AI & IoT Joint Seminar | Rigorous Approaches to Data Privacy...

Protecting user privacy is essential, yet traditional anonymization frequently fails against modern attacks. This talk presents a two-pronged research agenda for rigorous Differential Privacy (DP): algorithmic synthesis and principled auditing. First, I address the challenge of applying DP to relational data in Graph Neural Networks (GNNs). I introduce a node-level private learning algorithm that utilizes subgraph sampling to manage node dependencies, achieving significantly higher accuracy than prior methods. Second, to ensure the falsifiability of privacy claims, I propose an auditing framework that models privacy leakage as a "bits transmission" problem. By leveraging mutual information, this approach delivers tighter privacy loss estimates and improved efficiency compared to classical statistical audits. I conclude with future directions for scaling privacy protections to large AI models.

Zihang Xiang is a Postdoctoral Researcher in Electrical and Computer Engineering at UCLA, advised by Prof. Yuan Tian. He earned his Ph.D. in Computer Science from KAUST under the supervision of Prof. Di Wang.

His research advances principled data privacy by designing rigorous differentially private algorithms for complex machine learning cases and developing auditing frameworks to verify privacy claims. His work appears in top venues including IEEE S&P, USENIX Security, NDSS, SIGMOD.