A Zero-Shot Federated Unlearning Framework With Stability Verification

Federated unlearning is essential for ensuring compliance with data privacy regulations such as General Data Protection Regulation (GDPR) and California Consumer Privacy Act (CCPA) by enabling the removal of information from trained federated models. However, most existing federated unlearning methods rely on access to original client data or historical model updates, which fundamentally contradicts privacy-preserving principles. In practice, such access is often restricted, making zero-shot unlearning necessary, where the server removes data without any client interaction. The challenge is further heightened in heterogeneous, non-IID settings, where class entanglement and distribution drift can lead to unstable or incomplete unlearning. To address these limitations, we propose a novel zero-shot federated unlearning framework that achieves certified class-level unlearning without requiring client data or historical updates. Our approach integrates Wasserstein-constrained pseudo-data generation with knowledge distillation to isolate and remove target-class information while preserving retainedclass utility. To improve robustness under distributional shifts and pseudo-data uncertainty, we introduce a dynamic adaptation mechanism driven by internal feedback control. Specifically, the Wasserstein distance serves as both a geometry-aware constraint and a stability indicator, guiding the adaptive adjustment of temperature and attention suppression parameters through an information-theoretic control loop. Extensive experiments across multiple benchmarks demonstrate that the proposed framework effectively performs class-level unlearning while preserving model utility. Compared to state-of-the-art federated unlearning methods, our approach improves retained accuracy by 7.2 reduces computational cost by 65.3 convergence and stability guarantees. Source code is available at https://github.com/kayeewww/fuzv.