Li S.Z., Jain A.K. (eds.) Encyclopedia of Biometrics

Подождите немного. Документ загружается.

likelihood-ratio test for an unknown speech sample

can be formed between the match score of the speaker -

speciﬁc model and the UBM. The UBM may also

be used while training the speak er -speciﬁc model by

acting as a the prior model in Maximum A Posteriori

(MAP) parameter estimation.

Likelihood Ratio Test

To understand the development and use of a Universal

Background Model (UBM), the likelihood-ratio test

for which it is intended is ﬁrst described. Given an

observation, O, and a person, P, the task of veriﬁcation

is to determine if O was from P. This veriﬁcation task

can be restated as a basic

▶ hypothesis test between

: O is form person P

: O is not form person P

Using statistical

▶ pattern recognition techniques, the

optimum test to decide between these two hypotheses

is a likelihood ratio test (Strictly speaking, the likeli-

hood ratio test is only optimal when the likelihood

functions are known exactly. In practice this is rarely

the case.) given by

pðOjH

 y Accept H

< y Reject H



; ð1Þ

where p(OjH

), i ¼ 0,1 is the probability density func-

tion for the hypothesis H

evaluated for the mea-

surement Y, also referred to as the ‘‘likelihood’’ of the

hypothesis H

given the measurement (p(AjB)is

referred to as a likelihood when B is considered the

independent variable in the function.). The decision

threshold for accepting or rejecting H

is y. The basic

aim in developing a veriﬁcation system is to determine

techniques to compute this likelihood ratio function,

usually by ﬁnding method to represent and model

the two likelihoods, p(OjH

) and p(OjH

The ﬁrst step in a veriﬁcation system is to extract

from the observation features that convey person-

dependent information, such as vocal-tract related

spectral measurement when the observations are

speech samples in a speaker veriﬁcation system. The

output of this stage is typically a sequence of feature

vectors representing the observation, X ¼fx

1 ; :::;

These feature vectors are then used to compute the

likelihoods of H

and H

In statistical pattern recognition based veriﬁca-

tion systems, H

is represented by a model denoted

, that characterizes the distribution of features

derived from observations from the person P in the

feature space of x. For example, one could assume

that a Gaussian mixture model (GMM) distribution

best represents the distribution of feature vectors for

so that l

would be denoting the weights, means,

and covariance matrix parameters of a GMM. The

alternative hypothesis (see entry on

▶ GMM for more

details of this mode l). H

, is likewise represented by

a model l

. The likelihood ratio statistic is then

formed as

LRðXÞ¼

pðXjl

ð2Þ

See other articles in this book and chapters in [1, 2] for

more details on speaker veriﬁcation systems.

Alternative Hypothesis Modeling

While the model for H

is well deﬁned and can be

estimated using training samples from P, the model

for l

is less well deﬁned since it must potentially

represent the entire space of possible alternatives to

the person P.

From the area of speaker recognition, two main

approaches have been taken for this alternative hy-

pothesis modeling. Here terms ‘‘speakers’’ and ‘‘speech

samples’’ are used, but these apply equally to other

biometric measurements and features. The ﬁrst ap-

proach is to use a set of other speaker models to

cover the space of the alternative hypothesis. In various

contexts, this set of other speakers has been called

likelihood ratio sets [3], cohorts [4] and background

speakers [ 5]. Given a set of N background speaker

models {l

, ..., l

}, the alternative hypothesis model

is represented by

pðXjl

Þ¼FðpðXjl

Þ; ...; pðXjl

ÞÞ; ð3Þ

where FðÞis so me function, such as average or maxi-

mum, of the likelihood values from the background

speaker set. The selection, size, and combination

of the background speakers have been the subjec t of

much research (for example [4–6]). In general, it has

been found that to obtain the best performance with

this approach requires the use of speaker-speciﬁc back-

ground speaker sets. This can be a drawback in an

1350

Universal Background Models

application using a large number of hypothesized speak-

ers, each requiring its own background speaker set.

