Tightening up facial biometrics
Source: David Bradley
Facial biometrics for security applications is an important modern technology. Unfortunately, there is the possibility of "spoofing" a person's face to the sensor or detection system through the use of a photograph or even video presented to the security system. A team from China has now developed a counter-measure that could preclude face spoofing and make such biometric security systems far less prone to abuse. The team reports details in the International Journal of Computational Science and Engineering.
Fei Gu, Zhihua Xia, Jianwei Fei, Chengsheng Yuan, and Qiang Zhang of Nanjing University of Information Science and Technology, explain how anti-spoofing technology usually looks to illumination differences, colour differences, or textures differences to spot issues with the presented face to determine whether or not the face is a photo or video rather than a live human in front of the security camera. However, even these approaches are vulnerable.
In order to make a stronger anti-spoofing system, the team has proposed a method based on various feature maps and convolution neural networks for photo and video replay attacks. They explain that facial contour and specularly reflected features are taken into account when verifying a face so that depth and width can be determined, aspects of a living face that are not present in a photograph. Their proof of principle shows remarkable performance against multiple datasets and shows that the method can defend not only photo attack, but also video replay attack with a very low error rate.