Face Liveness Detection: The New Normal of Identity Verification
With the advancement of computer vision and deep learning, face recognition has become very efficient and is now widely used by financial institutions and other businesses for automatically identifying their customers. Unlike other biometric modalities such as fingerprints and voice, face recognition has the advantage that it requires very little user interaction. State-of-the-art face recognition can also detect and recognize faces at a distance with more than 99% accuracy and is getting progressively more advanced with time.
However, state-of-the-art face recognition is not yet perfect and is still subject to a variety of spoofing techniques such as wearing realistic face masks to impersonate others. These spoofing techniques are sometimes called presentation attacks (PAs). Such attacks are implemented to accomplish identity theft with detrimental impacts on individuals and societies.
To make matter worse, our faces are nowadays circulated publicly on the internet due to the use of social media. It is consequently very easy for hackers to find our faces and spoof face recognition system.
Hence, it is essential for face recognition systems to build the confidence that faces are not used by people other than their real owners. The technique for preventing spoofing is known as liveness detection. Here, we discuss different types of presentation attacks to face recognition and the countermeasure that is known as liveness detection — especially passive liveness detection — as well as its quality measurement and certification.
What is Liveness Detection?
For the use cases where customers are required to be present on-site, it is hard for a fraudster to fool face recognition systems since the systems are often accompanied by human operators. For example, it is extremely unlikely that someone wearing a face mask can go undetected through airport passenger screening. On the contrary, for use cases that are fully online such as online onboarding for e-commerce platforms, face recognition systems must determine whether presented faces come from ‘live’ persons, i.e., detecting the faces’ liveness.
In general, liveness detection is the mechanism to tell if the presented biometrics is real or fake by detecting signs of life. For faces, liveness detection techniques may require users to actively blink or perform specific sets of head movements whereas other techniques may passively detect liveness from users’ blood perfusion or light reflection. However, some attacks such as video playbacks can still show signs of life. Therefore, liveness detection must also identify if presented signs of life come from live persons at the point of capture.
Presentation Attack Techniques
Identity theft is now on the rise. For example, in the United States of America, 47% of Americans have reported experiencing financial identity theft in 2020 [3]. Cybercriminals are also getting more and more sophisticated at fooling face recognition systems, e.g., using deepfake techniques to impersonate other people. Sometimes, however, we do not even need advanced technology to fool a typical face recognition system with weak liveness detection, as demonstrated in the link below.
A simple presentation attack demonstrated by BiometricsDebunked
(Source: https://www.youtube.com/shorts/Sfhar1sGi7c)
A technique for spoofing a biometric recognition system is called a presentation attack (PA). As mentioned above, a PA can be performed easily, e.g., by using paper with a face image printed by an inkjet printer or a video replay displayed on an iPad. PAs could also involve more sophisticated techniques such as using silicone or resin to produce realistic face masks.
PAs can be classified into two types, namely 2D and 3D attacks. For example, 2D video replays on computer tablets are currently one of the most common attacks. Deepfake videos — which are becoming increasingly popular — can also be categorized as 2D attacks.
3D attacks are often more expensive and more difficult to perform. Face masks and sculptures made from silicone or resin are 3D methods that use depth or/and texture to fool face recognition systems.
PAs can also be further classified as static and dynamic. For example, print paper is a static PA that can be made dynamic by creating holes for the eyes and mouth of an attacker to artificially introduce signs of life by, e.g., blinking or showing micro-movements.
Types of Liveness Detection
Liveness detection, also known as presentation attack detection (PAD), can be categorized into 2 types, namely active and passive liveness detection.
Active Liveness Detection
Active liveness detection instructs a user to perform a specific set of facial movements in front of a camera, a screen and/or a sensor. These movements include:
Blinking,
— Smiling,
— Turning or tilting the head sideways or from side to side, and
— Looking and following a moving dot on the screen
In some active liveness detection algorithms, a set of movements are chosen at random from a larger pool to make a successful PA less likely.
Active liveness detection is often easier and cheaper to implement than passive liveness detection. However, requiring users to follow movement instructions degrades the user experience. The instructions also give clues to attackers about which facial movements need to be spoofed.
Passive Liveness Detection
Passive liveness detection does not require a user to perform any action while the program collects relevant facial data. It only relies on algorithms to identify facial artifacts as input, which can be either still images or short videos. This technique can be carried out by various methods, ranging from analyzing light reflection to analyzing skin texture or depth information of faces. This type of liveness detection is usually more expensive and more complex to implement than active liveness detection.
