Ethical Facial Recognition

Within the broader context of algorithmic facial recognition technologies, it’s essential to understand that everything the A.I. knows and learns usually originates from inputs and values provided within humanly selected sets. Particularly in the case of facial recognition, the faces that the A.I. is algorithmically trained to detect originate from a massive list of facial inputs. This dataset of face inputs is typically provided by the company that has created and maintained said technology. However, these facial inputs are often biased, heavily so, towards the facial features and skin tones of light-skinned men. This means that algorithms detect the faces of white men with considerably higher accuracy than the faces of women of color. This disparity of algorithmic bias, particularly within a technology that is readily and eagerly beginning to be utilized by law enforcement worldwide, is dangerous. The use of flawed technology shows a dystopian and disparaging marginalization of minority populations and non-male gender populations. Moreover,  adapting face recognition algorithms for law enforcement could potentially reverse many of the great progressive steps taken within the past half-century.

Race and Gender

When examining algorithms to spot biases, it’s essential to outline a series of criteria and factors to hone in on. In the case of facial recognition technology, the algorithms were examined from the perspectives of both racial bias and gender bias, within the criteria of algorithmic datasets, recognition accuracy, and the purposes that this facial analysis software might serve and how these biases may harm the demographics that are on the receiving end of such bias.

In one example in January of 2020, Detroit police arrested one Robert Julian-Borchak Williams after facial recognition software incorrectly identified him as a culprit in security footage of a watch store robbery, despite a human eye clearly being able to tell that he was not the man in the video. Being a man of color, Williams was victim to the algorithmic bias of facial analysis software, one largely the result of a lack of diversity in datasets that calibrate the computer technology behind the software [ source ]. These datasets are heavily weighted towards white men, marginalizing populations of other races and genders in terms of accuracy.

In a 2018 research paper by Joy Buolamwini, founder of the Algorithmic Justice League, it was found that while the facial analysis software had a 99.7% accuracy rate when it came to identifying white men, those numbers dropped to 98.6% for black men, 92.8% for white women, and an abysmal 66.7% for women of color; additionally, the facial recognition software held very little identification regard for those of other gender identities and would otherwise misgender transgender individuals.

The errors for these misclassifications of race and gender within individual faces has been found to largely be the result of inadequate representation within the datasets that underpin these programs - the foundings of Buolamwini’s thesis determined that approximately 80-85% of the faces used within the datasets were light-skinned, and over 75% of the faces were male, with darker-skinned female faces making up less than 5% of the examined datasets. [ source ]

The biases that these algorithms and datasets reinforce can be incredibly harmful to the marginalized populations at hand, particularly due to the purposes that facial recognition software serves and the perceived infallibility of machine learning and computing among human populations. Additionally, the racial and gender profiling of these algorithms represents a gross breach of privacy, as the scanning and storage of biometric footprints often goes completely without consent, and is often even tested on lower-income neighborhoods without their consent, further creating a divide and continuing to marginalize massive swathes of the population as the technology continues to benefit rich, white male populations and actively harms populations of color, women, and people of lower socioeconomic standing. [Coded Bias, 2020 film]

Context of Use of Facial Recognition Algorithms

With surveillance devices such as cameras being expanded and the advancement of computer vision, facial recognition has become more integrated into facets of modern life.

Countries such as China have embraced facial recognition technologies with the advancement of smart cities. [ source ] Smart cities are technologically modern urban areas that use different types of electronic methods, voice activation methods, and sensors to collect specific data. [ source ] In addition, to enforce aspects of what’s known as the “social credit system” to the western world. Which is a system primarily in urban Chinese communities that reward ideal social and societal activities and punish those with a lower score when not doing so. Those with lower scores can have negative consequences such as higher loan rates, denial of children to top universities and high school, and lower priority to certain medical treatments, to name a few.

In western societies such as the US, facial recognition has been confirmed in uses for cybersecurity, law enforcement, and physical security. [ source ] 

 In the case of cybersecurity, facial recognition is used internally for access to government websites and devices. For physical security, facial recognition is used to monitor watchlists and control access to government facilities. The last use is for domestic law enforcement, where its primary use is to generate investigation leads and identify potential criminals.

