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

Подождите немного. Документ загружается.

False Match Rate, such a s 0.001, which was the basis

for NIST ICE (Iris Challenge Evaluation 2006) report-

ing. The ICE iris database contained many very difﬁ-

cult and corrupted images, often in poor focus, and

with much eyelid occlusion, with motion blur, raster

shear, and sometimes with the iris par tly outside

of the image frame. As ROC curves require False

Matches, NIST used a much more liberal decision

criterion than is used in any actual deployments of

iris recognition. As seen in Figure 4, using liberal

thresholds that generate False Match Rates (FMR)

in the range of 0.001–0.00001, score normalization

adversely impacts on the ROC curve by increasing

the False nonMatch Rate (FnMR) . The Equal Error

Rate (where FnMR = FMR, indicated by the solid

squares) is about 0.001 without score normalization,

but 0.002 with the normalization. Similarly at other

nominal points of interest in this region of the ROC

curve, as tabulated within Figure 4, the cost of score

normalization is roughly a doubling in the FnMR,

because marginal valid matches are rejected due

to the penalty on fewer bits having been available

for compariso n. In conclusion, whereas Table 1, and

Figures 2 and 3 document the important beneﬁt of

score normalization when operating with very large

databases that require several orders of magnitude

higher conﬁdence against False Matches, Figure 4

shows that in scenarios which are much less demand-

ing for FMR, the FnMR is noticeably penalized by

score normalization, and so the ROC curve suffers.

Adapting Decision Thresholds to the Size

of a Search Database

Using the SQRT normalization rule, Figure 5 presents

a histogram of all 200 billion cross-compariso n simi-

larity scores HD

norm

among the 632,500 different irises

in the Middle Eastern database [3]. The vast majority

of these IrisCodes from different eyes disagreed in

roughly 50% of their bits as expected, since the bits

are equiprobable and uncorrelated between different

eyes [2, 1]. Very few pairings of IrisCodes could dis-

agree in fewer than 35% or more than 65% of their

bits, as is ev ident from the distribution. The form of

this distribution needs to be understood, assumi ng

that it is typical and predictive of any other database,

in order to understand how to devise decision rules

that compensate for the scale of a search. Without this

form of score normalization by the scale of the search,

or an adaptive decision threshold rule, False Matches

would occur simply because large databases provide so

many more opportunities for them.

The solid curve that ﬁts the distribution data very

closely in Figure 5 is a binomial probability density

function. This theoretical form was chose n because

comparisons between bits from different IrisCodes

are Bernoulli trials, or conceptually ‘‘coin tosses,’’ and

Bernoulli trials generate binomial distributions. If one

tossed a coin whose probability of ‘‘heads’’ is p in a

series of n independent tosses and counted the number

m of ‘‘heads’’ outcomes, and if one tallied this fraction

x ¼ m ∕n in many such repeated runs of n tosses, then

the expected distribution of x would be as per the solid

curve in Figure 5:

f ðxÞ¼

m!ðn  mÞ!

ð1  pÞ

ðnmÞ

ð3Þ

The analogy between tossing coins and comparing

bits between different IrisCodes is deep but imperfect,

because any given IrisCode has internal correlations

arising from iris features, especially in the radial direc-

tion [2]. Further correlations are introduced because

the patterns are encoded using 2D Gabor wavelet ﬁl-

ters, whose lowpass aspect introduces correlations in

amplitude, and whose bandpass aspect introduces cor-

relations in phase, both of which linger to an extent

that is inversely propor tional to the ﬁlter bandwidth.

The effect of these correlations is to reduce the value of

the distribution parameter n to a number signiﬁcantly

smaller than the number of bits that are actually

Score Normalization Rules in Iris Recognition. Table 2

Effect of score normalisation on the match quality required

with various amounts of iris visibility

Number

of bits

compared

Approximate

percent of iris

visible (%)

Maximum acceptable

fraction of bits

disagreeing

200 17 0.13

300 26 0.19

400 35 0.23

500 43 0.26

600 52 0.28

700 61 0.30

800 69 0.31

911 79 0.32

1,000 87 0.33

1,152 100 0.34

1140

Score Normalization Rules in Iris Recognition

compared between two IrisCode s; n becomes the num-

ber of effectively independent bit comparisons. The

value of p is very close to 0.5 (empirically 0.499 for

this database), because the states of each bit are equi-

probable a prior i, and so any pair of bits from different

IrisCodes is equally likely to agree or disagree.

