A look at the history of liveness detection

In 1950, renowned mathematician and computer scientist Alan Turing developed the famous Turing Test, a noteworthy milestone in the history of artificial intelligence. This test was designed to assess the ability of a computer to show intelligent human-like behaviour.

The idea behind the Turing Test was that if a machine could perform a talk convincingly enough to fool a human judge into believing it was interacting with another human being rather than a machine, then the machine could be considered human-like intelligence.

The Turing Test became a significant benchmark in artificial intelligence and stimulated the development of algorithms and systems capable of simulating human behaviour. Nevertheless, as technology advanced, it became clear that the Turing Test did not fully address the question of the authenticity of the interaction between a computer and a human.

When it was concluded, the concept of liveness detection came into play. Unlike the Turing Test, which assesses a machine’s ability to mimic human behaviour, liveness detection uses artificial intelligence algorithms to determine if a computer interacts with a real person rather than a simulation.

A pioneering view of liveness

Dorothy E. Denning became famous after Tuning when in 2001 published an article entitled “It’s “liveness,” not secrecy, that counts” in Information Security magazine. Denning introduced the term liveness in the context of biometric authentication. Rather than focusing only on protecting secrecy, Denning argued that a good biometric system should not rely only on keeping the biometric information used for authentication secret.

Furthermore, she emphasised that the validation process should focus on verifying the liveness, i.e. ensuring that the sample came from a live person and not a fraud.

Denning’s view was revolutionary for its time. At a time when most biometric authentication systems focused on protecting the secrecy and privacy of biometric data, Denning understood the importance of liveness detection as a critical element of ensuring authenticity and security.

Her perspective laid the foundation for creating more secure and reliable identity protection and fraud prevention systems.

But then, what is liveness detection?

Liveness detection is the ability of a system to determine the authenticity of a face, voice or other biometric feature. In other words, the technique verifies if the characteristic presented to the camera in a verification process comes from a real person or is an artefact.

Therefore, the liveness technique confirms the authenticity of the individual performing the process and identifies spoofing attacks (presentation attacks).

Liveness detection uses advanced algorithms and machine learning models to analyse and process the captured information. These algorithms identify specific patterns and features that indicate the sample authenticity, enabling accurate and reliable detection.

What is the difference between facial recognition and liveness detection?

Firstly, we need to say that they are not the same.

The main difference is that facial recognition answers “Are they the right users” whereas liveness answers the question “Are they real individuals?”.

Both terms are connected and can work together, but they are different. Liveness only involves detecting a human face in a frame or video, not identifying the person. On the contrary, facial recognition goes beyond simple liveness detection, trying to determine who the person is in a verification or authentication process. In this sense:

• In facial verification, a capture of the identity document is made, and the technology takes the user’s selfie to compare the photo on the ID card with the selfie, confirming that the individual is who they claim to be. Facial verification is normally used for onboarding processes or remote incorporation processes.
• Facial authentication compares the biometric data of a user’s previously registered with a selfie taken at that moment. In this case, the aim is to confirm the user’s identity to access an account or authorise a transaction.

How does liveness detection work? 

As commented, liveness is responsible for determining if the captured biometric data is real or being imitated to gain unauthorised access to the biometric system.

Liveness works differently depending on whether an image or video is analysed.

When an image is analysed in the liveness process:

1. The individual presents a biometric sample in front of the camera.
2. After that, the image is processed and ready for feature extraction. The output biometric template of the feature extraction process is a distinguishable sample of distinct features that allows the classifier to decide if the sample is real or fake with the help of pre-trained data. Genuine samples are processed for identification, while fake ones are automatically discarded.

When a micro video is analysed in the liveness process:

1. The user presents a biometric sample in front of the camera.
2. Then, the technology records a micro video from which a series of images are extracted, and predictions are made about each of them, with the neural network architectures making their own decisions (deep learning end to end).
3. The technology identifies if the user is a real person or an impostor.

In both cases, the purpose of liveness is to determine if the biometric sample presented is real or fake. If the sample is fake, it calls an impersonation or presentation attack.

Presentation attack detection (PAD) 

Presentation attacks (PA), or impersonation attacks, are purely biometric vulnerabilities not shared by other security solutions and require specific measures.

According to ISO/IEC 30107-1, the biometric feature used in an attack is a presentation attack instrument (PAI).

In these attacks, fraudsters use an artefact or instrument, usually artificial, such as a 2D facial impression, a video or a realistic 3D mask to gain access to the biometric system fraudulently.

Let’s review the most common classification:

Photo attacks 

In this kind of attack, a photo of the impersonated identity is displayed in front of the camera.

The photo can be printed or can be displayed on a screen. In the latter case, the attack is performed on a device’s screen, such as a mobile or tablet.

This type of attack also includes another more complex technique, photo masks. In this case, the instrument is a photo of the legitimate user with holes in the mouth, eyes and nose.

Video attacks

In In this attack, the artefact presented to the sensor of the biometric recognition system is a video played on a mobile phone, tablet or laptop.

They are also known as replay attacks.

Mask attacks

In mask attacks, a 3D reconstruction of the face of the person to be impersonated is made.

The mask can be done in various ways. The most common are:

• Printing a photo of the genuine user and glueing it onto a deformable structure such as a plastic bag.
• Print a frontal and side photo of the legitimate user and generate a 3D model.
• Take a 3D capture of the attacked person.

If you are interested in learning more about other attacks, such as synthetic theft and the famous deepfakes, you can read our article, Deepfakes: how to guarantee authenticity?

Liveness detection types 

Active liveness detection

Active liveness detection involves user collaboration to verify the authenticity of a facial sample in real time. This technique uses physical actions performed by the user to confirm that the sample is from a genuine person and not a static image, mask or pre-recorded video.

Some common methods used in active liveness detection include:

• Motion analysis: A flat object moves differently than a real face. This method includes the calculation of information about points in an image. The method is based on focus distance analysis, optical flow-based analysis and analysis based on scenic cues.
• Involuntary action recognition: Detection of random and unpredictable user actions, such as winks, which are difficult to replicate with a static image or pre-recorded video.
• Blink pattern analysis: Natural blink patterns can be analysed to detect the presence of a real person. For instance, missing or irregular blinking patterns could indicate a fake image.

There are other methods, such as 3D properties, that are not widely used but fall under this type of indicator.

 Passive liveness detection

Passive liveness detection is based on analysing visual and biometric features in the facial image without requiring the user’s collaboration. This technique uses algorithms and machine learning models to identify signals that indicate the authenticity of the sample. It occurs when the technology takes a selfie of the user

Some common methods used in passive liveness detection include:

• Texture analysis: A fake face produces a different texture pattern than a genuine one. Textural features are extracted from the face image to provide detectable information to help distinguish real identities from fake ones. For example, differences in skin texture.
• Depth detection: Image depth information is used to verify if the sample is three-dimensional and not a flat or printed image.
• Background motion analysis: The environment or background is observed to detect abnormal motion signals that may indicate an attack.
• Micro video capture from which a series of images are extracted.