FedFall: Federated Learning Based Framework for Fall Detection

Abstract

Falls are the world’s second-leading cause of unintentional injury death and disproportionately affect older adults. Wearable sensor methods combined with machine learning (ML) have improved automatic fall detection, but conventional centralised training requires transmitting raw, privacy sensitive signals to the cloud. We propose FedFall, a novel Federated Learning (FL) framework that trains fall detection models collaboratively across distributed, privacy constrained devices(e.g., wearable and ambient sensors). Experiments use the fall specific subset of the UMFall dataset, excluding everyday activities that lie outside this study’s scope. Five architectures are evaluated Deep Neural Network (DNN), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Bidirectional LSTM and a combination of CNN and LSTM. The DNN yielded the best performing results. While centralised baseline training started at just 41% accuracy, FL progressively improved performance, reaching 72% after successive communication rounds. This consistent convergence underscores FedFall’s effectiveness, enhancing detection accuracy while preserving the privacy of distributed sensor data.