Securing Graph Neural Networks in Machine Learning as a Service

This thesis focuses on the security issues associated with integrating Graph Neural Networks (GNNs) into Machine Learning as a Service (MLaaS) provided by cloud services. It explores the practical threats posed by MLaaS, revealing that prediction APIs can inadvertently disclose sensitive details like GNN model parameters and training graphs during inference. The thesis also examines the risks of serving GNNs in the cloud. To counter these threats, the thesis proposes a method for validating GNN model integrity within MLaaS. Additionally, it addresses situations where unauthorized graph data is used during local training of GNNs deployed in the cloud.