Facial recognition is becoming more accurate, but some systems show racist preferences, and some ways of using the technology are quite contradictory.
Can Big Brother identify your face from the street surveillance cameras and tell if you are happy, sad, or angry? Can this certification lead to your arrest with unpaid bail? What are the chances that the identification is incorrect and actually connects you to another person? Can you beat CCTV entirely with a trick?
On the other hand, can you enter a vault protected by a camera and facial identification software by carrying the face of an authorized person? What if you put a three-dimensional mask on the face of an authorized person?
Welcome to the world of facial recognition as well as facial recognition scams.
Tests of providers of facial recognition technologies
The US National Institute of Standards and Technology (NIST) has been testing the Face Recognition Vendor Test (FRVT) since 2000. The image databases used are mostly photos from law enforcement, but also include random photos such as those on Wikimedia, as well as low-resolution images from webcams.
FRVT algorithms are most often presented by commercial technology providers. Comparisons with the previous year show a large improvement in productivity and accuracy; according to suppliers, this is primarily due to the use of deep convolutional neural networks.
NIST-related face recognition test programs examined demographic effects, face morph detection, face identification on social media, and face identification in a video. A previous series of tests were conducted in the 1990s under a different name, namely, Face Detection Technology.
What exactly is facial recognition?
Face recognition is a method of identifying an unknown person or verifying the identity of a particular person only through his face. It is a branch of computer vision, but face recognition is specialized and is often accompanied by "social baggage" for some applications, as well as some vulnerabilities that can be used to defraud technology.
How does facial recognition work?
Early facial recognition algorithms (still used today in an improved and more automated form) rely on biometric data (such as eye distance) to convert measured facial features from a two-dimensional image to a set of numbers (vector function). or template) describing the face. The recognition process then compares these vectors with a database of celebrities that are compared to characteristics in the same way. One complication of this process is adjusting the faces to a normal view to account for the rotation and tilting of the head before retrieving the indicators. This class of algorithms is called geometric.
Another approach to face recognition is to normalize and compress two-dimensional facial images and compare them to a database of similarly normalized and compressed images. This class of algorithms is called photometric.
3D Face Detection uses 3D sensors to capture the facial image or reconstructs the 3D image from three 2D tracking cameras pointing at different angles. 3-D facial recognition can be significantly more accurate than 2-D recognition.
The analysis of the skin texture creates a map of the lines, patterns, and spots on a person's face through another characteristic vector. Adding skin texture analysis to 2D or 3D facial recognition can improve recognition accuracy by 20 to 25 percent, especially in the case of twins or very similar people. You can also combine all methods and add multispectral images (visible light plus infrared light) for even greater accuracy.
Face recognition has improved every year since the development of this area began in 1964. The error rate has halved on average every two years.
Facial recognition applications
Face recognition applications most often fall into three main categories: security, healthcare, and marketing/retail. Security involves law enforcement, and this class of face recognition can be as useful as it can more quickly detect a person's face-matching a passport photo, making it more accurate than the verifying officer. At the same time, it can be as sinister as in the script for the crime series "Under Surveillance" (Person of Interest), in which people are tracked through video surveillance and compared to databases with comparable photos. Non-enforcement security includes common applications such as face unlocking for mobile phones and access control to laboratories and vaults.
Health apps for face recognition include checking patients, detecting emotions in real-time, tracking the patient in a medical facility, assessing pain levels in "nonverbal" patients, detecting certain diseases and conditions, identifying staff and the medical institution. Face recognition applications for marketing and retail purposes include identifying members of loyalty programs, identifying and tracking known shoplifters, and recognizing people and their emotions to purposefully offer products of interest to them.
Contradictions, biases, and prohibitions in facial recognition
To say that some of these applications are contradictory would be an understatement. As a 2019 New York Times article discusses, face recognition technology is creating a whirlwind of controversy, from its use for stadium surveillance to racist software. Observation of the stadium? Face recognition was used at the end of the baseball season (Super Bowl) in 2001. The software then identified 19 people believed to be under unpaid bail, but none were arrested (not for lack of attempts to do so). ).
Racist software? There are several issues, starting with the 2009 face tracking software, which can track whites but not blacks, and continuing with the 2015 MIT survey, which shows that face tracking software at that time worked much better with white male faces than with female and/or black faces.