The binomial functional form that describes so

well the distribution of normalized similarity scores

for comparisons between different iris patterns is key

to the robustness of these algorith ms in large-scale

search applications. The tails of the binomial attenuate

extremely rapid, because of the dominating central

tendency caused by the factorial terms in (3). Rapidly

attenuating tails are critical for a biometric system to

survive the vast numbers of opportunities to make

False Matches without actually making any, when ap-

plied in an ‘‘all-against-all’’ mode of searching for any

matching or multiple identities, as is contemplated in

some national ID projects. The requirements of bio-

metric operation in ‘‘identiﬁcation’’ mode by exhaus-

tively searching a large database are vastly more

demanding than operating merely in one-to-one ‘‘ver-

iﬁcation’’ mode (in which an identity must ﬁrst be

explicitly asserted, which is then veriﬁed in a yes/no

decision by comparison against just the single nomi-

nated template).

If P

is the False Match probability for single one-

to-one veriﬁcation trials, then (1P

) is the probabili-

ty of not making a False Match in single comparisons.

The likelihood of successfully avoiding this in each of

N independent attempts is therefore (1P

)

, and

so P

, the probability of making at least one False

Match when searching a database containing N differ-

ent patterns is:

¼ 1 ð1  P

ð4Þ

Score Normalization Rules in Iris Recognition. Figure 4 Adverse impact of score normalisation in ROC regions where

high False Match Rates are tolerated (e.g., 0.00001 to 0.001 FMR). In these regions, the False nonMatch Rate is roughly

doubled as a result of score normalization.

Score Normalization Rules in Iris Recognition

1141

Observing the approximation that P

 NP

for

small P

<< 1, when searching a database of

size N an identiﬁer needs to be roughly N times better

than a veriﬁer to achieve comparable odds against

making False Matches. In effect, as the database

grows larger and larger, the probability of making

a False Match also grows almost in proportion.

Obviously the frequency of False Matches over time

also increases with the frequency of independent

searches that are conducted against the database. In

the Middle Eastern deployment [3] from which the

data of Figure 5 was taken, in which typically about

12,000 daily arriving passengers are each compared

with about a million stored IrisCodes, the total num-

ber of iris comparisons is about 12 billion per day. To

survive successfully so many opportunities to make

False Matches, the decision threshold policy must be

adaptive to both of these factors: the size of the database

and the frequency of searches throug h it (the quer y

rate). Fortunately, because of the underlying binomial

combinatorics, the algorithms with score normaliza-

tion generate extremely rapidly attenuatin g tails for the

norm

distribution. The consequence is that extreme-

ly small adjustments to tighten the decision threshold

yield order-of-magnitude increases in robustness

against False Matches, and therefore in large-scale

search capability. But before specifying those rules, it

is ﬁrst necesssary to understand the effect of the rota-

tion range over which repeated comparisons are done.

Factoring the Rotation Range into

Adaptive Decision Rules

Because cameras or heads may be tilted during iris

image acquisition, and indeed the eye itself can undergo

torsional (rotational) movements, it is necessary to

compare iris patterns at each of several relative orienta-

tions before deciding whether they match. Thus, when

searching a database of N enrollees and performing

each IrisCode comparison in each of k orientations,

the total number of comparisons effectively becomes

(k  N). Since the best match (smallest score) found in

each set of k comparisons is the score that is retained,

the new distribution for comparisons between differ-

ent eyes is biased towards lower scores, has a lower

mean, and has asymmetric tails compared with the

unrotated case seen in Figure 5. Using the same data-

base of IrisCodes but now performing all comparisons

Score Normalization Rules in Iris Recognition. Figure 5 Binomial distribution of normalised similarity scores in

200 billion comparisons between different eyes. Solid curve is (3).

1142

Score Normalization Rules in Iris Recognition

in each of k ¼ 7 orientations generates the new score

distribution seen in Figure 6.

