February 12, 2026 4:00pm - 5:00pm in Math Rm. 501
When
4 – 5 p.m., Feb. 12, 2026
Where
Title: Towards Trustworthy AI: Theoretical Foundations and Principled Algorithms
Abstract: Artificial intelligence (AI) is now deeply embedded in critical domains, making it essential to understand and ensure its safety. My research focuses generally on building trustworthy AI by strengthening model security, protecting data privacy, improving fairness and explainability, and ensuring reliable model evaluation. In this talk, I will present Model Privacy, the first statistical framework that rigorously characterizes model stealing attacks and defenses. This framework provides fundamental insights into how and why defenses succeed or fail, revealing the importance of breaking data independence in order to effectively protect model privacy. Beyond model stealing, our framework also offers a unified perspective for analyzing a broad class of teacher-student-style learning paradigms, opening doors to addressing a deeper question of how well models can learn from arbitrarily contaminated data.