Energy-Efficient Federated Learning Framework for Personalized Human Activity Recognition Using Smartphone Sensor Data

Abstract: Imagine what your smartphone already knows about you based on how you move. Each step, each climb up the stairs, each slouch into the chair – it is all recorded by your smartphone sensors. To unlock the immense possibilities of using this data for the identification of human actions in real-time, while keeping your privacy intact is an extremely promising endeavour . In this paper, EE-FL-HAR is presented, a system that allows training activity recognition models right inside the smartphone, without transferring any sensor data out of the device itself. Only the model updates are exchanged between devices through the network. The results were tested on a 2,004 sample dataset with 695 features and covering six different daily activities. The Random Forest model achieved 83.29% accuracy with the precision of 0.8348, recall of 0.8329 and F1-score of 0.8326. Federated learning simulation of 50 rounds approaches 93% of the accuracy of a centralized baseline, reducing the total energy consumption by 56%. Six diagnostic figures show what is happening in every step. The conclusion is clear: activity recognition and privacy do not need to be opposing forces anymore.

Keywords: Federated Learning, Human Activity Recognition, Sensor Suite of Smartphone, Energy Efficiency, Personalization, Privacy-preserving Machine Learning, Random Forest, Communication Efficiency.