Poltergeist: Acoustic Adversarial Machine Learning against Cameras and Computer Vision

Autonomous vehicles increasingly exploit computer-visionbased object detection systems to perceive environments and make critical driving decisions. To increase the quality of images, image stabilizers with inertial sensors are added to alleviate image blurring caused by camera jitters. However, such a trend opens a new attack surface. This paper identifies a system-level vulnerability resulting from the combination of the emerging image stabilizer hardware susceptible to acoustic manipulation and the object detection algorithms subject to adversarial examples. By emitting deliberately designed acoustic signals, an adversary can control the output of an inertial sensor, which triggers unnecessary motion compensation and results in a blurred image, even if the camera is stable. The blurred images can then induce object misclassification affecting safety-critical decision making. We model the feasibility of such acoustic manipulation and design an attack framework that can accomplish three types of attacks, i.e., hiding, creating, and altering objects. Evaluation results demonstrate the effectiveness of our attacks against four academic object detectors (YOLO V3/V4/V5 and Fast R-CNN), and one commercial detector (Apollo). We further introduce the concept of AMpLe attacks, a new class of system-level security vulnerabilities resulting from a combination of adversarial machine learning and physics-based injection of information-carrying signals into hardware.