Wi-Gait: Pushing the limits of robust passive personnel identification using Wi-Fi signals

Abstract

Personnel identification plays a crucial role in many security applications, where the knowledge factor, such as a personnel identification number (PIN), constitutes the most popular personnel identification element. Meanwhile, thanks to the pervasive Wi-Fi infrastructure, personnel identification enabled by wireless sensing is gaining increasing attention with the advantages of non-intrusiveness, privacy-preserving, and anti-counterfeiting. In particular, the popularity of the fine-grained Wi-Fi channel state information (CSI) allows us to identify people via gait recognition. However, existing systems still have multiple limitations: (1) heavily rely on the strong assumptions of walking conditions; (2) require the environment to remain unchanged, especially the device placement; (3) extract low-level gait features for personnel identification. To address the above issues, our paper proposes Wi-Gait, a gait-based personnel identification system, and the contribution is threefold. First, we customized a novel deep learning model to extract unique gait features and achieve high accuracy in personnel identification. Second, thanks to our designed model, we can remove the dependency on walking cofactors and device placement and make Wi-Fi gait-based identification more realistic. Third, we evaluated the performance using the most popular Wi-Fi gait dataset, i.e., Widar 3.0. Extensive experiments show an average identification accuracy of 92.9% for ten users under various complex conditions.