The second major approach to alternative hypoth-

esis modeling is to pool speech from several speakers

and train a single model. Various terms for this single

model are a general model [7], a world model, and a

universal background mode l [8]. Given a collection of

speech samples from a large number of speakers repre-

sentative of the population of speakers expected during

recognition, a single model, l

bkg

, is trained to represent

the alternative hypothesis. Research on this approach

has focused on the selectio n and composition of

the speakers and speech used to train the single

model [9, 10]. The main advantage of this approach

is that a single speaker-independent model can be

trained once for a particular task and then used for

all hypothesized speakers in that task. It is also possible

to use multiple background models tailored to speciﬁc

sets of speakers [10, 11].

Universal Background Models

Most modern speaker veriﬁcation system use a UBM for

modeling the alternative hypothesis in the likelihood

ratio test. Typically, GMMs are used for distribution

models and a speaker-speciﬁc model is derived by

using MAP estimation with the UBM acting as the

prior model (see article on GMMs for details on MAP

estimation). In the GMM–UBM system a single, speaker-

independent background model is used to represent

pðXjl

Þ. The UBM is a large GMM (2,048 mixtures)

trained to represent the speaker-independent distribu-

tion of features. Speciﬁcally, speech samples are selected

that are reﬂective of the expected alternative speech to

be encountered during recognition. This applies to

both the type and quality of speech, as well as the

composition of speakers. For example, for a veriﬁcation

system using telephone speech and only male speakers,

the UBM would be trained using telephone speech from

a pool of male speakers. In the case where such speciﬁc

prior knowledge of the gender composition of the alter-

native speakers is not known, speech samples from both

male and female speakers are used. Other than these

general guidelines and experimentation, there is no

objective measure to determine the right number of

speakers or amount of speech to use in training a UBM.

Given the data to train a UBM, there are many

approaches that can be used to obtain the ﬁnal

model. The simplest is to merely pool all the data

and use it to train the UBM via the EM algorithm.

One should be careful that the pooled data is balanced

over the subpopulations within the data. For example,

in using gender-independent data, one should be sure

that there is a balance of male and female speech.

Otherwise, the ﬁnal model will be biased towards

the dominant subpopulation. The same argument

can be made for other subpopulations such as speech

from different microphones. Another approach is to

train individual UBMs over the subpopulations in

the data, such as one for male and one for female

speech, and then pool the subpopulation models

together. This approach has the advantages that one

can effectively use unbalanced data and can carefully

control the composition of the ﬁnal UBM. Still other

approaches can be found in the literature (see for

example [10, 12]).

The concept of a UBM is also used for discrimina-

tive systems, such as Suppor t Vector Machines (SVM),

where explicit likelihood functions for the two hypoth-

esis are not used. In this case, the UBM refers to the

collection data from the general population used as

negative examples while training a person-speciﬁc dis-

criminate function [13].

Universal Background Models

1351

Acknowledgment

This work was sponsored by the Dep artment of De-

fense under Air Force Contract FA8721-05-C-0002.

Opinions, interpretations, conclusions, and recom-

mendations are those of the authors and are not nec-

essarily endorsed by the United States Government.

Related Entries

▶ Gaussian Mixture Models

▶ Session Effects on Speaker Modeling

▶ Speaker Matching

▶ Speaker Recognition, Overview

References

1. Benesty, J., Sondhi, M., Huang, Y. (eds.): Springer Handbook of

Speech Processing, vol. XXXVI. Springer, Berlin (2008)

2. Mu

ler, C. (ed.): Speaker Classiﬁcation I: Fundamentals, Features,

and Methods. Volume 4343/2007. Springer: Lecture Notes in

Computer Science, Berlin (2007)

3. Higgins, A., Bahler, L., Porter, J.: Speaker veriﬁcation using

randomized phrase prompting. Digital Signal Process. 1,

89–106 (1991)

4. Rosenberg, A.E., DeLong, J., Lee, C.H., Juang, B.H., Soong, F.K.:

The use of cohort normalized scores for speaker veriﬁcation. In:

International Conference on Speech and Language Processing,

Banff, Alberta, Canada, pp. 599–602 (1992)

5. Reynolds D.A.: Speaker identiﬁcation and veriﬁcation using

Gaussian mixture speaker models. Speech Commun, 17,

91–108 (1995)

