CloudyFL: A Cloudlet-Based Federated Learning Framework for Sensing User Behavior Using Wearable Devices

Abstract

Wearable devices have been widely utilized by the general public for tracking physical activities. Many complex machine learning models leverage wearable devices to address application problems, such as predicting pedestrian behaviors and health management. These models often incur heavy computing load and energy cost, which is challenging for wearable devices. However, aggregating the data from different wearable devices to a central server introduces privacy concerns. To address these challenges, we propose an architecture, CloudyFL, by deploying cloudlets close to wearable devices. In CloudyFL, each cloudlet forms a trusted zone covering a subset of nearby wearable devices. Models are trained in this trusted zone, and then, only the model parameters are transmitted to a centralized aggregator using a federated learning framework. We additionally propose an LSTM-based model for user behavior sensing, with a neural network design to adjust to the non-IID data distribution on multiple cloudlets. Experimental results show that our training model within the CloudyFL architecture can achieve a performance better than existing methodologies.