Webtool changes headshots just enough to trick face biometrics scrapers
Researchers say they have created software that protects social media posters from unwanted biometric surveillance. The evasion tool, called LowKey, reportedly fends off commercial facial recognition APIs by Amazon and Microsoft.
The computer scientists, from the University of Maryland and the U.S. Naval Academy, say Lowkey prevents facial recognition algorithms from matching harvested photos with new ones that might come from, for instance, surveillance cameras.
In a research paper yet to be peer reviewed, the authors claim LowKey can alter a posted image enough to throw face scrapers off the scent while maintaining a degree of resemblance to a person’s actual appearance that allows humans to make the connection.
Example images published in the paper demonstrate the perturbation. The effect resembles sub-par digital aging. Applying Gaussian smoothing to the image in pre-processing made the images look like those of people facing into a government wind tunnel.
Dialing up the effect on images of some well-known media personalities still left them recognizable but not in all cases.
A photo of actor Tom Hanks, one of the most famous people in the United States, is changed beyond the point of believability. A shot of actor and comedian Tina Fey devolves into something that could charitably be described as an off-brand bobblehead doll.
The result might feel basic, but the researchers say their code, in black-box setting, fools Amazon’s Rekognition and Microsoft’s Azure Face biometric products.
The proprietary Rekognition software reportedly saw through LowKey’s ruse just 2.4 percent of the time. The Azure Face code performed worse. It made accurate biometric connections with the LowKey gallery 0.1 percent of the time.
The researchers prioritized five conditions in building LowKey. First, the adversarial attack had to transfer effectively to unseen models.
Next, the perturbed images had to still be identifiable to human image eyes. There is no purpose in changing one’s image to fool both a scraper and one’s network of followers. A person could use a free entertainment tool that turns their image into, say, a Simpson’s-like character.
Third, LowKey had to be fast enough that run time does not outweigh the algorithm’s disguising features.
Protected images also had to work after being converted to JPG and PNG formats, as well.
Last, LowKey had to protect any size image.
Another notable aspect to the team’s work is that LowKey has been created to be used by social media posters themselves. Most tools are designed to be used by intermediaries between a person (and their photo) and the face scrapers.
Article Topics
biometric data | biometrics | biometrics research | data collection | data protection | facial recognition | social media
Comments