FedGSL: Federated Graph Structure Learning for Local Subgraph Augmentation

Geng Zhao, Yi Huang, Chi Hsu Tsai

Abstract: Federated Graph Neural Networks have recently attracted considerable attention for its application of graph-structured data distributed in different organizations in joint training of GNNs. Current FedGNN methods mainly focus on aggregating or sharing node features and local subgraphs held by each client under privacy protection. However, the critical premise that the graph structure held by the client is reliable is commonly not met. This is because, under partial semantics, an isolated client may only witness an insufficient amount of structural interactions, which is clearly suboptimal. The performance of FedGNN can be enhanced if the observed structure can gather additional structural data from the features or interactions of additional clients. The optimized structure will approximate the optimal graph structure. Inspired by this, we propose Federated Graph Structure Learning (FedGSL). FedGSL acquires multiple encrypted views and pseudo-views for each client while hiding features and the true graph, which we anonymously propagate and integrate into the learned graph structure. We concurrently aggregate the interactions of multi-hop structures to supplement the segmented topology for each individual semantic. Each client updates parameters in parallel after training local data, server-distributed desensitized and shared structures, and individualized embeddings. Extensive experimental findings reveal that FedGSL outperforms selected baselines on numerous widely used raw graph datasets and greatly optimizes the client’s local subgraph.