Exploiting class vulnerabilities: targeted mislabeling attack in object detection

Object detection is the task of determining the position, size, and class of objects in an image. Recently, object detectors based on Deep Neural Networks (DNNs) have become a de facto standard due to their superior detection performance. However, they inherit the vulnerability of DNNs to perturbations that are imperceptible to humans, and may fail to detect objects in images containing such perturbations. To prevent attacks based on this vulnerability, it is crucial to know what kind of false identifications can be caused by such perturbations. To this end, in this paper, we propose a novel attack method named Targeted Mislabeling Attack (TMA), which causes objects to be misidentified as a specific class, and utilize TMA to achieve a Universal Targeted Mislabeling Attack (UTMA) that generates universal perturbations independent of input images. Besides, we propose a novel metric to evaluate the success rate of TMA. The experimental results show that an object detector is vulnerable to the proposed attacks.