Machine Learning and Coding Schemes for Fault and Cyber-attack Detection of Unmanned Aerial Vehicles
Year: 2024
Type: dissertation
Abstract: The main objective of the thesis is to effectively detect and isolate faults and detect cyber-attacks and isolate faults from cyber-attacks for UAVs. \nIn our first result, we leverage SVMs as a supervised machine-learning technique and Kalman filtering as a proven model-based method. We show that these methods can effectively detect and isolate faults and cyber-attacks. This thesis emphasizes the importance of hybrid approaches to fault and cyber-attacks \nSVMs perform equally well for the FDI cyber-attack detection on the control inputs of the ground control system. We propose to train two-class SVM only with the healthy dataset while adding a small bias to control the input of randomly selected observations. To isolate faults from cyber-attacks, we propose to use two Chi-squared, one on the plant side and the other on the controller side that reliably isolates faults from cyber-attacks. \nIn our second result, we investigate the problem of stealthy caber-attack detection, where the attackers exploit full knowledge of the system to launch undetectable cyber-attacks. We study replay and covert cyber-attacks, and to make them detectable, we use coding techniques. \nWe extend the two-way coding for MIMO discrete-time systems and study its properties using Kalman filtering for stealthy cyber-attack detection. We show that for an attacker to remove the cyber-attack impact from coded output, it is necessary to know the coding. We propose a new coding scheme using randomly generated numbers on the controller side that effectively converts coded inputs and outputs to time-varying random numbers.
Author: Gharibian Saaki, Edward
Institution: Concordia University
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7001932422
