Data-Focused Methods for GPS Spoofing Detection
Mark Demore II, Patrick Sweeney, Joseph Curro, Brett Borghetti, Air Force Institute of Technology
Alternate Number 2
The United States Air Force (USAF) depends on the Global Positioning System (GPS) for countless essential systems for navigation and coordinated timing, but the signals it relies on are extremely insecure. GPS signals can be easily spoofed using commercially available hardware, such as Software-Defined Radios, and simple code. Existing detection methods for spoofing require specialized hardware and are not used widely. In order to address this problem, a lightweight, scalable detection solution is ideal. This work explores data-focused approaches in detecting GPS spoofing, with an emphasis on machine learning techniques, that make use of messages passed between avionics components in flight. Using data from four different flights, a variety of simulated spoofing attacks are applied that alter the reported latitude and longitude. Five deep learning methods, two classical machine learning methods, and one baseline threshold model are evaluated and compared across twelve different spoofing attack scenarios. A pipeline is established in creating spoofed data and feeding it into machine learning models, as well as evaluating performance and comparing results. Analysis of the results show promising efficacy of Convolutional Neural Networks (CNNs) and Multilayer Perceptrons (MLPs), but suggest that a threshold or physics-based model may be just as effective in the given scenarios.