Sample-Level Prototypical Federated Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• Federated learning (FL) enables collaborative model training on decentralized data, addressing privacy concerns.
• Non-identically and independently distributed (non-IID) data across clients is a major challenge in FL, especially in cross-silo settings with internal data heterogeneity.
• Existing FL methods often overlook intra-client data heterogeneity or require inaccessible domain indicators.

Purpose of the Study:

• To propose a novel federated learning approach that addresses the challenge of non-IID data within clients.
• To introduce a fine-grained personalization method that adapts models at the individual data sample level.
• To develop a federated learning solution that does not require explicit domain indicators.

Main Methods:

• Introduced Sample-Level Prototypical Federated Learning (SL-PFL), integrating prototypical learning into the FL framework.
• Developed a method for creating personalized models for each data sample, rather than a single model per client.
• Ensured the method can be trained effectively without ground-truth domain labels.

Main Results:

• SL-PFL demonstrated superior performance compared to existing FL methods using global or client-level personalized models.
• The proposed method achieved state-of-the-art results on various real-world regression and classification tasks.
• Effectiveness was validated across diverse domains including weather, computer vision, and healthcare.

Conclusions:

• SL-PFL effectively addresses the intra-client non-IID data challenge in federated learning.
• Sample-level personalization offers a significant advantage over global or client-level personalization in federated settings.
• The proposed method provides a practical solution for real-world federated learning applications with heterogeneous data.