Biometric recognition and authentication systems

How it works

Automatically recognising an individual by analysing an image of their face is a complex process for which there are a variety of approaches.

Facial recognition systems may be 3D or 2D, may use digital photographs or video to capture images. Visible light and infrared are both in use for this process. 

Approaches to facial recognition have changed rapidly in recent years, with the newer systems based on artificial intelligence, trained on massive numbers of face images. These systems can work quite accurately with high resolution (more than 100 pixels between the eye centres), full frontal images in good lighting. However, performance degrades as resolution reduces, pose angle increases or lighting varies.

Computers often start the recognition process by finding eye centres, reposing the facial image to a full frontal orientation and adjusting for shadows. Stable and distinctive "facial features" are extracted - these are abstract measures not related directly to distances between "landmarks" on the face, such as nose, mouth and ears.

Attacks

For a presentation attack, covert acquisition of a suitable face image of the target is likely to be fairly easy. The physical presentation of the resulting image in a presentation attack is likely to involve a photograph, a mask, or a computer/tablet display.

An attack may be more sophisticated than simply presenting a photograph of the target face or a static face mask. The attacker may enhance their artefact in a bid to overcome liveness or spoof detection protections.

Implementation notes

There have been a number of drivers that have led to improvements in performance for face recognition systems. The introduction of ePassports and automatic border control systems which utilise the face image stored in ePassports resulted in a significant improvement in performance in constrained environments, where the variations could be controlled to a great extent.

Protections

It is increasingly common for a face recognition system to offer some protection against presentation attacks, and any system which does not perform at least basic checks is likely to be easily compromised.

Liveness checks look for a simple, measurable test that there is a real person in front of the camera. This could, for example, require specific movements in response to challenges (blinking, nodding, shaking the head) or by continual assessment of small-scale movements of the head while the biometric measurement is taking place, using a video capture of the face.

With the advent of deep learning face recognition solutions, there is a parallel development of presentation attack detection, where a dedicated network is trained to identify known attacks, with the aim that such training will also detect new and unseen attacks.

Video sequences can also be used to protect against replay attacks. Although it is possible to simply capture a video sequence and replay it when prompted to authenticate, information can be added to the sequence that makes it either very difficult, or impossible, to re-use the sequence. As an example, a sequence of head movements can be used for both liveness detection and replay protection.

Summary

Face recognition historically has seen widespread application, particularly in border control and travel, where the existence of large databases of images (particularly passport images) are an important business driver.

In the last few years, the adoption of new approaches has resulted in a significant improvement in accuracy. As a result, FR is now being used in more and more applications, including those where a fairly high level of accuracy is required.

There are some significant challenges relating to presentation attacks - it is relatively easy to obtain an image or even video sequence of an individual which can be used as source data for an attack, but there are protections against such attacks and presentation attack detection techniques are evolving all the time.