Classifying Biometric Systems Users among the Doddington Zoo:

Application to Keystroke Dynamics

Denis Migdal

1 a

, Ilaria Magotti

and Christophe Rosenberger

2 b

Universit

e Clermont Auvergne, CNRS, Mines Saint-Etienne, Clermont Auvergne INP, LIMOS,

F-63000 Clermont-Ferrand, France

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

Keywords:

Doddington Zoo, Performance Evaluation of Biometric Systems, Keystroke Dynamics.

Abstract:

Doddington zoo deﬁnes four categories of users when using a biometric system related to their difﬁculty to be

recognized or attacked. In this paper, we propose an original work consisting in predicting for any biometric

modality the associated animal in the Doddington menagerie related to a user given few captured biometric

samples. Such a prediction could be useful for many applications, as for example, to adapt the behavior of

biometric systems to each user. In this work, we apply this methodology to keystroke dynamics as it is an

interesting behavioral biometric modality for user authentication. It consists in analyzing the way of typing

of a user in order to recognize him/her. We use a signiﬁcant keystroke dynamics dataset and we demonstrate

through experimental results the beneﬁt of the proposed approach.

1 INTRODUCTION

The performance of biometric systems varies for dif-

ferent reasons as detailed in (Phillips et al., 2000)

among human interactions (Blanco-Gonzalo et al.,

2017), environmental conditions (Tan et al., 2010) or

intrinsic variations related to users (Yager and Dun-

stone, 2008; Kirchgasser and Uhl, 2016). In this pa-

per, we focus on this last point. As it has been iden-

tiﬁed in 1998 by a pioneer article (Doddington et al.,

1998), the performance of biometric systems is far to

be similar for all users. A biometric system could

be efﬁcient for some users and generate many false

rejection for others. The biometric menagerie usu-

ally known by Doddington zoo, is a collection of an-

imal labels describing the performance behavior of a

user with biometric systems. It is an interesting ap-

proach usually used to improve biometric recognition

systems performance (Barron et al., 2008). In the bio-

metric menagerie, users are classiﬁed based on legit-

imate scores (comparison with samples belonging to

the user) and impostor ones (comparison with sam-

ples from other users considered as impostors) . In

fact, users are split into four categories: 1) Sheeps

are easy to recognize, 2) Goats are difﬁcult to rec-

ognize, 3) Lambs are easy to forge or counterfeit, 4)

https://orcid.org/0000-0002-4741-1849

https://orcid.org/0000-0002-2042-9029

Wolves are good to forge others. Being able to clas-

sify a user in the Doddington menagerie has many in-

terests, mainly for the deﬁnition of adaptive biomet-

ric systems. The biometric reference template of a

user can be updated considering the type of user (or

animal) (Mhenni et al., 2018). Synthetic biometric

datasets can be created by generating biometric sam-

ples from users considering these categories of ani-

mals (Lopes Silva et al., 2019). A multibiometric sys-

tem can be tuned in function of the animal associated

to the user (Poh, 2010). We believe that this user clas-

siﬁcation is particularly useful for behavioral biomet-

ric modalities. Indeed, the stability of user’s behavior

has a great impact on performance on such biometric

systems. In this work, we consider the keystroke dy-

namics as biometric modality in order to apply the

proposed method. Note that the proposed method

could be applied on any biometric modality.

Keystroke dynamics is a behavioral biometric

modality deﬁned in 1980 (Gaines et al., 1980). Its

principle consists in analyzing the behavior of a user

when typing on a keyboard. Times (pressure, ﬂight,

release) are measured by the operating system and can

be used as raw information on user’s behavior. This

biometric modality is very interesting for user authen-

tication as it does not require any additional sensor

and it is natural for users to type their password. Its

main drawback concerns the performance that cannot

Migdal, D., Magotti, I. and Rosenberger, C.

Classifying Biometric Systems Users among the Doddington Zoo: Application to Keystroke Dynamics.

DOI: 10.5220/0010577507470753

In Proceedings of the 18th International Conference on Security and Cryptography (SECRYPT 2021), pages 747-753

ISBN: 978-989-758-524-1

747

be as good as face or ﬁngerprint for example. We can

summarize here the contributions of this work. In this

paper, we propose an original method whose objective

is to classify a user in one of the 3 Doddington classes

(sheep, goat and lamb) given few biometric samples.

