Computer vision is a rapidly advancing technology that relies heavily on camera sensors to provide input for Machine Learning (ML) models to make decisions. It is confirmed to play a critical role in various futuristic applications, such as advancements in self-driving vehicles, autonomous & target-tracking drones, parking assistance, and collision avoidance systems. However, with the increasing prevalence of hardware-level sensor hacking, even camera sensors are susceptible to being compromised. This experimental paper proposes the idea of sensor hacking against Machine Learning capabilities of vehic-ular Computer Vision (CV) using Electromagnetic Interference (EMI). A mid-range EMI intrusion device is developed to disrupt computer vision systems' accuracy and supervisory capabilities. The evaluation examines the impact of sensor hacking on camera sensors crucial to obstacle identification models reliant on live feeds, comparing decision-making capabilities with and without sensor tampering to assess the overall effect. Our results show that EMI significantly affects camera sensor performance, reducing accuracy and frame rates in machine learning-based object detection systems. These findings underscore the vulnerability of camera sensors to sensor hacking and highlight the need for improved security measures to safeguard against such attacks in computer vision systems.
Undermining Live Feed ML Object Detection Accuracy with EMI on Vehicular Camera Sensors
2024-06-24
2404692 byte
Conference paper
Electronic Resource
English