Dynamic federated learning: coping with the challenge of label noise in real-time network data

Label noise is one of the critical challenges that cannot be overlooked in Federated Learning (FL). However, existing approaches primarily focus on label noise in static data, with limited exploration of real-time network data. In real-time network environments, client data exhibits dynamic growth, which may continuously introduce new label noise. This degrades data quality over time and can gradually contaminate originally clean clients. In such environments, federated denoising methods based on static data become less effective, as they lack the adaptability to dynamic settings and struggle to handle the continuous shifts in data distribution and noise levels. In this paper, we propose the FedRnd framework to investigate the problem of label noise in Federated Learning under real-time network environments. Given the dynamic nature of data growth, FedRnd primarily employs a localization strategy. After fuzzily classifying real-time data based on historical data and storing it into clean and noisy datasets, the method performs data augmentation on the clean dataset to enhance data quality. Simultaneously, LogitClip technology is introduced to clip model outputs, further suppressing model overfitting to noisy data. Additionally, a compactness-loss consistency score is designed to optimize global model aggregation. Extensive experiments conducted in simulated environments demonstrate that FedRnd exhibits superior performance compared to state-of-the-art methods in handling label noise within real-time network data. Our code is available at https://github.com/Donglin0730/FedRnd .