Moreover, passive liveness detection also gives a better user experience and takes a shorter time to perform. It also has the advantage of not giving any clue to attackers, making it potentially harder to spoof.
Passive Liveness Detection Techniques
In recent years, the fields of machine learning and deep learning have progressed considerably due to the development of high-performance computing, GPUs, and the increasing availability of high-quality data. This development has enabled precise and robust implementation of passive liveness detection, which sometimes has provided even better performance than the active counterparts. Below, we outline some of the state-of-the-art passive liveness techniques, which use different indicators and feature different advantages and disadvantages.
Light Reflection Technique
Using light reflection to analyze unique textual features on faces.
Pros:
— A user is not required to perform any specific movement.
Cons:
— The accuracy may decrease in environments with bright sunlight.
— The data-collection device must be held steadily.
— The technique may provide some clues to attackers.
Short Video Technique
Using short videos to sequentially analyze spatial information on faces. Temporal relationships between collected frames are taken into consideration.
Pros:
— A user is not required to perform any specific movement.
— Sequential information can enhance accuracy.
Cons:
— The technique is often computationally expensive.
— The process may take a considerable amount of time to capture videos.
— Low-light or noisy environments may degrade accuracy.
Still Image Technique
Using a collection of still images, where each image is analyzed independently. Temporal relationships are neglected.
Pros — · Low bandwidth requirement on the server.
· Good performance.
· Simple data collection and implementation.
· The technique does not require specialized client-side software or hardware.
Cons — · The technique often requires multiple still images to obtain sufficient accuracy.
3D Scan Technique
Using specialized hardware to collect depth information on faces.
Pros:
— The dept information can enhance accuracy.
— Good performance.
— Low bandwidth requirement on the server.
Cons:
— The technique often requires client-side hardware such as an infrared camera.
As of 2022, there is no consensus on which technique is the best for implementing passive liveness detection in all use cases. In practice, the technique of choice depends on many factors, including the tradeoff between user experience and accuracy, available budget, and the environment where data collection takes place.
Evaluation and Certification of Liveness Detection
The performance of a biometric recognition system is usually expressed by its false acceptance rate and false rejection rate. Following the metrics that are standardized by ISO, these are known as Attack Presentation Classification Error Rate (APCER) and Bona fide Presentation Classification Error Rate (BPCER), respectively.
Attack Presentation Classification Error Rate (APCER) — i.e., a false negative rate — describes a proportion of presentation attacks that are incorrectly classified as bona fide presentations.
Bona fide Presentation Classification Error Rate (BPCER) — i.e., a false positive rate — describes a proportion of bona fide presentations incorrectly classified as presentation attacks.
Currently, the most widely accepted certification on liveness detection is provided by iBeta. iBeta is an independent laboratory that is accredited by National Voluntary Laboratory Accreditation Program (NVLAP) to test and certify PAD algorithms, following ISO/IEC 30107–3. Currently, there are two levels of iBeta certification:
- Level 1 requires BPCER of less than 15% and APCER of 0%. PAs for level 1 conformance testing includes high-quality photos and videos.
- Level 2 requires APCER less than 1%. PAs for level 2 conformance testing includes latex masks, resin masks, inexpensive silicone masks, 3D animation software, and handmade masks from 2D photos.
Although liveness detection is now an essential component for online identity verification, its cost is currently an order of magnitude higher than simple face recognition without liveness. More importantly, the rapid development of deepfake technology presents a looming threat to online transactions. Hence, enabling trust in the digital world requires research and development of novel liveness detection techniques that drive its cost down while keeping up with the deepfake.
References
- D. E. Denning. “Why I Love Biometrics It’s liveness, not secrecy, that counts.” https://faculty.nps.edu/dedennin/publications/biometrics.pdf (accessed Oct. 02, 2022).
- I.A. Ghaffar, M.N.H. Mohd, “Presentation attack detection for face recognition on smartphones: A comprehensive review”, J. Telecommun. Electron. Comput. Eng, vol.9, no. 3–8, pp. 33–38, 2017.
- “Facts + Statistics: Identity theft and cybercrime.” Insurance Information Institute. https://www.iii.org/fact-statistic/facts-statistics-identity-theft-and-cybercrime (accessed Oct. 02, 2022).
- “IDLive® Face Passive Facial Liveness Detection” IDR&D. https://www.idrnd.ai/passive-facial-liveness/ (accessed Oct. 02, 2022).