We do not consider wolves in this work because we

want to propose solutions to enhance the performance

and not necessary the robustness face to attacks (it

could concern perspectives of this study). Most of

studies in the literature classify users with a posteri-

ori samples from a dataset (Teli et al., 2011). Another

contribution in this paper is to propose a validation

process for users classiﬁcation. In general, classiﬁ-

cation results are associated to performance (i.e. the

goat class should have the poorest results). It is clear

that the frontier between animal classes is difﬁcult to

establish. We assume in this work that a user classi-

ﬁed to a speciﬁc animal for a biometric modality re-

mains the same for different data. Under this assump-

tion, we can measure a recognition rate considering

the consensus between the classiﬁcation among data.

The paper is organised as follows. We present in sec-

tion 2, the state of the art concerning the classiﬁca-

tion of users among the Doddington zoo. In section 3,

we provide a brief background on keystroke dynam-

ics. Section 4 is dedicated to the proposed method for

associating a user with an animal from the Dodding-

ton zoo. Experimental results on a large dataset on

keystroke dynamics data are given in section 5. We

conclude this study in section 6.

2 LITERATURE REVIEW

The biometric menagerie has been deﬁned in 1998

(Doddington et al., 1998) with a ﬁrst study show-

ing the relationship between users and their perfor-

mance when using biometric systems. In 2009, Ross

et al. (Ross et al., 2009) proposed a user-dependent

multibiometric system by considering the animal in

the Doddington menagerie associated to a user in a

biometric dataset. They propose a classiﬁcation ap-

proach by considering all legitimate and impostor

scores in the dataset. Even if this classiﬁcation per-

mits to enhance the performance during the fusion, no

validation of the classiﬁcation is proposed. In 2011,

Teli et al. (Teli et al., 2011) investigated the biometric

zoos generalization across algorithms and data sets.

The question was to answer as for example if a sub-

ject classiﬁed as a goat for algorithm A on dataset X,

is also a Goat for algorithm B on data set Y? Experi-

ments have been conducted on a face database (FRVT

2006) with two matching algorithms. They propose

a framework for describing and testing for the exis-

tence of different levels of biometric zoo. Zeroth-

Order Zoo implies only that people may be labeled

as animals in a single experiment. A ﬁrst-order zoo

exists when personal identity is considered important

to others data within the same scenario. In this work,

we are addressing the ﬁrst-order zoo for different data

and matching algorithms to classify users. Morales et

al. in 2014 (Morales et al., 2014) proposed a predic-

tion method of ”good users” with keystroke dynam-

ics. They used the Kullback-Leibler divergence as a

quality measure to categorize users. They split users

from a keystroke dynamics dataset into 3 classes con-

sidering the value of the Equal Error Rate (EER) for

the validation process. This work is interesting even

if the proposed method permits to identify good bio-

metric samples more than good users (or sheeps). Re-

cently, Mehnni et al. (Mhenni et al., 2018) proposed

in 2018 to classify users in the Doddington menagerie

in order to adapt the template update strategy. This

approach has been applied to keystroke dynamics and

uses the notion of relative entropy for user classiﬁ-

cation. This adaptive process permits to improve the

veriﬁcation performance over time.

All these works are interesting and provide good

studies on users classiﬁcation in the Doddington

menagerie. Nevertheless, we have many remarks. In

most works, legitimate and impostor scores are used

for user classiﬁcation. Consequently, the process be-

comes very dependent of the used matching algo-

rithms (as identiﬁed by Teli at al. (Teli et al., 2011)).

The resulting fact is that the frontier between animal

classes is far to be clear. Considering scores is not

maybe a good idea. Studies in the literature on the

biometric menagerie are often dataset driven i.e. re-

searchers try to classify users in the dataset. Could it

be possible after acquiring few biometric samples to

predict the associated animal related to the user? The

validation of users classiﬁcation is not completely sat-

isfying. In machine learning applications, we expect

to measure a recognition rate but the ground truth is

here difﬁcult to establish. In many studies (Mhenni

et al., 2018), user classiﬁcation is used as posteriori

information to adapt the user recognition, the fron-

tier between classes has not to be precise. An im-

portant question remains. Are some users more dif-

ﬁcult to recognize/attack for any matching algorithm

and dataset? For morphological biometric modalities,

the question remains open. We think that it is easier

for behavioral modalities. It is well known that some

users are more difﬁcult to recognize for keystroke dy-

namics as for example (stability of typing, habits. . . ).

That is why we consider this biometric modality in

this work. Before presenting the proposed method for

user classiﬁcation in the biometric menagerie, we pro-

SECRYPT 2021 - 18th International Conference on Security and Cryptography

748

vide a brief background on keystroke dynamics as it

is the considered biometric modality in this paper.

