Smartphone Proximity Detection Using WiFi and BLE Fingerprinting

In light of events such as the recent pandemic and many potential applications in fields such as the social sciences, healthcare, and architecture, the detection of interactions or proximity between people has become increasingly important. In this context, this paper investigates the limitations of a machine learning-based approach that detects the proximity of two devices based on the WiFi and BLE fingerprints of their radio environments. More specifically, (i) we compare the use of a rudimentary set of two features and an extended, more complex set of features, (ii) we investigate the use of separate classifiers that treat WiFi and BLE features separately, and (iii) we investigate whether using only one of the two communication technologies for detection could provide better results. In addition, we also try to use techniques such as undersampling and oversampling or their combination to deal with the highly imbalanced set of examples. Our results show that the use of a more complex set of features that can be subjected to further feature selection procedures can provide a performance benefit of about 4.6 percentage points. In terms of the communication technologies used, our results also show that using BLE alone always gives significantly worse results than using WiFi alone or WiFi and BLE together. On the other hand, there is no clear winner between using WiFi alone or combining WiFi and BLE, as both provide comparable results. Finally, our results also show that using under/oversampling helps in scenarios where the classification task is somewhat more complex, but not in those where the diversity between instances is low; thus, the classification problem is simpler.