Dual asymmetric momentum improves federated class unlearning in edge systems

Abstract Federated learning is increasingly used in edge and on-device systems, where models may later need to reduce the influence of specific training data in response to governance or deletion requests. Doing this after training is difficult because full retraining is costly in communication and computation, especially under non-IID client heterogeneity. We propose FedDAM, a communication-efficient post-hoc method for federated class unlearning that freezes the trained backbone and main classifier and updates only a lightweight auxiliary head. FedDAM further separates retain and forget optimization through dual-asymmetric momentum, enabling faster forgetting while better preserving retained utility under a fixed unlearning budget. Experiments on CIFAR-10, CIFAR-100, and ImageNet-100 show consistent improvements over a unified-momentum auxiliary-head baseline under matched budgets. On CIFAR-100, FedDAM improves the retained-utility summary at matched forgetting by 9.4 percentage points, and similar gains persist on ImageNet-100 and under sparse sample-level and client-level removal settings. Additional analyses show that the method remains robust under different aggregation rules and offers a favorable utility-efficiency trade-off relative to conflict-mitigation adaptations and compressed full-model retraining. These results indicate that FedDAM is a practical approach for responsive post-hoc unlearning in resource-constrained federated systems.