Differentially Private Graph Neural Networks for Medical Population Graphs and The Impact of The Graph Structure

We initiate an empirical investigation of differentially private graph neural networks for medical population graphs. In this context, we examine privacy-utility trade-offs at different privacy levels on both real-world and synthetic datasets and perform auditing through membership inference attacks. Our findings highlight the potential and the challenges of this specific DP application area, which comes with an additional difficulty of graph structure construction that potentially complicates graph deep learning. We find evidence that the underlying graph structure constitutes a potential factor for larger performance gaps by showing a correlation between the degree of graph homophily and the accuracy of the trained model.