6. Matsui, T., Furui, S.: Similarity normalization methods for

speaker veriﬁcation based on a posteriori probability. In: Pro-

ceedings of the ESCA Workshop on Automatic Speaker Recog-

nition, Identiﬁcation and Veriﬁcation, Martigny, Switzerland,

pp. 59–62 (1994)

7. Carey, M., Parris, E., Bridle, J.: A speaker veriﬁcation system

using alphanets. In: Proceedings of the International Conference

on Acoustics, Speech, and Signal Processing, Toronto, Canada,

pp. 397–400 (1991)

8. Reynolds, D.A.: Comparison of background normalization

methods for text-independent speaker veriﬁcation. In: Proceed-

ings of the European Conference on Speech Communication

and Technology, Rhodes, Greece, pp. 963–967 (1997)

9. Matsui, T., Furui, S.: Likelihood normalization for speaker veri-

ﬁcation using a phoneme- and speaker-independent model.

Speech Commun. 17, 109–116 (1948)

10. Rosenberg, A.E., Parthasarathy, S.: Speaker background models

for connected digit password speaker veriﬁcation. In: Proceed-

ings of the International Conference on Acoustics, Speech, and

Signal Processing, Atlanta, Georgia, USA, pp. 81–84 (1996)

11. Heck, L.P., Weintraub, M.: Handset-dependent background

models for robust text-independent speaker recognition. In:

Proceedings of the International Conference on Acoustics,

Speech, and Signal Processing, Munich, Germany, pp. 1071–

1073 (1997)

12. Isobe, T., Takahashi, J.: Text-independent speaker veriﬁcation

using virtual speaker based cohort normalization. In: Proceed-

ings of the European Conference on Speech Communication

and Technology, Budapest, Hungary, pp. 987–990 (1999)

13. Campbell, W.M.: Generalized linear discriminant sequence

kernels for speaker recognition. In: Proceedings of the Interna-

tional Conference on Acoustics, Speech, and Signal Processing,

Orlando, Florida, USA, pp. 161–164 (2002)

Universal Latent Workstation

TOM HOPPER

Criminal Justice Information Services Division, FBI

(Retired)

Synonym

ULW

Definition

In the past, the commercial automated ﬁnger-

print identiﬁcation systems (AFIS) used by local,

state, and federal criminal justice agencies have not

accounted for cross-jurisdiction al searching of latent

crime scene ﬁngerprints. The mobile criminal might

avoid identiﬁcation by simply crossing a jurisdictional

boundary. This lack of interoperability stems from

the fact that each AFIS vendor has a unique set of

features to characterize and match ﬁngerprints. Search-

ing multiple AFIS in this environment involves redun-

dant encodings of the latent ﬁngerprint on separate

workstations, each using a different vendor’s feature set.

The universal latent workstation (ULW) simpli-

ﬁes cross jurisdictional searches by enabling an exam-

iner to search multiple AFIS with a single ﬁngerprint

feature encoding. In many cases, the examiner will

edit the features to optimize the search for a particu-

lar AFIS but they will not need to reenter the case.

The ULW is based on an open interface standard

developed in cooperation with the AFIS vendors,

1352

Universal Latent Workstation

effectively establishing a common language for crime

scene investigators to share late nt ﬁngerprint identiﬁ-

cation services. The FBI provides the ULW software,

training, and support to criminal justice agencies at

no charge.

Introduction

Fingerprints have a long history in the investigation of

crime and play a major part in criminal apprehensions

and convictions [1]. Identiﬁcations start with the dis-

covery, development, and capture of

▶ latent ﬁnger-

prints at the crime scene or on an object used in

connection with the commission of a crime. The

print is then scanned into a latent workstation where

the

▶ features are extracted and formatted into a

search record. The search record is then sent to an

AFIS to ﬁnd candidate matches for the latent print. A

ﬁngerprint examiner makes the ﬁnal comparison and

identiﬁcation.

