PGFed: Personalize Each Client's Global Objective for Federated Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Distributed Systems

Background:

• Conventional federated learning (FL) exhibits suboptimal performance on heterogeneous datasets.
• Personalized FL (PFL) offers client-specific models, but existing methods use implicit knowledge transfer.
• Implicit knowledge transfer in PFL may not fully leverage individual client data.

Purpose of the Study:

• To propose a novel PFL framework, Personalized Global Federated Learning (PGFed).
• To enable explicit and adaptive aggregation of empirical risks among clients.
• To enhance collaborative learning in heterogeneous FL environments.

Main Methods:

• PGFed allows clients to personalize their global objective by aggregating risks.
• First-order approximation estimates client risks, reducing communication overhead and privacy concerns.
• PGFedMo, a momentum-enhanced version, improves empirical risk utilization.

Main Results:

• PGFed consistently improves performance over state-of-the-art PFL methods.
• Experiments conducted on four diverse datasets demonstrate PGFed's effectiveness.
• The proposed method shows robust improvements across various federated settings.

Conclusions:

• PGFed offers a significant advancement in personalized federated learning.
• Explicit risk aggregation is more effective than implicit methods for PFL.
• The framework provides a scalable and privacy-preserving approach to PFL.