Personalized Federated Sensing for Heterogeneous Environment

Advancements in intelligent sensors are driving an increased demand for personalized intelligent services with privacy protection of the collected data. Traditional machine learning approaches store all distributed raw data at a central server and train a global model, posing potential risks of exposing privacy-sensitive data. Federated learning (FL) addresses this issue by training models across devices without the need to exchange raw data. However, FL faces challenges in personalization and is vulnerable to Byzantine attacks, which can compromise system performance. In addition, heterogeneous sensors may have varying hardware capacities (e.g., different sensor resolutions) as well as different data label distributions. To address these challenges, we propose a novel personalized federated sensing algorithm that trains local models divided into personalized and shared layers. Accordingly, sensors can create personalized models while preserving robustness by revealing only the shared part of the model. Extensive experiments validate its superiority over existing methods in heterogeneous settings.