Decentralized federated learning through proxy model sharing.

Area of Science:

• Computer Science
• Machine Learning
• Data Privacy

Background:

• Highly regulated institutions (finance, healthcare) face data sharing restrictions.
• Federated learning enables collaborative learning on decentralized data while protecting privacy.
• Existing federated learning methods often require centralized servers and struggle with model heterogeneity.

Purpose of the Study:

• To propose ProxyFL, a communication-efficient scheme for decentralized federated learning.
• To enable multi-institutional collaboration with enhanced data privacy.
• To address limitations of canonical federated learning, including model heterogeneity and communication overhead.

Main Methods:

• Introduced ProxyFL, a proxy-based federated learning scheme.
• Each participant maintains a private model and a publicly shared proxy model.
• Utilized proxy models for efficient, server-less information exchange and differential privacy analysis.

Main Results:

• ProxyFL allows for model heterogeneity, enabling diverse private model architectures.
• Demonstrated stronger privacy guarantees through differential privacy analysis.
• Achieved superior performance compared to existing alternatives with reduced communication overhead on image and histology datasets.

Conclusions:

• ProxyFL offers an efficient and private solution for decentralized federated learning.
• The method effectively supports collaboration in regulated domains with diverse data and models.
• ProxyFL represents a significant advancement in secure, collaborative machine learning.