Facial Recognition Software
Failing to Recognize Black Faces
Why Can’t it Accurately Identify Black Faces?
Georgetown’s Center on Privacy and Technology recently issued a study entitled The Perpetual Line-Up: Unregulated Police Face Recognition in America.
The study revealed that more than 110 million Americans, or more than half of all American adults, have been entered into a national database after the FBI used facial recognition software to scan the driver’s license photos of individuals in 26 states and counting.
Additionally, the FBI’s database includes biometric data beyond photographs, such as iris scans, voice prints, and fingerprints.
Privacy considerations aside, facial recognition software has been found time and time again to be inaccurate and unreliable. According to documents obtained by the Electronic Privacy Information Center in FOIA requests challenging the FBI’s Biometric Database, the software has an error rate of 20 percent.
This flawed database is widely available to law enforcement agencies at the local, state and national levels. Moreover, the facial recognition software is particularly inaccurate when it comes to identifying and distinguishing between and among the faces of African Americans and Asian Americans.
Watch Lists
Outside of the FBI, the LAPD has the second most extensive biometric database. The LAPD uses commercially available facial recognition software and hardware provided by for-profit companies such as DataWorks Plus LLC.
In Los Angeles and beyond, law enforcement agencies are compiling watch lists to cross-reference individuals with warrants or who are considered gang members with images collected and tabulated by drones and cameras equipped with facial recognition software purchased on the market.
When the watch lists are combined with the disproportionate use of drones in poor neighborhoods and the placement of cameras equipped with facial recognition software in Muslim and ethnic neighborhoods, a disturbing trend emerges.
People of color and Muslims are much more likely to be tracked with facial recognition software; when this is combined with the 20% error rate in proper identification, it becomes even more clear that the technology disproportionately harms communities of color. As such, the civil liberties rights of Muslim, Black and Asian Americans are being preemptively eroded before they have even come into contact with law enforcement agencies for a specific crime or infraction.
It bears noting that there are no federal laws governing facial recognition software. No warrants are required to use facial recognition databases. Moreover, in most cases officer’s do not even need to suspect that a crime has been committed before using such software to identify Americans.
While the legal, privacy and civil rights issues associated with the widespread use of facial recognition software is significant and interesting, I am more interested in how and why such software fails to accurately recognize the faces of Asian and African Americans.
What is the Source of Racial Disparities in Facial Recognition Software?
According to an article in The Atlantic, “Facial-Recognition Software May Have a Racial Bias Problem”, it appears that the facial recognition software on the market has the greatest difficulty identifying the faces of African Americans.
I found the article in The Atlantic, along with older articles in The Wall Street Journal on this topic “Regulator Warns Tech Companies of Big Data Bias” and “How Social Bias Creeps Into Web Technology”, striking.
They raised for me several questions, including:
1) What is the mechanism by which the tendency to misrecognize faces across racial categories is transmitted to algorithms?
2) Why are the faces of Black people the most difficult to identify correctly?
Surprisingly, it seems that the reasons for the facial-recognition bias in humans as well as in machines may be of similar origins: even down to why African American faces are the most difficult to identify.
I will tell you what it is not; algorithms don’t actually have a race. The software is not having a difficult time identifying black faces because it identifies as racially white and finds it difficult to distinguish black faces.
As to the larger question about the susceptibility of machine learning algorithms to replicate social biases, the predisposition is reflective of and constituted by the underlying social realities: realities that happen to include biases in a variety of forms.
First, some observations on algorithms and bias in this context.
What is an algorithm and are algorithms biased?
At the highest level of abstraction, an algorithm is the set of instructions, rules and operations that are designed to produce a targeted output based on an initial set of inputs.
Algorithmic outputs are a function of inputs, including programmatic instructions, vast data sets, a training environment and finally, the knowledge they accumulate from the real world environments into which they are released.
And while they are free to learn from the world around them, they do so within specific parameters.
From a nonscientific review of articles discussing bias in machine learning algorithms and facial recognition software, there are two primary ways in which bias in algorithms is constructed.
In the first construction, whether a given software is biased treats the algorithm as if it is sentient or as if it has some volition in the same sense that humans do. This sense of agency of course is missing from algorithms.
Secondly, when a question is posed regarding whether a given algorithm is biased, authors are often really asking whether the architects and programmers are biased. Understood in this way, neither construction is useful.
The social bias that is commonly attributed to humans found in the humanities and social sciences is often said to be a function of preferences, prejudices, predispositions, social conditioning, the amygdala, the limbic system etc.
The term carries with it — rightly or wrongly — some sense of proactive decision-making, ideas about agency and the feeling that someone could choose to act, think or feel differently. Especially if they are, we often claim, fortunate enough to meet new people, go to new places and learn about new ideas.
Surely, this cannot be the sense of the term ‘bias’ that applies to software. But many of the reports about this issue seem to do just that. They extend our everyday social understanding of bias and apply it to technology. And they do so by claiming that biases in the social/human sense are transmitted through biased programmers and machine learning architects: a claim that I do not find compelling. More importantly, this construction is not useful in identifying the source of the bias or its remediation and mitigation.
Programmers themselves are neither the original source of the social bias nor are they actively rendering the biased outcome. They are, however, taking large amounts of data, tabulating it, creating parameters and weighting factors to train learning algorithms that then produce biased outcomes.
Speculating about personal social biases isn’t useful because that is not where a corrective solution is going to be identified. We should avoid formulating the question as ‘are algorithms biased’.
