HandPass: Cross-Environment Static Spectral Signature Identification via Wi-Fi CSI

Abstract

Dynamic pattern identification and recognition (e.g., gestures, gait, and activities) via Wi-Fi Channel State Information (CSI) has been extensively explored. Despite the diversity of these approaches, most rely on closed-world evaluations, focus on recognizing inherent physiological attenuation and dispersions, or require continuous dynamic interaction. We empirically demonstrate the feasibility of extracting user static spectral signatures induced by near-field hand positioning across diverse domains, motivating us to develop HandPass-a novel contactless, crossenvironment static spectral identification framework. Specifically, following a comprehensive ablation study, we establish a tailored CSI preprocessing pipeline and propose the Masked Dual-Stream Disentangled Spectro Network (M-DDSNet). This architecture fuses amplitude and phase modalities while actively disentangling environmental multipath artifacts from the embedding of spectral signatures. Evaluations in cross-environment scenarios demonstrate a security gap (the margin between intra and interclass similarity) attaining 58.63% for known signatures and 46.91% for zero-shot identities. Moreover, an adversarial threat model assessment confirms that HandPass successfully purges over 40% of domain-specific noise from the signal’s latent representation. Ultimately, our model exhibits exceptional cross-environment generalization on unseen data, achieving a 100% True Accept Rate (TAR) and a 0.0% False Accept Rate (FAR).