Remote Attacks on Drones Vision Sensors: An Empirical Study

Abstract

Vision systems applied to drones, automatic vehicles, and robots have become an increasingly popular sensing method. However, vision sensors that make up these systems are vulnerable to malicious input attacks, which can lead to serious consequences. Privious work on attacking cameras of automatic vehicles shows that lasers can cause failure of camera-based functionalities, but it lacks analysis of the results and does not conduct experiments in the actual scenarios. In this article, a laser-based attack on cameras and binocular vision sensors of drones is presented. First, we propose a threat model that describes how an adversary attacks the drone then perform feasibility analysis of the attack from theory and practice. Next, we design multi-variable experiments in the lab to systematically study the effectiveness of the attack, and further analyze how each variable affects the results. To get intuitive and fine-grained results, multidimensional image similarity is used to measure the effects. In particular, experiments in the actual scenarios are carried out, and results show that the attack can make obstacle avoidance, target recognition and tracking completely failed. Finally, lightweight countermeasures based on hardware and software are proposed to improve sensor resilience against the attack.