In healthcare, where regulatory and ethical considerations prevent the aggregation of raw patient data, federated learning (FL) has emerged as the standard approach for training models across multiple institutions. However, the optimizer that combines client updates is not always selected as carefully as the privacy and integrity layers that surround it. This work compares five representative federated optimization strategies: Adam, SGD, RMSprop, AdaGrad, Adamax, and Adadelta, in a single, fixed secure-aggregation and blockchain-logging pipeline, without the influence of the federated optimization strategy. A set of 10 simulated hospital clients trains a ResNet-18 on OCTMNIST retinal images, local updates are secret-shared and securely aggregated, and a lightweight proof-of-work ledger records each round's transaction. Even with a fixed secure-aggregation and blockchain-logging pipeline, our results indicate that the choice of optimizer has a significant impact on the accuracy of convergence on IID and non-IID setting. SGD achieved the highest accuracy under both IID (91.65%) and non-IID (87.34%) partitioning on OCTMNIST, outperforming Adam, RMSprop, AdaGrad, Adamax, and Adadelta in both settings, with adaptive optimizers showing greater sensitivity to non-IID data heterogeneity.

Federated Learning, Blockchain, Secure Multi Party Computation, Healthcare, Optimizers