Investigating Federated Learning strategies for Pneumonia Image Classification

Deep Learning has emerged as a promising approach for building accurate and robust models in healthcare. This paper explores federated learning strategies for pneumonia classification using X-ray images and analyzes how it can address privacy issues. The performance of federated training has been compared and proved to be on par with traditional centralized training methodology. Three aggregation techniques, namely, Federated Averaging (FedAvg), Coordinate-wise median (COMED) and Geometric median (GEOMED) have been implemented and analyzed for two real-world scenarios. COMED and GEOMED have proved to be more resilient to the presence of outliers when compared to FedAvg. Layer shuffling protocol is proposed for implementing data encryption while sharing the model weights.