A Privacy-Preserving Federated Learning Framework for Ambient Temperature Estimation With Crowdsensing and Exponential Mechanism

Ambient temperature estimation plays a vital role in various domains, including environmental monitoring, smart cities, and energy-efficient systems. Traditional sensor-based methods suffer from high deployment costs and limited scalability, while centralized machine learning approaches raise significant privacy concerns. Recent crowdsensing-based systems leverage smartphone sensor data but face two major challenges: user privacy protection and unreliable participant contributions. To address these issues, this study proposes a privacy-preserving federated learning framework that integrates differential privacy with the exponential mechanism to ensure user anonymity during decentralized training. Furthermore, a novel utility-based filtering mechanism is employed to detect and exclude low-quality or adversarial data, enhancing model reliability. Advanced deep learning models, including long short–term memory (LSTM) and ensemble learning, are integrated to improve prediction accuracy in temporal and noisy environments. The dataset consists of mobile sensor data, including battery temperature, CPU usage, and environmental temperature measurements, collected from participants in real-world settings. The framework achieved high accuracy, with the LSTM model outperforming others (federated MAE: 1.292, MAPE: 0.0511) and performing comparably to centralized models (MAE: 1.179, MAPE: 0.0462) while ensuring privacy. The proposed framework showed comparable performance to centralized models while ensuring strong privacy guarantees. The integration of privacy-preserving mechanisms and robust data filtering enables a scalable and reliable solution suitable for practical deployment in large-scale ambient temperature estimation tasks.