3 BACKGROUND IN

KEYSTROKE DYNAMICS

Keystroke dynamics is a behavioral biometric modal-

ity. keystroke dynamics systems are low cost because

only a keyboard and an accurate timer are needed

to record the timings that will be used to recognize

the user. No other sensor must be purchased. In

term of usability, this solution is very good. Its main

drawback is a lower performance as it uses a behav-

ior less stable than morphological modalities. Try-

ing to adapt the processing to users is thus very im-

portant to enhance the performance of such systems.

It is possible to capture this behavior considering 1)

OS events (times), 2) video of the typing, using a

camera or a webcam to monitor the hands while the

user is typing on a keyboard and 3) audio sound of

the typing. In this paper, only the ﬁrst type of data

is employed, which represents the timing pattern of

keystroke (Idrus et al., 2013). After the capture phase,

an amount of unprocessed data is obtained. This in-

formation can be considered as a list of events in se-

quential order recorded from the moment in which the

user starts typing on the keyboard. The next phase

concerns the processing of the data collected during

the biometric capture, the data need to be organized

and modiﬁed in order to obtain a processed record

consisting in an ensemble of features (Idrus et al.,

2013). The time-based measure we consider in our

analysis can be described as: keystroke latencies can

be deﬁned as the differences of time between two keys

events (Giot et al., 2011) and can be determined by the

timing delay experienced by a process (Idrus et al.,

2013); keystroke duration represents for how long a

key is pressed. The notion of digraph has also to be

introduced: it is the time necessary to press two keys.

This notion has been extended to n-graphs, when con-

sidering n events.

As keystroke dynamics is a behavioral modal-

ity, the generation of the reference template requires

many samples to capture the behavior. The more data

we acquire on a user, the better will be the recogni-

tion results. The reference template of a user u

deﬁned by R

= {b

, .., b

} where b

corresponds

to a biometric feature vector, for keystroke dynamics,

it corresponds to collected times associated to the typ-

ing of a password and M corresponds to the number

of samples used during the enrollment. For usability

reasons, the number of captures should be limited, in

this paper, we use M = 3. Given a biometric probe

of assumed user u

, we consider 3 matching score

computations from the literature (Migdal, 2019):

, R

) = min

j=1:M

∑

k=1

− b

j,k

| (1)

, R

) =

∑

j=1

∑

k=1

− b

j,k

| (2)

, R

) =

∑

j=1

min

k=1:K

− b

j,k

| (3)

Where b

j,k

is the kth feature of the sample j from

user u

and K is the dimension of the biometric fea-

ture vector (depending on the number of characters in

the password). We consider 3 algorithms in this work

in order to estimate how invariant is our user classiﬁ-

cation to them.

4 PROPOSED METHOD

The proposed method has for objective to deﬁne an

operational approach to classify a user in one of the

class in the Doddington zoo. We expect to realize this

classiﬁcation using few biometric samples acquired

from the user. The proposed approach requires an ini-

tial step to achieve this goal. As mentioned in section

2, we will not use directly matching scores for the

classiﬁcation but AUC values.

4.1 Initial Step

Figure 1: User signature computation. Illustration on a

dataset composed of N biometric samples for each of the

P individuals (C is related to the number of possibilities for

the reference deﬁnition.

We suppose having a dataset ∆ = {b

, i = 1 : P, j = 1 :

N} where b

corresponds to the jth biometric sample

of size N for user u

in a dataset composed of P in-

dividuals. In order to generate the reference template

of user u

, one could use M samples. For mor-

phological biometric modalities, M could be equal

to 1, for behavioral ones, M should be higher (typi-

cally 3 or 5). To generate the reference template for

Classifying Biometric Systems Users among the Doddington Zoo: Application to Keystroke Dynamics

749

one user, there are





possibilities. For each pos-

sibility (i.e. choice of M samples among N) for the

generation of the reference template, we can com-

pute all legitimate matching scores with remaining

samples for the same user. We obtain for one user

(N − M) legitimate scores. We can also compute im-

postor scores by comparing the reference template of

the considered user with all samples from other users.

We thus obtain, (P − 1) × N impostor scores. Given

these scores, we can compute the False Match Rates

(FMR) and False Non Match Rates (FNMR) values

for each choice of the reference template. We can

compute the associated ROC curves and the Area Un-

der the Curve (AUC) value. This AUC value deﬁnes

the performance of the biometric system (describing

the ability to well recognize the considered user and

to differentiate him/her from others) when using this

reference template.

