Federated graph neural network

Abstract

Graph Neural Networks is a form of machine learning that has seen significant growth in popularity and use, owing to their natural affinity for capturing implicit representations that exist in real-world phenomena. Many of these real-world phenomena involve people-centric data, which are privacy-sensitive. Because of this, there are growing privacy concerns pertaining to the use of machine learning for privacy-sensitive data, resulting in regulations that discourage or even prevent centralized collection of people-centric data. In this project, we implement and introduce a possible alternative means of conducting Graph Neural Network machine learning on privacy-sensitive data by combining a form of de-centralized, privacy-preserving machine learning known as Federated Learning with Graph Neural Networks. Our approach is showcased through the augmentation of the GCN and GraphSAGE GNNs with FL. These augmented FL-GNN models are able perform privacy-preserving de-centralized learning through a server-client architecture that does not require the collection of user data to train a Graph Neural Network model.