1 Introduction
Federated learning (FL) decouples the ability to construct a machine learning model from the need to store the data in the cloud. By aggregating hundreds or thousands of clients’ local models without exposing their local data and training process to any third party, a global model is obtained [1], [2], [3]. The participants can be sensors, home gateways, micro servers, small cells, or smartphones, which are equipped with storage and computation capability. Motivating applications include training image classifiers [4], next-word predictors on users’ smartphones [2] and smart wearable healthcare [5], etc. Different from conventional distributed machine learning [6], an FL system consists of a large number of clients who may possess erroneous data (e.g., mislabeled data), which seriously hinders the global model from achieving a good performance [7], [8]. For instance, data collected by crowdsourcing [9] or web crawlers may contain mislabeled samples. One of our experiments in Section 6 shows that a two-class image classifier trained by FL with datasets crawled from image search engines suffered an accuracy loss from 91.6% to 88.2% due to the existence 9% mislabeled data.