Peer-to-Peer Federated Learning on Software-Defined Optical Access Network

Abstract

Recent advancements in AI have significantly propelled machine learning (ML) as a cornerstone of modern technological infrastructure, especially in enhancing machine-to-machine (M2M) communication. This necessitates the development of custom, advanced passive optical network (PON) devices tailored to sophisticated functionalities, heralding a shift towards decentralized data processing that aligns with global data protection and privacy standards. This paper introduces a novel software-defined architecture for TWDM-PON, designed to support Federated Learning (FL) traffic in a peer-to-peer (P2P) manner. Our proposed architecture integrates an advanced P2P-FL-OLT with an enhanced P2P-FL-DBA and P2P-FL-ONU with cache storage and a dedicated P2P transceiver, optimizing network resources and enhancing performance. Extensive simulations demonstrated that this architecture effectively manages increased P2PFL traffic, substantially improving system performance metrics such as throughput, packet delay, jitter, and packet loss ratio. The results affirm that our design not only meets but exceeds Quality of Service (QoS) requirements, making it a viable solution for future fiber optic networks facilitating distributed M2M machine learning applications.

Keywords

• Machine learning
• machine-to-machine communication
• software-defined networking
• TWDM-PON
• federated learning
• dynamic bandwidth allocation