Towards an Optimal Template Reduction

for Securing Embedded Fingerprint Devices

Beno

ıt Vibert, Christophe Charrier, Jean-Marie Le Bars and Christophe Rosenberger

Normandie Univ., UNICAEN, ENSICAEN, CNRS, GREYC, 14000 Caen, France

Keywords:

Fingerprint Template, Template Reduction, Genetic Algorithm.

Abstract:

Template protection is an important issue in biometrics for security and privacy reasons. One solution for

securing ﬁngerprint data is to store it on a Secure Element (a microcircuit chipset such as a smartcard). An

embedded On-Card-Comparison (OCC) module permits to compare two biometric templates and generates a

similarity score. The biometric template is usually composed of minutiae extracted from the ﬁngerprint image

because a Secure element is limited in terms of memory and computation capabilities. For these reasons,

a template reduction is necessary to quickly process ﬁngerprint comparison. In this paper, we propose a

new ﬁngerprint template reduction scheme by approximating the optimal choice of minutiae with a genetic

algorithm. We compared the proposed method with approaches from the literature using a ﬁngerprint dataset

and three matching algorithms. The experimental results show the beneﬁt of the proposed method especially in

order to estimate the optimal performance when reducing the ﬁngerprint template given a number of minutiae

to use.

1 INTRODUCTION

Nowadays, electronic transactions are part of our

daily life (e-commerce, smartphones, physical access

control . . . ). In order to guarantee the security of user

authentication, biometrics is often used. Many real

applications beneﬁt from this technology such as for

user access control or e-payment. According to (IHS,

2016), in 2020, the market of smartphones with a ﬁn-

gerprint sensor will reach 1.6 billions units. Never-

theless, a biometric data is very sensitive and can-

not be revoked in general (like a password). In or-

der to ensure its security and privacy, a biometric data

is usually stored in a Secure Element (SE). The Se-

cure Element could be a SmartCard, with an embed-

ded On-Card-Comparison (OCC) algorithm for com-

paring two biometric templates.

Figure 1: Enrollment and Veriﬁcation step.

Two steps are necessary when using a biomet-

ric system: 1) the enrollment and 2) the veriﬁcation

as described in Figure 1. The OCC algorithm com-

putes a comparison score between a captured biomet-

ric template and the reference one. The common ﬁn-

gerprint template is composed of a set of minutiae

corresponding to speciﬁc points as described in Fig-

ure 2. The number of minutiae varies considering the

used sensor but it is lower 80 in general.

Figure 2: Fingerprint minutiae.

As common practice, the biometric template

stored in the SE follows the ISO Compact Card stan-

dard (ISO, b) to ensure the interoperability between

biometric sensors and systems. This template is com-

posed of a set of minutiae represented by 3 octets

and 4 values (x

, y

, T

, θ

), i = 1 : N

where the coor-

dinates (x

, y

) correspond to the location of the minu-

Vibert, B., Charrier, C., bars, J-M. and Rosenberger, C.

Towards an Optimal Template Reduction for Securing Embedded Fingerprint Devices.

DOI: 10.5220/0006608903290336

In Proceedings of the 4th International Conference on Information Systems Security and Privacy (ICISSP 2018), pages 329-336

ISBN: 978-989-758-282-0

329

tiae in the image, T

corresponds to the minutiae type

(bifurcation, ridge ending . . . ), θ

to the minutiae ori-

entation (related to the ridge) and N

the number of

minutiae for the sample j of the user.

A SE has hardware and software constraints such

as the size of memory, the number of data we can send

with an APDU command (ISO, c) (ISO 7816 stan-

dard for the communication with a SE). These lim-

itations have an impact on the embedded algorithm

and the size of the ﬁngerprint template. The ISO/IEC

19794-2 standard recommends the maximal number

of minutiae for enrollment and veriﬁcation of the ISO-

CC template is 60 (ISO, a). However, in an opera-

tional OCC application, a ﬁngerprint template is usu-

ally limited to a speciﬁc number of minutiae which is

lower or equal to 50 to satisfy the memory space, the

APDU speciﬁcations and also the veriﬁcation time (in

general lower than 500 ms). In this case, it is neces-

sary to reduce the template size when the extractor

detected more minutiae.

