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

Подождите немного. Документ загружается.

thereby demanding the design of an effective (and

possibly more expensive) acquisition device.

4. Multi-sample systems. A single sensor may be used

to acquire multiple samples of the same biometric

trait to account for the variations that can occur in

the trait, or to obtain a more complete representa-

tion of the underlying trait. A face system , for

example, may capture (and store) the frontal pro-

ﬁle of a person’s face along with the left and right

proﬁles in order to account for variations in the

facial pose. Similarly, a ﬁngerprint system equipped

with a small size sensor may acquire multiple dab

prints of an individual’s ﬁnger to obtain images of

various regions of the ﬁngerprint. A mosaicing

scheme may then be used to stitch the multiple

impressions and create a composite image. One of

the key issues in a multi-sample system is deter-

mining the number of samples that have to be

acquired from an individual. It is important that

the procured samples represent the variability as

well as the typicality of the individual’s biome-

tric data. To this end, the desired relationship

between the samples has to be established before-

hand to optimize the beneﬁts of the integration

strategy. For example, a face recognition system

utilizing both the frontal- and side-proﬁle images

of an individual may stipulate that the side-proﬁle

image should be a three-quarter view of the face.

Alternately, given a set of biometric samples, the

system should be able to automatically select

the ‘‘optimal’’ subset that would best represent the

individual’s variability.

5. Multimodal systems. Multimodal systems establish

identity based on the evidence of multiple biomet-

ric traits. For example, some of the earliest multi-

modal biometric systems utilized face and voice

features to establish the identity of an individual .

Physically uncorrelated traits (e.g., ﬁngerprint and

Multibiometrics. Figure 1 Sources of information for biometric fusion.

970

Multibiometrics

iris) are expected to result in better improvement

in performance than the correlated traits (e.g.,

voice and lip movement). The cost of deploying

these systems is substan tially more due to the re-

quirement of new sensors and, consequently, the

development of appropriate user interfaces. The

identiﬁcation accuracy can be signiﬁcantly im-

proved by utilizing an increasing number of traits

although the

▶ curse-of-dimensionality phenome-

non would impose a bound on this number. The

number of traits used in a speciﬁc application will

also be restricted by practical considerations such

as the cost of deployment, enrollment time,

throughput time, expected error rate, user habitu-

ation issues, etc.

6. Hybrid systems. Chang et al. [7] use the term hybrid

to describe systems that integrate a subset of

the ﬁve scenarios discussed above. For example,

Brunelli et al. [8] discuss an arrangement in which

two speaker recognition algorithms are combined

with three face recognition algorithms at the match

score and rank levels via a HyperBF network. Thus,

the system is multi-algorithmic as well as multi-

modal in its desig n.

Levels of Fusion

Based on the type of information available in a certain

module, different levels of fusion may be deﬁned.

Sanderson and Paliwal [9] categorize the various levels

of fusion into two broad categories: pre-classiﬁcation

or fusion before matching, and post-classiﬁcation or

fusion after matching (see Fig. 2). Such a categoriza-

tion is necessary since the amount of information

available for fusion reduces drastically once the match-

er has been invoked. Pre-classiﬁcation fusion schemes

typically require the development of new matching

techniques (since the matchers used by the indiv idual

sources may no longer be relevant) thereby introdu-

cing additional challenges. Pre-classiﬁcation schemes

include fusion at the sensor (or raw data) and the

feature levels while post-classiﬁcation schemes include

fusion at the match score, rank and, dec ision levels.

1. Sensor-level fusion. The raw biometric data (e.g.,

a face image) acquired from an individual repre-

sents the richest source of information although it

is expected to be contaminated by noise (e.g ., non-

uniform illumina tion, background clutter, etc.).

Sensor-level fusion refers to the consolidation of

(1) raw data obtained using multiple sensors, or

(2) multiple snapshots of a biometric using a

single sensor.

2. Feature-level fusion. In feature-level fusion, the fea-

ture sets originating from multiple biometric algo-

rithms are consolidated into a single feature set by

the application of appropriate feature normalization,

transformation, and reduction schemes. The primary

beneﬁt of feature-lev el fusion is the detection of

corr elated featur e values generated by different bio-

metric algorithms and, in the process, identifying a

salient set of features that can improve recognition

accuracy. Eliciting this feature set typically requires

Multibiometrics. Figure 2 Fusion can be accomplished at various levels in a biometric system.

