering

Abstract

Genomic data contains highly sensitive personal information, yet cancer classification and collaborative medical research increasingly depend on sharing such data across institutions. Existing privacy-preserving approaches for deep learning, including differential privacy (DP) and standard federated learning (FL), leave critical vulnerabilities: DP degrades model utility through noise injection, while FL relies on a trusted aggregation server that remains susceptible to gradient-based attacks. This dissertation develops, evaluates, and optimizes a series of privacy-preserving deep learning frameworks specifically designed for genomic cancer classification. First, a systematic benchmarking study evaluates DP, FL, and fully homomorphic encryption (FHE) across three deep learning architectures (fully connected neural networks, convolutional neural networks, and Transformer models), both individually and in combination, using high-dimensional RNA-seq gene expression data across 29 to 37 cancer types from The Cancer Genome Atlas. This is the most comprehensive comparison of privacy-preserving genomic deep learning methods to date. Second, this dissertation introduces the FL-MP-CKKS-FHE framework, which, to our knowledge, is the first to combine multi-party CKKS fully homomorphic encryption with federated learning for deep learning model training. Unlike standard FL, this framework encrypts all gradient parameters before transmission using a threshold decryption mechanism, ensuring no single party, including the aggregation server, can access plaintext gradients. Third, three algorithmic optimization strategies are developed and evaluated: adaptive gradient scaling, single instruction multiple data (SIMD)-optimized ciphertext packing, and Top-K gradient compression. SIMD optimization achieved a 3.6x increase in training speed while improving classification accuracy. Top-K 10 percent compression reduced per-round training time by 84.77 percent while retaining 95.84 percent of baseline accuracy. Fourth, this study establishes a rigorous quantitative method using distribution-based analysis and four image similarity metrics to evaluate resistance to model inversion attacks. Finally, a configurable privacy-preserving platform is developed to automate DP and FL integration into existing deep learning models. Together, these contributions advance the feasibility of secure, privacy-preserving collaborative genomic intelligence.

Download

Share

COinS