Federated Active Learning Excels with Few Labels and Coordination

Federated Active Learning (FAL) aims to address data privacy and label scarcity challenges by allowing clients to collaboratively train models without sharing raw data, while actively selecting the most informative data points for labeling. This study focuses on FAL in low-budget scenarios where annotation decisions are critical. It uncovers a "heterogeneity reversal": surprisingly, homogeneous (IID) data presents a greater challenge for query selection in FAL, requiring more robust coordination to avoid redundant labeling, whereas heterogeneous data naturally promotes diversity in selected queries. Motivated by these insights, the researchers propose a novel FAL framework. This framework leverages federated representation learning to align client data into a shared embedding space. This alignment enables a central server to perform globally coordinated active selection over obfuscated client embeddings, while the actual data annotation remains local to each client. This new approach significantly outperforms existing FAL methods, even when those methods are given substantially larger annotation budgets, highlighting the critical value of centralized coordination under privacy constraints, especially in low-resource settings.