Priest: Adversarial Attack Detection Techniques for Signal Injection Attacks

Abstract

Machine learning is widely used for autonomous driving because it can recognize surrounding circumstances feasibly from sensor and determine appropriate actions. Most of these sensors are based on micro-electro-mechanical systems (MEMS), which enable autonomous vehicles to judge objects in conjunction with object-detection algorithms. However, recent studies have shown that MEMS are vulnerable to signal-injection attacks, in which the input images are manipulated to force the object detection algorithms to misclassify the results. These attacks can be critical in the wild because they deteriorate state-of-the-art detection techniques, dropping their detection rates until the objects would no longer be detected at all. In this paper, we propose Priest, a novel detection method against prior signal-injection attacks. Priest uses the similarity of pixel values between two consecutive images. Using only two images ensures a low computational cost. According to our performance analysis, Priest detects state-of-the-art signal-injection attacks in real-time with 99% accuracy on average, achieving practical autonomous driving security.

Keywords

• Autonomous vehicle
• adversarial attack detection