Maverick++: Collaboration-free Unlearning for Medical Privacy Preservation in Healthcare Federated Systems

Federated Learning (FL) enables decentralized model training while preserving patient privacy, making it essential for medical AI applications. However, regulatory frameworks such as GDPR, CCPA, and LGPD mandate “the right to be forgotten” , requiring patient data removal from trained models upon request. This has driven growing interest in Federated Unlearning (FU), but existing methods require the collaborative participation of all clients, which is often impractical and raises privacy concerns, particularly in privacy sensitive medical domain. This paper introduces Maverick++, the first collaboration-free FU framework designed specifically for medical imaging applications. Inspired by Lipschitz continuity, Maverick++ quantifies memorization through a novel model sensitivity metric and frames the unlearning task as a localized optimization problem. When an unlearning request is issued, only the target client performs local unlearning by applying a locally bounded optimization procedure. This process enables effective unlearning without requiring collaboration from other clients. Theoretical analysis and extensive experiments on three diverse imaging benchmarks, Colorectal Cancer Histology, Pigmented Skin Lesions, and Blood Cells, demonstrate that Maverick++ sustains classification accuracy on retain data within 1–2% of retraining from scratch, reduces accuracy on unlearn data to below 1%, and lowers the attack success rate of Membership Inference Attack (MIA) to under 10%. Furthermore, Maverick++ achieves speedups of 8×–45× over state-of-the-art baselines by eliminating the need for global client coordination. These results establish Maverick++ as a practical, efficient, and privacy-aware solution for enforcing data deletion requests for FL in medical imaging domain. The code is publicly available at https://github.com/OngWinKent/Maverick.