Anomaly Detection and Secure Position Estimation Against GPS Spoofing Attack: A Security-Critical Study of Localization in Autonomous Driving

Abstract

For advanced autonomous driving (AD) systems, localization is highly critical for safety. Recent results show that GPS is vulnerable to spoofing attacks, and it is not clear whether the current localization is secure enough against advanced GPS spoofing attacks. In this paper, a systematic study regarding the security of the localization under GPS spoofing is explored for safe and reliable AD. First, a novel and robust GPS adversarial attack design method is proposed to defeat the principle of the multi-sensor fusion algorithm and lead to wrong position. It can cheat the widely used Chi-squared detector in Kalman filter and cause the vehicle to drive off the road, posing greater challenge on safe driving. Second, a real-time Long Short-Term Memory (LSTM) attack detector is developed to detect the serious attack effectively. When the attack is detected, a multi-information fusion method based on the lateral direction localization from camera and map using Unscented Kalman filter is proposed to defend against the GPS attack and provide accurate position estimation for automated vehicles to drive on roads safely. The proposed method is validated in various scenarios in Carla simulator and a real-word driving dataset to demonstrate its effectiveness in timely GPS attack detection and secure position estimation. The results show that the LSTM-based detection method has best performance compared to the state-of-the-art detection approaches. The position estimation for attack defense is effective and robust in different driving scenarios, ensuring safe and reliable AD in closed-loop form.