For years, companies and scientists have build their sets with training data for systems for biometric facial recognition "wild" without the consent of those affected. Adam Harvey and Jules Laplace, who live as researchers and artists in Berlin, make this bustle in ethical and legal gray zones now at least more transparent. You have "Exposing.Ai" A search engine develops that users can see if their Flickr images were abused for such purposes.
The project recently online allows search with Flickr identifiers such as the username, the NSID assigned by the Picture Platform or a Photo ID. Only if exact agreement with corresponding data is found in the integrated monitoring databases, results are displayed. The operators of the platform store themselves no search data and do not give them further. The displayed photos will then be directly from Flickr.COM loaded, copies of it not kept.
Interested parties can also search for photos of themselves who have added third parties and put on Flickr by using a hashtag. It is conceivable about a short for visited events or private celebrations like "#MybirthDayParty". But the makers indicate considering that this shape of the search will last a little long: "Each photo can contain dozens of tags, which leads to millions of other data sets for the search."
High abuse and loss potential
"People have to realize that some of their intimate moments were made to weapons", Liz O’Sullivan, Technology Director at the Burger Rights Organization Surveillance Technology Oversight Project (STOP), the initiative against the New York Times. The activist has on the design of exposing.Ai worked. It was originally planned to use automated facial recognition for the search engine for the search engine. However, the team remedied again: the abuse and loss potential was too high.
Harvey had at a conference on the basis of findings from his leadership project "Megapixels" reported on the procedure of the Jagery according to facial shots to remove the sometimes high error rates of technology. Microsoft for example, for the Celeb database, simply served on pictures of celebrities as well as fewer known people on the Web, the Duke University for that "Multi-Tracking Register" DukemtMC recordings of students with telephoto lens made of a window of the institute. For "Brainwash" Had the causes even diverted image data of a video livestream from a cafe in San Francisco.
These databases have now been officially turned off the coarse division, the artist’s artist: "But you can not really get her out of the net." So the contents were still circulating in "Academic torrents" In peer-to-peer networks "around the world". It should be noted that parts of it take over from the Chinese army and now used for the vacuum of the Muslim minority in the Autonomous Region Xinjiang. Participants such as Megvii and universities had to stick to that, called the activist.
In addition to the data sets mentioned, this allows the group "Artificial intelligence and media philosophy" The university for design Karlsruhe as well as the wheat tree institute supported project a search approximately in Megaface with over 3.5 million photos, Diveface with over 115.000 recordings of Flickr, VGG Face, Pipa, IJB-C, FacesCrub, TownCentre, UCCS and Wildtrack. Although exposing.Ai searches so millions of data rates, there is the creators "Non-further training data records for face recognition that are constantly gathering from social media and news and entertainment pages". Art versions of the project could probably be extended accordingly.
A subsequent lot of images from already circulating copies of datasets is not possible, it is on the website. For still well-maintained training databases you are working on a function to call the operators to remove their own recordings based on the search results. Photos that can be eliminated with Flickr, followed on exposing.No anymore.