Privacy Preserving Federated Random Forest Approach for Human Activity Recognition

The rapid growth of sensor based wearable technology is continuously producing enormous amounts of data on health. These data have the potential to significantly change customized health monitoring and intervention. The traditional centralized machine learning approaches face significant challenges including serious privacy risks because user data must be shared with a central server, which may cause latency, and scalability issues. Recent studies have placed a strong emphasis on protecting user data privacy while improving Human Activity Recognition (HAR) accuracy. To address these challenges, this study employs Federated Learning (FL) using the Flower framework by incorporating the FedAvg algorithm and Random Forest classifier for HAR. One of the state-of-the-art datasets, KU-HAR has been utilized, which consists of smartphone sensor data, and is distributed among multiple clients in a federated setup using Flower framework. To improve feature representation, the raw sensor signals are subjected to the Fast Fourier Transform (FFT), which transforms time-domain data into the frequency domain prior to classifier training. Each client trains a local Random Forest model on its private data, and updates are aggregated using FedAvg to improve global model performance without sharing their raw data. Experimental results show that while traditional Random Forest-based HAR achieves around 92% accuracy, the FL setup with Random Forest reaches an accuracy of 89.5%, effectively preserving user privacy while maintaining high classification accuracy. The study shows the potential of Federated Learning with FedAvg and Random Forest for privacy-aware and decentralized HAR.