All biometric traits contain both anatomical and behavioral components. For example, the fundamental vibration frequencies in both gait and speech are primarily anatomical, although both contain learned (i.e., language) and temporarily-acquired (i.e., shoes) behavioral characteristics. Faces are primarily anatomical, but can change both directly (facial expression) and indirectly (drug use) by behavior. Jain, et al (Proc. IEEE, 1997) have stated that fingerprints can vary based on manual labor, finger placement and finger pressure -- all behavioral traits. DNA, although neither anatomical nor behavioral, should be admissible as a “biometric trait” if automated. To resolve these problems, the international standards committee on biometrics (ISO/IEC JTC1 SC37) uses the following definition for biometrics: "the automated recognition of individuals based on their biological and behavioural traits".
Response: The definition of biometrics have been changed in the revised version to the more general one as suggested.
Even if the purpose of a person authentication system is to determine or verify an individual’s identity, biometric technologies have applications far broader than authentication. In many very large biometric systems, such as US-VISIT, Texas driver’s licensing, and New York Social Service, the primary application is to determine not who someone is, but that someone is not on a list of some kind.
Response: Use of biometric as a negative identifier is highlighted in introduction part.
Francis Galton's 1888 work in Nature, although important, was not the first on the distinctive nature of fingerprints. The first published work on the distinctiveness of fingerprints and their use in forensic human recognition was by Henry Faulds ( “On the Skin-Furrows of the Hand”, Nature , Oct. 28, 1880). Use of fingerprints for civil recognition was discussed the following month in the same publication by William Herschel (“Skin Furrows of the Hand”, Nature, November, 25, 1880).
Response: References for Fauld and Herschel have now been added.
2. Biometric systems
“Biometric systems are essentially pattern recognition systems (Duda et al., 2000)” I don’t believe such a statement is attributable to Richard Duda. Duda, Hart, and Stork, Pattern Classification (Wiley, Interscience, 2001) makes no mention of biometrics in the index. Biometric systems are, by definition, essentially human recognition systems. These systems must recognize the single source for all of the patterns from a human face, varying over pose, illumination, facial expression, adornment, or aging, or a fingerprint, varying over placement, pressure, moisture and injury. We might say that biometric systems are essentially pattern generator recognition systems, except that this would over-emphasize the importance of the imbedded algorithms. Rather, we should emphasize biometric systems as human-computer interface systems, where the actions of the humans are as important to the outcomes as the actions of the computers.
Response: The reference to Duda et al. was included as a citation for pattern recognition and not for biometrics. It has been removed to avoid any confusion. The definition of an individual is beyond the scope of this article (as also pointed out in one of the later review comments), so it would be confusing to emphasize recognition of the individual as a whole.
Not all biometric systems require an “enrollment” operation that “associates it (biometric information) with some identity information (name, age, etc.).” For example, the biometric systems used at Walt Disney World since 1996 collect no identity information about the data subjects. Forensic biometric systems can recognize “unknown” samples against “unknown” samples to link events. The use of biometrics in entirely anonymous health care applications storing no identity information about the data subject has also been proposed.
Response: Manuscript has been modified to emphasize reviewer's suggestion. An enrollment stage is needed to link some identity information to the collected biometric data. If no identity information is explicitly provided by a user (i.e., name, age,..), system generates a unique identifier for the biometric (e.g. the record index). In the Disney system, the enrollment takes place implicitly the first time a ticket is used when a link is made between the fingerprint of the ticket holder and the ticket itself.
Biometric samples stored in a database are called “biometric references”. These references might be “templates” (“features” extracted from a sample), “models” (a structure, such as Neural Network or Hidden Markov Model parameters, created from “features”) or a sample from which features are yet to be extracted. (such as a facial image, compressed and stored on a passport)
Response: We agree that biometric reference is a more general term than biometric template and we mention it now. But it may unnecessarily confuse the readers of Scholarpedia if we attempt to differentiate between features and model parameters.
Biometrics was defined in the first sentence of this article as “automatic recognition”. Therefore, if a system is not operating in or preparing samples for operation in a “recognition” mode, it is not biometrics. For example, neither a camera nor a telephone handset is a biometric collection device until it is part of a system capable of recognizing people. The two modes of biometric recognition might be described as “applications in which recognition is expected, and applications where recognition is not”, or perhaps “applications that confer benefits on those recognized and those that confer benefits on those not recognized”. Biometrics is only about recognition and linking persons through recognition to previously established information. The technologies cannot determine if “this person is who he claims to be”, but only that a recognized person is linked to an identity claim of a previous encounter. Establishing “who a person is” is a legal (or philosophical) problem outside the scope of biometrics.
Response: We have clarified the meaning of "Identity" and defined the purpose of a biometric recognition system to verify the identity of an individual.
The table quoting error rates mixes FMR statistics both normalized and un-normalized by the number of comparisons made, indicates an operationally non-relevant place on the ROC curve for some technologies, and tests the different technologies in entirely different applications with different populations. The results of the ICE 2004 tests were not from any operationally relevant application, but the NIST Speaker Recognition Evaluation and FpVTE results were. The only meaningful comparisons between biometric technologies are those in which all the technologies are tested in operationally commensurate environments, with full ROC curves reported for each, as in Mansfield, et al (“Biometric Testing Final Report”, National Physical Lab, 2001). No doubt advances in the technology have been made in the last 7 years, but this report is the most recent we have comparing technologies in a commensurate environment. Even in these tests, statements regarding technology error rates are limited to the population and application environment of the test.
Response: We have provided necessary caveats (e.g., dependence on database) for the results reported. Since a number of different factors are involved in evaluating different technologies, it is debatable if a fair and objective comparison is ever possible. We have added Mansfield reference.
For the comparison among different modalities (table 1) I suggest:
1) To replace <0.1%FNMR @ 1%FMR> with <0.6%FNMR @ 0.01%FMR> which is closer to a typical operating point of a Fingerprint-based systems. The suggested data come from the same test (FpVTE 2003)
2) To replace the cited FRVT 2006 experiment with another more difficult experiment carried out in the same test (FRVT 2006). In fact for the other three modalities (Fingerprint, Voice and Iris) the database is more difficult and not taken in optimal/controlled conditions such as for face.
Response: We have incorporated the suggestions.