CyberLink’s FaceMe facial recognition engine has posted a top-three score in the anti-spoofing challenge at this year’s International Conference on Computer Vision (ICCV). The event was sponsored by ICCV and the Institute of Electrical and Electronics Engineers (IEEE), and was set up to determine whether or not the world’s leading facial recognition algorithms can distinguish real faces from high-quality masks.
This year’s competition was open to academics, researchers, and commercial vendors from all over the world. One hundred ninety-five teams enlisted during the development stage, though only 56 had solutions that were deemed good enough to make it to the final stage. Of those, only 18 cleared the minimum criteria laid out by the ICCV.
FaceMe made it to the top three with an anti-spoofing accuracy rate of 96.8 percent, a figure that was only 0.16 percent lower than the top-scoring algorithm. That number corresponds to a 3.215 percent average error rate.
Prior iterations of the face spoofing challenge have placed a greater emphasis on 2D facial recognition systems, and measured their ability to thwart video replay attacks. The 2021 ICCV challenge specifically looked at 3D systems, with a particular focus on high-fidelity masks. CyberLink noted that that is a category that many systems have struggled with in the past.
“With the increasing use of facial recognition, the risk for spoofing-attacks rises,” said CyberLink CEO Jau Huang. “Making facial recognition more reliable and secure is one of the top priorities for the providers of this technology.”
FaceLink has consistently improved its standing in multiple rounds of NIST testing, culminating with a top-six score back in January. The engine has been integrated into several access control terminals in the past few months, including Vypin’s eScreener kiosks and ACE Bioteck’s TC-800 Wallie Screen Health Screening System. The technology has also been deployed in ASUS’ Tinker Board 2 SBC to support a range of IoT applications.
July 29, 2021 – by Eric Weiss