The new distribution after k rotations of all Iris-

Codes in the search process still has a simple analytic

form that can be derived theoretically. Let f

(x) be the

raw density distribution obtained for the HD

norm

scores between different irises after comparing them

only in a single relative orientation; for example, f

(x)

might be the binomial deﬁned in (3). Then F

(x), the

cumulative of f

(x) from 0 to x, becomes the probabil-

ity of getting a False Match in such a test when using

norm

acceptance criterion x :

ðxÞ¼

ðxÞdx ð5Þ

or, equivalently,

ðxÞ¼

ðxÞð6Þ

Clearly, then, the probability of not making a False

Match when using decision criterion x is 1F

(x)

after a single test, and it is [1F

(x)]

after carrying

out k such tests independently at k different relative

orientations. It follows that the probability of a False

Match after a ‘‘best of k’’ test of agreement, when using

norm

criterion x, regardless of the actual form of the

raw unrotated distribu tion f

(x), is:

ðxÞ¼1  1  F

ðxÞ½

ð7Þ

and the expected density f

(x) associated with this

cumulative is:

ðxÞ¼

ðxÞ

¼ kf

ðxÞ 1  F

ðxÞ½

k1

ð8Þ

It is prudent to allow for at least a 20 deg orienta-

tion uncertainty, and so k ¼ 7 relative rotations in

2.81deg intervals were performed when arriving at

each best comparison HD

norm

score plotted in the

Figure 6 distribution. As expression (8) in the deriva-

tion shown earlier for postrotation probability density

Score Normalization Rules in Iris Recognition. Figure 6 Distribution of normalised similarity scores from the same set

of 200 billion iris comparison plotted in Figure 5, but now keeping only the best match after k ¼ 7 relative rotations to

compensate for unknown degrees of tilt in image acquisition. Solid curve is the f

(x) density function in (8), using the

binomial (3) as the f

(x) density prior to rotations.

Score Normalization Rules in Iris Recognition

1143

indicates, this means that the left tail of the distribu-

tion in Figure 6 is essentially k ¼ 7 times higher than

the one in Figure 5. The new functional form (8)

derived gives an excellent ﬁt to this new distribution,

as shown by the solid curve that closely matches the

Figure 6 data.

The cumulatives (up to various thresholds) under

the left tail of the distribution of normalized similarity

scores for different irises compared at k ¼ 7 relative tilts,

reveal the False Match Rates among the 200 billion iris

comparisons if the identiﬁcation decision policy used

each such threshold. These FMR rates are provided in

Table 3. Although the smallest observed match was

around 0.26, the Table has been extended down to 0.22

using the theoretical cumulative (7) of the extreme

value distribution (8) of multiple samples from the

binomial (3) plotted as the solid curve in Figure 6,

in order to extrapolate the theoretically expected

False Match Rates for such decision policies. These

False Match Rates, whether empirical or theoretical,

also serve as conﬁdence levels that can be associated

with a given quality of sample using the score normali-

zation rule (2). In this analysis, only a single eye

is presumed to be presented. Under the assumption

of independence between right and left eye IrisCodes,

which is strongly supported by the available data

(see Figure 6 of [2]), the conﬁdence levels in Table 3

could be multiplied together for matches obtained

with both eyes.

Now it is ﬁnally possible to state a general rule for

adaptively selecting a decision criterion threshold on

norm

normalized similarity scores, given the

empirical cumulatives shown in Table 3 under the dis-

tribution of postrotation scores plotted in Figure 6.

If one is performing iris identiﬁcations by exhaustive

search through an enrolled database of size N, using

k ¼ 7 relative rotations for every comparison and

normalizing raw Hamming Distance scores by the

amount of available data as per (2), then the recom-

mended strategy is this:

Calculate the total number of iris comparisons that

will be performed in a given period of time, i.e., the size N

of the enrolled database times the number of queries

against it during that time. Decide the risk tolerance for

False Matches during such a period of time, and ﬁnd the

corresponding entry in the second column of Table 3.

Then the ﬁrst column gives the recommended decision

threshold on HD

norm

scores.

Example: If every month 100,0 00 passengers are

each compared to 1 million enrolled IrisCodes by

exhaustive search, generating 10

comparisons per

month, and no more than one False Match can be

tolerated per month, then Table 3 indicate s that the

recommended single-eye decision criterion to use

would be around 0.265 for HD

norm

similarit y scores.

