Fed-DR-Filter: Using global data representation to reduce the impact of noisy labels on the performance of federated learning

Abstract

The label noise is a serious problem limiting the performance of federated learning. According to the performance evaluation for the trained federated models, data selection strategies or client selection strategies are used to solve this problem in previous studies. However, these methods require additional clean data to strengthen the election results, and they rely heavily on an initial model that is robust enough to not accumulate errors. To address these problems, we propose a novel data filtering method to deal with label noise in federated learning, which is called Fed-DR-Filter. Unlike previous methods, Fed-DR-filter focuses on identifying clean data by taking advantage of the correlation of the global data representations. The proposed solution transforms the private data into privacy-preserving data representations in each client, and identifies clean data based on the centralized data representations on the server. To evaluate the performance of Fed-DR-Filter, we conduct extensive experiments on three real-world datasets. The evaluation results show that our method outperforms the state-of-the-art approaches and is robust to various data distributions and noise levels.