Detection of Multiple Small Biased GPS Spoofing Attacks on Autonomous Vehicles Using Time Series Analysis

This research introduces an algorithm to identify GPS spoofing attacks in Autonomous Vehicles (AV). It uses data from onboard sensors such as speedometers and gyroscopes, which are integrated and analyzed using a Neural Network (NN). This network predicts the vehicle's future displacement and compares these predictions with GPS data to identify potential spoofing attacks such as turn-by-turn, stop, and overshoot incidents. Additionally, the same sensor data is evaluated using an analytical model based on the vehicle's dynamic equations to assess its position and speed against GPS information. To facilitate real-time detection, a threshold is pre-established from clean datasets, which determines the largest expected differences between sensor readings and GPS data. This threshold is then used for ongoing real-time assessments to detect spoofing activities. Moreover, the algorithm can detect multiple small biased attacks, incremental attacks that may not initially exceed the established threshold but eventually result in significant discrepancies in GPS and Inertial Measurement Unit (IMU) reported displacement and speeds. This detection is facilitated through time series analysis at 25 and 50 s intervals to build a profile of data errors and distribution to predict the probability of such attacks. To evaluate the algorithm's effectiveness, five different test datasets depicting four types of spoofing scenarios—turn-by-turn, overshoot, stop, and multiple small biased attacks—were created using data from the publicly accessible Honda Research Institute Driving Dataset (HDD). The analysis shows that the model accurately detects these types of attacks with average accuracies of 98.62$\pm$1%, 99.96$\pm$0.1%, 99.88$\pm$0.1%, and 95.92$\pm$1.7% respectively.