An Empirical Study of Federated Learning on IoT-Edge Devices: Resource Allocation and Heterogeneity

Area of Science:

• Computer Science
• Artificial Intelligence
• Distributed Systems

Background:

• Billions of IoT and edge devices generate vast data, driving ML applications.
• Privacy concerns and regulations limit centralized data collection for ML training.
• Federated learning (FL) offers a privacy-preserving alternative by training models on decentralized data.

Purpose of the Study:

• To empirically evaluate federated learning (FL) in realistic IoT-edge environments.
• To analyze FL's performance, computation, and communication costs.
• To investigate FL's effectiveness in challenging heterogeneous scenarios.

Main Methods:

• Extensive experiments conducted on a large network of IoT-edge devices.
• Systematic analysis of learning performance and operational costs.
• Focus on heterogeneous data distributions and device capabilities.

Main Results:

• FL demonstrates practical feasibility for on-device ML model training.
• Characterization of real-world FL performance and resource utilization.
• Insights into challenges and characteristics of FL in heterogeneous settings.

Conclusions:

• This study provides valuable insights into the practical implementation of FL.
• Findings promote the advancement and design of real-world FL systems.
• Highlights the importance of addressing heterogeneity for FL success.