Racial Discrimination in Face Recognition Technology

In the past decade, law enforcement, airports, and employers have employed facial recognition software as a form of biometric authentication. The adoption of facial recognition has provided quick and efficient verification systems, accuracy in identifying individuals, and is easily integrable into existing software. In Abu Dhabi, engineers are in the trial phase of testing facial recognition authentication at immigration checkpoints, paving a way for quick and efficient airport experiences. As a joint project with Integrated Health Information Systems, Singapore’s largest hospital group, SingHealth, has installed biometric entry systems. Using depth sensing, the new facial biometric technology can detect masks and prevent spoofing of virtual and photo-based presentations. It’s evident that facial recognition attempts to facilitate efficient decision-making algorithms for security and speed.

But are we compromising accuracy for speed? Facial recognition is implemented using machine learning and data analytics. Using computer-generated filters, the AI software transforms face and video feeds into numerical expressions to determine the similarity between a database of other known faces. This algorithm claims a high accuracy of over 90%, but research exposes racial bias due to testing inaccuracies, data privacy, and mass surveillance concerns [1]. In July 2020, an independent assessment by the National Institute of Standards and Technology confirmed that across 189 algorithms, there was a prevalent racial bias toward women of color [2]. In other words, the algorithms struggle to identify women and people of color as accurately as they identify white men.

This issue is attributed to the lack of diverse training data that these computer-generated filters are trained on. This refers to the fact that many AI systems are only as efficient as the data they are trained on. A limited or biased set of data used to train an AI algorithm will lead to algorithms that are limited or biased in their ability to identify and filter content. Thus, when datasets are composed of predominantly male white faces, a system may be less accurate in identifying people of color, as the data set lacks a diverse set of demographic groups. As a further complication, some facial recognition companies rely on external data sources to their systems, potentially compounding the biases already present in those datasets.

These inaccuracies have created an increasingly common story of injustice. In 2020, A 43-year-old black man from a Detroit suburb was arrested on alleged charges of stealing watches after police investigators employed facial recognition to identify him as the thief from surveillance cameras [3]. Despite not having been in that store for years, they kept him in custody for 30 hours. It was later found that he was wrongfully convicted as the surveillance technology was not 100% accurate. This isn’t an isolated incident, as an ever-growing number of people have been victimized by unregulated and flawed surveillance systems, particularly people of color. By using face recognition, the NYPD and ICE could potentially target marginalized groups, such as undocumented individuals and Muslims.

Since 2011, the NYPD has been using facial recognition technology for several years. However, in a report by the Georgetown Law Center on Privacy & Technology, it was found that the Face Recognition Unit in the NYPD had access to a database of more than 8 million photos, most of which were of law-abiding citizens. In opposition to this practice, critics contend that it violates civil liberties and unfairly targets communities of color who are more likely to be monitored and scrutinized by the police. Similarly, ICE has been confirmed to utilize a facial recognition tool developed by Clearview AI, a controversial facial-recognition software company that has amassed billions of images from social media platforms without users’ consent. This again raises privacy concerns and violates the rights of immigrants and other marginalized groups.

In the NSPE Code of Ethics for Engineers, a fundamental canon is to conduct oneself honorably, responsibly, ethically, and lawfully to enhance the reputation and usefulness of the profession [4]. It is difficult to conduct facial recognition lawfully within the code of ethics as it is subjective. This surveillance enforcement fuels a harmful and corrupt system. As stated by the Algorithmic Justice League, “face surveillance threatens rights including privacy, freedom of expression, freedom of association and due process” [5]. This breaks a fundamental code of ethics: deceptive acts. Yet, facial recognition is still a very useful tool. Thus, to combat such inequities without outright banning this technology, Representatives Ted Lieu, Sheila Jackson, Yvette Clarke, and Jimmy Gomez introduced the Facial Recognition Act, a bill that limits law enforcement’s use of facial recognition surveillance to ensure civil liberties are protected. The bill prohibits a single match from being the sole use of evidence to convict a person. If the FRT is enacted globally, law enforcement across the nation will be able to meet a universal standard that could halt racial biases while still utilizing such technologies.

These examples open the conversation about the need for diversity in spaces where such technologies are developed. Algorithms should be trained on a representative dataset and such data should be used with the consent of each individual. A standard on the image quality should be set, as well, to photograph darker-skinned people better. Researchers and engineers at Google have launched the Real Tone project to assist software developers in evaluating lighting algorithms and skin color research practices [6]. Lastly, companies can be held accountable by conducting regular assessments and ethical auditing.

References

1. Klare B., Burge M., Klontz J., Bruegge R., Jain A. (2012) Face Recognition Performance: Role of Demographic Information. IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY. https://s3.documentcloud.org/documents/2850196/Face-Recognition-Performance-Role-of-Demographic.pdf

2. Face recognition vendor test part 3: (2019). National Institute of Standards and Technology. https://doi.org/10.6028/nist.ir.8280

3. Harwell, D. (2021, April 13). Wrongfully arrested man sues Detroit police over false facial recognition match. Washington Post. https://www.washingtonpost.com/technology/2021/04/13/facial-recognition-false-arrest-lawsuit/?request-id=a31502fe-4491-4034-b136-fc5eac093aaf

4. Code of Ethics | National Society of Professional Engineers. (n.d.). Retrieved November 4, 2022, from https://www.nspe.org/resources/ethics/code-ethics

5. Spotlight - Coded Bias Documentary. (n.d.). Retrieved November 4, 2022, from https://www.ajl.org/facial-recognition-technology

6. Skin Tone Research Could Fix Facial Recognition Tech. (2022, October 22). ScreenRant. https://screenrant.com/google-skin-tone-research-facial-recognition-tech-fix/

7. Mascellino, A. (2022, November 4). SingHealth runs facial recognition pilots for visitor management at hospitals. Biometric Update |. https://www.biometricupdate.com/202211/singhealth-runs-facial-recognition-pilots-for-visitor-management-at-hospitals

8. Kamel, D. (2022, November 3). Abu Dhabi to launch facial recognition at Midfield Terminal - from check-in to boarding. The National. https://www.thenationalnews.com/business/2022/11/02/abu-dhabi-to-roll-out-facial-recognition-at-midfield-terminal-from-check-in-to-boarding/

9. Buolamwini, J., & Gebru, T. (2020, January 6). Racial discrimination in face recognition technology. Harvard School of Public Health - Science in the News. https://sitn.hms.harvard.edu/flash/2020/racial-discrimination-in-face-recognition-technology/

10. Garvie, C., Bedoya, A., & Frankle, J. (2018). The Perpetual Line-Up: Unregulated Police Face Recognition in America. Georgetown Law Center on Privacy and Technology. Retrieved from https://www.perpetuallineup.org/