Optimized IoT clustering and assignment in semi-synchronous federated learning

Area of Science:

• Edge Computing
• Federated Learning
• Machine Learning Optimization

Background:

• Current hierarchical semi-synchronous federated learning methods lack effective device clustering and data redistribution strategies.
• Suboptimal resource allocation and data processing lead to reduced model accuracy and system scalability in edge computing.
• Existing approaches fail to integrate model accuracy with efficient device grouping, necessitating innovative solutions.

Purpose of the Study:

• To enhance the performance and scalability of federated learning systems within edge computing environments.
• To optimize device clustering, server assignment, and data redistribution for hierarchical semi-synchronous federated learning.
• To improve both model training accuracy and resource utilization through advanced algorithms.

Main Methods:

• Utilized Graph Neural Network (GNN) for device grouping based on hardware and local datasets.
• Applied K-means algorithm for efficient device cluster formation.
• Implemented Hybrid Data Redistribution to equalize datasets within clusters.
• Employed Proximal Policy algorithm for resource allocation optimization based on real-time bandwidth and energy consumption.

Main Results:

• Achieved a 15% improvement in clustering metrics compared to existing algorithms.
• Demonstrated enhanced device assignment and data redistribution capabilities.
• Successfully addressed limitations in model accuracy and resource optimization within federated learning frameworks.

Conclusions:

• The proposed method significantly improves device clustering and data redistribution in hierarchical semi-synchronous federated learning.
• The integration of GNN, K-means, Hybrid Data Redistribution, and Proximal Policy enhances model accuracy and resource efficiency.
• This research provides a robust framework for advancing edge computing through optimized federated learning.