FLGAT: A federated graph attention framework for privacy-preserving community detection

Community detection plays a vital role in analysing real-world networks. However, traditional centralized community detection algorithms often violate data privacy constraints particularly when applied to sensitive or distributed data sources where aggregating raw graph data is either impractical or strictly prohibited. To address this challenge, the paper proposes a privacy-preserving community detection framework that is Federated Graph Attention Framework (FLGAT). In this framework, the global graph is partitioned into edge-disjoint subgraphs with a shared node set, each assigned to a distinct client in a federated environment. This ensures that raw graph data remains local to each client by preserving data privacy. Within each client, node embeddings are learned using a self-supervised GAT model trained with a hybrid loss function that combines contrastive InfoNCE loss with a modularity aware component. This encourages the embeddings to capture both semantic similarity and structural community. These local models are then collaboratively integrated using the Network Average Degree (NAD) weighted aggregation strategy, effectively integrating diverse local models based on graph structure quality. The globally learned embeddings are then utilized to construct a similarity graph that captures the relational patterns between nodes across the distributed network. Finally, the Louvain algorithm is applied to this similarity graph to detect communities by optimizing modularity. The FLGAT framework is assessed using three LFR benchmark networks and four real-world networks. It consistently outperforms existing approaches across key metrics modularity (Q), Normalized Mutual Information (NMI), and Adjusted Rand Index (ARI) while strictly adhering to privacy-preserving principles.