ROI-Guided Differential Privacy in Federated Learning for Enhanced Alzheimer's Disease Classification

Revised software release accompanying the JMI major revision. Major updates include: three-class AD/CN/MCI classification; OASIS and ADNI experiments; EfficientNet-B2, InceptionV3, and Swin-Tiny backbones; uniform Gaussian gradient perturbation; ROI-guided perturbation; layer-adaptive perturbation; layer-aware ROI perturbation; simulated secure aggregation; membership-inference attack evaluation; qualitative spatial importance and perturbation-scale visualization. Important privacy note: The current implementation does not provide certified (epsilon, delta)-differential privacy. DP-related method names are retained as experimental identifiers, and privacy is evaluated empirically using membership-inference attacks. Archived revised software release: https://doi.org/10.5281/zenodo.22093150