HyperFace: Emerging Strategies for Obfuscating Computer Vision Algorithms
HyperFace is a new type of camouflage to obfuscate computer vision algorithms. It is being designed to decrease the efficiency and accuracy of automated facial recognition. As an obfuscation strategy HyperFace’s primary goal is to introduce face-like noise into the visual wavelength signal domain by displaying maximally activated false-face regions that are not perceivable as faces to a human observer. As a countersurveillance item, HyperFace can be used as a decoy in combination with CV Dazzle to allow the wearer’s true face to become hidden in the background of higher-confidence face scores.
In any automated facial recognition system, the first step is to isolate the facial region using a face detection algorithm. The three most widely used approaches for 2D face detection in the visible spectrum are haarcascades, histogram of oriented gradients (HoG), and convolutional neural networks (CNN). The first iteration of the HyperFace project targets haarcascade detectors, which were chosen as starting point because of their ubiquity in consumer applications and previous research from CV Dazzle (2010). Future versions will target HoG and CNN detectors, which are much more effective at detecting faces in varying poses, illumination, expression, and skin tone. The first prototype from February 2017, shown below on a printed textile displayed by Hyphen Labs, has activated 9 false-positives.
In this example (Fig. 1) the true face is still detected and HyperFace only adds a small amount of computational excess for processing the image. But when combined with a strategy such as CV Dazzle (Fig. 2) the true face is now hidden and the face detection system is fooled.
Figure 1: HyperFace demonstration by Hypen-Labs. Face regions detected using haarcascade frontal face default detector. March 14, 2017.
Figure 2: HyperFace demonstration by Hypen-Labs with superimposed CV Dazzle pattern. Face regions detected using same haarcascade frontal face default detector settings as in Figure 1.
In addition to the technical goal of the HyperFace project, it also aims to engage the public in dialogue about the potential risks of computer vision by using a non-technical and analog approach. HyperFace, as a fashion accessory, becomes accessible to groups outside of academic research communities. Already, the project has seen a large amount of interest online and offline from large media and fashion outlets.
In summary, the HyperFace project aims to increase the difficulty of automated face recognition systems by introducing noise, then using this noise source to provide a camouflaging background for the true face to escape detection.
Stay in Touch
We'll send occasional announcements about conference details and follow-up initiatives.
International Program and Organizing Committee:
Paul Ashley, Anonyome Labs Benoît Baudry, INRIA, France Finn Brunton, New York University Saumya Debray, University of Arizona Cynthia Dwork, Harvard University Rachel Greenstadt, Drexel University Seda Gürses, Princeton University Anna Lysyanskaya, Brown University Helen Nissenbaum, Cornell Tech & New York University Alexander Pretschner, Technische Universität München Reza Shokri, Cornell Tech