At Alice Biometrics we are continually improving our algorithms against fraudulent attacks.
As part of this training, we “attack” our technology to make it more robust. In general, we try to cheat our system, that is, we carry out presentation attacks.
These tests help us strengthen our facial recognition system and protect against malicious actors.
If you want to know more about presentation attacks, keep reading.
Índice
Definition of presentation attacks and presentation attack detection
Presentation attacks are any attempt to interrupt a biometric process.
On the other hand, and according to the second edition of the Handbook of Biometric Anti-Spoofing, a PAD method or presentation attack detection consists of “any technique capable of automatically distinguishing between actual biometric features presented to the sensor and synthetically produced artifacts“.
In the specific field of facial recognition systems, PAD mainly focuses on countering presentation attacks in which attackers use some type of artifact, often a photograph, video or 2D mask / 3D.
The use of makeup and plastic surgery are also potential forms of presentation seizures, but are generally used less.
Photographs, videos, masks or makeup are examples of presentation attack instruments (PAI) or tools used in presentation attacks.
That is, presentation attacks are the problem, and presentation attack detection are the solution.
Presentation attack types
Broadly speaking, there are four ways to detect presentation attacks:
- Using sensors to detect characteristic patterns of living features. These are usually generic sensors (cameras) that provide RGB images and that are available on all mobile and portable devices
- Using specific hardware to detect evidence of life, that is, using other technologies to capture biometric information (such as depth sensors or thermal cameras) that cannot always be implemented
- Employing inherently robust recognition algorithms against attacks
- Using a method of challenge-response, in which a presentation attack can be detected. To do this, the user is asked to interact with the system in a specific way (for example, smiliging or winking)
The first three types correspond to passive liveness detection methods, while the last one is a form of active liveness detection ones.
⚠️⚠️⚠️ Remember that ….
Biometric algorithms are very different from each other (they depend on the developer), and the probability of success of an attack can vary greatly depending not only on the type of attack, but also on the characteristics of the facial recognition technology.
The gap between research and real world scenarios
The results of server facial recognition systems are very strong due to the incorporation of different types of PAD methods using powerful hardware (which, in return, takes up a lot of space).
On the other hand, a large part of current facial recognition systems are designed to work on mobile devices. Despite the advances in the hardware of these devices, they have to assume a loss of performance in comparison to the state of the art of research in favor of usability.
In recent years, research and industry have evolved considerably both in the field of biometric recognition and in the power of mobile devices. This is a major step forward in closing the performance gap between systems designed for mobile devices versus their server equivalents.
However, when we talk about detection of presentation attacks, the problem is no longer only in power.
In fact, in chapter 12 of Handbook of Biometric Anti-Spoofing, our colleague Artur Costa-Pazo identifies the main limitations of PAD systems: domain changes, limitations due to databases and usability. He adds an explanation of why some of these systems are far from perfect.
An example of this is that recently two of the most powerful devices on the market (Samsung Galaxy and iPhone) have been victims of presentation attacks using photos and videos. Hence, manufacturers’ own technology is not always the safest.
The challenges of PAD on mobile devices
There are three main challenges for PAD on mobile devices:
???? The first challenge has to do with how a system that has been aligned with a data set is able to maintain its performance when faced with situations different from those for which it has been designed or configured. This concept is called “generalization”.
In today’s widely used data-driven systems, increasing the training data set has been shown to have a positive impact on generalization capabilities. However, in the specific case of PAD systems, it is not only the quantity that matters, but also the variability in the types of attack and the way in which the attacks are carried out.
“Unless databases add a more representative variety of genuine attacks and accesses, loss of performance in real-world scenarios will continue to be a problem.”
???? Although it does not affect all PAD systems, the second major challenge in detecting presentation attacks on mobile devices is the need to create specific data sets to reproduce specific Face-PAD collaborative scenarios. Furthermore, such a successful dataset should include UI synchronization and catch, motion, and reaction annotations, and this is simply not feasible when considering larger datasets.
A possible alternative to this problem is to develop passive PAD methods (like ours), which allow the data capture to be fully automated, something that is already being done today.
???? Finally, the third challenge for PAD systems on mobile devices is usability. And here the power plays an important role, because according to the report of usability and biometrics of the NIST, the usability requirements. They are effectiveness, efficiency, satisfaction, ability to learn and memorization.
How we solve these challenges at Alice
Our team accumulates more than 10 years of experience in the research and development of facial biometric technologies and identity verification.
We strive to know the real problems to offer appropriate solutions and this is how we solve the challenges of PAD on mobile devices:
- First of all, we understand the challenges of PADs as dynamic elements. We are committed to constantly evaluating and analyzing our algorithms, throughout their life cycle. We focus on identifying potential attacks never seen to date and, if necessary, we capture new data that will allow us to improve the system against them.
- Our system is based on non-collaborative methods, which are totally transparent to the capture process. This allows us to gain agility when incorporating new data from different domains and with great variability, which improves our systems day by day.
- Usability is critical for our team. We continually review our processes to eliminate any type of friction that may arise in our clients’ onboarding processes (each service and industry is different, each company is different and this is how we treat our clients). In addition, our systems work in real time and in a totally fluid way.
Do you want to know more? Our colleagues will tell you all our secrets in a demo.
This publication has been financed by the Agencia Estatal de Investigación DIN2019-010735 / AEI / 10.13039/501100011033