Learning Privacy-Preserving Channel Charts

Channel charting (CC) employs dimensionality reduction to channel state information (CSI) collected in multi-antenna wireless systems to provide a low-dimensional repre-sentation of the radio environment. In CC, wireless users are assigned pseudo-locations on the channel chart, allowing for localization-related services to be delivered without requiring actual user locations to be estimated. Intuitively, the use of pseudo-location on a channel chart can be perceived as a privacy-preserving feature (user privacy). In practice, however, the study of user privacy requires a more careful examination of the assumptions under which channel charting operates. Besides user privacy, an additional concern may be the exposure of raw CSI measurements for the learning of channel charts which may disclose proprietary information held by hardware vendors to external entities (vendor privacy). Starting from these observations, in this paper, we provide a systematic study of the privacy threats associated with CC, with the aim of obtaining a nuanced understanding of the privacy implications inherent to CC. We address two main learning architectures and discuss two privacy mechanisms - differential privacy (DP) and homomorphic encryption (HE). In the studied scenarios, we evaluate via numerical examples the emerging trade-offs between the channel charting performance (quantified via Kruskal's stress) and the provided privacy level.