Within the US, every state and most of the larger

cities have an AFIS with arrest records from within

their jurisdiction. At the national level the FBI has the

integrated automat ed ﬁngerprint identiﬁcation system

(IAFIS) with arrest prints for the entire country,

currently about 55 million records. A latent print

search that is not identiﬁed in the state AFIS must be

searched in the national system to determine if it

matches arrest prints from another state. ULW was

developed to facilitate searching a single latent ﬁnger-

print in multiple AFIS.

Fingerprint Features

Fingerprints comprise ridges and furrows in a ﬂow like

pattern. The characteristics or information content of

the ridge structure is traditionally categorized into

three levels of detail. Level 1 includes the general

ridge ﬂow and pattern classiﬁcation (Fig. 1). A partic-

ular ridge ﬂow pattern is not unique to an individual

but can be used in ﬁltering out or excluding a portion

of a data base from further consideration [2]. If the

crime scene print is a whorl, for example, there is no

need to search arches and loops; however this strategy

can occasionally cause a miss at the classiﬁcation

boundaries [3].

Level 2 details are the ridge path characteristics

including bifurcations, end ings, and dots. The normal

parallel ﬂow of the ridges is occasionally disrupted

when a ridge ends or bifurcates into two ridges

(Fig. 2). These disruptions, called minutia, occur at

random locations and are the primary basis for identi-

ﬁcation. Level 3 features are the sweat pores and edge

texture that make up the ﬁner details along the ridges.

When level 3 features are visible, they provide the

examiner with additional points of comparison to

reach an identiﬁcation or exclusion decision [4]. Cur-

rent research is bringing level 3 detail into the auto-

mated AFIS search as well [ 5].

Universal Latent Workstation. Figure 1 Fingerprint ridge flow pattern classifications: Arch (left), Loop (center), Whorl

(right).

Universal Latent Workstation. Figure 2 Fingerprint

features: ridge ending and bifurcation (left), pores, dots,

and distinctive ridge edge detail.

Universal Latent Workstation

1353

Fingers are pliable and ▶ distortions in the ﬁnger-

print may cause some shifting in the positions of the

minutiae. By focusing on the topologi cal structure of

the minutiae, the examiner can accurately assess the

similarity between two ﬁngerprints without adverse

effects due to variance in minutia locations. The num-

ber of intervening ridges between corresponding mi-

nutia pairs will be the same even if one of the

impressions is stretched (Fig. 3). Additionally, an ex-

aminer may follow a ridge path forward from a minu-

tia to see if another event affects the same ridge.

Search Preparation with ULW

The clarity of a ﬁngerprint image determines how

difﬁcult it will be to locate and mark the minutiae.

Many latent prints are low contrast partial impressions

on a substrate or surface with texture and graphics that

make it difﬁcult to follow the ridge structure. On these

low quality prints, automated feature extraction may

not locate all the correct minutiae and will likely mark

several false minutiae. In this case, the examiner will

have to mark the minutiae in the a reas where the auto

▶ encoder could not follow the ridges. The ideal work

environment for the examiner would be to drop the

false minutiae from the auto encoding and display only

the correct minutiae for editin g and veriﬁcation. Un-

fortunately, the software does not know which minu-

tiae are false; otherwise they would not have been

included in the ﬁrst place. ULW addresses this problem

by assigning a conﬁdence score to each minutia based

on the local image quality. ULW will then only display

minutiae above a threshold conﬁdence score and the

examiner can adjust the threshold to their own prefer-

ence (Fig. 4).

Counting the number of intervening ridges

between neighboring minutiae has similar issues. It

is a very tedious task to do manually and on a clear

image the software will provide accurate results. As

the image quality degrades, a procedure is needed

to split the task between the software and the exami-

ner. ULW assigns a conﬁdence score to each ridge

count based on the clarity of the ridges traversed.

In ridge count veriﬁcation mode, ULW will present

the ridge counts in score sequence starting with

the lowest. Once the examiner reviews several con-

secutive ridge counts without ﬁnding any errors it is

likely that all the remaining ridge counts are correct

as well.

