Facial authentication systems are penetrating various places in our daily life, such as unlocking smartphones and PCs. However, a research team at Tel Aviv University in Israel announced that they have developed an AI that generates a master face rather than a master key. It is reported that it is possible to break through face recognition by disguising more than 40% of the faces with only 9 faces.
The AI developed this time uses the adversarial generation network StyleGAN announced by Nvidia in December 2018 to create a master face that can break through various face authentication. It is said that LM-MA-ES (Limited-Memory Matrix Adaptation) is used as an approach to optimize high-dimensional black box as well as style GAN.
The AI developed by the research team was tested in a convolutional neural network CNN-based face recognition algorithm (SphereFace, FaceNet, Dlib) using LFW (Labeled Faces in the Wild), a face image dataset released as an open source by MIT University.
AI first finds the most generalized features included in various face data and generates faces with as many tilts as possible. A new face was created by tilting as many faces as possible against the face data that could not be broken through with the face created earlier. By repeating this process, it is possible to break through face authentication with a small number of faces.
Although there was a difference by AI accuracy and the test face recognition algorithm, the research team is reporting that it can break through 40-60% face recognition with just 9 face photos. In addition, AI can properly identify the most generalized features and generate face images at an early stage, so even if the amount of learning is increased, the system precision did not improve much.
In the paper, the research team said that this result showed that it is vulnerable even when there is no identity information for face authentication. Related information can be found here.