Mobile User Trajectory Anomaly Detection via Unsupervised Channel Charting

Joint communications and sensing is a key feature in beyond 5G cellular systems. In this letter, we considered the mobile user trajectory anomaly detection problem using merely channel state information (CSI) and its associated timestamp. For this, we propose an unsupervised trajectory anomaly detection framework, which consists of (i) collaborative collection of CSI and timestamp by distributed base stations (BSs); (ii) CSI feature extraction and unsupervised trajectory representation learning via multi-point channel charting (MPCC) and (iii) Anomaly detection algorithm via Fréchet similarity in the trajectory representation space. We show that MPCC using Laplacian eigenmaps with power angular profile feature and timestamp yields differentiated and robust trajectory representations. Simulations in an urban outdoor scenario demonstrate that the proposed CSI-based method achieves the same detection accuracy as GNSS-based trajectory anomaly detection, while preserving user privacy by using pseudonymous identifiers and location-free trajectory representations.