Related Entries

▶ Iris Encoding and Recognition using Gabor Wavelets

References

1. Daugman, J.G.: The importance of being random: Statistical

principles of iris recognition. Pattern Recognit. 36, 279–291

(2003)

2. Daugman, J.G.: How iris recognition works. IEEE Trans. Circuits

Syst. Video Technol. 14, 21–30 (2004)

Score Normalization Rules in Iris Recognition. Table 3

False match rates with HD

norm

score normalisation:

dependence on criterion (200 Billion cross comparisons)

HD Criterion Observed False Match Rate

0.220 0 (theor: 1 in 5  10

)

0.225 0 (theor: 1 in 1  10

)

0.230 0 (theor: 1 in 3  10

)

0.235 0 (theor: 1 in 9  10

)

0.240 0 (theor: 1 in 3  10

)

0.245 0 (theor: 1 in 8  10

)

0.250 0 (theor: 1 in 2  10

)

0.255 0 (theor: 1 in 7  10

)

0.262 1 in 200 billion

0.267 1 in 50 billion

0.272 1 in 13 billion

0.277 1 in 2.7 billion

0.282 1 in 284 million

0.287 1 in 96 million

0.292 1 in 40 million

0.297 1 in 18 million

0.302 1 in 8 million

0.307 1 in 4 million

0.312 1 in 2 million

0.317 1 in 1 million

1144

Score Normalization Rules in Iris Recognition

3. Daugman, J.G.: Probing the uniqueness and randomness of

IrisCodes: Results from 200 billion iris pair comparisons. Proc.

IEEE 94(11), 1927–1935 (2006)

4. Daugman, J.G.: New methods in iris recognition. IEEE Trans.

Syst. Man Cybern. 37 , 1167–1175 (2007)

Sealed Local Biometric Identity

Verification Systems

A sealed local biometric identity veriﬁcation system is

a strong example of a trusted biometric identity veriﬁ-

cation client. Sealed biometric systems are built around

the concept that the entire biometric acquisition and

matching process is ‘‘sealed’’ within a demonstrably

secure environment (see the following ﬁgure). No bio-

metric data can ever enter the system except via the

integrated biometric sensor, and no biometric data

ever leaves the system, period. In fact there is no way

to externally access the biometric data at all. The sealed

system securely stores cryptographic credentials for

each entity that registers its need to verify the user’s

identity, and releases those credentials only when that

entity has cryptographically identiﬁed itself, and the

correct user has been biometrically veriﬁed to be

present.

Figure 1 illustrates the concept of a sealed local

biometric identi ty veriﬁcation system.

▶ Biometric Sample Acquisition

Second Level Detail

This refers to the occurrence of fundamental events

within the general ridge ﬂow which disturb the regu-

lar and parallel ﬂow of ridges. These events can

manifest themselves as ending/starting ridges, diverg-

ing/merging ridges ( bifurcations), and dots and com-

binations of these events such as short ridges, eyes,

spur, and islands. Thes e events are also referred to

as Galton characteristics, points or major ridge path

deviations.

Minutiae reﬂect events in a persistent system of

papillary ridges that once developed to their ﬁnal

form during gestation remain unchanged throughout

life.

Minutiae in the latent keep their basic properties

such as relative location, direction, and relations to

other points even under adverse condi tions (Fig . 1).

This differentiates ﬁnge rprints from other types of

forensic evidence. Minutiae and their formation in

Sealed Local Biometric Identity Veriﬁcation Systems. Figure 1

Second Level Detail

1145

sequence are the backbone of ﬁngerprint identiﬁca-

tion. Additional features that are also referred to as

level two are incipient ridges, scars, and creases.

▶ Fingerprint Matching, Manual

Secure Biometric Token Operating

System

▶ Tamper-proof Operati ng System

Secure Biometrics

▶ Biometric and User Data, Binding of

Secure Element

Synonyms

Security processor; Secure token

Definition

A closed, tamper-resistant, well-trusted, and usually

certiﬁed embedded system with a very lightweight

microcontroller and some FLASH storage. Typical se-

cure elements are single-chip entities designed to

achieve a black box nature for maximum security.

Smart cards, SIM cards, and NFC controllers are the

most prevalent examples . Most secure elements inter-

act with a host using a wired or wireless communica-

tion channel, but may be standalone systems as well.

