Back to search

University of South Florida

Analyzing and Extending Machine Learning Frameworks on High Risk Domains

Abstract

dc:description

Machine learning (ML) has become a transformative force in high-risk domains such as genomics and cybersecurity, where accurate predictions and robust defenses are essential. This dissertation advances ML frameworks in these areas by developing methods to enhance predictive power in health applications and assess vulnerabilities in machine learning systems. In the genomics field, the work addresses challenges in Non-Invasive Prenatal Testing (NIPT) of monogenic disorders by proposing a deep learning model that reconstructs the fetal genome using maternal plasma cell-free DNA (cfDNA) and parental whole-genome sequencing (WGS) data. This model achieves high accuracy in single nucleotide variation (SNV) prediction, surpassing current methods with an overall accuracy of 89.3% ± 1.8%. Next, we introduce GenomGPT, a large language model specifically adapted for precision classification in metagenomic DNA sequences. This model leverages advanced tokenization and attention mechanisms, achieving a 94.7% classification accuracy in mixed-species samples, significantly enhancing species differentiation in complex genomic analyses and applications such as environmental monitoring and forensic analysis. In this study, we also investigated the predictive potential of cell-free DNA (cfDNA) as a biomarker for Gestational Diabetes Mellitus (GDM) by analyzing cfDNA profiles across pregnancy trimesters in two independent cohorts. By focusing on cfDNA dynamics, we developed a model capable of predicting GDM in the first trimester, achieving an area under the curve (AUC) of 0.919. This high predictive accuracy underscores the value of cfDNA as an early biomarker for GDM, offering a non-invasive approach to risk stratification. Our results suggest that cfDNA profiling could facilitate timely intervention and personalized treatment strategies, ultimately improving outcomes for both mother and child. In cybersecurity, this dissertation focuses on the vulnerabilities of GPT-based classifiers to adversarial attacks, a critical concern for ML in high-stakes applications. We propose the Stepping Autobreaker, a non-targeted adversarial attack that induces minimal, imperceptible perturbations to mislead GPT-based classifiers with over 95% attack successful rate. This method, evaluated across image classification tasks, reveals the susceptibility of Transformer-based models to adversarial manipulations and underscores the need for robust defenses in real-time applications. This dissertation contributes novel ML methodologies for genomic prediction and security analysis in high-risk domains. It emphasizes the importance of accuracy and robustness in health and security applications, offering insights for more secure and effective ML frameworks across critical domains.

Degree

thesis:*
Name thesis:degree_name
Doctor of Philosophy (Ph.D.)
Level thesis:degree_level
Dissertation
Discipline thesis:degree_discipline
Computer Science and Engineering
Grantor
University of South Florida
Year dc:date
2024

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Hu, Chengbin
Contributors dc:contributor
  • Yicheng Y. Tu, Ph.D.
  • Kaiqi K. Xiong, Ph.D.
  • Ankur A. Mali, Ph.D.
  • Seungbae S. Kim, Ph.D.
  • Hadi H. Gard, Ph.D.

Subjects

dc:subject × 6

Identifiers

dc:identifier.*
Repository record dc:identifier
https://digitalcommons.usf.edu/etd/10633

Chain of custody

source
Harvested from
University of South Florida
Base URL
digitalcommons.usf.edu/do/oai/
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Hu, Chengbin. Analyzing and Extending Machine Learning Frameworks on High Risk Domains. Dissertation thesis, University of South Florida, 2024. https://digitalcommons.usf.edu/etd/10633