Multibiometrics

971

the use of dimensionality reduction methods and,

therefore, feature-level fusion assumes the availability

of a large number of training data. Also, the feature

sets being fused are typically expected to reside in

commensurate vector space in order to permit the

application of a suitable matching technique upon

consolidating the feature sets.

3. Score-level fusion. In score-level fusion the match

scores output by multiple biometric matchers

are combined to generate a new match score (a

scalar) that can be subsequently used by the veriﬁca-

tion or identiﬁcation modules for rendering an iden-

tity decision. Fusion at this level is the most

commonly discussed approach in the biometric lit-

erature primarily due to the ease of accessing and

processing match scores (compared to the raw bio-

metric data or the feature set extracted from the

data). Fusion methods at this level can be broadly

classiﬁed into three categories: density-based

schemes, transformation-based schemes, and classi-

ﬁer-based schemes.

4. Rank-level fusion. When a biometric system oper-

ates in the identiﬁcation mode, the output of the

system can be viewed as a ranking of the enrolled

identities. In this case, the output indicates the set of

possible matching identities sorted in decreasing

order of conﬁdence. The goal of rank level fusion

schemes is to consolidate the ranks output by

the individual biometric subsystems to derive a

consensus rank for each identity. Ranks provide

more insight into the decision-making process of

the matcher compared to just the identity of the

best match, but they reveal less information than

match scores. However, unlike match scores, the

rankings output by multiple biometric systems are

comparable. As a result, no normalization is needed

and this makes rank level fusion schemes simpler

to implement compared to the score level fusion

techniques.

5. Decision-level fusion. Many commercial off-the-

shelf (COTS) biometric matchers provide access

only to the ﬁnal recognition decision. When such

COTS matchers are used to build a multibiometric

system, only decision level fusion is feasible. Meth-

ods proposed in the literature for decision level

fusion include ‘‘AND’’ and ‘‘OR’’ rules, majority

voting, w eighted majority voting, Bayesian decision

fusion, the Dempster–Shafe r theor y of evidence,

and behavior knowledge space.

Summary

Multibiometric systems are expected to enhance the

recognition accuracy of a personal authentication sys-

tem by reconciling the evidence presented by multiple

sources of information. Typically, early integration

strategies (e.g., feature-level) are expected to result in

better performance than late integration (e.g., score-

level) strategies. However, it is difﬁcult to predict the

performance gain due to each of these strategies prior to

invoking the fusion methodology. While the availability

of multiple sources of biometric information (pertain-

ing either to a single trait or to multiple traits) may

present a compelling case for fusion, the

▶ correlation

between the sources has to be examined before deter-

mining their suitability for fusion [10]. Combining

uncorrelated or negatively correlated sources is expected

to result in a better improvement in matching perfor-

mance than combining positively correlated sources

[11]. However, deﬁning an appropriate diversity mea-

sure to predict fusion performance has been elusive

thus far. Other topics of research in multibiometrics

include (1) protecting multibiometric templates;

(2) indexing multimodal databases; (3) consolidating

biometric sources in highly unconstrained non-ideal

environments; (4) designing dynamic fusion algo-

rithms to address the problem of incomplete input

data; and (5) predicting the matc hing performance

of a multibiometric system.

Related Entries

▶ Fusion, Decision-Level

▶ Fusion, Feature-Level

▶ Fusion, Rank-Level

▶ Fusion, Score-Level

▶ Fusion, Sensor-Level

▶ Multiple Classiﬁer Systems

▶ Multiple Experts

References

1. Bigun, E.S., Bigun, J., Duc, B., Fischer, S.: Expert Conciliation

for Multimodal Person Authentication Systems using Bayesian

Statistics. In: Proceedings of First International Conference

on Audio- and Video-Based Biometric Person Authentication

(AVBPA), pp. 291–300. Crans-Montana, Switzerland (1997)

2. Hong, L., Jain, A.K., Pankanti, S.: Can Multibiometrics Improve

Performance? In: Proceedings of IEEE Workshop on Automatic

972

Multibiometrics

Identiﬁcation Advanced Technologies (AutoID), pp. 59–64. New

Jersey, USA (1999)

3. Ross, A., Nandakumar, K., Jain, A.K.: Handbook of Multibio-

metrics. 1st ed. Springer, New York, USA (2006)