▶ Transportable Asset Protection

Secure Sketch

▶ Encryption, Biometric

Security and Liveness, Overview

ANDY ADLER

,STEPHANIE SCHUCKERS

Carleton University, Ottawa, ON, Canada

Clarkson University, Potsdam, NY, USA

Definition

The security of a biometric system may be understood

to be its resistance to active attacks. Such attacks

may be classiﬁed as Presentation attacks (spooﬁng),

in which the appearance of the biometric sample is

physically changed or replaced; Biometric processing

attacks, in which an understanding of the biometric

algorithm is used to cause incorrect processing and

decisions; Software and networking vulnerabilities

based on attacks against the computer and networks

on which the biometric sy stems run; and Social and

presentation attacks, in which the authorities using the

systems are fooled. This article presents an overview of

the techniques used for classifying and assessing these

threats. Additionally, newer biometric schemes, such as

cancelable biometrics and bio metric encryption, that

are designed to counter these security threats are

reviewed.

Introduction

Security must be deﬁned in the context of an attacker.

However, biometric systems, even when not under

active attack, should always be assumed to operate in

an (at least somewhat) hostile environment – after all,

why should one test identity if all can be trusted? The

ability of a biometric system to stand up to ‘‘zero-

effort’’ attackers is measured by the false accept rate

(FAR). Such measures are not typically considered to

measure biometric security, but rather biometric per-

formance. This article gives a broad overview of the

1146

Secure Biometric Token Operating System

security and liveness issues in biometric systems where

biometric security is understood to be the resistance of

a system to attackers prepared to take active measures

to circumvent the system. This article considers four

broad types of active attack:



Presentation attacks (spooﬁng): The appearance of

the biometric sample is changed either physically

or by replacement with a fabricated sample. For

physical biometrics, attackers may then change

makeup, facial hair and glasses, or abrade and cut

ﬁngerprints in order to avoid being recognized; to

attempt to be recognized as another person, a

spoofed ﬁngerprint or false iris contact lens may

be constructed and placed over the corresponding

body part.



Biometric processing attacks: A detailed understand-

ing of the biometric algorithm is used to cause

incorrect processing and decisions. The possible

attacks depend on the details of the biometric

algorithm (

▶ Biometric vulnerabilities, Overview).

Some examples are enrolling specially crafted

noisy images that artiﬁcially lower thresholds; re-

generation of sample images from stored templates;

and

▶ side-channel attacks based on ‘‘leaked’’ sys-

tem information such as from match scores or

timing of processing.



Software and networking vulnerabilities: Since bio-

metric systems are implemented on server compu-

ters, they are vulnerable to all cryptographic, virus,

and other attacks, which plague modern computer

systems [1]. Examples are databases security, denial

of service (DoS) attacks, and overriding the bio-

metric decision with compromised software. These

issues are not covered in detail in this article, since

they are not unique to biometric system security.



Social and presentation attacks: Security systems de-

pend on a chain of trust. Links in this chain between

systems are especially vulnerable [2]. Presentation

attacks involve the use of fraudulent identity docu-

ments, which may be legitimately issued; social

attacks focus on convincing an operator to override

or allow fraudulent exceptions. This article points

out the importance of these issues like software

vulnerabilities, but does not cover them in detail.

This article gives an overview of the security issues

in biometric systems, including classiﬁcations, security

performance measures, liveness and antispooﬁng and

novel biometric protection schemes.

Biometric Security Classifications

Several authors have developed classiﬁcation schemes,

which provide a taxonomy of biometric security chal-

lenges. Maltoni et al. [3], classify biometric system

vulnerabilities as follows:



Circumvention is an attack by which one gains

access to the protected resources by a technical

measure to subvert the biometric system. Such

an attack may subvert the underlying compu-

ter systems (overriding matcher decisions, or repla-

cing database templates) or may involve a replay

of valid data.



Covert acquisition (contamination) is the use of

biometric infor mation captured from legitimate

users to access a system. Examples are spooﬁng

via capture and playback of voice passwords and

lifting latent ﬁngerprints to construct a mold. This

category can also be considered to cover regener-

ated biometric images (

▶ Template Security). For

example, a ﬁngerprint image can be regenerated

from the template stored in a database (and these

data can be captured covertly [4]). Covert acquisi-

tion is worrisome for cross-application usage (eg.,

biometric records from a ticket for an amusement

park used to access bank accounts).



