Context-Aware Online Model Splitting and Device Association for Semi-Decentralized Federated Learning in Internet of...

Area of Science:

• Artificial Intelligence
• Wireless Networks
• Machine Learning

Background:

• Federated learning (FL) enables privacy-preserving AI model construction but increases local computing load.
• Split learning (SL) adapts to device capabilities but can be limited by single central servers.
• Existing methods face challenges with distributed AI training efficiency and resource management.

Purpose of the Study:

• To propose a semi-decentralized hybrid federated split learning (SD-HFSL) framework for efficient AI model training.
• To address limitations of single central servers by aggregating split models across multiple edge servers.
• To optimize training efficiency by considering computation, communication resources, and latency.

Main Methods:

• Developed a semi-decentralized hybrid federated split learning (SD-HFSL) framework.
• Formulated an online optimization problem for local model splitting and device association.
• Proposed a context-aware online training algorithm using contextual multi-armed bandits (CMAB) for latency estimation and iterative optimization.

Main Results:

• The SD-HFSL framework demonstrated improved convergence performance under limited resources.
• The CMAB-based algorithm effectively estimated latency and optimized training.
• Experimental results showed reduced training latency and enhanced test accuracy compared to benchmarks.

Conclusions:

• The proposed SD-HFSL framework and CMAB algorithm offer an effective solution for efficient, privacy-preserving AI training in wireless networks.
• Latency optimization is crucial for improving training efficiency in distributed learning environments.
• The approach successfully balances computational and communication constraints for better AI model performance.