Hyperface: a fabric that makes computer vision systems see faces everywhere

Adam Harvey, creator of 2012's CV Dazzle project to systematically confound facial recognition software with makeup and hairstyles, presented his latest dazzle iteration, Hyperface, at the Chaos Communications Congress in Hamburg last month.

Hyperface is a pattern designed to fool the most widely used "retail surveillance" technologies, in which shops use cameras and facial recognition to identify customers in order to gather intelligence on their shopping habits, including the psychological blind-spots that might make it easier to sell to them. It's intended to go on things that are not faces, to confuse computer vision systems about whether and where faces are present in a scene.


The pattern is studded with elements that facial-recognition algorithms identify as being parts of a face, causing the systems to make bad guesses about which face it's seeing and the sentiment being displayed by that face.


Adversarial facial recognition — in which the person being recognized does not want to be recognized by the system — is a lot harder than cooperative facial recognition (like Instagram filters that add graphic overlays to your face in realtime, for your own pleasure), because it assumes that the detected party isn't trying to confound the system. As with many analysis systems — from Google's Pagerank mining of inbound links to estimate the importance of online documents to stylometry that tries to identify anonymous texts by examining documents created by people who weren't using software to hide their writing styles — the fact that this works looking backwards on the stuff that predates its creation does not mean it can survive an arms race against its subjects once they discover its existence.

Hyperface will be more formally presented at Hyphen Labs' NeuroSpeculative AfroFeminism at this year's Sundance.

To emphasise the extent to which facial recognition technology changes expectations of privacy, Harvey collated 47 different data points commercial and academic researchers claim to be able to discover from a 100×100 pixel facial image – around 2.5% of the size of a typical Instagram photo. Those include traits such as "calm" or "kind", criminal tendencies like "paedophile" or "white collar offender", and simple demographics like "age" and "gender".


Research from Shanghai Jiao Tong University, for instance, claims to be able to predict criminality from lip curvature, eye inner corner distance and the so-called nose-mouth angle.

"A lot of other researchers are looking at how to take that very small data and turn it into insights that can be used for marketing," Harvey said. "What all this reminds me of is Francis Galton and eugenics. The real criminal, in these cases, are people who are perpetrating this idea, not the people who are being looked at."

Hyperface [Adam Harvey]

Retail Surveillance / Retail Countersurveillance [Adam Harvey/CCC]


Anti-surveillance clothing aims to hide wearers from facial recognition
[Alex Hern/The Guardian]