Visiting Researcher Talk: Dr Xiaoning Liu| 9 Febuary 2026

Talk Title:
Oblivious Inference on Graph Neutral Network 

Speaker:

Dr xiaoning liu

About the Speaker: 

Dr Xiaoning Liu is an ARC DECRA Fellow and a Senior Lecturer at the School of Computing Technologies, RMIT University, Australia. Her research pivots on data privacy and security related to machine learning, cloud computing, and digital health. Her current focus is on the on the design of practical secure multiparty computation protocols with applications in privacy-preserving machine learning. In the past few years, her work has appeared in prestigious venues in computer security, such as USENIX Security, NDSS, and ESORICS, IEEE Transactions on Dependable and Secure Computing (TDSC), IEEE Transactions on Information Forensics and Security (TIFS). She is the recipient of the Best Paper Award of ESORICS 2021 and the RMIT HDR Research Prize 2023. She has served on the technical program committee of USENIX Security, EuroS&P, and the program co-chair of LAMPS at CCS 2025. Her research has been supported by Australian Research Council and CSIRO.

Description: 

Graph Neural Networks (GNNs) extend the benefits of deep learning to graph data. In practice, their applications span from common utilities such as recommendation systems and fraud detection, to advanced domains such as drug discovery and physics simulation. Due to the increasing popularity of GNNs, commercial Machine Learning as a Service (MLaaS) platforms have integrated graph learning development tools for launching GNN services on the cloud, e.g., AWS integrated DGL. Despite the convenience and low cost of model development and deployment, such graph-based MLaaS is also facing critical security challenges. In particular, graph data contains sensitive information and GNN models are considered as IP of model owners, which demand strong protection when being deployed in MLaaS. In this talk, I will present our recent work named OblivGNN. It is designed to efficiently enable encrypted inference over encrypted graph data and GNN models. Specifically, it is the first to support both transductive and inductive settings of GNNs while ensuring confidentiality and data obliviousness of both graphs and models. Along the line, I will also pinpoint open problems and future directions in this area.