Scalable Federated Learning with Differential Privacy and Graph-Based Communication

Abstract

Federated learning (FL) offers a promising paradigm for training machine learning models on decentralized data sources without directly exchanging the data itself. However, traditional FL approaches often suffer from significant communication overhead and potential privacy vulnerabilities. This research proposes a novel scalable federated learning system that addresses these limitations through the integration of graph-based communication topologies and differential privacy mechanisms. The core idea is to represent the relationships between participating clients as a graph, enabling selective communication and aggregation of model updates based on the graph structure, thereby reducing unnecessary communication and enhancing privacy. The system utilizes a graph convolutional network (GCN) to learn client embeddings reflecting their connectivity and data characteristics, which then guide the communication process. Differential privacy is incorporated by adding calibrated noise to the aggregated model updates, providing a provable privacy guarantee. Simulation results demonstrate that the proposed approach achieves significant improvements in communication efficiency and privacy compared to standard FL methods, particularly in scenarios with complex client relationships and stringent privacy requirements. This work contributes a practical and scalable solution for deploying federated learning in diverse applications where data privacy and communication efficiency are paramount.