Personalizing Federated Learning based on Adapting Representation Learning for Anomaly Detection in Home Health Monitoring Systems⋆

Abstract

Abstract: The proliferation of Internet of Things (IoT) wearable devices has enabled real-time
human activity recognition for home health monitoring, including fall detection. However, deploying
machine learning models in these scenarios faces challenges from data privacy concerns and the nonIndependently and Identically Distributed (non-IID) nature of sensor data across users. Federated
learning (FL) addresses privacy by enabling collaborative training without centralized data collection,
but existing FL frameworks struggle with data heterogeneity. We propose FL based on Adaptive
Local Training (FedALT), where each client’s model consists of an Autoencoder (AE) for learning
latent representations and a Predictor (PD) for classifying activities, including abnormal actions such
as falls. The key innovation is the Adaptive Local Training (ALT) mechanism, which dynamically
adjusts the contribution of global and local AE models during local initialization using only the
unsupervised RE loss, thereby mitigating bias caused by non-IID label distributions. Experiments
on three wearable sensor datasets, i.e., MobiAct, MobiFall, and HAR, demonstrate that FedALT
consistently outperforms existing FL frameworks, particularly in classifying minority and abnormal
action classes under heterogeneous data distributions.
Keywords: Anomaly Detection, Federated learning, IoT.