Russian face recognition system protection wins international competition | Central regions, Technology & innovation

Central regions | Technology & innovation

Russian face recognition system protection wins international competition

30 Jul '21
One of the world leaders in the field of machine learning, a resident of the Skolkovo Foundation (VEB.RF Group), VisionLabs has been ranked first in the Face Anti-spoofing (Presentation Attack Detection) Challenge for the third year in a row. The international competition was held as part of the largest computer vision conference ICCV 2021. 195 teams from major universities and companies, including Tencent, ByteDance and others, took part in it.

This year's competition focused on verifying a living person and detecting spoofing attacks (a cyber attack in which a fraudster impersonates a trusted source in order to gain access to sensitive data or information) using 3D masks. Today, this task is considered one of the most difficult - depending on the quality of the mask, even a person may not always correctly identify it in a photograph.

The organizers increased the complexity of the competition by using photos in the dataset with representatives of several races, various 3D masks (transparent, plastic and silicone) and additional attributes like glasses or a wig that hid the details. The data for the most complex attack - with a silicone mask - was only revealed 10 days before the end of the competition to exclude manual marking.

Also, one of the conditions was a ban on the use of external data and pre-trained neural networks. This limitation allowed the selection of the winner based on new ideas and optimization of the model architecture, rather than the amount of data for training. Among all the participants, the VisionLabs team has created the most effective Liveness algorithm, which, with an accuracy of 97%, distinguishes a living person from his substitution in a 3D mask.

The results of the competition were summed up according to the following metrics:

APCER is the percentage of undetected attacks in which the participants' algorithms identified the fake as a real person;

BPCER - the percentage of incorrectly identified people, where the participants' algorithms identified real people as fake;

ACER is the average of APCER and BPCER.
Dmitry Markov, CEO of VisionLabs: “Creation of liveness algorithms with high accuracy even for such complex spoofing attacks as 3D masks helps to make computer vision-based solutions more widespread and affordable. We are proud of our research team, which continues to develop this topic from year to year and ensures the high security of VisionLabs products. The developed anti-fraud algorithms will strengthen our industrial solutions for the financial industry, retail, transport, where payments by face and remote verification are especially in demand. ”
In 2019, Liveness VisionLabs algorithms became the best in recognizing fakes on multimodal data from depth cameras, and in 2020 - identifying attacks using printed photos, video from a phone or tablet screen, 3D masks on a set of RGB frames.
Sergey Khodakov, Director of Operations, IT Cluster, Skolkovo Foundation: “The Russian school of video analytics is traditionally one of the strongest in the world, which is confirmed by such significant events. By winning the largest international competition, VisionLabs has once again confirmed its status as one of the most professional developers of video analytics systems. The company's solutions and algorithms are constantly being improved to combat the actions of intruders and make modern payment methods based on biometric data more secure. ”
VisionLabs algorithms, among others, are used in the Moscow metro for biometric identification of passengers. At the end of the second quarter of 2021, they showed the best result in terms of accuracy among competitors. By the end of the year, it is planned to launch payment for travel through Face Pay on the basis of the face recognition system.