4. Jain, A.K., Ross, A.: Multibiometric Systems. Communications of

the A CM, Special Issue on Multimodal Interfaces 47(1), 34–40

(2004)

5. Kuncheva, L.I.: Combining Pattern Classiﬁers - Methods and

Algorithms. Wiley (2004)

6. Ho, T.K.: Multiple Classiﬁer Combination: Lessons and Next

Steps. In: H. Bunke, A. Kandel (eds.) Hybrid Methods in Pattern

Recognition, Machine Perception and Artiﬁcial Intelligence,

vol. 47, pp. 171–198. World Scientiﬁc (2002)

7. Chang, K.I., Bowyer, K.W., Flynn, P.J.: An Evaluation of Multi-

modal 2D+3D Face Biometrics. IEEE Transactions on Pattern

Analysis and Machine Intelligence 27(4), 619–624 (2005)

8. Brunelli, R., Falavigna, D.: Person Identiﬁcation Using Multiple

Cues. IEEE Transactions on Pattern Analysis and Machine Intel-

ligence 17(10), 955–966 (1995)

9. Sanderson, C., Paliwal, K.K.: Information Fusion and Person

Veriﬁcation Using Speech and Face Information. Research

Paper IDIAP-RR 02-33, IDIAP (2002)

10. Poh, N., Bengio, S.: How Do Correlation and Variance of Base-

Experts Affect Fusion in Biometric Authentication Tasks?. IEEE

Transactions on Signal Processing 53(11), 4384–4396 (2005)

11. Kuncheva, L.I., Whitaker, C.J., Shipp, C.A., Duin, R.P.W.: Is

Independence Good for Combining Classiﬁers? In: Proceedings

of International Conference on Pattern Recognition (ICPR),

vol. 2, pp. 168–171. Barcelona, Spain (2000)

Multibiometrics and Data Fusion,

Standardization

JUNG SOH

,FARZIN DERAVI

,ALESSANDRO TRIGLIA

LEX BAZIN

University of Calgary, Calgary, AB, Canada

University of Kent, Canterbury, Kent, UK

OSS Nokalva, Inc., Somerset, NJ, USA

Fujitsu Services, London, UK

Synonyms

Biometric fusion standardization; Multibiometric

fusion standardization

Definition

Multibiometrics is the automated recognition of indi-

viduals based on their biological or behavioral charac-

teristics and involves the use of biometric fusion. Some

applications of biometrics require a level of technical

performance that is difﬁcult to obtain with a single

biometric measure. Preventing illegitimate multiple

applications by the same individual for national iden-

tity cards and checking security for air travel are exam-

ples of such applications. In addition, provision is

needed for people who are unable to give a reliable

biometric sample for some biometric modalities. Use

of multiple biometric measurements from substan-

tially independent biometric sens ors, algo rithms, or

modalities typically gives improved technical perfor-

mance, increases system ﬂexibility and reduces security

risks. This includes an improved level of performance

where not all biometric measurements are available

such that decisions can be made from any number of

biometric measurements within an overall policy on

accept/reject thresholds. At the current level of under-

standing, combining results from different biometric

sources at the matching score level typically requires

knowledge of both genuine and impostor distribu-

tions of such scores. Such distributions are highly

application-dependent and generally unknown in

any real system. Research on the methods not requir-

ing previous knowledge of the score distributions is

continuing and research on fusion at both the image

and feature levels is still progressing. Preliminary work

on ISO/IEC international standardization of multibio-

metrics has culminated in a Technical Report, while in

the United States substantial progress has been made

on standards to support multibiometrics.

Overview of Multibiometric Systems

In general, the use of the terms ▶ multimodal or

▶ multibiom etric indicates the presence and use of

more than one

▶ biometric modality, sensor, instance,

and/or algorithm in some form of combined use for

making a speciﬁ c biometric identiﬁcation or veriﬁca-

tion decision. The methods of combining multiple

samples, matching scores, or matching decisions can

be very simple or mathematically complex.

Multimodal biometrics were ﬁrst proposed, imple-

mented and tested in the 1970s. Co mbining measures

was seen as a necessary future requirement for biomet-

ric systems. It was widely thought that combining

multiple measures could increase either security by

decreasing the false acceptance rate or user conve-

nience by decreasing the false rejection rate. These

Multibiometrics and Data Fusion, Standardization

973

early systems, however, did not seem to advance into

practical applications.