If we apply this process for all users and all

choices of reference templates, we obtain a matrix

Γ = {AUC

, i = 1 : P, k = 1 :





}. In order to il-

lustrate the amount of computations, we give some

ﬁgures with the dataset we use in this work (dataset

of P = 110 individuals described by N = 10 biomet-

ric samples). If we use M = 3 samples for the ref-

erence generation, we have





= 120 choices. For

each choice, we compute 10 −3 = 7 legitimate scores

and 109∗10 = 1090 impostor ones. The Γ matrix has

consequently 120 lines (corresponding to the number

of possible reference templates) and 110 columns (re-

lated to the number of users in dataset ∆). The Γ ma-

trix describes the difﬁculty of recognizing each user

for the different choices of the reference template.

Let’s consider now a column of this matrix (corre-

sponding to the AUC

values for user i). If these val-

ues are in average low, it means that user u

is easy to

recognize and well differentiated from other users (a

sheep in the Doddington menagerie). On the contrary,

if values are in average high, user u

can be a goat or

a lamb. To decide among these two classes, we could

consider the variations of the AUC values for each

choice of the reference template. If there are some

variations, it means that user u

is a lamb otherwise

it is a goat. To implement this strategy, we compute

for each user a signature composed of E[AUC] (mean

of AUC values) and σ[AUC] (standard deviation of

AUC values) describing its performance behavior for

the classiﬁcation. We obtain a signature for each of

the P users. Figure 1 summarizes the whole process.

Once we have a signature for each user, we need to

deﬁne the decision frontier to Doddington classes. We

adapted the proposed process in (Ross et al., 2009)

to AUC values. In this paper, they considered the

70th percentile of low legitimate scores as sheep and

the 10th percentile of higher impostor scores as lamb.

Others are classiﬁed as goats. In our work, we use

the 70th percentile of low E[AUC] values (associated

to threshold T

) as sheep and the 10th percentile of

higher σ[AUC] (associated to threshold T

) as lamb.

Others are classiﬁed as goats. The classiﬁcation is

thus achieved with a simple decision rule:

Class =







sheep i f E[AUC] < T

lamb i f σ[AUC] > T

goat otherwise

(4)

Note the values of the decision thresholds T

and T

are related to the used matching algorithm. The pro-

posed signature for each user has also the advantage

to be normalized.

4.2 Prediction Step

The user class prediction is quite simple and

consists in ﬁrst computing the user signature

(E[AUC], σ[AUC]). The predicted class is obtained

by applying equation 4. In order to be used in real

conditions and to especially avoid the computing of

all impostor scores, it is possible to select K biomet-

ric samples as a sub-sampling. It is possible to use a

simple clustering approach with the matching score as

distance to generate K clusters and keep the biomet-

ric samples the closest to the obtained K centroids.

The user will have to give few biometric samples for

computing legitimate scores and the previous K se-

lected biometric samples are used as impostor ones.

The user signature can quickly be generated for the

prediction.

5 EXPERIMENTAL RESULTS

5.1 Experimental Protocol

The ﬁrst step is to select a biometric dataset. We use

in this paper the GREYC-NISLAB keystroke dataset

(Idrus et al., 2013). The collection of data has taken

place in two locations: France and Norway. Sub-

jects came from 24 different countries. A total of

110 individualshas taken part in the experiment (70

in France and 40 in Norway).Users have been asked

to type 5 static passphrases, which were chosen be-

cause of their popularity. We refer to these pass-

words as P1, P2, ..., P5 in next discussions. There-

fore, the database contains 11000 samples in total (5

passwords * 2 classes of hand * 110 users * 10 en-

tries).The great beneﬁt of this dataset is to have the

biometric samples for many users on different data

SECRYPT 2021 - 18th International Conference on Security and Cryptography

750

(here passwords). Under our assumption, a user clas-

siﬁcation should be stable among the data for a given

matching algorithm at least. For the experiments, we

tested 2 scenarios for the choice of the reference tem-

plate with M = 3 (i.e. 3 samples for the reference

generation). In scenario 3/5, e choose the 3 samples

for the reference among the 5 ﬁrst samples (10 possi-

bilities). This choice has for objective to take into ac-

count the chronology of the data acquisition. It could

be important for a behavioral biometric modality. In

scenario 3/10, we choose the 3 samples for the ref-

erence among all samples (120 possibilities). In this

scenario, we do not consider the chronology of data

for the reference generation.

