Federated Learning for Unsupervised Anomaly Detection in ADLs of Elderly in Single-resident Smart Homes

One of the main concerns regarding the facilitation of the elderly well-being monitoring system is to preserve the participants' privacy, enable the older adults to live longer independently, and support caregivers. Human Activity Recognition (HAR) in smart homes allows us to foresee the residents' needs by identifying changes in behaviour that might link to possible health conditions. We propose a Federated learning (FL) model within the health monitoring application to generalize for diverse participant populations and to achieve comparable performance without disclosing the raw data to the traditional centralized approach, which raises privacy issues. In this study, we evaluate an unsupervised variational autoencoder (VAE) in centralized, individualized, compared to federated learning settings to learn the features of normal patterns of daily activities and build an anomaly detector based on reconstructed error resulting as outcomes by the trained model. Further, we validated our proposed approach on real-world datasets collected over three years from six single-resident elderly households. The individual and centralized-based learning models were used as a baseline to compare with FL. Our results show that the personalized FedAvg models achieve RMSE of about 1%, while the Global FL models achieve RMSE of approximately. 4%. The centralized model achieves RMSE of about 0.5%, and the RMSE of individual models based on local training ranges between 1% to 6%. The FL models are relatively comparable to the centralized baseline model.