Pick-Object-Attack: Type-specific adversarial attack for object detection

Many recent studies have shown that deep neural models are vulnerable to adversarial samples: images with imperceptible perturbations, for example, can fool image classifiers. In this paper, we present the first type-specific approach generating examples object detection, which entails detecting bounding boxes around multiple objects in and classifying them at same time, making it a harder task than against classification. We specifically aim attack widely used Faster R-CNN by changing predicted label particular an image: where prior work has targeted one specific (a stop sign), generalise arbitrary objects, key challenge being need change labels of all instances type. To do so, propose novel method, named Pick-Object-Attack. Pick-Object-Attack successfully adds perturbations only object, preserving other detected image. terms perceptibility, induced method very small. Furthermore, examine effect attacks on detection downstream task, captioning; show modify types leads obvious changes captions, from our constrained much less apparent.