Collaborative Split Learning-Based Dynamic Bandwidth Allocation for 6G-Grade TDM-PON Systems

TL;DR

This paper introduces a new dynamic bandwidth allocation method for 6G networks using collaborative split learning to improve traffic prediction and reduce communication overhead.

Contribution

A decentralized collaborative split learning framework for dynamic bandwidth allocation in TDM-PON systems is proposed.

Findings

CSL-DBA achieves 99% traffic prediction accuracy across various traffic scenarios.

The method reduces communication overhead by approximately 60% compared to traditional federated learning.

It offers minimal inference latency and scalable learning performance for 6G-grade TDM-PON systems.

Abstract

Dynamic Bandwidth Allocation (DBA) techniques enable Time Division Multiplexing Passive Optical Network (TDM-PON) systems to efficiently manage upstream bandwidth by allowing the centralized Optical Line Terminal (OLT) to coordinate resource allocation among distributed Optical Network Units (ONUs). Conventional DBA techniques struggle to adapt to dynamic traffic conditions, resulting in suboptimal performance under varying load scenarios. This work suggests a Collaborative Split Learning-Based DBA (CSL-DBA) framework that utilizes the recently emerging Split Learning (SL) technique between the OLT and ONUs for the objective of optimizing predictive traffic adaptation and reducing communication overhead. Instead of requiring centralized learning at the OLT, the proposed approach decentralizes the process by enabling ONUs to perform local traffic analysis and transmit only model updates to the OLT. This cooperative strategy guarantees rapid responsiveness to fluctuating traffic conditions. We show by extensive simulations spanning several traffic scenarios, including low, fluctuating, and high traffic load conditions, that our proposed CSL-DBA achieves at least 99% traffic prediction accuracy, with minimal inference latency and scalable learning performance, and it reduces communication overhead by approximately 60% compared to traditional federated learning approaches, making it a strong candidate for next-generation 6G-grade TDM-PON systems.

Figures