Personalized Federated Learning Based on Dynamic Parameter Fusion and Prototype Alignment.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Distributed Systems

Background:

• Federated learning (FL) enables collaborative model training without sharing raw data.
• Generalization remains a challenge in FL, particularly with non-independent and identically distributed (Non-IID) data across clients.
• Existing FL methods often struggle to balance global model performance with individual client needs.

Purpose of the Study:

• To propose FedDFPA, a novel personalized federated learning framework.
• To address the generalization limitations of FL under Non-IID data.
• To enhance both personalization and collaboration in federated learning systems.

Main Methods:

• Developed FedDFPA, integrating dynamic parameter fusion and prototype alignment.
• Implemented a class-wise dynamic parameter fusion mechanism for adaptive merging of global and local classifier parameters.
• Introduced a prototype alignment mechanism using global and historical data to improve semantic consistency and feature stability.

Main Results:

• FedDFPA demonstrated significant improvements in average test accuracy compared to state-of-the-art algorithms.
• Achieved a 3.59% accuracy improvement in practical heterogeneous settings.
• Achieved a 4.71% accuracy improvement in pathological heterogeneous settings.

Conclusions:

• FedDFPA effectively mitigates generalization issues in federated learning with Non-IID data.
• The dual-mechanism design successfully balances personalization and collaboration.
• The framework offers a robust solution for personalized classification in distributed environments.