Semi-Supervised Open-Set Through-Wall User Identification via MIMO Radar Respiratory Signals

With the rapid evolution of 6G communications and the Internet of Things (IoT), wireless sensing–based identity recognition is emerging as a key enabler for intelligent security and smart environments. However, existing popular mmWave radar–based methods often suffer from severe signal attenuation and unreliable sensing performance in non-line-of-sight (NLOS) environments. To overcome these limitations, this paper proposes Resp-TWOSID, a through-wall semi-supervised open-set respiration-based identification framework leveraging low-frequency MIMO radar. By utilizing low-frequency signals, Resp-TWOSID reliably captures fine-grained respiratory patterns through thick masonry walls up to 26 cm, enabling robust user identification. Additionally, conventional identification methods typically rely on fully supervised learning, which demands extensive data collection and labeling, and cannot recognize unknown users during deployment. To address these challenges, Resp-TWOSID introduces the dynamic open-set semi-supervised learning with self-supervised consistency (DOS3L-SSC) algorithm. DOS3L-SSC explicitly models the unknown category as the (C+1)-th class and employs a dynamic first-in-first-out (FIFO) queue to select reliable unknown samples for sustained open-set learning continually. It further enforces joint open-set and closed-set consistency learning to stabilize predictions under data perturbations and introduces a self-supervised feature-consistency loss to enhance representation robustness while efficiently exploiting hard unlabeled samples. Extensive experiments demonstrate that Resp-TWOSID consistently outperforms state-of-the-art approaches, achieving superior accuracy for known-user identification and high-confidence rejection of unknown users. These results highlight its strong potential for deployment in security-critical and privacy-sensitive 6G sensing environments.