Beyond Centralized Learning: Privacy-Preserving Federated Learning for Explainable Brain Tumor MRI Classification

* We built a merged and de-duplicated four-class brain tumor MRI dataset containing 10,534 images (glioma, meningioma, pituitary, and no tumor). Seven centralized backbone models were benchmarked, and EfficientNet-B3 with a lightweight attention module was selected as the base model for all federated experiments.

* Using the selected model’s weights, we conducted a three-client federated learning simulation and compared FedAvg and FedProx under both IID and Non-IID (Dirichlet, α = 0.5) data partitions. Two different local-epoch schedules were also evaluated, resulting in eight configurations under identical experimental conditions.

* We incorporated differential privacy into both aggregation strategies using gradient clipping and calibrated Gaussian noise. Two noise multipliers and two local-epoch schedules were tested. The best of the resulting 16 configurations achieved performance within 0.01 accuracy points of the best non-private result.

* Finally, we applied Grad-CAM, patch occlusion, attention maps, and t-SNE embeddings to the centralized and best federated models. These results were compared with state-of-the-art studies to assess whether the privacy-preserving models maintained clinical interpretability alongside high predictive performance.