Some automatic methods have been proposed in

the literature such as the INCITS (Grother and Sala-

mon, 2007) standard (called ”Barycenter” on this

study) which keeps only the minutiae closest to the

CORE point. However, existing template standards

are diverse, and mostly provide minutiae type instead

of the quality of minutia point to assist the matching

algorithm. Therefore, techniques for reducing the size

of minutiae template without the quality information

of a minutia point should be considered. Few works

in the existing studies paid attention to this issue. The

ISO organization (ISO, b) proposed a method based

on peeling off minutiae (we call it ”Truncation” in

this study). Three other methods have been proposed

by Vibert et al. (Vibert et al., 2015). The ”evolu-

tive Barycenter” is based on the method proposed by

the NIST. A loop is used to re-compute the centroid

when one minutiae is peeled off, until the number of

minutiae expected is reached. The ”Truncation Ran-

dom Permutation” is based on the ISO organization

method. With this method, the template of minutiae

is shufﬂed, only the number of minutiae expected is

kept on the ﬁnal reduced template. ”K-Means” is used

as another approach where only minutiae closest than

each cluster is kept on the ﬁnal reduced template.

Yet, the main drawback of all of the above men-

tioned methods is that there is no guarantee to reach

the optimal reduced template. For us, an optimal re-

duced template maximizes the similarity score with

the original template for a selected OCC matching al-

gorithm. In other words, all the obtained templates

approximated more or less the optimal template with-

out actually reaching it. The objective of this paper

is to approximate as close as possible the optimal re-

duction of a minutiae template. This approximation

provides a landmark to determine whether it is worth

to look for better practical reduction methods. To our

knowledge, this is an original contribution of this pa-

per.

The proposed approach is presented in section 2.

Section 3 provides the experimental protocol. Evalua-

tion results are discussed in Section 4. Section 5 con-

cludes this study and gives the associated perspective.

2 PROPOSED METHOD

The general framework of this study is a two-step

work: 1) minutiae acquisition and 2) performance

evaluation. The acquisition involves two tasks which

are illustrated in Figure 3. We ﬁrst use an extractor to

generate full-size minutiae template and then perform

selection operations considering the desired minutiae

number to obtain the reduced template. The quality of

the reduction is linked to security and usability since

it has an impact on performance, especially on FAR

(False Acceptance Rate) and FRR (False Rejection

Rate).

In this paper, we propose a Minutiae Reduction

with Genetic Algorithms scheme (namely MRGA) to

estimate the optimal reduction of any minutiae tem-

plate. Given a template containing N minutiae, we

want to determine the optimal reduced template con-

taining N

max

minutiae (under the constraint N

max

N), i.e., providing the best performance. To be sure

to determine this optimal template, we should test



max



possibilities (number of combinations of N

max

elements among N) that is not possible. To achieve

this goal, the proposed method is based on the use of

genetic algorithm (GA).

A genetic algorithm is a method for stochastic

search introduced in the 70s by John Holland (Hol-

land, 1975) and by Ingo Rechenberg (Rechenberg,

). Genetic algorithms allow to determine the optimal

value of a criterion by simulating the evolution of a

population until the survival of the best individuals

(Wall, 1996). The survivors are obtained by selection,

transformation or crossing of the previous generation.

We estimate that the optimal search function is a non-

linear multidimensional function, usually character-

ize by several minima. Therefore, the search strategy

should ﬁnd the global minimum, and avoid remaining

trapped in local minima. The objective is to obtain

a reduced minutiae template having the best perfor-

mance compared to the original template applying an

OCC algorithm. A genetic algorithm is deﬁned by

ﬁve essential elements:

1. Genotype: This is a set of characteristics rep-

ICISSP 2018 - 4th International Conference on Information Systems Security and Privacy

330

Figure 3: Two tasks involved in template reduction.

resenting each individual in the population. In

our case, the initial population consists of 500

individuals composed of N elements, N being the

desired number of minutiae within the reduced

biometric template. As we want to get a template

with minutiae existing in the initial template, the

population will be constituted by random draws

of N minutiae in the initial template containing M

minutiae.

