Personalized Federated Human Activity Recognition through Semi-supervised Learning and Enhanced Representation
Lulu Gao, Shin’ichi Konomi · 2023
Owing to the widespread utilization of activity recognition, numerous deep learning models have been proposed to facilitate recognition in diverse contexts. Federated learning (FL) is an emerging learning paradigm that enables the collaborative learning of a shared model without compromising the data privacy of users. However, the assumption of FL is to rely on the annotated data on clients, which is difficult to acquire the annotations for human activity recognition (HAR) on all clients due to the lack of expertise or resource. Moreover, a general model is not suitable for each person due to data heterogeneity, resulting from the different physical characteristics and various contextual information. To this end, we propose a semi-supervised learning method for personalized federated HAR, in which clients have completely unlabeled data, while the server has a small amount of labeled data contributed by volunteers. Clients conduct unsupervised learning on autoencoders with locally unlabeled data to collaboratively learn a general representation model. The server conducts supervised learning on an activity classifier with labeled data stored on the server. After that, the shared model is personalized using individually pseudo-labeled data on each client side, wherein both confidence and uncertainty are taken into account concurrently, with the aim of achieving a balanced selection for assigning pseudo-labels to samples. We conduct extensive experiments with two different real-world HAR datasets, demonstrating the effectiveness of the proposed methods.