Federated Forests With Differential Privacy for Distributed Wearable Sensors

Training machine learning models on wearable sensor data is useful for applications like human activity recognition (HAR), but the sensitivity of such data often precludes centralized data collection. Federated learning offers a decentralized solution, enabling model training across distributed data sources. To further protect privacy, differential privacy (DP) can be integrated into the federated learning process. In this paper, we introduce a novel algorithm for training tree ensemble models (forests) within a federated learning framework under differential privacy constraints. Our approach allows each client to independently train a differentially private decision tree with randomized splits, which is sent to a central server. The server aggregates these trees into an ensemble that classifies data points via majority voting. We evaluate our DP federated forest algorithm on a HAR task, demonstrating that an effective privacy-utility tradeoff can be achieved with an appropriate selection of the privacy budget.