Label-noise-robust Two-Stage Federated Learning in Sensor-Based Human Activity Recognition

Sensor-based human activity recognition is a key technology for special population care, precise industrial control, and other applications. However, in practical engineering scenarios, it encounters challenges such as inadequate data samples on individual client sides, the necessity for data privacy preservation, and noisy data labels, all of which considerably impede the training process and the performance of human activity recognition models. In this paper, we propose Label-noise-robust two-stage federated learning architecture for federated human activity recognition based on sensors (LN-FHAR). The proposed framework is structured into two distinct stages. During the client selection stage, we differentiate between high and low-quality clients by computing the average loss per class for each client. In the noise-robust training stage, we introduce prototype regularization to mitigate the issue of client drift. Furthermore, reliable neighbors are leveraged to collaboratively identify clean samples for low-quality clients, ensuring that only clean samples are utilized for training. Additionally, we propose a clients’ data-aware model aggregation method, which assigns weights based on both the quantity and quality of client samples. Experimental results on PAMAP2 and OPPORTUNITY synthetic noise datasets demonstrate the marked superiority of LN-FHAR when compared to eight other baselines across multiple label noise and non-IID scenarios.