This type of problem has led to outright bans on face recognition software in certain places
or for specific purposes. In 2019, San Francisco became the first major U.S. city to ban police and other law enforcement agencies from using face recognition software. Microsoft has called for federal regulations in the field of facial recognition, and MIT found that Amazon Rekogmination has more problems identifying women than men in facial images, and more problems with black women than whites.
In June 2020, Microsoft announced that it would not sell or sell its face recognition software to the police. Amazon banned police from using Rekogmination for a year, IBM abandoned its face recognition technology. However, a total ban on face recognition will not be easy, given its widespread penetration into the iPhone (via Face ID) and other devices, software, and technology.
Not all face recognition software suffers from the same biases. NIST's 2019 demographic effects study, followed by MIT's work, found that algorithmic demographic biases vary widely among facial recognition software developers. Yes, there are demographic effects on the degree of false matching and the frequency of false discrepancies in facial identification algorithms, but they can vary by an order of magnitude and decrease over time.
Hacking facial recognition and anti-fraud techniques
Given the potential threat to confidentiality due to a facial recognition application and the desire of many to gain access to valuable resources protected by authentication, there have been many efforts to hack or falsify the technology. You can present a printed face photo instead of a live face or screen image, or a 3D printed mask to defraud. You can play a video for the CCTV system. To avoid video surveillance, you can try fabrics and make-up and/or infrared emitters to trick the software into not recognizing your face.
As a result, efforts are being made to develop techniques to deal with all of these attacks. To detect printed images, vendors use a vitality test, such as waiting for an object to blink or performing a motion analysis, using infrared light to distinguish a living face from a printed image. Another approach is to analyze the microtexture, as human skin is optically different from fingerprints and mask materials. The latest anti-counterfeiting techniques are based on deep convolutional neural networks.
This area is constantly evolving. There is a fierce war between attackers and software developers against counterfeiting, and academic research is being conducted on the effectiveness of various attack and defense techniques.
Providers of facial recognition technology
According to the Electronic Borders Foundation, MorphoTrust, a division of Idemia (formerly known as OT-Morpho, or Safran), is one of the largest providers of facial recognition and other biometric identification technologies in the United States. The company has developed systems for state vehicle departments, federal and state law enforcement agencies, border control services and airports, and ministries. Other well-known providers are 3M, Cognitec, DataWorks Plus, Dynamic Imaging Systems, FaceFirst, and NEC Global.
NIST's facial recognition test lists the algorithms of many more manufacturers around the world. There are also a number of open-source algorithms, as well as several large cloud services that offer facial recognition.
Amazon Rekognition is a photo and video analysis service that can identify objects, people, texts, scenes, and actions, including face analysis and special labels. The Google Cloud Vision API is a pre-trained image analysis service that can detect objects and faces, read printed and handwritten text, and create metadata in your photo catalog. Google AutoML Vision lets you train custom photo templates.
At the same time, the Azure Face API is a facial recognition that perceives faces and attributes in images, identifies faces that match a person in your private repository that can accommodate up to 1 million people and performs emotion recognition. Face API can run in the cloud or in peripheral containers.
Data set for recognition training individuals
There are dozens of data sets that can be downloaded and used to train the recognition algorithm. Not everyone is the same. They differ in image size, the number of people represented, the number of images per person, photo conditions, lighting. Law enforcement also has access to non-public data sets such as current photos and driver's license images.
Some of the larger face databases in the United States are Labeled Faces in the Wild with approximately 13,000 unique images; FERET used for the first NIST tests; the Mugshot database used for the current NIST FRVT test; the SCFace surveillance camera database, also available with face markers; and Labeled Wikipedia Faces with about 1,500 unique identities. Some of these databases contain multiple images of a person. Researcher Ethan Myers offers some compelling tips for choosing a set of data for a specific person.
In summary, facial recognition is improving and providers are learning to detect most counterfeits, but some applications of the technology are controversial. The percentage of facial recognition errors is halved every two years, according to NIST. Vendors have improved their anti-counterfeiting techniques by including convolutional neural networks.
Meanwhile, there are initiatives to ban the use of facial recognition in surveillance, especially by the police. However, banning face recognition altogether will be difficult, given its widespread use.
