Unveiling access point signal instability in WiFi-based passive sensing

In the last few years, there has been a growing interest on the usage of passive WiFi signals for detecting and classifying human behavior activities. In most generality, the idea underlying most state of the art approaches is to monitor changes in the Channel State Information (CSI) or equivalent spectral metrics, caused by the different multi- path conditions (and their temporal variations) induced by human activity. The (implicit) underlying intuitive assumption is that, in a static environment, i.e. with no moving humans, the multi-path conditions do not vary, and hence the CSI gathered at the passive receiver side should remain substantially stable in time. Goal of this paper is to document, with an extensive set of experimental measurements, a perhaps surprising result: with modern multi-antenna 802.11n/ac access points, such an intuitive assumption does not seem to hold anymore. Rather, the measured CSI time pattern, which differs for different access point brands, appears to be affected by access points' implementation artifacts. Indeed, our measurements unveil frequent fluctuations of the CSI pattern among multiple "modes". Albeit still preliminary and, at least in part, still requiring a more thorough understanding (that only the reverse engineering of the access points' implementation may ultimately provide), our measurements unveil a phenomenon which may play havoc with several CSI-based human behavior detection approaches proposed in past works, and especially with those based on CSI measured on data frames, opposed to those based on relatively more stable (but still to be reconsidered in sight of our results) Beacon frames.