| Summary: | Today, with the advent of IoT and the resulting fragmentation of wireless technologies, they bring not only benefits, but also concerns. Daily, several individuals communicate with each other using various communication methods. Individuals use a variety of devices for innocuous day-to-day activities; however, there are some malicious individuals (dishonest agents) whose aim is to cause harm, with the exfiltration of information being one of the biggest concerns. Since the security of Wi-Fi communications is one of the areas of greatest investment and research regarding Internet security, dishonest agents make use of side channels to exfiltrate information, namely Bluetooth. Most current solutions for anomaly detection on networks are based on analyzing frames or packets, which, inadvertently, can reveal user behavior patterns, which they consider to be private. In addition, solutions that focus on inspecting physical layer data typically use received signal power (RSSI) as a distance metric and detect anomalies based on the relative position of the network nodes, or use the spectrum values directly on models classification without prior data processing. This Dissertation proposes mechanisms to detect anomalies, while ensuring the privacy of its nodes, which are based on the analysis of radio activity in the physical layer, measuring the behavior of the network through the number of active and inactive frequencies and the duration of periods of silence and activity. After the extraction of properties that characterize these metrics,an exploration and study of the data is carried out, followed by the use of the result to train One-Class Classification models. The models are trained with data taken from a series of interactions between a computer, an AP, and a mobile phone in an environment with reduced noise, in an attempt to simulate a simplified home automation scenario. Then, the models were tested with similar data but containing a compromised node, which periodically sent a file to a local machine via a Bluetooth connection. The data show that, in both situations, it was possible to achieve detection accuracy rates in the order of 75 % and 99 %. This work ends with some ideas of resource work, namely changes in the level of pre-processing, ideas of new tests and how to reduce the percentage of false negatives.
|
|---|