FedPoll: A Novel Robustness and Communication-Efficient Aggregation Method

Federated learning enables decentralized model training across multiple clients, preserving data privacy by not sharing raw data with a central server while exploiting the variety of data available across clients. Despite its advantages, challenges such as frail privacy, non-independent and non-identical data distributions among clients, and other issues remain unsolved. This paper introduces FedWork, a flexible framework for benchmarking federated learning methods. FedWork supports various datasets, aggregation methods, and neural network architectures. We also propose FedPoll, a novel aggregation method that reduces communication overhead and enhances privacy. FedPoll includes two variants: FedPoll-Nearest and FedPoll-MaxMin. Our evaluations show that in non-i.i.d. environments, FedPolI-MaxMin achieves 95% accuracy on the MNIST dataset with partial client participation, outperforming FedAvg, FedAvg-QSGD, FedProx and FedPoll-Nearest, which achieve around 40% with fluctuations. Despite a slight increase in computation time, FedPoll-MaxMin demonstrates significant potential for improving federated learning in challenging settings. Moreover, in a more challenging federated learning scenario, FedPoll-MaxMin achieved acceptable accuracy with ResNet18 on the CIFAR-10 dataset, while FedAvg failed to converge, and SCAFFOLD achieved lower accuracy with greater fluctuations.