Lightweight Transformer in Federated Setting for Human Activity Recognition

Human activity recognition (HAR) is a machine learning task with applications in many domains including health care, but it has proven a challenging research problem. In health care, it is used mainly as an assistive technology for elder care, often used together with other related technologies such as the Internet of Things (IoT) because HAR can be achieved with the help of IoT devices such as smartphones, wearables, environmental and on-body sensors. Deep neural network techniques like convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have been used for HAR, both in centralized and federated settings. However, these techniques have certain limitations: RNNs cannot be easily parallelized, CNNs have the limitation of sequence length, and both are computationally expensive. Moreover, the centralized approach has privacy concerns when facing sensitive applications such as healthcare. In this paper, to address some of the existing challenges facing HAR, we present a novel one-patch transformer based on inertial sensors that can combine the advantages of RNNs and CNNs without their major limitations. We designed a testbed to collect real-time human activity data and used the data to train and test the proposed transformer-based HAR classifier. We also propose TransFed: a federated learning-based HAR classifier using the proposed transformer to address privacy concerns. The experimental results showed that the proposed solution outperformed the state-of-the-art HAR classifiers based on CNNs and RNNs, in both federated and centralized settings. Moreover, the proposed HAR classifier is computationally inexpensive as it uses much fewer parameters than existing CNN/RNN-based classifiers.