Abstract: As Graph Neural Networks (GNNs) increasingly analyse sensitive relational data, protecting the privacy of underlying nodes and edges has become a critical challenge, often characterized by trade-offs between privacy guarantees, model utility, and computational efficiency. In this talk, I will present two independent yet foundational pillars for building trustworthy graph learning systems, bridging the gap from local privacy foundations to collaborative secure environments. First, our research addresses the challenge of "noise explosion" in deep architectures through CARIBOU, a convergent privacy framework for multi-layer GNNs. By developing contractive message passing, CARIBOU ensures that private GNNs maintain high utility under standard Differential Privacy (DP) guarantees, even as model depth increases. Second, we transition to the challenges of secure collaboration with GRACE, an efficient protocol suite for GNN training on vertically split data. By leveraging Secure Multi-party Computation (MPC) and sparse matrix decomposition, GRACE significantly reduces the communication overhead by 62%-78% associated with full graph protection. Finally, I will discuss how these advancements scale to real-world security analytics and outline future directions for securing emerging graph applications.
Bio: Yu (Jennie) Zheng is a Postdoctoral Researcher in the Department of Electrical Engineering & Computer Science at the University of California, Irvine (UCI). Before that, she earned her Ph.D. in Information Engineering from The Chinese University of Hong Kong (CUHK) in 2024. Her research focuses on the intersection of privacy-enhancing technologies and artificial intelligence, with an emphasis on developing efficient solutions for privacy-preserving AI and secure computation.
Abstract: As Graph Neural Networks (GNNs) increasingly analyse sensitive relational data, protecting the privacy of underlying nodes and edges has become a critical challenge, often characterized by trade-offs between privacy guarantees, model utility, and computational efficiency. In this talk, I will present two independent yet foundational pillars for building trustworthy graph learning systems, bridging the gap from local privacy foundations to collaborative secure environments. First, our research addresses the challenge of "noise explosion" in deep architectures through CARIBOU, a convergent privacy framework for multi-layer GNNs. By developing contractive message passing, CARIBOU ensures that private GNNs maintain high utility under standard Differential Privacy (DP) guarantees, even as model depth increases. Second, we transition to the challenges of secure collaboration with GRACE, an efficient protocol suite for GNN training on vertically split data. By leveraging Secure Multi-party Computation (MPC) and sparse matrix decomposition, GRACE significantly reduces the communication overhead by 62%-78% associated with full graph protection. Finally, I will discuss how these advancements scale to real-world security analytics and outline future directions for securing emerging graph applications.
Bio: Yu (Jennie) Zheng is a Postdoctoral Researcher in the Department of Electrical Engineering & Computer Science at the University of California, Irvine (UCI). Before that, she earned her Ph.D. in Information Engineering from The Chinese University of Hong Kong (CUHK) in 2024. Her research focuses on the intersection of privacy-enhancing technologies and artificial intelligence, with an emphasis on developing efficient solutions for privacy-preserving AI and secure computation.