In addition to the ﬁngerprint features the search

record can include variables to limit the search t o a

subset of the ﬁle. Descriptors such as sex, eye color, and

weight range can be entered if available. Also, the

search can be focused on a geographic region by enter-

ing state codes. In practice, the variables that are most

commonly used are the ﬁnger number and pattern

classiﬁcation. Focusing the search can improve accu-

racy with a large system. When searching a latent print

with limited information a small percentage of the ﬁle

prints will tend to get high scores in the same range as

the true match. Reducing the size of the search popu-

lation will reduce the high scoring non-matches and

improve the chances that the true match makes it into

the candidate list [6].

Searching Multiple AFIS

The commercial AFIS used each by local, state,

and Federal criminal justice agencies have different

rules for marking the minutiae and other features on

a latent search. Figure 5 shows a diagram of two

Universal Latent Workstation. Figure 3 Exceptionally

high clarity latent print. The bifurcation (a) and the ridge

ending (b) are on the same ridge and the ridge count

between them is four intervening ridges. You can use

these relationships to compare this latent to the matching

print in Fig. 1.

1354

Universal Latent Workstation

different vendor’s encoding for the same print. The

minutiae placement rules for vendor on the left require

that the ridge ending be marked in the center of the

valley just beyond the actual end of the ridge. The

bifurcation in the center of the print is not marked

because it is in a high curvature region. The other

vendor, on the right, expects the ridge ending to be

marked right on the ending of the ridge and also

expects minu tiae in high curvature areas to be marked.

In addition to the differences in the placement of

minutiae, some vendors require ridge counts to four

neighboring minutia, some require ridge counts to

eight neighbors and others do not use ridge counts

[7]. In practice, the differences are important for accu-

racy but do not present a problem for the examiners.

ULW can create a latent search record for any of

the major AFIS vendors or translate a record from

one format to another. In some states ULW is used to

search both the state system as well as the FBI IAFIS.

The examiner ﬁrst marks the features for the state

system and if that search does not ﬁnd a match, they

can then translate the state search into an IAFIS search.

IAFIS requires eight ridge counts per minutia so ULW

would calculate the additional ridge counts. The

Universal Latent Workstation. Figure 5 Two different

vendor encodings for the same print.

Universal Latent Workstation. Figure 4 ULW screen display for marking minutia. This latent print matches the arch

print in Fig. 1 .

Universal Latent Workstation

1355

examiner would then verify and edit the new ridge

counts. Any ridge counts from the state search that

had already been veriﬁed would have a conﬁdence

score of 100 so they would not be veriﬁed a second

time. Even if the state search was conducted with a

latent workstation provided by the state AFIS vendor,

the search record can still be imported into ULW and

then converted using the same process in order to

conduct an IAFIS search.

The search records are formatted in accordance

with the ANSI/NIST standard: Data Format for the

Interchange of Fingerprint, Facial, & Other Biometric

Information [8]. This is a tagged ﬁeld format and

a block of ﬁelds have been reserved for each vendor.

The development of the standards and its use for lat-

ent searching has been a cooperative effort by state

and local agencies, the AFIS vendors, NIST, and the

FBI. The ULW was ﬁrst developed as a proof of concept

for standards based latent searching and is now in

use across the country. There are currently about

200 agencies using ULW to conduct almost 10,000

searches a month. Approximately, 10% of the cases

that make it up to a National level search produce

identiﬁcations.

Summary

Any effort to combat crime and the threat of terror is

dependent on cooperation and sharing information

across agencies. Sharing latent identiﬁcation services

within the US is complicated by the hierarchical net-

work of AFIS systems and the variation in latent search

feature sets. The FBI developed ULW to address thes e

issues and facilitate information sharing across the

Nation’s city, county, and state borders.

Related Entries

▶ Biometric Standards for Law Enforcement Applications

▶ Fingerprint Features

▶ Large Scale System Design

References

1. History of Fingerprinting. http://onin.com/fp/fphistory.html

2. Federal Bureau of Investigation: The Science of Fingerprints

(Classiﬁcation and Uses), 12–84 edn. US Government Printing

Ofﬁce, Washington DC (1984)

3. Maltoni, D., Maio, D., Jain, A.K., Prabhakar, S.: Handbook of

Fingerprint Recognition. Springer, New York (2003)

4. Ashbaugh, D.R.: Quantitative-Qualitative Friction Ridge Analy-

sis: An Introduction to Basic and Advanced Ridgeology. CRC

Press, West Palm Beach, FL (1999)

5. Chen Yi, Demirkus, M., Jain, A.K.: Pores and ridges: high-

resolution ﬁngerprint matching using level 3 features. Trans.

