面向时空开销均衡的轻量级联邦遗忘学习框架

In the era of booming data-driven artificial intelligence technologies, societal concerns about data privacy protection are escalating. Federated Learning (FL) has emerged as a promising distributed machine learning paradigm, with its core objective centered on addressing data privacy challenges through decentralized collaborative training mechanisms. While FL effectively mitigates data leakage risks and complies with regulations like the General Data Protection Regulation (GDPR) and California Consumer Privacy Act (CCPA) that mandate local data storage, it fails to satisfy the“right to be forgotten”granted to users under these regulations. To achieve compliance with the“right to be forgotten”, Federated Unlearning (FU) has been proposed. Currently, federated unlearning has become a promising paradigm to fulfill users’ “right to be forgotten”, enabling the elimination of users’data contributions from global models while meeting data privacy protection requirements. However, existing federated unlearning methods typically either sacrifice storage space to reduce unlearning time or sacrifice unlearning time to reduce storage space, failing to balance time overhead and storage overhead. Beyond introducing significant time overheador storage overhead, current FU methods generally support only specific unlearning levels—such as client-level or class-level unlearning—limiting their practicality in real-world scenarios. For instance, in a smart healthcare setting, multiple hospitals may collaboratively train disease diagnosis models via federated learning while preserving data privacy. However, if a hospital needs to withdraw from the collaboration due to policy changes, identifies mislabeled data, or encounters sensitive disease categories in the model, it would require client-level, sample-level, and class-level unlearning, respectively. Therefore, in this paper, we propose FedUR, a lightweight federated unlearning framework that supports three unlearning levels: client-level, sample-level, and class-level. FedUR achieves clean unlearning effect while preserving global model performance and effectively balancing time overhead and storage overhead. To accommodate the three unlearning levels and optimize resource trade-offs, FedUR decouples the federated unlearning process into two stages: adaptive unlearning and optimized recovery. In the stage of adaptive unlearning, the target client employs stochastic gradient ascent (SGA) on target data to achieve unlearning while minimizing storage overhead. If the unlearning model violates constraints, projected gradient descent (PGD) is applied to prevent degradation into a random model. In the stage of optimized recovery, the server leverages knowledge distillation with an outsourced labeled dataset to restore the performance of the global model, achieving faster recovery compared to traditional post-training or methods that integrate the unlearning process into federated learning. Extensive experiments on real-world datasets compare FedUR against five state-of-the-art FU methods across three unlearning levels using four metrics: accuracy (reflecting the post-unlearning model performance), backdoor attack success rate (reflecting the unlearning effect), time overhead, and storage overhead (reflecting the efficiency). Results demonstrate that FedUR achieves the highest accuracy (reflecting the superior model performance), lowest backdoor attack success rate (reflecting the cleanest unlearning effect), minimal storage overhead, and second-lowest time overhead (reflecting the high efficiency). While the FUG method is the fastest, it shows a 1%-10% accuracy gap compared to FedUR, and its unlearning effect is incomplete. Overall, FedUR optimizes model performance, unlearning effect, and efficiency. These findings validate FedUR's effectiveness and efficiency, enabling robust unlearning, preserving model performance, and balancing time overhead and storage overhead. It addresses the typical performance degradation and spatiotemporal trade-off challenges associated with federated unlearning.