2. Initial Population: This is a set of individuals

randomly drawn from the original template. Each

individual consists of N elements. Each element

corresponds to a unique minutiae present in the

original template.

3. Evaluation Function: It measures the quality

of an individual. If we consider individual I

evaluate, we compare it with the original template

using an OCC and we get an S(I

) score. The

evaluation function is based on the OCC compu-

tation, since it is fast to compare two biometric

templates and it has good performances. This

algorithm returns us a similarity score between

the two templates. The higher is the similarity

score is, better is the tested individual i.e., the

reduced template.

4. Operations on Genotypes: the genes of the indi-

viduals are modiﬁed by the use of three function-

alities:

• Selection: Individuals that do not match

the environment (insufﬁcient score) are not

selected. To do this, we apply the elite mode

(the 5 individuals with the highest score are

kept in the next generation).

• Crossing: the genes resulting from the crossing

of two individuals is a combination of the

genes of its parents. To obtain the individual

resulting from individuals I

and I

, we look

at the elements present in the two individuals

without the duplicated ones and randomly

select the ﬁrst N elements. We thus obtain an

individual (son) mixing the genes of the two

individuals (parents). As a result, the resulting

minutia template contains minutia from both

parents.

• Mutation: Random genes are modiﬁed in order

to adapt to the environment. We randomly

draw an individual, then we cross this indi-

vidual with an elite individual. The resulting

individual I

= mutation (I

) = cross(I

, I

) with

the random individual. It makes it possible to

obtain an individual having genes from an elite

individual crossed with genes of a random one.

5. Termination: This is the end-of-evolution crite-

rion depending on the score of individuals or the

number of generations. If an individual keeps the

same score for 10 generations or 500 generations

have been made, the algorithm ends.

We summarize here the work-ﬂow of the execu-

tion of a genetic algorithm:

1. Deﬁnition of initial population of 500 individuals,

2. Evaluation of Individuals,

3. Generation of the population at current genera-

tion:

• Selection of 5 elite individuals;

• 30 % of the population (here 150 individuals)

is obtained by mutating elite individuals with

random ones;

Towards an Optimal Template Reduction for Securing Embedded Fingerprint Devices

331

• 30 % of the population (here 150 individuals)

is obtained by crossing elite individuals;

• Selection of random individuals to complete the

population.

4. Return to step 2 if the stopping criterion is not sat-

isﬁed.

The next section presents the results obtained with

MRGA method in comparison with some state-of-

the-art methods.

3 EXPERIMENTAL PROTOCOL

To evaluate the performance of the MRGA method,

we need to make some choices about biometric

database, minutiae extractor, comparison algorithms

and performance metrics. We detail these aspects in

the following sub-sections.

3.1 Database

In this study, the FVC2004DB1 ﬁngerprint database

from the Fingerprint Veriﬁcation Competition (FVC)

(www, ) is used. This database is composed of 800

images from 100 individuals with 8 samples from

each user. The image resolution is 640 × 480 pix-

els acquired with an optical sensor ”V300” by Cross-

Match. Figure 4 shows some examples of ﬁngerprints

in this database.

3.2 Minutiae Extractor

The minutiae templates used in the experiment have

been extracted using the NBIS tool, and more speciﬁ-

cally MINDTCT (Watson et al., 2007) from the NIST.

We used this extractor because it is widely used in

academic research.

3.3 Matching Algorithms

In this study, we used three matching algorithms:

1. Bozorth3 (Watson et al., 2007). This comparison

algorithm uses only the locations and orientation

of the minutiae to match the ﬁngerprints. We get

a similarity score as output of the algorithm.

