Adversarial Detection: Attacking Object Detection in Real Time

TL;DR

This paper introduces the first real-time online adversarial attack on object detection models, demonstrating high success rates in dynamic environments, which poses significant security concerns for robotic applications.

Contribution

It presents three novel real-time attack methods that can fabricate bounding boxes for nonexistent objects with high success rates, advancing security testing of robotic perception systems.

Findings

Achieved about 90% success rate within 20 iterations

Developed three real-time attack methods

Demonstrated vulnerability of object detection in dynamic settings

Abstract

Intelligent robots rely on object detection models to perceive the environment. Following advances in deep learning security it has been revealed that object detection models are vulnerable to adversarial attacks. However, prior research primarily focuses on attacking static images or offline videos. Therefore, it is still unclear if such attacks could jeopardize real-world robotic applications in dynamic environments. This paper bridges this gap by presenting the first real-time online attack against object detection models. We devise three attacks that fabricate bounding boxes for nonexistent objects at desired locations. The attacks achieve a success rate of about 90% within about 20 iterations. The demo video is available at https://youtu.be/zJZ1aNlXsMU.