Evaluating Federated Learning Simulators: A Comparative Analysis of Horizontal and Vertical Approaches

Area of Science:

• Machine Learning
• Decentralized Systems
• Data Privacy

Background:

• Federated learning (FL) trains models on local devices, avoiding centralized data collection for enhanced privacy and security.
• FL relies on effective communication and collaboration between devices for scalability and robustness.
• Existing FL frameworks require evaluation for diverse topologies and approaches.

Purpose of the Study:

• To explore and evaluate various decentralized and centralized topologies in federated learning.
• To compare the performance of four major end-to-end FL frameworks: FedML, Flower, Flute, and PySyft.
• To investigate the combination of horizontal and vertical FL for common challenges like synchronization and communication overhead.

Main Methods:

• Focused on horizontal and vertical federated learning systems.
• Utilized a logistic regression model aggregated by the FedAvg algorithm.
• Conducted experiments on MNIST and Fashion-MNIST image datasets to assess framework efficiency and performance.

Main Results:

• Evaluated the efficiency and performance of FedML, Flower, Flute, and PySyft frameworks.
• Identified trade-offs in performance across different federated learning simulation frameworks.
• Provided initial findings on combining horizontal and vertical FL strategies.

Conclusions:

• Federated learning offers a privacy-preserving alternative to centralized machine learning.
• Framework selection involves performance and communication trade-offs, especially when combining FL approaches.
• Further research can optimize FL strategies for practical, large-scale deployments.