Smartphone apps for exposure notification and contact tracing have been shown to be effective in controlling the COVID-19 pandemic. However, Bluetooth Low Energy tokens similar to those broadcast by existing apps can still be picked up far away from the transmitting device, making them ineffective for COVID-19–related proximity detection applications in some scenarios. In this thesis, we present two new classes of methods for detecting whether or not two devices are in immediate physical proximity, i.e. 2 or fewer meters apart, as established by the U.S. Centers for Disease Control and Prevention (CDC). One method uses Wi-Fi RSSI fingerprints, and the other uses magnetometer traces. Both of these types of data can be recorded by almost all modern smartphones. Our ultimate goal is to enhance the accuracy of smartphone-based exposure notification and contact tracing systems.

We first design a set of binary machine learning classifiers that take as input pairs of Wi-Fi RSSI fingerprints. These classifiers distinguish between pairs of RSSI fingerprints recorded 2 or fewer meters apart and pairs recorded further apart but still in Bluetooth range. We empirically verify that a single classifier cannot generalize well to a range of different environments with vastly different numbers of detectable Wi-Fi Access Points (APs). However, specialized classifiers, tailored to situations where the number of detectable APs falls within a prescribed range, are able to detect physical proximity significantly more accurately. As such, we design three classifiers for situations with low, medium, and high numbers of detectable APs. We characterize their balanced accuracy for proximity detection to be between 66.8% and 77.8%.

Next, we design a second set of binary machine learning classifiers, which take as input pairs of 10-second traces of smartphone magnetometer readings. These classifiers distinguish between pairs of trace segments for which the two recording devices are 2 or fewer meters apart for at least 75% of the segment duration and pairs for which the two devices are further apart but still in Bluetooth range. We first evaluate these classifiers’ performance on traces from the MagPIE dataset, a dataset for evaluating magnetometer-based localization algorithms; we characterize their balanced accuracy for homogeneous-device proximity detection to be between 89.3% and 93.3%. We show that our classifiers can generalize well to different buildings whose traces are not present in their training data. We introduce a simple method of compensating for different magnetometer biases in heterogeneous devices, and evaluate our approach with this added mitigation by training and evaluating classifiers on different disjoint subsets of traces from 4 different smartphone models. We characterize their balanced accuracy for heterogeneous-device proximity detection with non–tilt-compensated traces to be between 93.8% and 96.9%; these results indicate that our classifiers can generalize well to devices whose traces are not present in their training data.