FedDSHAR: A dual-strategy federated learning approach for human activity recognition amid noise label user

Abstract

Federated learning (FL) has recently achieved successes in privacy-sensitive health-care applications like medical analysis. Most previous studies suppose that collected user data are well-annotated, however, it is a strong assumption in practice. For instance, human activity recognition (HAR) task aims to train a model which predicts a certain person’s activity based on sensor data series collected from a given period of time. Due to diverse and incomplete annotation approaches, user-side data inevitably contain significant label noise, which greatly degrade model convergence and performance. In this work, we propose a novel FL framework FedDSHAR, which partitions the user-side data into the clean data subset and noisy data subset. Two strategies are utilized on two subsets to further exploit extra effective information from data, where strategic time-series augmentation is adopted on the clean subset and the semi-supervised learning scheme is used for the noisy subset. Extensive experiments conducted on three public real-world HAR datasets demonstrate that FedDSHAR outperforms six state-of-the-art methods, particularly in addressing extreme label noise in real-world distributed noisy HAR scenarios. Our code is available at https://github.com/coke2020ice/FedDSHAR.