This is an old revision of the document!
When Biometrics Fail, Shoshana Amielle Magnet
https://www.dukeupress.edu/when-biometrics-fail
Defining Biometrics (p21)
A biometric attribute is defined as “a physical or psychological trait that can be measured, recorded, and quantified” (P. Reid 2004:5). The process of acquiring the information about the physical or psychological trait — whether a digital fingerprint, iris scan, or distinctive gait— and then storing that information digitally in a biometric system is called enrollment (P. Reid 2004:6; Nanavati, Thieme, and Nanavati 2002:17). A template is the digital description of a physical or psychological trait, usually containing a string of alphanumerical characters that expresses the attributes of the trait (P. Reid 2004:6). Before the biometric data are converted to a digital form, they referred to as raw data (Nanavati, Thieme, and Nanavati 2002:17). Raw biometric data are not used to perform matches —they must first be translated into a biometric template before they can be utilized—a process that is achieved with the help of a biometric algorithm. Biometric algorithms are frequently described as recipes for turning a person’s biological traits into a “digital representation in the form of a template” (P. Reid 2004:6). This recipe is usually proprietary, and it is what a biometric technology company sells, arguing that their recipe for creating a template from raw biometric data is better than another company’s recipe.
Vendors represent biometric technologies as able to answer two questions. The first question refers to identification and asks, Who am I? Often described as a 1:N matching process, the presentation of a biometric template created in real time (called a live biometric) is checked against a database of stored biometric templates. Used more commonly is security and law enforcement applications, this process allows for one person to be checked against as list of persons (P. Reid 2004:14). The second question that biometric technologies are imagined to be able to answer concerns verification: Am I who I say I am? Referred to as a 1:1 matching process, verification checks the presentation of the live biometric with the person’s template stored in the database to determine if they match. If the live biometric is the same as the stored biometric, there is a match and the identity is verified. Verification is held up in biometric discourse to be a much quicker process than identification, since it must check only one live biometric against one stored biometric, rather than checking a particular biometric against an entire database. Biometric technologies that rely of verification are more commonly used in physical and informational access applications, including secure building and computer network access (14)
Scientific Representations of Biometrics (p.32)
The number of reported failures in biometric technologies demonstrates the fallibility of any technology that takes as its starting assumption the consistency of bodily identity.
[…]
It is necessary to investigate those instances when biometric technologies fail and to ask what their failures tells us. The moments in which they fail are useful to identify the assumptions upon which they rely and the cultural context they encode.
(p.39) These experiments represent only a few of the many biometric studies that classify individuals based on bodily identity, each using slightly different methods or a different data set. The initial labeling of the images by gender or race in order to train the computer system must be done manually by the scientists themselves. The following statement is standard, demonstrating that so-called physiognomies differences are key: “The ground truth label for gender and ethnic origin were determined by visual inspection after the images where collected” (Gutta et al. 2000:198). That is, scientists themselves decide on the gender and race of an individual before using algorithms to train their computers to do the same. […] The ways that biometric technologies identify racialized and gendered bodies differently is known to scientists. From comparisons of the impact of biometric scanners that could identify some “races” better than others (Tanaka et al. 2004) to attempts to teach biometric systems to classify gender (Ueki et al. 2004), these studies help to explain biometric failures and their connection to race and gender identities. The papers just reviewed rely on racist research that is long debunked, while the host of empirical work in the past century on the complexity of bodily identity.