The use of fusion and related methods has been a

key tool in the successful implementation of large-scale

automated ﬁngerprint identiﬁcation systems (AFISs),

starting in the 1980s. Until recently, multiple modal-

ities have not been used in AFIS; however, many meth-

ods of fusion have been successfully implemented

using ﬁngerprints alone. Some of the ways that fusion

has been implemented in AFISs include:

1. Image (i.e., sample) fusion in creating a single

‘‘rolled’’ image from a series of plain impressions

on a livescan device.

2. Template fusion in the use of multiple feature ex-

traction algorithms on each ﬁngerprint image.

3. Multiinstance fusion in the use of ﬁngerprints from

all ten ﬁngers.

4. Multipresentation fusion in the use of rolled and

slap (plain) ﬁngerprints.

5. Algorithm fusion for the purpose of efﬁciency

(cost, computational complexity, and throughput

rate); generally matchers are used as a series of

ﬁlters in order of increasing computational com-

plexity. These are generally implemented as a mix

of decision- and score-level fusion.

6. Algorithm fusion for the purpose of accuracy (de-

creasing false accept rate and/or false reject rate,

lessening sensitivity to poor-quality data); matchers

are used in parallel, with fusion of resulting scores.

The use of fusion has made AFISs possible, because

of fusion’s increase in both accuracy and efﬁciency.

To further understand the distinction among the

multib iome tric categori es, Table 1 illustrate s the

basic distinctions among categories of multibiometric

implementation. The key aspect of the category that

makes it multi-‘‘something’’ is shown in boldface.

Multimodal biometric systems take input from single

or multiple sensors that capture two or more biometric

characteristics of different modalities. For example, a sin-

gle system c ombining face and iris information for bio-

metric recognition would be considered a ‘ ‘ multimodal’’

system regardless of whether face and iris images were

captured by different imaging devices or the same device.

It is not required that the various measures be mathemat-

ically combined in any way. For example, a system with

ﬁngerprint and voice recognition would be considered

‘‘multimodal’’ even if the ‘‘OR’’ rule was being applied,

allowing users to be veriﬁed using either of the modalities.

Multialgorithmic biometric systems receive a sing le

sample from a single sensor and process that sample

with two or more algorithms. This technique could be

applied to any modality. Maximum beneﬁt (theoreti-

cally) would be derived from algorithms that are based

on distinctly different and independent principles

(such algorithms may be called ‘‘orthogonal’’).

Multiinstance biometric systems use one (or possi-

bly multiple) sensor(s) to capture samples of two or

more different instances of the same biometric character-

istic. For example, systems capturing images from multi-

ple ﬁngers are considered to be multiinstance rather than

multimodal. However, systems capturing, for example,

sequential frames of facial or iris images are considered to

be multipresentation rather than multiinstance.

Multisensorial biometric systems sample the same

instance of a

▶ biometric characteristic with two or

more distinctly different sensors. Processing of the

multiple samples can be done with one algorithm, or

Multibiometrics and Data Fusion, Standardization. Table 1 Multibiometric categories illustrated by the simplest case

of using 2 of something

Category Modality Algorithm Biometric characteristic (e.g., body part) Sensor

Multimodal 2 (always) 2 (always) 2 (always) 2 (usually)

Multialgorithmic 1 (always) 2 (always) 1 (always) 1 (always)

Multiinstance 1 (always) 1 (always) 2 instances of 1 characteristic (always) 1 (usually)

Multisensorial 1 (always) 1 (usually)

1 (always, and same instance) 2 (always)

Multipresentation 1 1 1 1

Exception: a multimodal system with a single sensor used to capture two different modalities (e.g., a high resolution image used to

extract face and iris or face and skin texture)

Exception may be the use of two individual sensors each capturing one instance (e.g., possibly a two-ﬁnger ﬁngerprint sensor)

It is possible that two samples from separate sensors could be processed by separate ‘‘feature extraction’’ algorithms, and then through

a common comparison algorithm, making this ‘‘1.5 algorithms,’’ or two completely different algorithms

974

Multibiometrics and Data Fusion, Standardization

some combination of multiple algorithms. For exam-

ple, a face recognition application could use both a

visible light camera and an infrared camera coupled

with a speciﬁc frequency (or several frequencies) of

infrared illumination.