When it comes to algorithms, we ought to focus on the data sets that are being used, the rules behind the programs and the instructions that are being provided. This approach, it seems to me, is much more useful in determining the origins of the cross-race effect facial-recognition software.
In short, we should focus on the outcomes and instead of bias per se; we should be focused on skewed data sets and statistical bias rather than social bias.
Are algorithmic outcomes biased?
We should instead ask: are algorithmic outputs and outcomes “reliable”, or “biased”, or “skewed.” Formulating the question in this manner means we can abandon the conventional use of the term “bias” found in the humanities and social sciences. And replace it with the term’s definition and common usage in the fields of data science and statistics.
So why does facial-recognition software exhibit what appears to be the same difficulty accurately recalling or recognizing the faces of people?
The answer lies, in part, in how and why humans fail to recognize the faces of people from other races. A phenomenon often referred to as the ‘cross-race effect’ or the ‘own-race bias’.
Humans and the Cross-Race Effect/Own-Race Bias
For decades, psychologists and sociologists have demonstrated that humans have difficulty identifying and distinguishing the faces of people from other races. Contrary to the jokes of many comedians, this tendency is not limited to just white people: it is common among all races and cultures.
The “cross-race effect” or “own-race bias” is a psycho/social phenomenon whereby people of all races frequently misidentify or fail to remember the faces of people who are of a different race.
While there is widespread agreement among scholars that the cross-race effect or own-race bias exists, there is little agreement about why it exists. Legal scholars and defense attorneys have steadily objected to eyewitness testimony in criminal prosecutions as unreliable in general, and particularly troublesome in those cases where the witness and the defendant are of different races.
As early as 1984, professor of law Sheri Lynn noted that “legal observers have long recognized that cross-racial identifications by witnesses are disproportionately responsible for wrongful convictions.” (Sheri Lynn Johnson, Cross-Racial Identification Errors in Criminal Cases, 69 CORNELL L. REV. 934, 935–36 (1984).
The cross-race effect/own-race bias is so well established that in some jurisdictions, psychologists are permitted to give expert testimony explaining the own-race bias/cross-race effect to the jury.
The one consensus that seems to emerge regarding the origins of the cross-race bias is that it is partially tied to exposure. We are less able to distinguish between faces from races that are unfamiliar to us and with which we do not have much meaningful contact in sufficient enough numbers (Tanaka, Kiefer, and Bukach, 2003).
As the authors note, the tendency to recognize faces from our own race is a function of the “racial experience such that people have more exposure and practice recognizing faces from their own race relative to faces of other races [ on a tangential note, it would be interesting to see studies conducted on populations of cross-race adoptions, or people of one race who grew up largely in communities outside of their race].
The Selection/Exposure Bias: Extending the Human Cross-Race Bias to Algorithms
Facial recognition software is unreliable, and particularly unreliable when it comes to the faces of non-white populations and the faces of women.
It is possible that the data set used in the software have a paucity of images of women, African Americans and Asians. The data sets used to train the algorithms behind facial recognition software are not available. According to the New Scientist, among the four largest commercial providers of this technology, none have provided the details behind their data sets.
It is possible that the data sets suffer from a selection bias that under-represents the faces of African American and Asian American faces. In the same way that humans have difficult recognizing faces from races with which they have minimal contact, we can hypothesize that algorithms trained on data sets that include minimal Asian and African American faces also suffer from a similar deficiency.
A selection bias at the level of the inputs included in facial recognition algorithms is likely to return skewed and unreliable results.
There may also be deficiencies in the instructions and the order of operations that guide the selection of an individual photograph in response to a query. The instructions and operations behind the commercial software are also unavailable and considered proprietary.
If a selection bias is in fact the cause of the software’s high rates of failure, there is a business question about the validity of the software being sold to law enforcement agencies and the legality of any resulting criminal apprehension or prosecutions of individuals based on said flawed software.
An Undesirable Solution
Barring laws that restrict and regulate the collection, use, and dissemination of biometric data collected by facial recognition software, the only solution is to decrease the rate of error.
If the political will to curtail the collection of vast troves of biometric data does not exist, then we must ensure that the datasets that are used to develop the software include more faces of African American and Asian faces.
Ultimately, this is also about poorly designed technology that has legal ramifications for downstream law enforcement consumers. This is also about the potential lawsuits that may be brought against companies that provide this software. They may very well who find themselves indemnifying police departments that are sued for racially-discriminatory civil rights violations through the use of such technology.
The biggest problem with facial-recognition software that can’t recognize black people is not that it can’t distinguish black features, and therefore the company has to deal with a public relations problem.
Rather, the biggest problem is that the software doesn’t work as intended. Companies that produce and sell facial-recognition software must explain to investors and customers that they have developed and are selling a product with a giant asterisk next to its core function.
*Software works as intended provided that the face is white and male.
Khullani Abdullahi is currently working on a book about the stories that lie within the conflict zones between the modern FBI, artificial intelligence, surveillance technologies, civil liberties, the movement and bodies of civilians, formal spies and subcontractors she is calling “Citizen Spies” and the emerging Electronic Panopticon (E-Panopticon).
By using the philosophy of French philosopher, Michel Foucault, emergent technologies, and the law, she hopes to build a new way of understanding the political coercion and control of contemporary political, actual and virtual bodies.
Send a DM on Twitter at @Khullani_