A Model-Heterogeneous Federated Unlearning Method via Negative Knowledge Distillation

Federated Unlearning (FU) enables the removal of specific data from a trained model, supporting essential privacy requirements like the "right to be forgotten". However, a major limitation of current studies is their reliance on homogeneous model architectures across clients, which severely restricts their practicality. In view of this challenge, we propose FedNKD, a class-level unlearning method tailored for model-heterogeneous federated settings. Specifically, a Negative Knowledge Distillation (NKD) mechanism based on a public dataset is introduced to enable information aggregation across heterogeneous clients while effectively erasing knowledge of the target classes. In NKD, the server first constructs tailored target distributions by attenuating the logits corresponding to forgotten classes within the aggregated global logits and renormalizing over the retained classes. These refined distributions are then employed to distill knowledge back to clients, enabling effective guidance without requiring direct access to local models. Extensive experiments on SVHN and CIFAR-10 under model-heterogeneous settings demonstrate that FedNKD achieves retrain-level unlearning effectiveness, nearly eliminating knowledge of the target classes and mitigating membership inference attacks, while maintaining superior accuracy on the retained classes.