Vulnerability of Privacy Visor Used to Disrupt Unauthorized Face Recognition

Abstract

This work studies a vulnerability in privacy visors, new wearable devices that aim to prevent unauthorized face recognition from being performed. Although the use of a privacy visor assumes that the detectors' targets are uncovered bare faces, it is not hard to detect the privacy visor itself. To quantify the effects of the disruption and the vulnerability, we conducted experiments involving two major face-recognition algorithms, namely a method based on convolutional neural networks and a method that aims to identify coordinates of facial landscapes. Our experiments were able to demonstrate that using a privacy visor can reduce the mean face-recognition rates for both algorithms. However, they are less effective if faces with privacy visors are used in training. Faces with privacy visors is detected at a rate of 42.28 % on average.