Here are the details of the talk: Dr. Junyuan Hong from UT Austin will be presenting an online guest lecture in our class today. The details are below. Dr. Junyuan Hong from UT Austin Title: Foundation Models Meet Data Privacy: Risks and Countermeasures Abstract: Foundation models have become more and more important for Artificial Intelligence but also pose risks when they are fostered from massive data. In training large models, personal information in data could be encoded and raises concerns in data privacy. Therefore, how to measure the risks and design countermeasures is a crucial topic and gains more attention especially with the popularity of Generative AI. In this lecture, I will present three of our recent efforts towards answering the question. First, we developed a general method to measure the privacy risks in model gradients efficiently and effectively. Second, we presented a new security risk that finetuning can seduce text-to-image models to leak more training data. Third, we novelly made Large Language Models privacy-preserving prompt engineer for close-source ones (e.g., ChatGPT) that were not born to protect data privacy. Bio: Junyuan is a postdoctoral fellow in the Institute for Foundations of Machine Learning (IFML) at UT Austin. He obtained Ph.D. degree from Michigan State University and M.S. and B.S. from University of Science and Technology of China. His research interests lie in trustworthy AI, data privacy and AI for healthcare. He has accomplished a series of work that enhanced the fairness, robustness, security and privacy of federated learning. He is also a core member in ILLIDAN Lab team that won the 3rd place in U.S. PETs prize challenge and was covered by the White House. Recently, his researche is centered on the Trustworthy Generative AI with manageable costs and its application in modeling and treating dementia disease. Time: 11AM ET Zoom link: https://virginiatech.zoom.us/j/86594075180?pwd=NDNJUExiS1RUOWpHV0dsUDFlRWFKQ... Meeting ID: 865 9407 5180 Passcode: 010981 --