Abstract
Recent studies proved that deep learning approaches achieve remarkableresults on face detection task. On the other hand, the advances gave rise to anew problem associated with the security of the deep convolutional neuralnetwork models unveiling potential risks of DCNNs based applications. Evenminor input changes in the digital domain can result in the network beingfooled. It was shown then that some deep learning-based face detectors areprone to adversarial attacks not only in a digital domain but also in the realworld. In the paper, we investigate the security of the well-known cascade CNNface detection system - MTCNN and introduce an easily reproducible and a robustway to attack it. We propose different face attributes printed on an ordinarywhite and black printer and attached either to the medical face mask or to theface directly. Our approach is capable of breaking the MTCNN detector in areal-world scenario.