Dynamic Bandwidth Slicing in Passive Optical Networks to Empower Federated Learning.

Area of Science:

• Telecommunications Engineering
• Machine Learning
• Network Management

Background:

• Federated Learning (FL) offers decentralized machine learning by training models locally on devices.
• Passive Optical Networks (PON) are key infrastructure for high-speed communication.
• Integrating FL with PON presents opportunities for 6G but requires efficient bandwidth management for FL traffic.

Purpose of the Study:

• To explore the integration of Federated Learning (FL) with Passive Optical Networks (PON) for 6G environments.
• To address the challenge of managing bandwidth for complex FL traffic within PON.
• To introduce and evaluate a novel Dynamic Bandwidth Allocation (DBA) approach for FL traffic in PON.

Main Methods:

• Developed a novel Dynamic Bandwidth Allocation (DBA) algorithm tailored for Federated Learning (FL) traffic.
• Simulated the proposed DBA approach within a PON framework to analyze its performance.
• Compared the novel DBA approach against state-of-the-art solutions using key network performance metrics.

Main Results:

• The proposed DBA approach efficiently allocates PON bandwidth for FL traffic generation.
• Demonstrated the benefits of utilizing multiple upstream grant allocations for FL flows in PON.
• Simulations showed superior performance compared to existing solutions, particularly in reducing upstream delay.

Conclusions:

• The novel DBA approach effectively enhances bandwidth management for FL traffic in PON.
• The integration of FL and PON, supported by efficient bandwidth allocation, is a promising enabler for 6G services.
• This research paves the way for real-time, data-intensive services crucial for future 6G networks.