Pattern Anal. Mach. Intell. 29(1), 15–27 (2007)

6. Wilson, C., Garris, M., Watson, C.: Matching performance for

the US-V ISIT IDENT system using ﬂat ﬁngerprints, NISTIR

7110, May 2004, ftp://sequoyah.nist.gov/p ub /nist_i nternal_

reports/ir_7110.pdf

7. Federal Bureau of Investigation: CJIS electronic biometric

transmission speciﬁcation, appendix J, descriptions and ﬁeld

edit speciﬁcations For type-9 logical records. www.fbibiospecs.

org

8. ANSI/NST-ITL 1–2007: Data Format for the Interchange of

Fingerprint, Facial, & Other Biometric Information. http://ﬁn

gerprint.nist.gov/standard/

Unnecessary Data Collection

A service provider typically requires data from an end-

user in order to provide the appropriate type or level of

service to the end-user. Data that are not relevant for

this purpose are deemed unnecessary, and should be

labeled as optional. In principle, the service provider

must be able to explain why each and every data item

collected is necessary, while the end-user has the right

to such an explanation.

▶ Privacy Issues

Unsupervised

Unsupervised is the class information of data which are

not available. Algorithms are designed purely based on

the attributes of data and data distribution.

▶ Fusion, Rank-Level

▶ Linear Dimension Reduction

1356

Unnecessary Data Collection

Unsupervised Rank Level Fusion

The rank level fusion methods can be generally cate-

gorized under the headings of supervised and unsuper-

vised. The rank level fusion method that does not

require any training data to achieve the fusion of

ranks can be categorized as unsupervised method.

Thus, using unsupervised rank level fusion methods,

one can combine the ranks from different matchers

without any ‘‘teacher’’ or training data. The Highest

rank method and Borda count methods are the exam-

ples of unsupervised rank level fusion methods.

▶ Fusion, Rank-Level

Unvoiced Sounds

The unvoiced speech is generated by constriction of the

vocal tract narrow enough to cause a turbulent airﬂow,

which results in noise, e.g ., in fricatives like /f/, /s/, or

breathy voice (where the constriction is in the glottis).

Unvoiced plosives like /p/, /t/, /k/ fall into this category,

too.

▶ Speech Production

Usability

The extent to which a product can be used by speciﬁed

users to achieve speciﬁed goals. Usability testing

employs techniques to collect empirical data during

the observation of users using the product for a speciﬁc

task in order to rectify usability deﬁciencies of a prod-

uct. The ISO document 924111 discus ses three factors

that compose usabilit y: effectiveness, efﬁciency, and

satisfaction. The IEEE Standard Computer Dictionary

further describes usability as the ‘‘ease with which a

user can learn to operate, prepare inputs for, and

interpret outputs of a system or component.’’

▶ Ergonomic Design for Biometric Systems

▶ Hand Geometry

User Acceptance

MAREK REJMAN-GREENE

Home Ofﬁce Scientiﬁc Development Branch,

Sandridge, St Albans, Herts, UK

Definition

User acceptance: the demonstrated willingness within a

user group to employ information technology for the

tasks it is designed to support [1].

Introduction

The effective use of many of the applications that include

a biometric component requires users to follow speciﬁed

procedures, and hence, calls for cooperation from the

user. Such cooperation is predicated on the acceptance of

the biometric technology. In recent years, there have

been a number of studies addressing the nature and

extent of such acceptance, although our understanding

is still partial and further research is needed. Among the

often cited reasons for willingness to use a biometric

system is trust in the technology as well as in the organi-

zation holding biometric reference data – although

many of the major studies have not explored user accep-

tance in sufﬁcient depth to elicit the reasons for a lack of

acceptance among a minority of the population.