2. Minutia Cylinder-Code (MCC) Algorithm

(Cappelli et al., 2010). The representation of

MCC associates a local structure with each minu-

tia. This structure contains the spatial and di-

rectional relationships between minutia and its

neighborhood (ﬁxed radius). Each structure is

Figure 4: Example of ﬁngerprint images in the

FVC2004DB1 database.

invariant in translation, rotation, distortions and

small errors of extraction of characteristics. A

double measure of similarity is computed and

consolidated to provide an overall score for com-

parison.

3. Commercial OCC. We do not have information

on how this algorithm works since it is a commer-

cial one. As output of the algorithm, we do not

have a score but simply a decision result of type

”Accepted” or ”Declined”.

3.4 Evaluation Metrics

In order to assess the performance of a biometric sys-

tem, we can use the Receiver Operating characteristic

Curve (ROC). This curve plots the False Match Rate

(FMR) (i.e., accepted impostor attempts) on the x-

axis against the corresponding False Non-Match Rate

(FNMR) (i.e., rejected genuine attempts) on the y-

axis. This curve is parametrically plotted as a func-

tion of the decision threshold. An example of a ROC

ICISSP 2018 - 4th International Conference on Information Systems Security and Privacy

332

curve is presented in Figure 5. The area under the

curve (hatched zone) should be as low as possible to

minimize recognition errors. The associated measure

is called AUC (Area Under the Curve) and is often

considered as a global performance criterion. We use

this value in this paper to quantify the efﬁciency of all

trial minutiae selection methods.

Figure 5: Deﬁnition of the ROC curve: evolution of the

False Match Rate face to the False Non Match Rate.

The ﬁrst sample from each individual is chosen

as reference template while other seven samples are

used for tests. With a matcher, we generate a group

of intra-class matching scores (GMS) and a group of

inter-class matching scores (IMS) for each dataset.

The evaluation result is indicated by the global AUC

value computed from the two groups of matching

scores. This computation is done for every group of

reduced templates. A curve of AUC values obtained

for different sizes of the reduced template is plotted

to compare one reduction algorithm with others. For

each of the AUC value, we calculated the associated

conﬁdence interval (CI), which allows us to have an

additional precision for our results. Since we use a

small number of data for the evaluation, the conﬁ-

dence interval gives us additional information on the

accuracy of the results. Another important criterion

in our study is the computation time needed for the

reduction of a biometric template.

4 EXPERIMENTAL RESULTS

First of all, we have to build the baseline performance

by using the original template.

4.1 Performance Evaluation of MRGA

Table 1 shows the AUC value for the original template

that will serve as baseline performance to compare

the performance of other methods with. We notice a

much better performance of the commercial OCC. We

then compute the AUC value and its associated con-

ﬁdence interval for each reduction method with N

max

varying from 30 to 50 in steps of 4. We observe that

the NIST comparison algorithm as well as the MCC

scheme have much worse performance than the com-

mercial OCC. This is not so many surprising since

commercial algorithms are supposed to have high per-

formance level.

Table 1: AUC values for FVC2004DB1 database with Bo-

zorth, MCC and the commercial OCC.

Bozorth MCC commercial OCC

11.1% ± .18 18.4% ± .17 3.77% ± .09

Figure 6 presents the evolution of the ﬁtness score

until the stopping criterion is reached. After 70 gener-

ations, the reduction of the template scheme provides

a solution close to the optimal reduced template. The

blue line represents the average ﬁtness score obtained

during the generations. The black line is the best ﬁt-

ness score obtained for each generation. We can ob-

serve that the optimal reduced template is obtained

after few generations.

Figure 6: Evolution of the Fitness score.

