OblivGNN: Oblivious Inference on Transductive and Inductive Graph Neural Networks

Watch a 13-minute conference presentation from USENIX Security '24 exploring OblivGNN, a comprehensive solution for secure Graph Neural Network (GNN) inference in Machine Learning as a Service. Learn how researchers from Monash University, CSIRO's Data61, RMIT University, and The University of Melbourne developed a system supporting both transductive and inductive inference services while protecting sensitive graph data and model intellectual property. Discover how function secret sharing enables low communication and computation overhead, and understand the novel secure update protocol for inductive settings that updates graphs without revealing modified sections. Examine experimental results using Cora, Citeseer, and Pubmed datasets that demonstrate comparable accuracy to Additive Secret Sharing baselines while achieving up to 38% faster runtime and 10x to 151x reduced communication costs.