Dual-channel meta-federated graph learning with robust aggregation and privacy enhancement

Graph neural networks (GNNs) are effective for graph-based node classification tasks, such as data mining and recommendation systems. Combining federated learning(FL) with GNN enables multiple participants to collaboratively train powerful models without sharing private data. However, subgraph-level FL faces challenges, including missing cross-client edges and non-IID data distributions. Additionally, ensuring security in non-fully trusted environments is a critical concern. To address these issues, we propose RMFGL (Robust Meta Federated Graph Learning), a framework for subgraph-level node classification. RMFGL integrates cross-client information through pre-feature aggregation and leverages model-agnostic meta-learning (MAML) to optimize meta-parameters with minimal federated updates. For robustness, we employ a GCN architecture with dual-channel attention aggregation, while Multi-key Fully Homomorphic Encryption (MKFHE) ensures privacy during training. Experimental results on Cora, CiteSeer, PubMed and Coauthor-CS datasets show that RMFGL achieves up to a 2x accuracy improvement with minimal fine-tuning compared to baseline methods and outperforms state-of-the-art techniques. Notably, RMFGL significantly enhances robustness against malicious clients, with up to 100x improvement in stability, while maintaining strong performance with non-IID data.