Federated learning(FL) enables collaborative model training without sharing raw data, yet faces a growing challenge: efficiently removing the influence of a specific client’s data to comply with privacy regulations such as the “right to be forgotten.” In this paper, we introduce a novel knowledge distillation-based federated unlearning framework that enables the global model to “forget” a targeted client’s data while preserving overall performance. Our method constructs a fresh student model by distilling knowledge from the original global model, training it only on data from remaining clients, and strategically replaying representative samples to mitigate disaster unlearning. Experiments conducted in MNIST and CIFAR-10 in various non-IID settings demonstrate that the proposed approach not only achieves verifiable unlearning – validated through membership inference attacks – but also reduces communication overhead and maintains high accuracy across sequential tasks. This work provides a practical, efficient, privacy-preserving solution for federated unlearning in dynamic and heterogeneous environments.
Qiao et al. (Mon,) studied this question.