FedUP: One-Shot Federated Unlearning via Centroid-Guided Plug-in Filters

Federated Learning (FL) allows models to learn from decentralized data without centralizing raw information. However, legal mandates like the 'right to be forgotten' require that models be able to 'unlearn' specific data contributions. Existing Federated Unlearning (FU) methods typically force a choice between two extremes: server-side methods that are fast but often damage non-target knowledge, or client-side methods that are accurate but require slow, multi-round retraining. FedUP aims to resolve this dilemma by providing a one-shot, reversible, and precise unlearning mechanism.

FedUP introduces a pluggable filter architecture that acts as a 'knowledge funnel.' Instead of modifying the global model's parameters, which is computationally expensive and irreversible, FedUP freezes the original model and trains a small, lightweight filter at the server side. To facilitate this, clients compute differentially private (DP) class centroids from their retained data—summarizing their knowledge without exposing raw data—and upload these to the server. The server then uses these centroids to train the filter to block the target data while preserving the performance of the rest of the model. Because the filter is an external component, the unlearning process is inherently reversible: removing the filter restores the model to its original state.

FedUP significantly reduces unlearning latency from minutes to seconds by bypassing iterative client-server retraining. Experiments across diverse image and text tasks demonstrate that the framework achieves superior unlearning precision compared to existing baselines. By avoiding direct parameter updates, FedUP effectively minimizes non-target knowledge loss, ensuring that the model's performance on retained data remains stable. Furthermore, the inclusion of Gaussian noise in the centroid generation process provides a formal differential privacy guarantee, protecting individual data points during the unlearning request.