For a speciﬁc application in an operational envi-

ronment, there are numerous system design considera-

tions, and trade-offs that must be made, among factors

such as improved performa nce (e.g., identiﬁcation or

veriﬁcation accuracy, sy stem speed and throughput,

robustness, and resource requirements), acceptabi lity,

ease of circumvention, ease of use, operational cost,

environmental ﬂexibility, and population ﬂexibility

[1]. Especially for a large-scale human identiﬁcation

system, there are additional system design considera-

tions suc h as operation and maintenance, reliability,

system acquisition cost, life cycle cost, and planned

system response to identiﬁed susceptible means of

attack, all of which will affect the overall deployability

of the system [1].

Levels of Combination

As a basis for the deﬁnition of levels of combination in

multibiometric systems, Fig . 1 shows a sing le-biome tric

process. A biometric sample captured by a biometric

sensor (e.g., a ﬁngerprint image) is fed into the feature

extraction module. Using sign al processing methods,

the feature extraction modu le converts a sample into

features (e.g., ﬁngerprint minutiae), which form a rep-

resentation apt for matching. Usually, multiple features

are collected into a feature vector. The matching mod-

ule takes the feature vector as input and compares it to

a stored template (a type of biometric reference, as

deﬁned in [2]). The result is a match score, which is

used by the decision module to decide (e.g., by applying

a threshold) whether the presented sample matches

with the stored template. The outcome of this decision

is a binary match or non-match. Generalizing the

above process to a multibiometric one, there are sever-

al levels at which fusion can take place: (1) decision

level; (2) match score level; (3) feature level; and (4)

sample level.

Decision-level fusion takes place only after the

results of matching from all biometric components

are available. The decision module outputs match or

non-match as a binary decision value. If a biometric

system consists of a small number of biometric com-

ponents, assigning logical values to match outcomes

allows fusion rules to be formulate d as logical func-

tions. For two decision-level outputs, two most com-

monly used logical functions are logical AND and

OR. For many decision-level outputs, various voting

schemes can be used as fusion rules, the most common

of which is majority voting. The logic al AND and OR

functions can be considered as voting schemes.

In score-level fusion, each system provides match-

ing scores indicating the proximity of the feature vec-

tor with the template vector. These scores can then be

combined to improve the matching performance. The

match score output by a matcher contains the richest

information about the input biometric sample in the

absence of feature-level or sensor-level information.

Furthermore, it is relatively easy to access and combine

the scores generated by several different matchers. Con-

sequently, integration of information at the match score

level is the most common approach in multimodal

biometric systems. From a theoretical point of view,

biometric processes can be combined reliably to give a

guaranteed improvement in matching performance.

Any number of suitably characterized biometric pro-

cesses can have their matching scores combined in

such a way that the multibiometric combination is

guaranteed (on average) to be no worse than the best

of the individual biometric devices. The key is to

Multibiometrics and Data Fusion, Standardization. Figure 1 Generic single-biometric process. (Reproduced from

Figure 2, Section 5.1 of ISO/IEC TR 24722:2007 Information technology – Biometrics – Multimodal and other

multibiometric fusion).

Multibiometrics and Data Fusion, Standardization

975

identify correctly the method which will combine these

matching scores reliably and maximize the improve-

ment in matching performance. The mechanism (for

this sort of good combination of scores within a multi-

biometric system) must follow at least two guidelines:

(1) each biometric process must produce a score and

make it available to the multibiometric combiner;

and (2) in advance of operational use, each biometric

process must make available to the multibiometric

combiner, its technical performance (such as score dis-

tributions) in the appropriate form (and w ith sufﬁcient

accuracy of characterization). Both veriﬁcation (1:1) and

identiﬁcation (1:N) systems can support fusion at the

match score level.

In the context of veriﬁcation, there are two distinct

approaches to formulate a score-level fusion problem:

(1) classiﬁcation; and (2) combination [3]. In the clas-

siﬁcation approach, a feature vector is constructed

using the matching scores output by the individual

matchers; thi s feature vector is then classiﬁed into

one of two classes: ‘‘Accept’’ (genuine user) or ‘‘Reject’’

(impostor). Generally, the classiﬁer used for this pur-

pose (e.g., decision tree, neural network, support vec-

tor machine, k-nearest neighbor, random forest, etc.) is

capable of learning the decision boundary, given some

training data, irrespective of how the feature vector is

generated [4]. Hence, the output scores of the different

modalities can be non-homogeneous (distance or similar-

ity metric, different numerical ranges, etc.) and no proces-

sing is required prior to presenting them to the classiﬁer .

