Robust Federated Learning for Non-IID Medical Images with Trust-Guided Aggregation

Federated learning (FL) allows training medical professionals cooperatively without exchanging patient information. Nevertheless, the heterogeneity in the real world tends to cause client drift and unjust models. Our suggested FL framework is comprised of confidence-guided FL, dynamic regularization (FedDyn), local batch normalization (FedBN), and label-aware aggregation. Our approach, which is validated on MedMNIST benchmarks and the HAM10000 (high-resolution) dataset, is much more stable and fair than baselines. Through extensive ablation experiments, FedDyn is shown to avoid accuracy decline, and our trust-based aggregation can drop inter-client variance by a factor of more than 50, providing fair diagnosis between different hospitals.