Detection of Multiple Small Biased GPS Spoofing Attacks on Autonomous Vehicles

Abstract

This paper introduces an algorithm designed to detect GPS spoofing attacks in Autonomous Driving Systems (ADS). The algorithm combines data from in-vehicle sensors, including the speedometer and gyroscope. This data is then fed into a Deep Neural Network (DNN) to predict the vehicle’s displacement between timestamps. These predictions are continuously compared with GPS-provided displacement and velocity to detect the turn-by-turn, stop, and overshoot GPS spoofing attacks. Also, the same data is processed through an analytical model that uses the vehicle’s dynamic movement equations to compute its position and velocity and to juxtapose them with GPS-provided position and velocity to detect the aforementioned spoofing attacks. A threshold is predetermined by using clean datasets to calculate the maximum possible deviation between in-vehicle sensor and GPS provided position/displacement and speed. Subsequently, spoofing attack detection uses this threshold for its continuous, real time comparison. Additionally, the proposed algorithm is able to detect a sequence of multiple biased attacks that fall below the mentioned threshold but can create a large deviation between the position/displacement and speed provided by GPS and Inertial Measurement Unit (IMU). To validate the effectiveness of the work, datasets simulating four distinct spoofing scenarios such as turn-by-turn, overshoot, stop and multiple biased attacks were generated using real-world data from [1]. The analyses shows that the Data-driven model successfully identifies the turn-by-turn, stop, overshoot and multiple biased spoofing attacks with the accuracies of $100 \%$, $\mathbf{9 9. 9 5 \%}, \mathbf{9 9. 9 1 \%}, \mathbf{9 7. 8 3 \%}$, respectively.

Topics & Concepts