Provably Communication-Efficient and Privacy-Preserving Federated Graph Neural Networks
Researchers have developed a new framework called CE-FedGNN for training graph neural networks (GNNs) on distributed datasets. This method addresses the challenges of privacy and communication costs associated with federated learning on relational data. CE-FedGNN minimizes the need to share raw data or node embeddings by infrequently exchanging aggregated representations, while also incorporating differential privacy to protect the released information. AI
IMPACT This research offers a more efficient and privacy-preserving method for training GNNs on distributed data, potentially enabling new applications where data cannot be centralized.