A Survey of Privacy-Preserving Techniques in Federated Graph Neural Networks
Lingyu Wang · Theory and Practice of Science and Technology · 2025
Federated Graph Neural Networks (FedGNNs) are increasingly adopted in high-sensitivity domains—such as finance, healthcare, and government—imposing stringent requirements on privacy, performance, and communication efficiency. This survey systematically reviews four core privacy-preserving techniques (homomorphic encryption, differential privacy, secure multi-party computation, and trusted execution environments) and introduces an evaluation framework that balances privacy, performance, and efficiency. Within this framework, we compare encryption-based aggregation, embedding perturbation, graph-structure anonymization, uncertainty-aware distillation, and decentralized optimization. To handle dynamic graphs and multimodal data, we propose a sliding-window temporal encoding and cross-modal alignment strategy, and design a spatiotemporal multimodal defense pipeline with adaptive aggregation and privacy-budget scheduling for robust, efficient protection. Finally, we identify future challenges—standardized evaluation, adversarial-defense benchmarks, and industrial deployment—and outline design principles to guide secure FedGNN implementations in sensitive environments.