Solitude: A Decentralized Graph Privacy-Preserving Learning Framework Based on Local Differential Privacy

The researchers propose Solitude, a privacy-preserving learning framework based on graph neural networks (GNNs), providing formal privacy guarantees via edge local differential privacy. The framework calibrates the noise introduced into the decentralized graph by leveraging properties such as the sparsity common to real graphs, can protect the privacy of both node features and edges, and can be combined with any GNN. Experiments on a large number of benchmark datasets show that, under a given privacy budget, Solitude protects user data privacy while preserving the generalization ability of GNNs.