In the combination approach (see Fig. 2), the individual

matching scores are combined to generate a single scalar

score, which is then used to make the ﬁnal decision [5]. To

ensure a meaningful combination of the scores from

the different modalities, if necessary, the scores may be

ﬁrst transformed to a common domain prior to com-

bining them. This is known as score normalization.

Score normalization methods attempt to map the

scores of each biometric process to a common domain.

Some reasons why scores need to be normalized prior to

fusion include [3]: (1) the matching scores at the output

of the individual matchers may not be homogeneous.

For example, one matcher may output a distance (dis-

similarity) measure while another may output a proxim-

ity (similarity) measure; (2) the outputs of the individual

matchers need not be on the same numerical scale

(range); and (3) the matching scores at the output of

the matchers may follow different statistical distribu-

tions. Some approaches are based on the Neyman-

Pearson lemma [6], with simplifying assumptions. For

example, mapping scores to likelihood ratios allows

them to be combined by multiplying under an

Multibiometrics and Data Fusion, Standardization. Figure 2 A framework for score-level fusion using

combination approach. (Reproduced from Figure 5, Section 5.3.2 of ISO/IEC TR 24722:2007 Information technology –

Biometrics – Multimodal and other multibiometric fusion).

976

Multibiometrics and Data Fusion, Standardization

independence assumption. Other approaches may be

based on modifying other statistical measures of the

match score distributions. The parameters used for

normalization can be determined using a ﬁxed training

set or adaptively based on the current feature vector.

In feature-level fusion, biometric information is

fused after feature extraction but before matching.

The simplest form is to integrate the feature vectors

(or sets if there is no implicit correspondence) of com-

ponent biometrics and to apply feature classiﬁcation

methods to the combined feature vector. Where features

from contributing multibiometrics are not indepen-

dent, good feature-level combination should, in

some circumstances, allow dependencies to be more

fully exploited than by solely using score-level combina-

tion. This should give better overall performance. How-

ever, fusion at this level is difﬁcult to achieve in practice

because of the following reasons: (1) the feature

vectors of multiple modalities may be incompatible

(e.g., minutiae set of ﬁngerprints and Eigen-coefﬁcients

of face); (2) the relationship between the feature spaces

of different biometric systems may not be known;

(3) concatenating two feature vectors may result in a

feature vector with very large dimensionality leading to

the ‘‘curse of dimensionality’’; and (4) a signiﬁcantly

more complex matcher might be required in order to

operate on the concatenated feature vector [7].

Notwithstanding these challenges, fusion at the

feature level has been attempted in several contexts.

Chang et al. [8] demonstrate feature-level fusion of

face and ear modalities showing signiﬁcant improve-

ments in performance. Kumar et al. [9] integrate the

palm-print and hand geometry features of an individ-

ual in order to enhance matching performance. In

their experiments, fusion at the match score level was

observed to be superior to fusion at the feature level.

However, Ross and Govindarajan [2] combine the

hand and face modalities of a user (multibiometrics)

as well as the R, G, B channels of the face image

(multisensorial) of a user at the feature level and dem-

onstrate that a feature selection scheme may be neces-

sary to improve matching performance at thi s level.

State of International Standardization of

Multibiometrics

At the time of this writing (late 2008), no standards on

multibiometrics have been developed within ISO/IEC.

However, the ISO/IEC subcommittee on biometrics

(ISO/IEC JTC 1/SC 37) had instead produced a Technical

Report on multibiometrics [10] (hereafter referred

to as the ‘‘Technical Report’’), which contains descrip-

tions and analyses of current practices on multibio-

metric fusion and provides useful information for

future development of international standards in

this area.

