A comprehensive experimental comparison between federated and centralized learning.

Area of Science:

• Machine Learning
• Data Science
• Computational Science

Background:

• Federated learning enables collaborative model training without centralizing sensitive data, crucial for domains like medical research facing privacy and legal hurdles.
• Existing comparisons between federated and centralized learning are largely theoretical, lacking extensive experimental validation of their performance and learning behaviors.

Purpose of the Study:

• To conduct a comprehensive experimental comparison between federated learning and centralized learning strategies.
• To evaluate the performance and learning behavior of various classifiers across diverse datasets and data distributions.
• To explore the impact of sample and class distribution imbalances on federated learning effectiveness.

Main Methods:

• Experimental evaluation of multiple classifiers using various datasets.
• Analysis of performance under different sample and class distributions across clients.
• Assessment of federated learning's robustness to data imbalances, dimensionality, and model complexity.

Main Results:

• Federated learning demonstrates comparable performance to centralized learning across a wide range of settings.
• Federated learning effectively handles various data imbalances, including skewed distributions and multiclass problems.
• Federated learning shows robustness to high data dimensionality and complex models, though sensitive to location-based batch effects.

Conclusions:

• Federated learning presents a promising alternative to data sharing, offering similar performance to centralized approaches while preserving data privacy.
• The experimental findings support the application of federated learning in scenarios where data centralization is challenging or prohibited.
• Federated learning's resilience to data heterogeneity and complexity makes it suitable for real-world applications with distributed datasets.