Optimizing Federated Learning Using Graph Neural Networks

Résumé fourni par la source

Federated Learning (FL) ensures data security and privacy, has become a viable decentralized method for training machine learning models across numerous clients without exchanging raw data. Conventional FL techniques, however, have drawbacks such expensive transmission, inconsistent client input, and ineffective model aggregation. To find these problems, the chapter proposes GNN-based optimization framework which improves the federated learning by detecting complications with client relationships. The methodology represents clients as graph nodes, with edges encoding pairwise similarities based on network proximity, model performance, and data distributions. During the model update phase, node embeddings that dynamically affect aggregation weights are learned using a Graph Neural Network (GNN). This technique mitigates the challenges posed by non-IID (non-independent and identically distributed) data in federated settings. This framework includes three primary mechanisms Graph Representation of clients, in which clients are showed as nodes in a graph with edges signifying model similarity. Graph convolutional networks is used to learn optimal weight assignments and refine global updates, which is represented as GNN-Based Aggregation and to transmit important chosen nodes which minimize the redundant updates can be optimized by Communication Optimization. The findings show the comparison of conventional FedAvg algorithm with FL techniques, the accuracy has increased 17%. Additionally it reduces communication overhead by 30% and increased the FLs scalability and efficiency by convergence value between 25–40%. This chapter presents the baselines in various scenarios and regulates the statistical variations effectively among clients. The evaluation results include GNNs with FL, which improves learning, performance, and resource efficiency.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.