Location-independent adversarial patch generation for object detection

Object detection models are at the core of various computer vision tasks and have shown excellent performance on public datasets, but they also inherit disadvantage neural networks that vulnerable to adversarial example attacks. Adversarial patches specific forms examples that, as in previous studies, can only make objects (such pedestrians traffic signs), not all objects, disappear. In addition, a patch must be placed every object deceive detector. To solve above problems, we propose location-independent generation method attack range detected with single patch. By attacking confidence loss detector, creatively assign greater weight foreground region, which makes its decrease faster effectively guides convergence direction training process. Furthermore, glue randomly images them less sensitive location during training. Experimental results indicate generated using our proposed restricted areas image provide minimum recall 29.5%.