Graph neural networks: a survey on the links between privacy and security

Abstract

AbstractGraph neural networks (GNNs) are models that capture the dependencies between graph data by passing messages between graph nodes and they have been widely used to process graph data that contains relational information. Example application areas include social networks, recommendation systems, and life sciences. However, like all neural networks, there are underpinning security and privacy concerns associated with GNN deployments in practice. For example, attackers can perturb a graph’s data to undermine a model’s effectiveness, or they can steal the model’s data and/or parameters, thus threatening the privacy of the model. In this survey, we provide a comprehensive review of recent research efforts on security and/or privacy in GNNs. We also systematically describe the distinctions and relationships between security and privacy, as well as providing an outlook on future directions of research in this area.