centroid-guided

FedUP proposes a one-shot federated unlearning framework that uses lightweight, pluggable filters guided by differentially private class centroids to efficiently remove specific knowledge without multi-round communication, achieving low latency and inherent reversibility.