Exploiting Wireless Communications for LocalizationBeyond Fingerprinting

Resumen

The field of Location-based Services (LBS) has been steadily growing over the last decade. There are several factors contributing to this growth, including growing interest in tracking of fitness and other activities, robotics and eHealth. Each field has different requirements for technology, precision of positioning, computational resources, as well as for privacy and security. There are several available solutions in these fields however there are unresolved challenges and issues. This dissertation evaluates currently employed privacy related measures in Indoor Positioning Systems (IPS). The main focus is on the most widespread wireless technology around us, Wi-Fi. Wi-Fi networks are ubiquitous and the majority of our devices possess the capability to use these networks. As such the question of privacy comes to mind: are the networks all around us safe, and is it possible to compromise our privacy without our knowledge. We address non-cooperative user tracking by exploiting regular wireless com-munications. User Equipment (UE) uses management frames (specifically Probe Requests) of Wi-Fi for scanning the environment for Wi-Fi Access Points (AP) in the proximity. These management frames are not encrypted and contain enough information required for fingerprinting devices. That is despite the widely adopted practice of Media Access Control (MAC) address randomization. Following presence detection, the next step is the introduction of an algorithm capable of estimating the occupancy of a room based on the passive sniffing of Wi-Fi management frames. This is done by measuring the Received Signal Strength Indicator (RSSI) of the received packets from a known device close to the edges of the room. Furthermore, the measured RSSI can be used as a threshold for filtering devices that are out of range. Previous research shows that management frames of Wi-Fi are a threat to location privacy. The threat ranges from just presence detection of users, to possibly compromising the exact location of users in case of the employment of more sniffers. Room occupancy can be determined without the need for user information, and as such, it can be useful in energy regulations of smart buildings. The second point of this thesis is the exploration of methods to reduce computa-tional requirements of machine learning (specifically neural networks), and positioning algorithms. This is split into two parts: reduction of memory requirements of neural networks and usage of interpolation techniques to improve fingerprinting methods in IPS. The memory requirements of neural networks were reduced by changing the data type of weights from a single precision floating point format to just half precision. This simple change resulted in a reduction of memory requirements by 50 %. In this case, the accuracy drop was negligible. However, reducing the data type further, without any retraining provided mixed results. Overall, the findings were that single precision is unnecessarily accurate for neural networks, and half precision is a good choice for weights, without any sacrifice in accuracy. The second part of the optimization was focused on interpolations of RSSI based Radio Map (RM) for IPS. As means of evaluation, the time required for data collection, accuracy, and inference of position using Nearest Neighbors (NN) were used. For interpolation, Gaussian Process Regression (GPR), linear interpolation, and a combination of both were used. From the accuracy point of view, the interpolation did not really increase accuracy. However, the takeaway is that the grid of 0.5 m is much too dense and actually reduces the accuracy while also increasing the inference time. However, it was clear that it is possible to drastically reduce the inference time by using RM with a single sample per Reference Position (RP), as it dramatically lowers the amount of reference samples for NN and all of the samples are created from information contained in many samples. Finally, during this thesis several datasets of probe requests were created, as well as data analysis scripts and packet sniffer firmware for ESP32 microcontroller. All of these are publicly available for future research, and to improve the reproducibility and replicability of the work. The thesis provides insight into non-cooperative tracking of users, presence detection and occupancy estimation, as well as into optimizations of machine learning algorithms and indoor positioning algorithms employing RSSI fingerprinting.