Though this has been shrouded in controversy itself related to both privacy and ethical concerns. [ source ] These issues even prompted Amazon to ban its facial recognition software from law enforcement use. [ source ] In addition, Facebook was required to pay some 650 million USD for using images not publicly released from its users to train its facial recognition software. [ source ]

There have also been several reports of how facial recognition has led to false arrests. In a study conducted in 2019, it was found that facial recognition algorithms were less effective against both Asian and Black individuals. [ source ]

Dataset Creation

Storage of facial information cannot be treated the same as other forms of data since the face ties an identity directly to the information. By nature, these images are not protected by the anonymity of numbers and statistics. Most analytics can be distilled into population trends and averages, and this can distance specific user data from the individual itself. With facial data, this defeats the point of collecting it in the first place. Facial recognition is about picking out individuals and linking matching data out of the population, not the other way around.

This becomes more of an issue depending on where such information is located. Local and private storage solutions are much more secure compared to cloud-based storage options, but certain features are impossible without sharing it to centralized servers. Most mobile devices use local storage for fingerprints and faces, but Windows and other account-based logins are all stored on remote servers. Many companies will take the data and store it on their own servers so that they can use it to train new AI and optimize and improve facial recognition. This presents the issue that even if the company can be trusted, security breaches can happen, and the user no longer has autonomy over the data.

In the case of dataset creation and collection, consent is not just about the collection of data, but the different usages as well. Most people are okay with their information being collected for ads and general use if it helps optimize the experience and services, but their face is a different matter because of the inherent connection to identity and the lack of anonymity. Companies should be clear about whether data is collected for ad purposes, ai training, or anything else. Asking the user for consent under the general blanket of “sending analytics back” to improve functionality should not be enough since the nature of facial recognition data is so much more sensitive. Other than just the initial creation of the dataset, the creation of the algorithm itself can also be used in unforeseen ways. Facebook can suggest people to tag in posts based on its algorithm, even if the auto-tagging feature isn’t enabled [ source ]. It’s entirely possible that the exact same algorithm can be used to track down the identity of someone who hasn't given their consent at all, even if the original poster has.

Most of the time, companies make sure to include everything legally in a EULA, so legally it is there, but since the nature of facial recognition data is much more sensitive, it should be a separate agreement that is concise and apparent instead of buried behind a wall of text and documents. While collecting data, companies like Google obviously make sure to tell people they are being recorded, or their picture is being taken, but oftentimes they do not explain what it is for exactly. For example, they get the consent of homeless people by enticing them with cash, so consent isn’t the issue necessarily, but the lack of transparency for what the data collection is used for is lacking [ source ]. Most data can be comfortably assumed that it is collected for ad revenue, especially with Google,but with facial recognition image data sets, it is not really possible without profiling or without targeting an individual since every individual's face is unique. By nature, facial recognition applications target a specific user, and it is very easy to over-step personal boundaries when combining such technology with different services. Being transparent about the exact usage possibilities can help alleviate consent concerns. 

Ethical Face Recognition

Face recognition technologies have shown their usefulness by streamlining authentication at airports [ source ] and giving law enforcement another tool for fighting crime. [ source ] However, face recognition algorithms are not perfect, as we have shown, with significantly high false positives for communities with darker skin tones. [ source ] This means that advertising face recognition as the “ideal” tool for reducing crime or even using face recognition data for prosecuting suspects should be limited until the technology works perfectly. The use of face recognition algorithms for law enforcement has drawn significant backlash from engineers who developed such algorithms. [ source ] This backlash has forced companies such as Amazon to prevent law enforcement from using the face recognition algorithms they create. [ source ] But, the backlash cannot counter the inevitable increase in the use of face recognition for law enforcement. There is an urgent need for developing rules and guidelines for ethical use and the development of face recognition technologies. There are two targets for developing policies for ethical face recognition algorithms: dataset creation and the use of the algorithms.

Dataset development is critical for creating accurate algorithms for face recognition, as we discussed above. The prime example of unethical data collection is Clearview AI. Clearview AI is a company that sells its face recognition algorithms to law enforcement and continues to do so when there is a strong backlash against the biased nature of its algorithms. [ source ] The company created its dataset by scrapping photos of people from the internet. This means that it got photos on wedding websites, Twitter profile photos, Flickr photos, or even photos from news     websites, all without the people’s consent in those photos. This behavior has been judged illegal in Australia [ source ], and Clearview AI has been ordered to delete photos of any residents of Australia it captured. The techniques used by Clearview AI are deeply unethical. Ethical face recognition is when every single image used for training algorithms is used with the consent of the people in the picture. Moreover, commercial use of algorithms should mean that companies must pay the people whose images are used in datasets. Furthermore, all datasets unethically collected should be deleted.