In a wider sense, user acceptance of IT technologies

has been the subject of research for more than 20 years,

as ever more complex systems were developed for use

in personal, corporate, and government applications

[2]. Unfortunately, many of the documented case stud-

ies are limited to self-reporting by survey respondents

rather than the obser vation of usage (e.g., obtaining

take-up metrics), and often research is addressed to

systems where the use of a technology is voluntary. The

validation of models such as TAM, the Technology

Acceptance Model of Fred Davis [3] (and its succes-

sors), remains to be demonstrated as each new tech-

nology tests their assumptions.

In his original paper, Dav is identiﬁed two deter-

minants of an individual’s intent to use an IT system:

perceived usefulness and perceived ease of use. Sub-

sequent developments of the model have also includ-

ed aspects such as subjective norm (the subject is

inﬂuenced by the positive attitudes of people in

User Acceptance

1357

positions of respect), personal experience, and de-

monstrability of the success of the new technology

in achieving a goal, as well as the extent to which the

use of the application enhances the subject’s status

or image [2].

Even though there is considerable empirical sup-

port for this ty pe of modeling in the deployment of

other IT components, such modeling has yet to be

validated for user acceptance of biometrically enabled

applications. Nevertheless, the literature on user accep-

tance of biometric systems offers numerous indicators

of intent to use – or not use, prominent among which

are concerns about privacy and data misuse, health

risks, usability issues, and uncertainty about system

reliability.

For some applications, especially where convenience

is the principal reason for using biometrics, user accep-

tance may not be sufﬁcient. In these instances, the

challenge is to recast the metrics for success in terms

of the users’ satisfaction with the application (and,

hence, with the biome tric elements). Furthermo re, we

can envisage systems where the use of the biometric is

so integrated with an application that it is no longer

seen to be exceptional, for example, in multiplayer,

multimedia games where veriﬁcation by facial image

and/or speech blends into the player’s immersive

environment.

Historical Context

Biometric systems have been deployed for over

30 years, yet the literature to support a uniﬁed view

on their acceptance has been sparse. Early trials of the

technolog y adde d a short questionnaire for partici-

pants in scenario tests, asking about acceptance and

comfort in use, w ithout a more detailed exploration of

the reasons for any concerns.

A 1995 review of user acceptance studies summarizes

the position at that time [4]. This notes the ground-

breaking 1988 study for the State of California’s De-

partment of Motor Vehicles on retinal scanning and

ﬁngerprint technologies. In this extensive study, con-

siderably more par ticipants declined to use the eye

method than the live scan ﬁnge rprint equipment –

perhaps related to rumors circulating at the time

about the risk of disease transmission through eye

ﬂuids.

Reference is also made in this review to the results

of the 1991 Sandia National Laboratories’ pioneering

performance evaluation of ﬁve biometric modalities

[5]. Nearly 100 employees and contractors took part

in the trial in ‘‘an ofﬁce-like environment.’’ Of these,

76 returned a questionnaire that sought their views

on matters that have appeared time and again in

subsequent studies: which devices required most con-

centration and most proﬁciency? w hich were most

frustrating to use? and which gave health and safety

or privacy concerns or were most intimidating? The

hand geometry system was rated as the most acceptable

in most of the categories, while a retinal scanning

system and one of the two voice systems were viewed

as the least acceptable.

Biometric technologies can be applied in a wide

range of systems: from personal use in mobile handsets

and laptops, and access control at doors of homes and

cars, through to the management of time and atten-

dance at work and the control of migration at national

borders. With such a diversity of uses, user acceptance

is likely to vary considerably according to the proposed

application and its context, since concerns that have

been raised in studies, such as ‘‘invasion of privacy,’’

trust in the quality of the database management, per-

ceptions of health risks [6] etc., are likely to differ in

prominence for each context.

More recently, as large-scale applications are

deployed, the trend has been to use telephone surveys,

even though respondents are unlikely to have experi-

enced the range of biometric modalities about which

they are questioned, nor they had the opportunity to

view this type of technology [7, 8]. Although these

surveys (often with over 1,000 participants) can pro-

vide valuable insights into the perception of biometric

technologies across wider populations, focus group

studies will be needed to elicit the rationale behind

the statistics and comment on speciﬁc issues based

upon users’ experience of the equipment (and its ap-

plication) at ﬁrst hand [6, 9].

