In federated learning (FL), classifiers (e.g., deep networks) are trained on
datasets from multiple centers without exchanging data across them, and thus
improves sample efficiency. In the classical setting of FL, the same labeling
criterion is usually employed across all centers being