- Conference Article
- 10.1109/iccvw69036.2025.00717
Federated Active Learning for Target Domain Generalisation
- Oct 19, 2025
- Razvan Caramalau + 2 more +2
Federated learning (FL) has recently enabled deep learning models to scale over distributed client data. In this context, a new area of research has emerged, extending to various vision tasks from a domain generalisation (DG) paradigm known as Federated Domain Generalisation (FDG). Common challenges in this framework, such as data heterogeneity, obtaining domain-invariant representations, and addressing privacy concerns, have been addressed in recent years. However, we have observed that all approaches consider the entire client source data, which is often heterogeneous, corrupted, or redundant. To tackle these challenges, we propose FEDALV, the first Federated Active Learning (FAL) framework for DG. Prior to implementing our sampling strategy, we establish and solve the generalisation problem with our FDG pipeline, FEDA, leveraging the free energy alignment between each source domain and the target. Furthermore, this measure facilitates FEDALV in the active learning setting by sampling the most informative source samples only from clients aligned with the target domain. Our extensive empirical study demonstrates the superiority of our method in accuracy and efficiency compared to multiple contemporary methods. FEDALV manages to achieve the performance of the full training target accuracy while sampling as little as 5% of the source client's data.
Read more