Acceptance, as reported in surveys, will depend on

the context in which the question is put and the trust

that the respondents place in the integrity of the survey

organizers. As a consequence, reports that claim that a

speciﬁc biometric technology is more or less acceptable

to a population need to be examined critically. There

should always be suppor ting evidence of a validated

methodolog y in the context of one or more types of

1358

User Acceptance

application. As a minimum, a study should offer the

following in order to interpret its results:

1. A sam ple questionnaire and the associated tele-

phone scripts, especially if the questions relating

to biometrics form part of a more general survey

(scripts may describe certain features of the tech-

nology, while remaining silent about crucial aspects

of their operation);

2. The dates of beginning and completion of the sur-

vey, as media comments during the survey period

may affect the results;

3. The method of selection of participants, the num-

bers successfully contacted, together with an indi-

cation of reasons for nonresponse and any

weighting factors applied to the results; and

4. Any incentives offered

It is exp ensive to undertake combined scenario trials

and faci litated focus groups, along with extensive ques-

tionnaire surveys, even though this strategy is likely

to give the best indication of the user acceptance of

systems to be deployed in the future [9]. However,

their relevance begins to diminish with time as better

biometric equipment and user interfaces are devel-

oped, and the media debate moves on to other aspec ts

of the deployment.

The Importance of User Acceptance

Although there has been limited research to prove the

point, it is an established axiom in the ﬁeld that conﬁ-

dent and cooperative users of biometric technology are

a prerequisite for successful applications. Sasse [10]

notes the close link with usability of systems, since

systems that are difﬁcult to use, or where users have

unresolved concerns, will result in more veriﬁcation

errors, longer throughput times, and therefore, lead to

additional costs for the operator of the system. Re-

duced usability has also been associated w ith a reduc-

tion in the trust users place in a service.

Reference is also made by Sasse to the Biovision

Technology Roadmap [11], which concluded that three

factors lead to the acceptance of biometric technology:

trust in the security enhancement offered by this form

of authentication; greater convenience in use com-

pared with alternative systems; and trust in those hold-

ing the biometric data, maintaining the security of that

data and not using data for other purposes. In addition

to the absence of these factors, other reasons that lead

to systems being less accepted include fears of health

effects of long-term use of the equipment and the

reliability of the identiﬁcation process.

Principal Large Scale Studies

In 2001–2002, ORC Internat ional under took two

nationwide telephone surveys (commissioned by the

National Consortium for Justice Information and Sta-

tistics) to assess attitudes toward the use of bio-

metrics by government and the private sector [7].

As the ﬁrst was conducted directly after the events of

9/11, a follow up survey was needed a year later to

ensure that responses were not colored by the extensive

media discussion of the response to terrorist attacks.

The authors of the studies commented, however, that

very little had changed in the intervening time. Over

1,000 US adults were sampled in each year and the

full results of the survey, together with sample ques-

tionnaires and telephone scripts, are available in the

public domain.

The study found that only half of the sampled

population was aware of the existence of biometrics,

indicating that public acceptance would require fur-

ther education and marketing initiatives. Personal ex-

perience of its use was also very low: on ly 5% of the

sample (57 users) had ever used a biometric system,

with the majority reported as feeling comfortable with

the experience.

Support for the use of biometrics in major govern-

ment systems (passport veriﬁcation, entry to government

buildings, and for airport check-in) was above 80% in

2002, although these ﬁgures represented a decrease

in acceptance from the immediate post 9/11 period.

For applications designed to counter terrorism, about

two-thirds of the survey participants trusted that there

would be no unwarranted extension of their use (‘‘func-

tion creep’’). Nevertheless, there was strong support for

privacy safeguards in accordance with Fair Information

Practices. Other government applications that scored

highly were those in support of law enforcement and

reduction of social security fraud.

There was also clear support for private sector use

of biometrics, although expressed with markedly less

enthusiasm. Respondents voiced support for biometric

User Acceptance

1359