Artificial Intelligence has become deeply integrated into diverse systems, transforming industries and reshaping our daily lives. However, this widespread adoption also introduces critical privacy risks across the training data, AI models, and AI-powered systems. This talk will explore privacy challenges through these three aspects. First, I will introduce the first high-fidelity attack that exposes the privacy vulnerabilities of training data in pre-trained models and commercial AI services. Next, I will present a novel physical impersonating attack that highlights the privacy risks inherent in AI-based authentication systems. Additionally, I will discuss the first data-free framework designed to eliminate trigger-based model watermarks in diffusion models that aim to protect their intellectual property. Finally, I will conclude with a forward-looking perspective on addressing privacy risks in emerging generative AI techniques, such as Large Language Models and Stable Diffusion Models.

---

UMBC Center for AI