hFedLAP: A Hybrid Federated Learning to Enhance Peer-to-Peer
Ismail M. Elshair, Tariq Jamil Saifullah Khanzada · Engineering Technology & Applied Science Research · 2024
The concept of Federated Learning (FL) is a branch of Machine Learning (ML) that enables localized training of models without transferring data from local devices to a central server. FL can be categorized into two main topologies: Aggregation Server Topology (AST) and Peer-to-Peer (P2P). While FL offers advantages in terms of data privacy and decentralization, it also exhibits certain limitations in efficiency and bottleneck. However, the P2P topology does not require a server and allows only for a small number of devices. To overcome these limitations, this study proposes a hybrid FL Aggregation of P2P (hFedLAP) that mitigates some of the limitations of AST by combining it with P2P. This fusion model helps to remove the bottleneck and combines the advantages of both topologies. In the proposed hFedLAP model, clients are organized into 49 groups, each consisting of 51 clients, including one in each group serving as a client and an admin node in a P2P setup. In these groups, communication is restricted to admin nodes, supporting a maximum of 2,495 devices. Platform accuracy is maintained by implementing measures to prevent new devices with inadequate accuracy levels from joining until they attain the minimum required accuracy. The experimental results of hFedLAP were compared with AST and P2P using the MNIST dataset, showing that hFedLAP outperformed AST and P2P, achieving remarkable accuracy and scalability, with accuracy levels reaching 98.81%.