FUCH: Non-IID Federated Unlearning via Classification-layer Hybrid-order Constraints

Abstract

Federated unlearning (FU) aims to support the ''right to be forgotten'' by reducing the influence of target samples without full retraining. However, when applied to deep neural networks (DNNs) in practical non-IID scenarios, existing FU methods suffer from low unlearning efficiency and deviate from the retraining reference's predictive behavior. Specifically, they require costly full-model updates or multi-round communication, failing to disentangle the complex coupling of target-sample influence and retained knowledge. In this work, we revisit the necessity of full-model updates for sample-level FU and investigate the formation of this complex coupling. We reveal key empirical insights: (1) updating solely the final classification layer can effectively approximate the predictive behavior of full retraining; and (2) disentangling target-sample influence under non-IID settings requires separately resolving distinct intra-client and cross-client knowledge couplings. Based on these insights, we propose FUCH, a classification-layer-based hybrid-order-constrained FU framework. To improve unlearning efficiency, FUCH freezes the feature extractor and performs a single-round update on the final classification layer. Furthermore, to disentangle the complex knowledge coupling, FUCH employs a coupling-aware hybrid-order mechanism combining local second-order and global first-order information to constrain the unlearning updates. Extensive experiments across diverse non-IID benchmarks validate our empirical insights and demonstrate that FUCH achieves efficient unlearning while significantly outperforming SOTA baselines in closely approximating the retraining reference's predictive behavior.