DBLoc: A Lightweight and Universal BFI-Enabled Deep Learning Framework for Wi-Fi Localization

Wi-Fi-based passive indoor localization has gained prominence owing to its high accuracy and ease of deployment in global positioning system (GPS)-denied environments. However, channel state information (CSI)-based systems face challenges, including high data acquisition requirements, significant computational overhead, and limited transferability. In this article, we introduce DBLoc, a Wi-Fi localization system that leverages beamforming feedback information (BFI), a novel attribute provided by modern Wi-Fi hardware. BFI's clear-text transmission and stable characteristics make it an ideal choice for localization tasks. We prove that BFI provides a lightweight alternative to CSI, significantly reducing both data acquisition and storage requirements. Compared with traditional deep learning frameworks using convolutional networks, DBLoc employs a pruning-based residual architecture to reduce computational overhead, achieving an inference cost of only 175.7 MFLOPs, thus optimizing performance within an edge-deployment budget. To enable transferability that surpasses current meta-learning approaches, DBLoc incorporates a virtual-domain-based meta-learning algorithm, ensuring robust performance with minimal target-domain data. In addition, a spatial-encryption mechanism is proposed to safeguard the BFI-based model from eavesdropping. Extensive evaluations demonstrate that DBLoc achieves a median localization error of approximately 0.5 m while significantly reducing localization accuracy for unauthorized attackers.