A privacy-enhancing and lightweight framework for device-free localization-based AIoT system

Abstract

With the growing demand for location-based services in smart cities, Artificial Intelligence of Things (AIoT)-enabled device-free methods have gained attention for their ability to address privacy and usability challenges. WiFi-based target localization, leveraging channel state information, offers advantages such as ease of deployment and obstacle penetration but faces privacy and computational challenges in centralized training. To address these issues, we propose a privacy-enhancing and lightweight federated device-free localization framework (PLDFL). The PLDFL integrates local differential privacy in federated learning to safeguard user data, uses the Fisher Information Matrix for model pruning to reduce model complexity, and employs three-dimensional convolutional neural network (3DCNN) for efficient feature extraction. Experimental results on real-world data validate its effectiveness in achieving accurate, private, and lightweight device-free localization.