RefinedFed: A Refining Algorithm for Federated Learning

Federated learning (FL) is a machine learning approach where the goal is to train a centralized model using a large number of clients that host private datasets. FL trains a smaller version of the model at each dataset site and then aggregates all the models at the server. In practice, clients (i.e., dataset holders) that participate in the learning process may possess corrupted or noisy datasets resulting in low accuracy models. Additionally, malicious clients may poison the data or carry out model discovery attacks.In this paper, we propose a refining algorithm called RefinedFed, to eliminate corrupted, low accuracy, and noisy models that can negatively impact the centralized model by reducing its accuracy or cause other malicious activities. Furthermore, RefinedFed reduces the uplink communication cost with the centralized server, which in return results in faster aggregation on the server side. Based on our preliminary experiments on the MNIST dataset, we observed that RefinedFed improved the global model accuracy from 84% to 91% while consuming less time for aggregation.