Obtained performances using the trial matching

algorithms are summarized in Table 2. We display the

obtained results using the MCC algorithm too, even if

it has been used in the evaluation function of the ge-

netic algorithm. This allows us to observe if we have

a performance close to the initial template. We ob-

serve that the proposed MRGA method provides per-

formance almost similar to the initial template, as it

could expected. We can conclude that the reduced

templates obtained with the MRGA method obtain a

performance very close to the original templates for

all the matching algorithms.

Towards an Optimal Template Reduction for Securing Embedded Fingerprint Devices

333

Table 2: Difference between AUC values for the initial template and MRGA method for different values N

max

of minutiae on

FVC2004DB1 for all OCC algorithm.

OCC Initial Template 30 34 38 42 46 50

Bozorth3 11.1% ±.18 +2.9%±.46 +2.71%±.32 +2.1%±.28 +0.9%±.28 +0.2%±.26 +0.1%±.24

MCC 18.4% ±.17 +0.8%±.17 +0.6%±.16 +0.4%±.15 +0.3%±.12 0%±.12 +0.1%±.09

Commercial 3.77% ±.09 +1.73%±.25 +1.13%±.24 +0.55%±.21 +0.03%±.20 0%±.18 0%±.14

(a) Bozorth

(b) MCC

Figure 7: Comparison of the MRGA method with the best reduction methods applied on FVC2004DB1 database for the three

ﬁngerprint comparison algorithms.

Table 3 presents the average time needed to re-

duce a biometric template when the MRGA method

is used. This reduction has been made on Matlab run-

ning on a computer (PC) with a Intel Core I7, quad-

core with a frequency at 2.8Ghz and 16GO of RAM.

These times have to be considered for a relative com-

parison. These computation times could be easily re-

duced if the computation is made on server in C++,

that could be least 10× more efﬁcient.

Table 3: Average time to perform the reduction for different

values N

max

of minutiae when MRGA method is used.

max

30 34 38 42 46 50

Time 38 min 35 min 28 min 20 min 18 min 13 min

ICISSP 2018 - 4th International Conference on Information Systems Security and Privacy

334

Considering computation times, this method can-

not be used in an operational framework but only for

validation purposes. Nevertheless, we could imagine

to use it if the computations are done on a server. Dur-

ing enrollment, the reduced biometric could be com-

puted online and sent in a secure way to the secure

element.

4.2 Comparison with State-of-the-art

Methods

We want to evaluate in which way some state-of-

the-art methods, Truncation, Barycenter, Evolutive

Barycenter, Truncation Random Permutation and K-

Means, are close to the ”optimal” reduced template.

Figure 7(a) for the Bozorth3 algorithm, ﬁgure 7(b)

for the MCC algorithm and ﬁgure 7(c) for the com-

mercial OCC show the comparison results.

We could note that the trial methods are more or

less close to the MRGA method especially for large

number of minutiae. This allows us to better un-

derstand whether a change in methods is possible to

achieve the best possible reduction. We can observe

from Figure 7(b) with MCC matching algorithm, that

K-means method provides the best reduced template

when the number of minutiae ranges from 38 to 50.

For Bozorth3 matching algorithm (Figure7(a)), K-

means is the best method but, we could improve it

to reduce the gap with the MRGA method. Consid-

ering the commercial OCC, the three state-of-the-art

methods are very close to the performance of the ini-

tial template showing the beneﬁt of MRGA.

Figure 8 shows the initial template, in blue on each

ﬁgure, and the associated selected minutiae (red dot)

when we applies: Truncation, Barycenter, Troncation

Random Permutation, Evolutive barycenter, K-Means

and the MRGA method. We observe that, the best

method for reducing the minutiae template, K-Means

and MRGA, have a good minutiae spatial distribution

on the ﬁngerprint image. If we analyze with more

details the MRGA reduction template we could ob-

serve, it is a combination of Barycenter approach and

K-Means.

5 CONCLUSION

When we want to have secure device to store your bio-

metric data, you use a SE. When we want to increase

the security of the matching algorithm, we have to

store very efﬁcient biometric data on it. SE are lim-

ited in term of memory size, this is why we need to

have methods which permit to reduce the size of the

biometric template.

