Machine Learning Applications for Drones Security

Abstract

Drones, also known as Unmanned Aerial Vehicles (UAVs), are rapidly gaining popularity due to their attractive functions and steadily falling costs. Nowadays, drones are used for a variety of activities, including inspections, perimeter control, remote surveillance, and emergencies. Unfortunately, drones are an example of a classic dual-use technology that, while offering many advantages, could also be used for malicious purposes. For example, they could be used to record video or take pictures of restricted-access areas, halt airport traffic, or even carry weapons against specific targets. With such possible dreadful applications, many drone detection, countermeasures, and authentication techniques are widely investigated in both the academia and industry. This dissertation presents a research project focused on studying different techniques to detect drones, estimate their payload and authenticate them. We first present PiNcH, a drone detection method based on analyzing the wireless traffic exchanged between the drone and its remote controller. Our results prove that PiNcH can efficiently and effectively: (i) identify the presence of the drone in several heterogeneous scenarios; (ii) identify the current state of a powered-on drone, i.e., flying or lying on the ground; (iii) discriminate the movements of the drone; all the three scenarios with over 95% accuracy. Finally, PiNcH enjoys a reduced upper bound on the time required to identify a drone with the requested level of assurance. In the second project, we present Noise2Weight that leverages the acoustic fingerprint of the drone when carrying different payloads to infer on the specific weight of the payload leveraging its unique motor noise. we characterize how the differences in the thrust needed by a drone to carry different payloads affect the speed of the motors and the blades and, in turn, introduces significant variations in the resulting acoustic fingerprint. We show that it is possible to achieve a minimum classification accuracy of 98% in the detection of the weight of the payload carried by the drone, using an acquisition time of only 0.25 s. Finally, we present Drone-Mag that investigates fingerprinting drones via their unique unintentional magnetic emissions. Drone-Mag exploits the inherent non-idealities and imperfections present in drones’ electronic integrated circuits (ICs) that are introduced during their manufacturing process; hence they are hard to mimic or replicate. Drone-Mag is a passive, non-interactive, and privacy-preserving authentication solution and does not require software or hardware modifications to available drones. In particular, we addressed three main tasks: (i) Identification of 6 different drone brands—achieving a 100% classification accuracy; and, (ii) Authentication of 10 identical (same brand and model) drones—achieving a 99.9% classification accuracy; (iii) Rogue drone detection using autoencoders, with over 99% accuracy.

Date of Award	2023
Original language	American English
Awarding Institution	HBKU College of Science and Engineering