Keywords

Data Safety, Watermarking, Data Plagiarism, Reliability

Abstract

With the rapid advancement of artificial intelligence, its applications have become indispensable in our daily lives. Emerging advanced AI models, from general-purpose Large Language Models (LLMs) such as ChatGPT and Deepseek to domain-specific systems such as autopilot for autonomous vehicles, are changing the way we work and live by increasing efficiency and fostering innovation. However, these advancements also bring significant risks, as they may compromise data ownership or yield unreliable outcomes. In response, we focus on two critical areas: data safety and model reliability. To ensure data safety and promote lawful data processing, we investigate anti-neural watermarking techniques designed to detect unauthorized use of user data for training neural models. Additionally, we address the emerging threat of ``neural plagiarism'', whereby diffusion models can effortlessly replicate copyrighted content while bypassing protection methods such as trademarks, signatures, and even invisible watermarks. To enhance model reliability, we present robust techniques, such as adversarial mixup that can improve models' robustness, and we facilitate research on legal and regulatory compliance for autonomous driving algorithms in real-world applications. Extensive experiments show that our methods work effectively in real-world scenarios, supporting the research community in building safe and reliable AI applications.

Completion Date

2025

Semester

Spring

Committee Chair

Wang, Liqiang

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Department

Computer Science

Identifier

DP0029420

Document Type

Dissertation/Thesis

Campus Location

Orlando (Main) Campus

Download

Share

COinS