VisionLabs Company Wins International Competition for Protecting Facial Recognition Systems from Hacking
One of the global leaders in machine learning for facial recognition is VisionLabs, a Skolkovo (VEB.RF) resident company which took first place at the Face Anti-spoofing (Presentation Attack Detection) Challenge. The international contest is part of the largest global computer vision event, ICCV 2021. One hundred and ninety-five teams from the most prominent universities and companies took part, including Tencent, ByteDance, etc.
Facial recognition has come a long way in recent years, and its use across sectors is increasing with each passing year, whether for security, convenience, payment, etc. It is now the second most deployed biometric authentication model worldwide after fingerprints. Unfortunately, this also means a new sector for hackers and fraudsters to exploit because facial recognition software used for daily tasks is invariably a source of valuable personal data. Despite the advances, computer vision still falls short of human vision, which is quicker to perceive and understand. A human eye and brain take milliseconds to see and process a vast amount of data about a given scene and the objects that are part of it.
Things get a little more complicated when it comes to computer vision, and facial recognition is one of the most challenging tasks for it. We take for granted our ability as humans to look at a photo and take in its contents; yet, factors such as object detection, depth, and event categorization make the entire process complicated for computers.
Biometric data is not kept secret in the same way that a password is, so the attacker's goal is to fool the system by mimicking the user's biometric traits. There are several ways in which fraudsters attack a facial recognition system. The most common method is a so-called "presentation attack” (a.k.a. spoofing), which is performed without needing access to the system. The attacker exploits biometric vulnerabilities by using a 3D mask, a photo, a synthetic fingerprint, etc., to mimic the actual user and access their data.
A 3D mask attack is when the fraudster wears a 3D reconstruction of the user's face and presents it in front of the camera. Another method is to play a video of the actual user and show it to the sensor/camera. A photo attack involves presenting a user's picture to the camera to fool the biometric security system. Other techniques include fraudsters changing their features using makeup to mimic the real user.
Spoofing has proven effective in bypassing facial recognition systems. This makes it imperative for facial recognition developers to design their systems to detect spoofing attacks, and VisionLabs has done just that.
This year's competition focused on the ability to verify a real person and detect a spoofing attack using 3D masks. This task is difficult not just for computer vision, but also for the human eye, assuming the mask is of good quality. The organizers increased the difficulty of the competition by using a dataset of photos with people from different ethnicities with different 3D masks (see-through, plastic, and silicon) and additional attributes such as glasses or a wig that hide facial details. In addition, data on the most challenging type of attack, which utilizes a silicon mask, were provided only ten days until the end of the competition.
External data and pre-taught neural networks were also banned. The main purpose of the aforesaid restriction was to narrow down the winning conditions to optimization of the model architecture and not the quantity of data for programming. Out of all the teams, the VisionLabs team created the most effective algorithm, Liveness, which is 97% accurate and can differentiate between a real person and a substitute in a 3D mask.
Liveness facial detection can distinguish between real and fake faces in an image. Several factors come into detecting a spoofing attack, such as teaching computer vision algorithms to spot a reflection on a phone screen containing the user's image, the edges of a mask, and so on.
The competition results were as follows:
- APCER – the percentage of undetected attacks in which the participants' algorithms recognized a fake as a real person;
- BPCER – the percentage of incorrectly identified persons, where the participants' algorithms recognized a real person as fake;
- ACER – average indicator of APCER and BPCER.
Dmitry Markov, CEO of VisionLabs: "The creation of high accuracy liveness algorithms even for such difficult spoofing attacks such as 3D masks, helps make computer vision-based solutions more widespread and affordable. We are proud of our team of researchers who continue to develop this theme from year to year and ensure the high security of VisionLabs products. In addition, the algorithms developed for fraud prevention strengthen our industrial solutions for the financial sector, retail, and transport, where facial recognition payment and distance verification is required."
In 2019, the VisionLabs Liveness algorithms became the best for recognizing fake multimodal data with depth cameras, but in 2020 attack recognition using printed photos, videos from a phone screen or a tablet, 3D masks on a set of RGB frames.
Sergey Khodakov, the director of operations at the Skolkovo Foundation IT Cluster: "The Russian school of video analytics is traditionally one of the strongest in the world, and such events confirm this. Winning at the largest international competition, VisionLabs once again confirmed its status as one of the most professional developers of video analytics systems. In addition, the company's solutions and algorithms continually improve and allow us to fight fraud and make modern payment methods with biometric data more secure."
The VisionLabs algorithms, among other things, are used in the Moscow metro for the biometric identification of passengers. According to the second quarter of 2021, they showed the best result in terms of accuracy among their competitors. By the end of the year, a fare payment system is planned to be launched via Face Pay using facial recognition.