According to the Technical Report, there are many

ways of combining multibiometric processing and

performing

▶ biometric fusion, not all of which can

be made part of a biometric fusion standard. It is likely

that future biometric fusion standardization activity

within ISO/IEC will be of ﬁve types:

1. Record formats: The deﬁnition and standardization

of data to be exchanged be tween processes and

stored on various media. The biometric data inter-

change formats speciﬁed in SC 37/WG 3 standards

are examples of this type of standard. Another

example is the Fusion Information Format national

standard developed in the US [11]

2. Interfaces: Deﬁnition of standard APIs for pro-

cesses, the record formats used by the processes,

and the initialization procedure of the processes in

a system. The BioAPI standard [12] developed in

SC 37/WG 2 is an example of this type of standard,

which mig ht have to be modiﬁed in order to sup-

port multibiometrics and fusion data. The US ver-

sion of the BioAPI standard [13] has been amended

[14] to support biometric fusion

3. Application proﬁle: A standard containing a list of

references to provisions of one or more other stan-

dards, which are speciﬁed as optional in those

standards but are made mandatory by this standard

in order to facilitate interoperability in a particular

set of use cases. The SC 37/WG 4 project on ILO

(International Labor Organization) Seafarer ID

proﬁle [15] is an example of this type of standard

4. Conformance testing: A description of the criteria

and test data that allows for the assurance that

systems have complied with the standards. These

types of standards are under development in SC 37

for the biometric record formats

5. Per formance testing: Online testing of biometric

systems is complicated by the implied existence of

multiple and sequential sensors. A testing protocol

that develops procedures for doing this should be

established

Multibiometrics and Data Fusion, Standardization

977

The use of multibiometric systems has been consid-

ered for two major and differing use cases. The ﬁrst is

high-securi ty biometric use where the combination of

biometrics provides a stronger assurance of impostor

rejection for a relatively small, trained population.

The second is in the context of large-scale ID systems,

such as travel document systems, where the multibio-

metric combination may provide for the reduction of

rejection rates and easier system usage for a very large,

untrained population. In the context of the large-scale

ID systems, there can be many solution providers

providing components to the overall system. For ex-

ample, the creator of the electronic biometric docu-

ment may not be the same vendor that creates the

physical document, and neither may be the vendor

that performs the biometric test(s) (veriﬁcation

or identiﬁcation) during the document’s usage. This

situation can clearly beneﬁt from a biometric fusio n

standard when the document contains multiple

biometrics.

In the context of biometric fusion, one can propose

the following multibiometric system interoperability

requirements:

1. Standard multibiometric systems may be required

to be designed and certiﬁed (or evaluated) based on

common performance requirements. These perfor-

mance requirements should be independent of the

biometric modalities in use. This includes perfor-

mance measures such as failure to enroll, failure to

acquire, false rejection rate, false acceptance rate,

system throughput, and the resistance to active

impostor attacks

2. Standard multibiometric systems may be required

to be designed so that a single biometric subsystem

can be separately upgraded. All biometric device

characteristics change over time as research and

development improves accuracy and lowers cost.

The development of each biometric system however,

proceeds on its own timeline. Therefore only if

separate upgrading is possible it will be convenient

to upgrade a multib iometric system in the ﬁeld

3. A standard multibiometric system may be required

to be able to accept historical information for a

given user, such as scores and processing times.

With this information, the system can be optimized

in both security and throughput to take advantage

of the type of biometric modality that is favored by

the particular user

4. Standard multibiometric systems may be required

to be compatible with existing standard-based sys-

tems that use a single biometric characteristic. To

achieve support of the system requirements, exist-

ing biome tric technical interfaces, such as the

BioAPI standard, may need to be revised to provide

support for fusion while allowing the use of inde-

pendently developed BioAPI Biometric Service

Providers (BSPs) each implementing a single bio-

metric modality

State of Standardization of

Multibiometrics in the United States

In the United States a signiﬁcant amount of work had

been done, at the time of this writing, to produce

the ﬁrst two national standards on multibiom etrics.

The two standards that have been developed so far are:

1. An amendment [14] to the US version of the

BioAPI standard [13] that adds support for bio-

metric fusion (published)

2. A data format carrying information in support of

score-level fusion [11] (published)

The changes made to BioAPI 1.1 to support fusion

were inspired by the Technical Report. The Technical

Report describes a multibiometric process as a combi-

nation of processes, and describes the inputs and the

outputs of each process. This general model is now

reﬂected in BioAPI 1.2.

In BioAPI 1.2, the concept of Biometric Informa-

tion Record (BIR) was generalized, and now includes

the following types of biometric data (sometimes

called ‘‘processed levels’’):

1. Raw biometric data (e.g., a raw image produced for

audit purposes during a capture operation)

2. Intermediate biometric data (e.g., an image ready

to be passed as input to a feature extraction process)

3. Processed data (e.g., features)

4. A matching score (as produced by some matching

algorithm)

5. A matching decision (a match/non-m atch output)

6. Personal data (any non-biometric data speciﬁc to

the subject that can be passed as input to a biomet-

ric operation)

Correspondingly, the BioAPI operations that handle

biometric data (either accepting one or more BIRs as

978

Multibiometrics and Data Fusion, Standardization

input or producing a BIR as output) have been exten-

ded in BioAPI 1.2 to support the new types of BIRs.

