A two-phase semantic fusion framework for privacy-preserving federated heterogeneous graph learning

Graph-structured data plays a critical role in high-stakes applications such as financial fraud detection and medical diagnosis. However, the non-Euclidean topology and inherent privacy sensitivity of graph data make it difficult for traditional federated learning (FL) to effectively balance model performance and privacy protection. We present FedHGPP (Federated Heterogeneous Graph Privacy Protection), a two-phase semantic fusion framework that enables privacy-preserving collaborative graph learning across institutional data silos. The FedHGPP framework employs a two-phase differential privacy perturbation mechanism (DPRR+EM) in the local preprocessing step, injecting calibrated noise into high-order structural statistics of the original graph. We allocate the total privacy budget strictly following the sequential composition theorem of differential privacy, ensuring rigorous edge-level $$\varepsilon $$-DP guarantees for the framework. During federated training, clients upload homomorphically encrypted node embeddings, while the server performs secure cross-graph attention aggregation in the ciphertext domain to prevent leakage of intermediate representations. Experiments on three real-world datasets show that under edge-level $$\varepsilon $$-DP constraints ($$\varepsilon $$=2.0), FedHGPP achieves an average HR@10 of 0.476 and NDCG@10 of 0.254. It outperforms state-of-the-art federated heterogeneous graph learning methods with formal privacy guarantees on all tested datasets, narrowing the average performance gap with the privacy-free centralized upper bound (Central-HAN) to below 13 % for NDCG@10 and within 10 % for HR@10.