Non-targeted Adversarial Attacks on Object Detection Models

Adversarial attacks involve introducing imperceptible noise into images to induce incorrect outputs from models, serving as a means to assess model security. Object detection, as a crucial task in the field of computer vision, has garnered significant attention regarding its security. NMS mechanisms in object detection models suppress low-confidence detection boxes. So, most current adversarial attacks on object detection are for targeted attacks. However, some object detection datasets designed for specific scenarios contain only one class, rendering targeted attacks unsuitable. If the original image's category is very different from the target category, it will be hard to attack. This is true even if a large perturbation is made. This paper introduces a non-targeted adversarial attack method capable of effectively compromising object detection models. We conduct experiments on a power dataset from the national power grid, demonstrating promising results. Furthermore, we prove that utilizing the proposed UnTargeted Attack (UTA) method from this paper enables the generation of more stealthy perturbations with fewer iterations.