Adam Harvey has updated CV Dazzle
with all new patterns. (Previously.) The new patterns are aimed at defeating newly developed face detection algorithms from a variety of vendors including OpenCV, VeriLook, and Apple.
His main suggestions are:
- Use contrasting makeup, light colors on dark skin, and vice versa.
- Obscure the nose bridge.
- Partially obscure an eye.
- Modify the contrast, tonal gradients, and spatial relationships of light and dark areas of the face.
- Try obscuring the shape of the head.
- Make the face appear asymmetrical.
Philip McCarthy‘s Pareidoloop overlays semitransparent polygons until a human face is found. (github link) It’s inspired by Roger Alsing‘s Evolution of Mona Lisa, where he applied the same technique recreate the Mona Lisa.
(google code link)
A more interesting variation of this technique would be throw the images through a classifier trained to recognize various religious images, print them on toast, and the post the toast to eBay Virgin Mary toast style.
Adam Harvey played with different makeup patterns to defeat facial detection software. Called CV Dazzle – named after the dazzle camouflage of World War I – it basically asymmetric white and black eye black that is intended to confuse the location of the eyes in relation to the mouth.
While interesting, it isn’t a complete privacy system. Even wearing CV Dazzle, the wearer’s image is still stored. I recall seeing online a hoodie that had high intensity IR LEDs sewn around the opening in an attempt to blind cameras, which I would think either works extremely well, or extremely poorly (i.e. providing more light to the wearer’s face, and thus a sharper image) depending on the orientation of the LEDs. Because of this, it seems like going with an old fashioned mask is better solution, unless you’re afraid of being arrested on a trumped up and dubious
anti-mask law, in which case perhaps CV Dazzle isn’t such a bad idea in order circumvent mass surveillance.