Federated Spectral Graph Neural Networks

Federated graph neural networks aim to collabo-ratively train a global graph neural network across multiple clients while preserving privacy constraints. Existing federated graph neural networks fall into the category of federated learning on spatial graph neural networks which rarely consider the challenges of data noise and biased data distribution. To this end, we present a new class of Federated Spectral Graph Neural Networks (FedSGNN for short) to enable federated learning on spectral graph neural networks which can better handle the data noise and biased data distribution problem. Under the obser-vation that graph convolutions are essentially low-pass filters which have limited capability to handle noisy data, FedSGNN uses Bernstein polynomial approximation as the polynomial filter function to capture useful information from multiple frequency domains to combat data noise. Experimental results show that the first-order Bernstein polynomial approximation meets the requirement of federated graph learning with noise. Moreover, experiments on four benchmark datasets show superior results compared to existing federated graph neural networks.