Channel-Aware Federated Unlearning With Gradient Perturbation in Open Radio Access Networks

Abstract

Federated learning in Open Radio Access Networks (O-RAN) enables privacy-preserving distributed training, yet regulatory mandates such as the General Data Protection Regulation (GDPR) require verifiable erasure of client data influence from trained models. Existing federated unlearning schemes assume ideal channels, rendering them impractical under real wireless conditions. We propose Channel-Aware Federated Unlearning (CAFU), which reconceives fading channel noise as a constructive gradient perturbation resource, eliminating the need for full model retraining. Orchestrated via a dedicated xApp on the near-Real-Time RAN Intelligent Controller (near-RT RIC) using real-time channel state information, CAFU provably achieves $\varepsilon $ -certified forgetting with substantially reduced communication overhead compared to baselines, while preserving post-unlearning model utility.