Researchers Say Facial Recognition Not Foolproof

At the Usenix security conference, researchers from the University of North Carolina, using digital 3D facial models based on photos taken from social networks such as Facebook, have shown on a smartphone's screen they can successfully beat facial recognition software.

Collecting publicly available photos of 20 volunteer subjects from image engine searches and social networks such as Facebook, LinkedIn and Google+, much like any online stalker would do, the researchers said they could find between three to 27 photos of each volunteer online. Although most of the participants were security researchers keen on protecting their privacy online, the researchers said they were able to dig up at least three photos for each subject.
"We could leverage online pictures of the [participants], which I think is kind of terrifying," study author True Price told Wired. "You can't always control your online presence or your online image."

The UNC researchers created 3D models of the subjects' faces, added facial animations and adjusted the eyes to look directly at the camera. If the photo didn't show the subject's whole face, they improvised by recreating the missing parts and adding in textures and shadows.

They then tested the virtual reality face models on five authentication systems that are readily available via consumer software vendors such as the iTunes Store and Google Play Store, including Mobius, KeyLemon, TrueKey, BioID and 1D. Researchers found that they could trick four out of five of these systems used to lock smartphones and safeguard sensitive data with success rates between 55 percent and 85 percent.

After taking indoor headshots of each subject and rendering them for virtual reality to test them against the five systems, the researchers found they were able to dupe all five systems in every case tested.

"Our work outlines several important lessons for both the present and the future state of security, particularly as it relates to face recognition systems," the researchers noted in a paper titled, 'Virtual U: Defeating face liveness detection by building virtual models from your public photos.'

"First, our exploitation of social media photos to perform facial reconstruction underscores the notion that online privacy of one's appearance is tantamount to online privacy of other personal information, such as age and location," the paper said. "The ability of an adversary to recover an individual's facial characteristics through online photos is an immediate and very serious threat, albeit one that clearly cannot be completely neutralized in the age of social media. Therefore, it is prudent that face recognition tools become increasingly robust against such threats in order to remain a viable security option in the future."

The researchers note that it is crucial for facial authentication systems to be able to reject synthetic faces with low-resolution textures, given the fast-paced developments being made in virtual reality and computer vision technologies that are quickly becoming more "commonplace, cheap and easy-to-use." They recommend that several additional features be added to bolster these systems' security including light projection patterns, detection of minor skin tone fluctuations related to pulse and the use of illuminated infrared sensors.

Suggested Articles

One forecast from Cameron Chell: the best AI designers of the future won’t come from top universities

Survey of 30 chipmakers offers a good sign for research and development of self-driving vehicles, analyst says

Research dollars for AV are expected to remain, if slowed, especially for companies that see self-driving as a key to their success