Adaptive Federated Unlearning in Heterogeneous Edge Networks via Contrastive Fine-Tuning

Abstract

Federated unlearning (FU) has become a critical capability of privacy-sensitive edge intelligence, which enables mobile and Internet of Things (IoT) devices to revoke previously contributing data from collaboratively trained models. While previous studies have explored unlearning via retraining or approximate mechanisms, most have assumed idealized homogeneous data distributions in federated learning (FL). This caused them to overlook the substantial impact of data heterogeneity, which is a fundamental challenge in federated edge networks. To address this limitation, we propose an adaptive federated unlearning framework tailored for heterogeneous edge environments. Through a representation-level analysis, we characterize how data heterogeneity shapes forgetting behaviors and identify two key factors that hinder unlearning performance: directed representation drift and heterogeneity-induced hard samples. Guided by these insights, we develop contrastive federated unlearning (CFU), a two-stage framework that integrates local contrastive fine-tuning with global model calibration. CFU employs contrastive unlearning to induce targeted representation shifts, whereas lightweight fine-tuning mitigates the adverse effects of heterogeneity. The unlearning process is further safeguarded through localized forgetting and data-free knowledge distillation. Extensive experiments on both client and sample-level unlearning tasks demonstrate that CFU consistently outperforms state-of-the-art approximate unlearning methods across diverse real-world datasets, achieving a 2–4× efficiency improvement over retraining.