FedUP Enables One-Shot Federated Unlearning with Centroid-Guided Filters.

Researchers have introduced FedUP, a novel framework for one-shot federated unlearning, designed to address the challenges of data removal in decentralized AI systems. This method is crucial for compliance with privacy regulations like the "right to be forgotten" without compromising model utility or incurring high latency. FedUP employs lightweight, pluggable filters that act as a "knowledge funnel," selectively screening out target data while maintaining the original model's performance. The core innovation lies in training these filters at the server side using differentially private (DP)-protected class centroid samples. This approach eliminates the need for multi-round client-server communication and complex retraining, drastically cutting unlearning latency from minutes to mere seconds. Furthermore, FedUP's pluggable architecture offers inherent reversibility, allowing for the seamless restoration of "forgotten" knowledge by simply removing the filters. Extensive experiments across various image and text tasks demonstrate FedUP's effectiveness in reducing non-target knowledge loss and achieving superior unlearning precision and efficiency.