Provably Communication-Efficient and Privacy-Preserving Federated Graph Neural Networks

In three linesCE-FedGNN is a federated framework for graph neural networks that reduces communication by infrequently exchanging aggregated node representations instead of per-round embeddings. A moving-average estimator handles cross-client dependencies and staleness. The framework provides privacy guarantees via metric-DP and achieves O(1/√T) convergence with O(T^3/4) communication complexity.

Summary generated by Claude — human-verified