基于更新残差的差分隐私联邦遗忘学习机制

Federated unlearning (FU) enables the removal of specific client data and its influence from an already-trained model, thereby supporting “the right to be forgotten". Although retraining the model from scratch is a straightforward approach, its high computational cost often renders it impractical. Most existing methods achieve FU by either removing the historical contributions or gradually adjusting the model using gradient ascent. However, storing historical gradients or performance-enhancing training faces high storage and communication overhead, resulting in inefficient FU. Additionally, the privacy risks associated with the FU require further investigation. Therefore, this paper proposes FedUR, a differentially private federated unlearning mechanism based on update residuals, which achieves an effective balance among privacy protection, model utility, and unlearning efficiency. Specifically, FedUR enhances privacy by rendering the unlearned model indistinguishable from a retrained model according to the differential privacy paradigm. Moreover, the proposed approach innovatively quantifies the historical impact of unlearning clients on the global model through update residuals, and enables rapid FU by removing all historical weighted update residuals without relying on the model recovery training process, thus markedly reducing storage overhead. Furthermore, FedUR integrates importance sampling and periodic weighted aggregation strategies to mitigate the adverse effects of data heterogeneity on model performance while also lowering storage and communication overhead. Experimental results demonstrate that FedUR maintains robust model utility and high unlearning efficiency while providing strong privacy protection.