Adapt to Adaptation: Learning Personalization for Cross-Silo Federated Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Conventional federated learning (FL) faces challenges with non-IID datasets due to its one-model-fits-all approach.
• Personalized FL aims to address data distribution shifts in decentralized learning environments.

Purpose of the Study:

• To propose APPLE, a personalized cross-silo FL framework that adaptively determines client model benefit.
• To introduce a flexible training control method for balancing global and local objectives in personalized FL.

Main Methods:

• Developed APPLE, a personalized FL framework with adaptive client benefit learning.
• Implemented a flexible training control mechanism for global and local objective balancing.
• Conducted empirical evaluations on benchmark and medical imaging datasets under non-IID settings.

Main Results:

• APPLE demonstrated state-of-the-art performance compared to existing personalized FL approaches.
• Evaluated convergence and generalization behaviors of the proposed framework.
• Validated effectiveness across diverse datasets and non-IID scenarios.

Conclusions:

• APPLE effectively mitigates challenges posed by non-IID data in federated learning.
• The proposed framework offers superior performance and flexibility in personalized FL.
• The study provides a robust solution for personalized federated learning applications.