Privacy is a key challenge in modern learning systems that increasingly rely on sensitive data. Differential privacy (DP) provides rigorous protection for individual records, but its guarantees do not always match what is actually sensitive in a learning task. In this talk, I will present a line of work on fine-grained privacy, with guarantees designed around the information that truly needs protection. I will discuss how to strengthen privacy when sensitive information spans multiple records, and how to avoid unnecessary utility loss when only part of the data is sensitive. This perspective leads to new privacy notions and practical learning algorithms with formal guarantees, including privacy-preserving methods for modern LLM alignment.
Speaker Info:
Yangfan Jiang is a Ph.D. candidate in Computer Science at the National University of Singapore, advised by Prof. Xiaokui Xiao. His research focuses on privacy and security for modern AI over sensitive data, including algorithms that enable large language models and data analytics methods to learn from private data with formal privacy and security guarantees.