Collusion and Coercion are biometric system vul-

nerabilities from legitimate system users. The dis-

tinction is that, in collusion, the legitimate user is

willing (perhaps by bribe), while the coerced user is

not (through a physical threat or blackmail). Such

vulnerabilities bypass the computer security sys-

tem, since the biometric features are legitimate. It

may be possible to mitigate such threats by auto-

matically detecting the unusual pattern of activity.

Such attacks can be mounted from both adminis-

trator and user accounts on such a system; attacks

from user accounts would ﬁrst need to perform a

privilege escalation attack [1].



Denial of Service (DoS) is an attack that prevents

legitimate use of the biometric system. This can

take the form of slowing or stopping the system

(via an overload of network requests) or by degrad-

ing performance. An example of the latter would be

enrolling many noisy samples that can make a

system automatically decrease its decision thresh-

old and thus increase the FAR. The goal of DoS is

often to force a fall back to another system (such as

Security and Liveness, Overview

1147

operator override) that can be more easily circum-

vented, but DoS may be used for extortion or

political reasons.



Repudiation is the case where the attacker denies

accessing the system. A corrupt user may deny her

actions by claiming that her biometric data were

‘‘stolen’’ (by covert acquisition or circumvention)

or that an illegitimate user was able to perform the

actions due to the biometric false accept. Interesting-

ly, biometric systems are often presented as a solu-

tion to the repudiation problem in the computer

security literature [1]. One approach to help prevent

repudiation would be to store presented images for

later forensic analysis. However, this need must be

balanced against user privacy concerns [ 5].

Another class of biometric vulnerabilities are those

faced by the system user, developed by Ratha et al. [6].

These issues impact on the user’s privacy and can lead

to identity theft or system compromise.



Biometrics is not secret: Technology is readily avail-

able to capture images of faces, ﬁngerprints, irises

and make recordings of voice or signature – with-

out subject consent or knowledge [2, 7]. From this

perspective, biometrics is not secret. On the other

hand, from a cry ptography or privacy [5] perspec-

tive, biometric data are often considered to be

private and secret. This distinction is important,

as our understanding of computer and network

security is centered around the use of secret codes

and tokens [1]. For this reason, cry ptographic pro-

tocols that are not robust against disclosure of

biometric samples are ﬂawed.



Biometrics cannot be revoked: A biometric feature

is permanently associated with an individual, and

a compromised biometric sample will compro-

mise all applications that use that biometric

feature. Such compromise may prevent a user

from re-enrolling [2]. Note, however, that this con-

cern implies that biometrics is secret, as opposed

to the previous consideration. On e proposed solu-

tion is Cancelable biometrics, although the vulnera-

bility of such systems is not well understood.



Biometric features have secondary uses: If an indi-

vidual uses the same biometric feature in multiple

applications, then the user can be tracked if the

organizations share data. Another aspect of this

problem is secondary use of ID cards. For example,

a driver’s license is designed with the requirements

to prove identity and driver certiﬁcation to a police

ofﬁcer, but it is used to prove age, name and even

citizenship. Similarly, biometric applications will

be designed with a narrow range of security con-

cerns, but may be used in very different threat

environments.

Biometric systems form part of larger security

systems and their risks and vulnerabilities must be

understood in the context of the l arger system require-

ments. An excellent review of the security of biometric

authentication systems is [7]. Each assurance level

from ‘‘passwords and PINs’’ to ‘‘Hard cr ypto token’’

is analyzed to determine which biometric devices are

suitable. Since biometric systems are complex and

represent many interconnected subsystems, there are

many potential points for attack. Vulnerabilities in

Biometric Systems are considered in the article

▶ Bio-

metric Vulnerabilities: Overview.

Liveness and Spoofing

Clearly, biometric systems are vulnerable to artiﬁcial

changes to the biometric features. Such changes can be

of two types: to avoid detection as an enrolled user or

watch list candidate and to masquerade as another

legitimate user. The former is easier and can sometimes

be as simple as using glasses, makeup, or abrasions and

cuts to the ﬁnger. Masquerading or spooﬁng attempts

to gain unauthorized access at the biometric sensor

with artiﬁcial biometric features of authorized users,

called ‘‘spoofs.’’ This is widely publicized for ﬁnger-

print where it is possible to spoof a variety of ﬁn-

gerprint sensors through relatively simple techniques

using casts of a ﬁnger with molds made of materials,

including silicon, Play-Doh, clay, and gelatin (gummy

ﬁnger). Such spoof molds can be scanned and veriﬁed

when compared with a live enrolled ﬁnger [8–11].