5.2 Data Visualisation

Before analyzing data, we propose in this section to

visualize data. We use the S

matching algorithm de-

ﬁned in Equation 1 in this section. As illustration, we

display the signature of 3 users for the 5 passwords in

Figure 2. Signatures from the same user are displayed

with the same color. Because the decision thresholds

are slightly different for each passwords, we use the

average one. It is not surprising to notice that sheeps

are more stable, this is an important characteristic of

these users. There are some variations for lambs and

goats which is logical as these users are less stable

by deﬁnition. If we consider the average signature of

these 3 users, the associated animal in the biometric

menagerie is rather clear. Figure 3 presents the dis-

tribution of the dispersion measure of users signature

for the 5 passwords. The dispersion measure com-

putes the average distance of users signature with its

average value. As it can be seen, the dispersion is low

showing in general a good stability.

Figure 2: Illustrations of 3 user signatures for the 5 pass-

words.

5.3 Validation Process

The validation process is very important for user clas-

siﬁcation within the Doddington menagerie. The

Figure 3: Distribution of the dispersion measure of users

signature.

main problem is that there is no ground truth like in

many machine learning problems. In most of stud-

ies in the literature (Poh, 2010; Ross et al., 2009;

Mhenni et al., 2018), the efﬁciency of user classiﬁ-

cation is demonstrated by obtaining different perfor-

mances for each predicted class (i.e. sheeps have bet-

ter performance than goats). In this work, we can use

an additional information and concerns the biomet-

ric samples for different users on different data. We

have no ground truth but under our assumption, a user

classiﬁed as a goat should be affected in this class for

all data. In the GREYC-NISLAB dataset, we have

5 passwords typed by the same users. We can thus

compute a consensus value between all passwords for

user classiﬁcation. For each user, we apply the pro-

posed method for the 5 passwords. A majority vote is

then applied to deﬁne the consensus class with a con-

ﬁdence index (CI). As for example, a user could be

affected to the goat class with a conﬁdence CI=60%

(meaning that for 3 passwords among 5, the user has

been affected to this class). We then propose a global

metric called Global Consensus Rate (GSR) as:

GSR =

∑

i=1

CI(i) (5)

Where P is the number of individuals in the dataset

(here P=110). Note that CI is normalized by the num-

ber of available data (here 5 passwords).

5.4 Results

Table 1 provides the GSR values for the 3 matching

algorithms (deﬁned by equations 1 to 3) and the two

scenarios (for the choice of the reference). We can see

ﬁrst that the GSR value is quite stable for all match-

ing algorithms. This is an important result, it conﬁrms

our assumption that user classiﬁcation is not related to

the used matching algorithm. Second, the two testing

scenarios permit to obtain very similar results. Con-

sidering we have processed keystroke dynamics data

that are less stable than morphological biometric data,

reaching GSR ' 80% is a good result.

Classifying Biometric Systems Users among the Doddington Zoo: Application to Keystroke Dynamics

751

Table 1: Value of the consensus value (GSR) for the 3

matching algorithms and for the 2 scenarios.

Matching Scenario 3/5 Scenario 3/10

79,8% 78,9%

80,4% 80,4%

79,4% 78,7%

We tried to improve the previous results by optimizing

the decision thresholds. The question we wanted to

answer is to know if it was possible to deﬁne common

values of T

and T

for the 5 passwords. We tested dif-

ferent threshold values between the minimal and max-

imal values for the 5 passwords. Table 2 presents the

obtained results by optimizing the thresholds. Note

that we used the testing scenario 3/10 as we saw pre-

viously that there was no difference with the other.

We obtain a nice gain of the GSR value showing that

it is possible to enhance slightly the performance of

the proposed method.

Table 2: Value of the consensus value (GSR) for the 3

matching algorithms with optimized thresholds.

Matching algorithm GSR value

82.4%

83,6%

82.7%

6 CONCLUSION AND

PERSPECTIVES

In this work, we addressed the problem of user clas-

siﬁcation in the biometric menagerie. Such a method

could have many applications in biometrics mainly to

adapt the processing in function of the behavior of the

user while using a biometric system. The proposed

approach is based on the deﬁnition of a signature re-

lated to the stability and performance associated to a

user. The proposed framework makes it possible to

predict user class in an operational mode by a sim-

ple decision rule. Obtained results on a keystroke dy-

namics dataset composed of biometric data for dif-

ferent passwords permits to measure the consensus of

the prediction. We obtained quantitative results up-

per than 82%. Perspectives of this study concern the

application of the proposed method on other biomet-

ric modalities. We believe that the Doddington zoo is

particularly interesting for behavioral ones. We also

intend to apply the prediction results to enhance/adapt

the performance of biometric systems.

