Federated Learning for Human Activity Recognition on the MHealth Dataset

Sanchez S.; Machacuay J.; Quinde M. · Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (2023) · DOI: 10.1007/978-3-031-42505-9_19

Resumen

Human activity recognition (HAR) tasks are essential for developing context-aware features due to their capacity to generate higher-level contexts and personalisable services. The traditional centralised training approach requires clients to share their data with a server that trains the HAR algorithms. This data transfer raises privacy concerns on the clients’ side regarding data leaks that can be used to obtain personal data. This research work studies Federated Learning (FL) as a privacy-preserving alternative to developing HAR algorithms. The centralised and FL approaches are implemented to train a Multi-Layer Perceptron neural network using three different activation functions (Tanh, ReLU and Sigmoid). The experiments use the MHealth dataset to train and test the HAR algorithms. The results are analysed to compare the performance of the centralised and FL approaches, the activation functions outcomes, and the communication rounds needed between the server and the clients to reach competitive results for the FL models. The outcomes are positive and encourage further work in using of FL for HAR.

Instituciones: Universidad de Piura