Masquerade is also possible in the scenario of dismem-

bered ﬁngers; cadaver ﬁngers can be scanned and ver-

iﬁed against enrolled ﬁngers [9]. It is also possible to

spoof other common biometric modalities: for iris and

face, using pictures or high resolution video, for iris

with contact lenses, with voice with recordings [8, 11].

There are several approaches to increase the difﬁ-

culty of spooﬁng: multiple biometric features, liveness,

and the use of biometrics in combination with a

1148

Security and Liveness, Overview

challenge response, passwords, tokens, smart cards.

The goal of liveness testing is to determine if the

biometric feature being captured is an actual measure-

ment from the authorized, live person, who is present

at the time of capture. Typically, liveness is a secondary

measure after biometric authentication, which must be

met in order to achieve a positive response. Liveness

and antispooﬁng methods are covered in detail in the

following summaries [12–14]. The need for protec-

tions from spooﬁng may be assessed on an application

basis, although there is a need to address the spooﬁng

vulnerability throughout the industry, as the reputa-

tion of biometric systems as a securit y measure must

be considered. Liveness detection adds an additional

layer of security that also can increase the users trust in

biometric technology.

Characteristics of Liveness Approaches: The follow-

ing characteristics for evaluating biometric systems

need to be considered in implementing a liveness

algorithm.



Ease of use: liveness approaches vary in the ease

of use. For example, a ﬁngerprint deformation

approach, which requires a speciﬁc rotation proce-

dure, may be considered more difﬁcult to use [15].

A ﬁngerprint sensor using spectroscopy where

liveness inherent to the biometric feature may be

considered easier to use.



Collectability: liveness approaches vary in the ease

of collection. For example, electrocardiogram,

which requires two points of contact on opposite

sides of the body or pulse oximetry where the

ﬁnger must be enclosed to protect from ambient

light [16].



User acceptance: liveness approaches that may have

low user acceptance are teh ones that are more

likely to be linked with medical conditions (eg.,

electrocardiogram, DNA).



Universality: clearly, all authorized users must be

live when enrolling. However, the liveness method

may be difﬁcult to measure in some subjects.

For example, perspiration in ﬁngerprint images

may be difﬁcult to measure in subjects with very

dry skin.



Uniqueness: For liveness approaches that are inher-

ent to the biometric feature, this is essential.

However, it is not clear that, for example, electro-

cardiogram or gait is unique to large data sets

of individuals. Thus, these biometric/liveness

approaches may be appropriate for applications

with a smaller number of individuals.



Permanence is important to liveness approaches

taht are inherent to the biometric feature and

where the biometric/liveness features may vary

over time. This will impact on the biometric and

liveness error rates.



Spoof-ability describes whether the liveness mecha-

nism designed to protect against spooﬁng can be

spoofed. For example, it may be possible to fool

pulse oximetry-based liveness, using a clear spoof

that allows transmission of the light need ed to

make the pulse oximetry measurement.

The terms liveness and antispooﬁng are not com-

pletely synonymous. Measurements that rule out spe-

ciﬁc spoofs do not absolutely measure liveness. For

example, a liveness measure to detect pupil movement

will detect attempts based on a simple photograph

of a face. However, a modiﬁed spooﬁng method,

such as cutting a hole in the picture and putting a

real pupil behind it, may result in a successful spoof

attempt. Such a spoof is partially alive (to fool the

liveness) and partially a spoof (fabricated user biomet-

ric feature).

Encoded Biometric Schemes

Classical biometric systems require access to enrolled

templates in unencoded form. This differs from tradi-

tional computer security systems where a raw pass-

word need never be stored. Instead, a cryptographic

hash (one-way function) of the password is stored, and

each new test password is hashed and compared with

the stored version. Since such cryptographic techni-

ques provide important protections, there is great in-

centive to develop analog ous methods for biometric

systems. Encoded biometric schemes are designed to

avoid these problems by embedding the secret code

into the template, in a way that can be dec rypted only

with an image of the enrolled individual [17, 18].

Since the code is bound to the biometric template, an

attacker should not be able to determine either the

enrolled biometric image or secret code, even if he

had access to the biometric software and hardware.

Such technology would enable enhanced privacy

Security and Liveness, Overview

1149