A comparative study of federated learning algorithms on Tiny ImageNet

Federated Learning (FL) has become a promising framework for collaborative model training across distributed clients without sharing raw data, addressing privacy concerns in sensitive fields. However, performance often deteriorates when client data are non-independent and identically distributed (non-IID), leading to unstable convergence and accuracy degradation. This study assesses the effectiveness of three representative algorithms of federated learning – Federated Average (FedAvg), Federated Proximal (FedPro) and Federated Adaptive Optimization (FedOpt) on the Tiny ImageNet dataset using convolutional neural networks including Resnet-18, MobileNetV2, and ResNet-50. The experiments focus on assessing algorithm stability, convergence rate, and accuracy under both IID and non-IID data partitions. Results indicate that FedAvg exhibits instability under non-IID conditions, while FedProx improves robustness and FedOpt achieves the fastest and most consistent convergence. Across models, deeper architectures such as ResNet-50 reach higher accuracy but incur greater computational cost, whereas lightweight networks like MobileNetV2 offer better efficiency with modest accuracy. Overall, the findings provide a comprehensive evaluation of federated algorithms in complex vision tasks and highlight the trade-off between accuracy, stability, and efficiency in non-IID learning environments.