FEDERATED GRAPH NEURAL NETWORKS FOR PRIVACY-PRESERVING DATA SCIENCE IN HETEROGENEOUS ENVIRONMENTS: A SYSTEMATIC REVIEW

The proliferation of distributed data across healthcare, finance, and smart city infrastructures has necessitated the development of collaborative learning paradigms that honor stringent privacy regulations. This systematic review explores the evolution of Federated Graph Neural Networks (Fed-GNN) as a robust solution for privacy-preserving data science in heterogeneous environments. We analyze recent advancements (2024–2026) in addressing "topology-aware" optimization and personalized aggregation for non-identically distributed (non-IID) graph data. This paper provides a comprehensive taxonomy of formal privacy mechanisms, specifically evaluating the trade-offs between Dynamic Adaptive Partitioned Homomorphic Encryption (DAPHE) and noise-injection strategies in differential privacy. By synthesizing benchmarks from high-impact frameworks such as FedGraphHE and FedDQ, we identify critical bottlenecks in communication efficiency and structural privacy. Finally, this review highlights significant research gaps in decentralized Byzantine resilience and proposes a two-phase roadmap for future research extensions in dynamic graph learning and lightweight edge-node deployment.