(a) Truncation (b) Troncation Random

Permutation

(e) K-Means (f) MRGA

Figure 8: Illustration of a reduced ﬁngerprint template for

each method of the state-of-the-art and the MRGA method

on the initial template (in blue on each ﬁgure).

Few methods have been proposed on the state-of-

the-art and we want to estimate in what way we are

close to the best reduction. The purpose of this pa-

per is to ﬁnd a solution for estimating the best re-

duced minutiae template for different templates size.

Thus, we could increase the security of embedded ﬁn-

gerprint systems. On the state-of-the-art, no method

which permit to estimate this ”optimal” template that

is why we proposed the MRGA approach. We have

used one database FVC2004DB1 and three matching

algorithms to test the performance of this method. We

have shown that the MRGA is for all matching algo-

rithms the best method and is close to the initial tem-

plate in term of performance. We could consider we

obtain the upper value performance and template in

Towards an Optimal Template Reduction for Securing Embedded Fingerprint Devices

335

comparison with the initial template. We observe than

some methods are closer to the MRGA method when

we have a high number of minutiae on the reduced

template. In opposition, when we have a small num-

ber of minutiae, we have many ways to improve the

reduction and obtain better performance and security.

With this study, we proposed an approach to evalu-

ate the upper performance for peeling off the minutiae

template. This method could be used on server when

enrollment part is done, to increase the efﬁciency of

the matching algorithm.

As perspectives, we want to analyze and proposed

new methods to reduce the initial template to be closer

to the ”optimal” reduced template provided by the

MRGA method.

REFERENCES

Fvc2004db1. http://bias.csr.unibo.it/fvc2004/databases.asp.

ISO/IEC 19794-2. information technology - biometric data

interchange format format - part 2 : Finger minutiae

data, 2011.

ISO/IEC 19795-2. information technology - biometric per-

formance testing and reporting - part 2 : Testing

methodologies for technology and scenario evalua-

tion, 2007.

Cappelli, R., Ferrara, M., and Maltoni, D. (2010). Minu-

tia cylinder-code: A new representation and matching

technique for ﬁngerprint recognition. IEEE Transac-

tions on Pattern Analysis and Machine Intelligence,

32(12):2128–2141.

Grother, P. and Salamon, W. (2007). Interoperability of the

iso/iec 19794-2 compact card and 10 iso/iec 7816-11

match-on-card speciﬁcations 11.

Holland, J. H. (1975). Adaptation in natural and artiﬁcial

systems: an introductory analysis with applications to

biology, control, and artiﬁcial intelligence. U Michi-

gan Press.

IHS (2016). Fingerprint sensor market growth continues up-

ward trajectory, ihs says. https://technology.ihs.com/

571358/ﬁngerprint-sensor-market-growth-continues-

upward-trajectory-ihs-says.

ISO. ISO/IEC 7816-1 to 15: Identiﬁcation cards - Inte-

grated circuit(s) cards with contacts(Parts 1 to 15).

ISO/IEC, http://www.iso.org.

Rechenberg, I. Evolutionsstrategie94. frommann-holzboog-

verlag, stuttgart (germany), 1994. German; includes

also, 1581.

Vibert, B., Charrier, C., Le Bars, J.-M., and Rosenberger, C.

(2015). Comparative study of minutiae selection al-

gorithms for iso ﬁngerprint templates. In SPIE/IS&T

Electronic Imaging, pages 94090C–94090C. Interna-

tional Society for Optics and Photonics.

Wall, M. B. (1996). A genetic algorithm for resource-

constrained scheduling. PhD thesis, Massachusetts

Institute of Technology.

Watson, C. I., Garris, M. D., Tabassi, E., Wilson, C. L., Mc-

cabe, R. M., Janet, S., and Ko, K. (2007). User’s guide

to nist biometric image software (nbis). Technical re-

port, NIST.

ICISSP 2018 - 4th International Conference on Information Systems Security and Privacy

336