LAFUL: Lesion-Aware Federated Unlearning via Channel-Wise Gradient Masking and Feature Distillation

초록

The right to be forgotten, mandated by modern privacy regulations, poses unique challenges to federated learning in medical imaging, where models must erase patient-specific information while preserving diagnostic utility. Existing federated unlearning (FU) methods struggle with this trade-off due to the high privacy sensitivity and spatial complexity of medical data. To address these challenges, we propose LAFUL, a lesion-aware federated unlearning framework designed for privacy-preserving medical image analysis. LAFUL exploits the spatial sparsity and semantic focus of pathological lesions to identify and selectively mask lesion-sensitive gradient channels, effectively removing private information while retaining anatomy-relevant representations. To mitigate distributional shifts caused by targeted gradient removal, a feature-level knowledge distillation module aligns intermediate representations using an unlabeled proxy dataset. Experiments on two public benchmarks - intracranial hemorrhage detection and skin lesion classification - show that LAFUL achieves near-retrained accuracy with over 9 × efficiency improvement, providing an interpretable and practical solution for scalable FU in healthcare. © 2025 IEEE.

키워드