Personalized Federated Few-Shot Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Decentralized Systems

Background:

• Personalized federated learning (PFL) enables decentralized model training for individual clients using private data.
• Existing PFL methods require sufficient client data, limiting their effectiveness in few-shot scenarios.
• Few-shot learning (FSL) typically relies on centralized data, making it unsuitable for decentralized PFL.

Purpose of the Study:

• To address the challenge of enabling PFL in scenarios with limited training samples per client.
• To develop a method that allows effective personalized model training in decentralized environments with few-shot data.
• To enhance PFL performance when clients have minimal data available.

Main Methods:

• Proposed personalized federated few-shot learning (pFedFSL) to tackle data scarcity in PFL.
• pFedFSL learns a personalized and discriminative feature space for each client.
• Identifies optimal models and collaborating clients without exposing local data.

Main Results:

• pFedFSL effectively learns personalized models even with limited client data.
• The method creates feature spaces where intra-class samples are closer and inter-class samples are farther apart.
• Experimental results show pFedFSL outperforms existing methods on benchmark datasets.

Conclusions:

• pFedFSL successfully enables personalized federated learning in few-shot scenarios.
• The approach enhances model personalization and discriminative power under data constraints.
• This work provides a practical solution for PFL with limited client data.