REFERENCES

Barron, U. G., Corkery, G., Barry, B., Butler, F., McDon-

nell, K., and Ward, S. (2008). Assessment of reti-

nal recognition technology as a biometric method for

sheep identiﬁcation. Computers and electronics in

agriculture, 60(2):156–166.

Blanco-Gonzalo, R., Sanchez-Reillo, R., Liu-Jimenez, J.,

and Sanchez-Redondo, C. (2017). How to assess user

interaction effects in biometric performance. In 2017

IEEE International Conference on Identity, Security

and Behavior Analysis (ISBA), pages 1–6. IEEE.

Doddington, G., Liggett, W., Martin, A., Przybocki, M.,

and Reynolds, D. (1998). Sheep, goats, lambs and

wolves: A statistical analysis of speaker performance

in the nist 1998 speaker recognition evaluation. Tech-

nical report, National Inst of Standards and Technol-

ogy Gaithersburg Md.

Gaines, R. S., Lisowski, W., Press, S. J., and Shapiro, N.

(1980). Authentication by keystroke timing: Some

preliminary results. Technical report, Rand Corp

Santa Monica CA.

Giot, R., El-Abed, M., and Rosenberger, C. (2011).

Keystroke dynamics authentication. In Biometrics,

page chapitre 8. InTech.

Idrus, S. Z. S., Cherrier, E., Rosenberger, C., and Bours, P.

(2013). Soft biometrics for keystroke dynamics. In

International Conference Image Analysis and Recog-

nition, pages 11–18. Springer.

Kirchgasser, S. and Uhl, A. (2016). Biometric menagerie

in time-span separated ﬁngerprint data. In 2016 Inter-

national Conference of the Biometrics Special Interest

Group (BIOSIG), pages 1–7. IEEE.

Lopes Silva, P., Luz, E., Moreira, G., Moraes, L., and

Menotti, D. (2019). Chimerical dataset creation proto-

col based on doddington zoo: A biometric application

with face, eye, and ecg. Sensors, 19(13):2968.

Mhenni, A., Cherrier, E., Rosenberger, C., and Amara, N.

E. B. (2018). Adaptive biometric strategy using dod-

dington zoo classiﬁcation of user’s keystroke dynam-

ics. In 2018 14th International Wireless Communi-

cations & Mobile Computing Conference (IWCMC),

pages 488–493. IEEE.

Migdal, D. (2019). Contributions to keystroke dynamics for

privacy and security on the Internet. PhD thesis, Nor-

mandie Universit

Morales, A., Fierrez, J., and Ortega-Garcia, J. (2014). To-

wards predicting good users for biometric recognition

based on keystroke dynamics. In European Confer-

ence on Computer Vision, pages 711–724. Springer.

Phillips, P. J., Martin, A., Wilson, C. L., and Przybocki, M.

(2000). An introduction evaluating biometric systems.

Computer, 33(2):56–63.

Poh, N. (2010). User-speciﬁc score normalization and fu-

sion for biometric person recognition. Advanced Top-

ics in Biometrics, 16:401–418.

Ross, A., Rattani, A., and Tistarelli, M. (2009). Exploiting

the “doddington zoo” effect in biometric fusion. In

2009 IEEE 3rd International Conference on Biomet-

SECRYPT 2021 - 18th International Conference on Security and Cryptography

752

rics: Theory, Applications, and Systems, pages 1–7.

IEEE.

Tan, B., Lewicke, A., Yambay, D., and Schuckers, S.

(2010). The effect of environmental conditions and

novel spooﬁng methods on ﬁngerprint anti-spooﬁng

algorithms. In 2010 IEEE International Workshop on

Information Forensics and Security, pages 1–6. IEEE.

Teli, M. N., Beveridge, J. R., Phillips, P. J., Givens, G. H.,

Bolme, D. S., and Draper, B. A. (2011). Biometric

zoos: Theory and experimental evidence. In 2011

International Joint Conference on Biometrics (IJCB),

pages 1–8. IEEE.

Yager, N. and Dunstone, T. (2008). The biometric

menagerie. IEEE transactions on pattern analysis and

machine intelligence, 32(2):220–230.

Classifying Biometric Systems Users among the Doddington Zoo: Application to Keystroke Dynamics

753