For example, an important evolution was the addition

of a function similar to VerifyMatch but producing a

BIR of type ‘‘score’’ instead of the traditional ‘‘achieved

FAR’’ (a simple integer value) and ‘‘result’’ (a simple

boolean value). In BioAPI 1.2, the new Fuse function

can take as input two or more ‘‘score’’ BIRs to produce

either a new ‘‘score’’ BIR or a ‘‘decision’’ BIR (score-

level fusion); it can also take as input two or more

‘‘decision’’ BIRs to produce a new ‘‘decision’’ BIR

(decision-level fusion) or two or more ‘‘processed’’

BIRs to produce a new ‘‘processed’’ BIR (feature-level

fusion). The Fuse function can be implemen ted either

by a regular BSP that also performs capture and/or

matching, or by a specialized ‘‘fusion BSP.’’ The new

functions also accept as input any number of addition-

al BIRs of any type (including ‘‘personal’’) carrying

information that can be used by the BSP to increase

the quality of the output of the function.

The Fusion Information Format standard [11] spe-

ciﬁes a container of information about the statistical

distribution of similarity scores for a particular bio-

metric technology (including a matcher). The standard

supports the representation of both genuine and im-

postor distributions, and provides four different ways

of representing tho se distributions: location and scale

parameters, empirical cumulative distribution func-

tion, B-spline function ﬁt of the empirical cumulative

distribution function (CDF) , and interpolant of the

CDF. This information is intended to be provided as

input to a software component that performs score-

level fusion within a multib iometric system. When

such a software component is invoked to fuse the

scores produced by two or more comparison subsys-

tems, the knowledge of the score distributions of each

comparison subsystems will allow it to correctly pro-

cess the corresponding input scores.

Summary

The building blocks for a biometric fusion standard

would be mainly of two types: data records and

processes. The three key factors driving the imple-

mentation of the fusion algorithms will be: interop-

erability, performance, and industry consensus.

Sample-, feature-, score-, and decision-level fusion

have been identiﬁed from the preliminary work on

biometric fusion standardization. However, support-

ing sample- or feature-level fusion will be a challenge.

The nature of feature-lev el fusion requires the deﬁ-

nition and creation of a feature speciﬁc to a particu-

lar biometric characteristic and capture/extraction

system, as well as a matching algorithm for the

fused feature. Requiring vendors to suppor t fea-

ture-level fusion across many biometric modalities

may not be practical, given the current level of

industry consensus reachable in today’s marketplace.

On the contrary, decision-level fusion is rather sim-

ple mathematically, so a fusion standard might not

seem to be required for this level of fusion. Yet the

initialization, security speciﬁcation, and using multi-

ple biometric decisions make it an inherently compli-

cated process, and there is signiﬁcant beneﬁt to be

gained by including decision-level fusion in stan-

dards. Nonetheless, the ﬁrst beneﬁciaries of the fusion

standardization activity are most likely score-level

fusion systems. For this reason, the ISO/IEC Techni-

cal Report provides a wealth of descriptions of score-

level fusion, and the current US multibiometrics

standardization work focuses mainly on supporting

score-level fusion.

Acknowledgment

The terms and deﬁnitions taken from ISO/IEC TR

24722:2007 Information tech nology – Biometrics –

Multimodal and other multibiometric fusion, sections

4.1, 5.1, 5.3.1, 5.3.2, 5.3.3, 5.4, 7.2, 7.3 and the introduc-

tion, are reproduced with permission of the Internation-

al Organization for Standardization, ISO. This standard

can be obtained from any ISO member and from the

Web site of ISO Central Secretariat at the following

address: www.iso.org. Copyright remains with ISO.

Related Entries

▶ Fusion, Decision-Level

▶ Fusion, Feature-Level

▶ Fusion, Rank-Level

▶ Fusion, Sensor-Level

▶ Multi-Algorithm Systems

▶ MultiBiometrics

▶ Multi-Instance Systems

▶ Multi-Modal Systems

Multibiometrics and Data Fusion, Standardization

979