Graduation Year

2024

Document Type

Dissertation

Degree

Ph.D.

Degree Name

Doctor of Philosophy (Ph.D.)

Degree Granting Department

Industrial and Management Systems Engineering

Major Professor

Trung Le, Ph.D.

Committee Member

Tapas K. Das, Ph.D.

Committee Member

Ankit Shah, Ph.D.

Committee Member

John Licato, Ph.D.

Committee Member

Jun Kong, Ph.D.

Keywords

Data-centric AI, Federated Learning, Generative AI, Robust AI

Abstract

Artificial Intelligence (AI) systems have demonstrated remarkable performance across various domains. However, their robustness remains a critical concern, particularly in terms of data and model reliability. This dissertation aims to address the challenges associated with building robust AI systems by focusing on two key aspects: data robustness and model robustness. Data robustness poses significant challenges, including data shift, concept shifting, limited and imbalanced datasets, and interoperability issues in IoT systems for data collection. Existing methods fall short in handling dynamic business objectives and evolving data landscapes effectively. To bridge these gaps, we propose an IoT framework that ensures interoperability, seamless communication, and scalability for efficient data collection. Furthermore, we employ data-centric approaches and generative AI techniques to enrich data quantity and quality, enabling the AI system to adapt to data shift and concept shifting. Model robustness is hindered by noisy data, which impedes the model's ability to generalize well to unseen examples. Additionally, AI models are susceptible to adversarial attacks designed to deceive the system. Prior research has not adequately explored the potential of data-centric approaches to enhance model robustness and lacks methods that simultaneously address data security and model generalization. To overcome these challenges, we propose a hybrid approach that combines data-centric and model-centric techniques to improve model generalization and resilience to noisy data. Moreover, we leverage federated learning to decentralize model training, enhancing data security and privacy while harnessing diverse data sources. By addressing the identified research gaps and implementing the proposed solutions, this dissertation contributes to the development of robust AI systems that can effectively handle data and model challenges.

Download

Share

COinS