The second method for ensuring ethical use of facial recognition is limiting the use of such algorithms. Users of face recognition models, such as law enforcement officials, need to recognize the biased nature of the algorithms. More importantly, face recognition algorithm results should not be used when prosecuting people. The data from face recognition algorithms should stand in court if we can prove that the algorithms are completely unbiased and do not function like black boxes similar to current algorithms. There is a desperate need for transparency to understand how models detect and identify people. 

However, the essential method for ensuring ethical face recognition is knowing that the technology is not ready for use. The algorithms ato understand algorithms’ datare black boxes, the datasets are biased towards lighter skin tones, and law enforcement officials are not trained on understanding data from algorithms. A “percentage match” with a criminal is not enough. It is essential to know what was matched. Wearing the same hat as a wanted person should not be an arrestable offense. Face recognition algorithms can make our communities safer, but the technology is not ready, and we should not develop new algorithms that assist the discriminatory nature of law enforcement in the United States. [ source ]

Hollister, Sean. “Google Contractors Reportedly Targeted Homeless People for Pixel 4 Facial Recognition.” The Verge, 2 Oct. 2019. Accessed Nov. 2021. 

Johnson, Khari. “Facebook Drops Facial Recognition to Tag People in Photos.” WIRED, 2 Nov. 2021. Accessed Nov. 2021. 

Botkin-Kowacki, Eva. “Humans Are Trying To Take the Bias Out of Facial Recognition Programs. It's Not Working - Yet.” News@Northeastern, 22 Feb. 2021. Accessed Nov. 2021. 

Kantayya, Shalini, director. Coded Bias. Netflix, 2020. 

Buolamwini, Joy Adowaa. “Gender Shades: Intersectional Phenotypic and Demographic Evaluation of Face Datasets and Gender Classifiers.” Massachusetts Institute of Technology, Massachusetts Institute of Technology, 2017, pp. 21–95. 

Keegan, Matthew. “In China, Smart Cities or Surveillance Cities?” U.S. News, 31 Jan. 2020. Accessed Nov. 2021. 

Rieger, Marc Oliver, et al. “What Do Young Chinese Think about Social Credit? It's Complicated.” Mercator Institute for China Studies, 26 Mar. 2020. Accessed Nov. 2021. 

Brandom, Russell. “Most US Government Agencies Are Using Facial Recognition.” The Verge, 25 Aug. 2021. Accessed Nov. 2021. 

Van Noorden, Richard. “The Ethical Questions That Haunt Facial-Recognition Research.” Nature, 18 Nov. 2020, Accessed Nov. 2021. 

Dastin, Jeffrey. “Amazon Extends Moratorium on Police Use of Facial Recognition Software.” Reuters, 18 May 2021, Accessed Nov. 2021. 

Singer, Natasha, and Cade Metz. “Many Facial-Recognition Systems Are Biased, Says U.S. Study.” The New York Times, 19 Dec. 2019, Accessed Nov. 2021. 

Yamanouchi, Kelly. “Delta Rolling out Facial Recognition Technology in Domestic Terminal at Atlanta Airport.” Atlanta Journal Constitution , 26 Oct. 2021, Accessed Nov. 2021. 

Cooper, Anderson. “Police Departments Adopting Facial Recognition Tech amid Allegations of Wrongful Arrests.” 60 Minutes, 16 May 2021, Accessed Nov. 2021. 

Hao, Karen. “A US Government Study Confirms Most Face Recognition Systems Are Racist.” MIT Technology Review, 20 Dec. 2020, Accessed Nov. 2021. 

Vincent, James. “Amazon Employees Protest Sale of Facial Recognition Software to Police.” The Verge, 22 June 2018, Accessed Nov. 2021. 

Press, Associated. “Clearview AI Sued in California over ‘Most Dangerous’ Facial Recognition Database.” Chicago Sun Times, 11 Mar. 2021, Accessed Nov. 2021. 

Lomas, Natasha. “Clearview AI Told It Broke Australia's Privacy Law, Ordered to Delete Data.” TechCrunch, 3 Nov. 2021, Accessed Nov. 2021. 

Voigt, Rob, et al. “Language from Police Body Camera Footage Shows Racial Disparities in Officer Respect.” Proceedings of the National Academy of Sciences, vol. 114, no. 25, 2017, pp. 6521–6526., https://doi.org/10.1073/pnas.1702413114. Accessed Nov. 2021.