Human Activity Recognition with Wearables using Federated Learning

Abstract

The increasing use of Wearable devices opens up the use of a wide range of applications. Using different models, these devices can be of great use in Human Activity Recognition (HAR), where the main goal is to process information obtained from sensors located in them, especially in eHealth. The high volume of data collected by various smart devices in contemporary ML scenarios, leads to higher processing consumption and in many cases results in compromised privacy. These shortcomings could be overcome by using Federated Learning (FL), a learning paradigm that allows for decentralized training of models such that user’s personal data does not need to ever leave their devices, which substantially reduces to possibility of a breach. This paper analyses the behaviour and performances of FL when applied to the context of HAR. The obtained results show that FL can achieve comparable performances to those of centralized Deep Learning, while facilitating improved data privacy and diversity, as well as fostering real-time continuous learning.