Hybrid FedProx Optimization for Privacy Preserving Diabetic Retinopathy Classification

Abstract

Diabetic Retinopathy (DR) is a leading cause of vision impairment worldwide. Early detection through automated analysis of retinal images is critical. Nonetheless, the centralized training of deep learning models presents privacy concerns owing to the sensitive nature of medical data. This paper investigates the use of the Federated Proximal algorithm for collaborative DR classification across distributed data sources, addressing the challenges of data heterogeneity. We enhance Federated Proximal (FedProx) with momentum-based optimization techniques including Adam, Shuffling SGD, SGD momentum and RMSProp in order to accelerate convergence and improve generalization. Experimental results demonstrate that incorporating momentum techniques with Federated Proximal significantly improves accuracy and reduces loss on diabetic retinopathy dataset with shuffling SGD achieving the best performance with accuracy of 92.63% and 0.22 loss. These results highlight the potential of integrating proximal regularization and momentum strategies for privacy preserving, robust and effective DR detection in federated learning environments.