FedLASE: Performance-balanced system-heterogeneous FL via layer-adaptive submodel extraction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Distributed Systems

Background:

• Federated Learning (FL) enables privacy-preserving distributed model training.
• System heterogeneity in edge devices poses challenges for global model deployment.
• Existing submodel extraction methods show performance discrepancies across varying resource levels.

Purpose of the Study:

• To introduce FedLASE, a Layer-Adaptive Submodel Extraction framework.
• To address performance imbalances in Federated Learning due to system heterogeneity.
• To improve convergence and overall system performance in heterogeneous FL environments.

Main Methods:

• Quantifying layer importance based on parameter significance.
• Hierarchically extracting critical parameters within each layer.
• Strictly adhering to client resource constraints during submodel extraction.

Main Results:

• FedLASE achieves balanced performance across heterogeneous FL clients.
• The framework demonstrates improved convergence rates compared to existing methods.
• Experimental results show FedLASE's superiority and robustness in diverse heterogeneous scenarios.

Conclusions:

• FedLASE effectively overcomes limitations of current submodel extraction techniques.
• The proposed layer-adaptive approach preserves structural integrity and enhances FL performance.
• FedLASE offers a robust solution for deploying FL in real-world heterogeneous systems.