Margin-aware prototype learning for client withdrawal in federated unlearning

Federated client withdrawal requires removing targeted clients' influence from a collaboratively trained model while preserving utility for remaining participants. Existing approaches face a hard trade-off. To achieve efficiency, they often rely on stored historical information, such as states or gradients, which incurs substantial memory overhead and makes it difficult to precisely isolate and remove a single client's influence from aggregated historical updates. While methods that guarantee complete removal, such as retraining from scratch, are computationally prohibitive in practice. To address this dilemma, we introduce Margin-Aware Prototype Learning (MAPLE), a novel framework that achieves both high efficiency and efficacy without relying on storage-intensive historical data. MAPLE decouples the unlearning task into two synergistic components: (i) at the local level, Margin-aware Label Reassignment (MLR) adaptively perturbs labels on the withdrawing client's data, producing targeted forgetting signals that are most intense for low-confidence samples near the decision boundary; (ii) at the global level, Prototype-driven Constraints (ProCons) use compact, class-wise prototypes from remaining clients as lightweight geometric anchors in the feature space, preserving shared knowledge via contrastive objective. Extensive experiments demonstrate that MAPLE achieves an unlearning quality nearly identical to the gold standard of complete retraining while being orders of magnitude faster and requiring negligible memory overhead, consistently outperforming state-of-the-art approaches.