Facial Recognition Technology - Innovatrics

Facial recognition is a technology capable of matching a human face from a digital image or a video frame against a database of faces to confirm an individual’s identity. Although less accurate than fingerprint recognition, it is often favored because of its contactless nature. It is mostly used in personal security, law enforcement, or digital onboarding in finance.

How Facial Recognition Works

Former face recognition techniques were Voila-Jones Haar cascade or a histogram of oriented gradients. However, modern facial recognition technology is based on a specific neural network called convolutional neural network. To match the face templates, convolutional neural networks process each image through several steps:

Face detection
Face alignment
Template extraction
Template matching

A convolutional neural network converts every face pattern into a numerical code with every template expressed as a numerical vector. The closer two vectors are to each other, the more likely there is a face match between them.

Facial Recognition Accuracy Is Improving

Facial recognition algorithms are not as accurate as fingerprint or iris algorithms. However, given the development of convolutional neural network, the accuracy of facial recognition algorithms has been on the uptick.

The accuracy of each algorithm is measured by two error classes:

False positivity
False negativity

False positivity occurs when the software wrongly considers photos of two different individuals to be the same person. On the other hand, false negativity means the software failed to recognize the same person. There is always a tradeoff between false positivity and false negativity, therefore, it can sometimes be difficult to set up the thresholds in biometrics.

Face recognition accuracy is rapidly increasing. The world’s leading face recognition algorithms have reduced error rates by orders of magnitude (from 10% to below 1%) in just 3 years. Between 2014 and 2018, the accuracy of facial recognition technology increased 20-fold. These continued improvements over the next several years will likely bring a myriad of new use cases. See Innovatrics results here.

Nist Results Facial Recognition Error Rate

Using Biometric Templates on Personal Devices

Biometric authentication is now used on personal devices instead of PIN or built-in within application as a verification or client onboarding tool. However, the main reason mobile phone manufacturers use it is due to its convenience. There is no need for password protection and management. Also, both fingerprint and face recognition are much faster than traditional authentication methods. There is also more security, because no one can simply look over the shoulder to steal your password. Finally, as there is match-on-device technology and the data is never provided to third parties nor is it stored in the cloud, the user can be sure no one else is tampering with it.

Facial Recognition Technology Standards

The National Institute of Standards and Technology (NIST) is the leading international biometric testing organization and industry standard developer. NIST collaborates with other federal agencies, law enforcement, enterprise, and academic partners in order to:

Research measurement, evaluation, and interoperability to advance the use of biometric technologies including face, fingerprint, iris, voice, and multi-modal techniques
Develop common models and metrics for identity management, critical standards, and interoperability of electronic identities
Support the timely development of scientifically valid, fit-for-purpose standards
Develop the required conformance testing architectures and testing tools to test implementations of selected standards

The most prestigious industry test is the NIST Face Recognition Vendor Testing Program (FRVT), which accepts submissions from any organization worldwide. NIST FRVT assesses the capabilities of facial recognition algorithms for one-to-many identification and one-to-one verification. See Innovatrics NIST FRVT results.

Algorithmic Biases

NIST has recently run a large-scale test focused on identifying biases in facial recognition technology. NIST found out that even the best algorithms still displayed a higher false positive rate among West and East African and East Asian individuals, while Eastern Europeans had the lowest false positive rate. The reason is demographic imbalances in the training data used to train these engines. Changing these imbalances is not easy – some are even built-in physics such as dark skin reflecting less light, and therefore, providing less detail to analyze. However, Innovatrics facial recognition algorithms are well trained for darker skin tones, according to NIST results, with real world confirmation on its Guinea election project.

In general, when preventing algorithmic biases, these methods can be used:

Better data labelling to ensure all groups of people are present
External dataset auditing – unbiased datasets make unbiased algorithms
Using de-biasing variational auto-encoders, which are able to automatically discover and mitigate hidden biases among training data

Biometric facial recognition-Algorithmic_biases

Applications and Use Cases of Facial Recognition

Nowadays, facial recognition technology has a broad spectrum of applications worldwide – from individual onboarding and personal security to the identification of individuals in crowds and gatherings. The most common use cases of this technology include: