AN APPROACH TO DATA COLLECTION IN AN ONLINE

SIGNATURE VERIFICATION SYSTEM

Andreea Salinca, Sorin Mircea Rusu and Ştefan Diaconescu

Research & Development, SOFTWIN, Bucharest, Romania

Keywords: Data Collection, Database, Biometry, Dynamic Signature, Online Signature, Authentication System.

Abstract: Online handwritten signature verification is one of the branches of behavioral biometry that is gaining

popularity in protecting sensitive information. Our paper addresses a key issue in evaluating performances

of online signature authentication systems: data collection. Acquiring a real dataset with handwritten

signatures is a major step in the system verification. We will present our collecting techniques in the process

of acquisition of dynamic handwritten signatures (more than 5000 genuine signatures and more than 2000

skilled forgeries have been collected from a total of 113 people) useful in the improvement of evaluation

results for the authentication system.

1 INTRODUCTION

Biometric authentication has been widely used as a

trusted security solution in protecting sensitive asset.

These systems are not based on what a person

possesses (password, PIN, token), but on the basis of

what the person “is”. Unlike physiologic biometry,

the authentication with holographic signature is non

intrusive. For hundreds of years, the signature has

been approved by all cultures and civilizations with

a major social implication, and it has been accepted

as legal evidence.

This paper presents several signatures databases

acquisitions (PHILIPS, SVC’2004 Development Set,

MCYST Signature Subcorpus, BIOMET Signature

Subcorpus, BioSecure Signature Subcorpus DS2 and

DS3) and includes a survey on acquisition devices,

procedures of acquiring genuine signatures and

several types of forgeries, and the main results

obtained in their evaluation and in international

evaluations like SVC’2004 - First International

Signature Verification Competition (Bernadette,

Chollet, and Petrovska-Delacrétaz, 2009).

The signatures database collected for SVC’2004

contains samples from only 60 people, and for

privacy reasons, the signatures are not “real”. There

were two sets of signatures: one containing only

coordinate information and the other containing also

pen pressure and orientation. The team from Sabanci

University of Turkey obtained the best result in the

evaluation of a Dynamic Time Warping based

system for both sets (the first set had an ERR of

2.84% and the second set had an ERR of 2.89%)

(Yeung et al., 2004). A recent article presents

evaluation results of online signature obtained in

BioSecures Signature Evaluation Campaign

(BSEC’2009), depending on the signatures quality

captured on fixed (DS2) or mobile platforms (DS3)

from a total of 382 people. The main task was to

evaluate the algorithms’ results in different

acquisition conditions of signatures as well as the

complexity of information contained in signatures.

The results revealed that system robustness depends

on the quality of signatures (on DS2 an ERR of

2.2% for skilled forgeries and an ERR of 0.51% for

random forgeries, and also, on DS3 an ERR of

4.97% for skilled forgeries and an ERR of 0.55% for

random forgeries) (Houmani et. al, 2011).

Comparing to previous works in collecting

signatures databases which have focused on

improving the evaluation methods, we will focus on

post-processing the data already collected in order to

improve the evaluation performance of a system.

The first section briefly describes our data

acquisition system and the acquisition application.

Further on, we present the procedures we used to

accomplish the acquisition process and our strategies

for post-processing the dataset. Finally, we present

experimental results proving the impact of data

collecting strategies over the system performances

on evaluating the authentication system.

251

Salinca A., Mircea Rusu S. and Diaconescu ¸S..

AN APPROACH TO DATA COLLECTION IN AN ONLINE SIGNATURE VERIFICATION SYSTEM.

DOI: 10.5220/0003908702510254

In Proceedings of the 8th International Conference on Web Information Systems and Technologies (WEBIST-2012), pages 251-254

ISBN: 978-989-8565-08-2

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

2 ONLINE SIGNATURE

ACQUISITION SYSTEM

The online acquisition system contains three main

sub-systems: the acquisition device, the acquisition

application and the signatures database (Figure 1).

Figure 1: Acquisition system architecture.

The acquisition device is an electronic pen which

captures the bio-kinetic information of a signature. It

writes on a pattern paper and it does not need a

special tablet (Rusu, Dinescu, Diaconescu, 2011).

When a person signs, the acquisition device (Figure

2) captures hand movements using the embedded

MEMS (micro-electro-mechanical systems)

accelerations sensors into a set of 4 time-based

series. Moreover, it captures the graphic form of the

signature through an optical navigation sensor

(ONS) located inside the biometric pen.

Figure 2: Dynamic signature acquisition device.

2.1 Acquisition Application

The actors of our acquisition system were subjects,

operators, acquisition device, acquisition application

and acquisition database. In the process of data

collecting the subjects were assisted by operators.

The acquisition application is designed as a web-

based application. It offers a real-time visual

feedback of the signature. Besides the biometric

signature acquisition and storage in the system’s

relational database, another functionality provided

by the application was of managing a set of genuine

signatures and a set of forgeries for every user. The

application manages additional features for operators

and users using the electronic pen for signing.

3 ACQUISITION PROCESS

The creation of a dynamic handwritten signatures

database must correspond to a real environment use

of the authentication system. The data collection has

to be large enough to cover particular cases (e.g.

signatures which are very hard to forge). The

subjects were asked to use their ”daily life”

signatures and to try to sign in the most natural way.

Our signatures database was built from 113 persons

of different ages and genders. The number of

collected signatures was over 7000. The acquisition

process of the online handwritten signatures

database contains three phases:

 Phase 1 – This is the preparation phase before

the proper acquisition phase. The operator

enters the names of the signers in the

acquisition application and prints the signing

forms. We use two types of signing forms: one

for genuine signatures and one for forgeries.

 Phase 2 – Each of the subjects has to give a set

of 50 samples of genuine signatures and 20

skilled forgeries for another user. The

signatures are collected in 6 different sessions

which are presented below. The subjects

providing signatures are assisted by multiple

operators. Every operator has to assist an equal

number of subjects.

 Phase 3 – Post-processing of the database

occurs in this phase. Several procedures are

applied to the dataset in order to eliminate some

signatures that could be affected by different

types of errors. We will prove the significant

role of this phase in the evaluation of the

authentication system.

3.1 Acquisition Sessions

The acquisition sessions from the second phase of

the acquisition process represent the acquisition

itself of the set of dynamic signatures. In 6 different

sessions, each signer supplied a total of 50 samples

of genuine signatures and 20 skilled forgeries. There

was a few days distance in between the sessions, to

make the process more real. In this purpose, the

subjects giving genuine signatures were encouraged

WEBIST2012-8thInternationalConferenceonWebInformationSystemsandTechnologies

252

by the operators to give their real signature that

expresses their individualities. Also, subjects giving

forgeries were advised to practice the signature

before signing not only in terms of graphics, but also

in terms of speed and accelerations. They have been

trained to reproduce the dynamics of the signature

even by the person that they tried to forge.

Table 1 summarizes the sessions in the

acquisition process of the signatures database. In the

acquisition process, several measures were taken in

order to enhance data collecting such as day breaks

between acquisition sessions or practicing forgeries.

Table 1: Sessions of collecting the database.

Dataset

Number of

signatures

Signatures Type

No. of

Samples/S

ubject

Session 1 1130 Genuine 10

Session 2 1130 Genuine 10

Session 3 1130 Genuine 10

Session 4 1130 Genuine 10

Session 5 2260

Genuine &

Forgery

Session 6 1130 Forgery 10

3.2 Post-processing Data Collection

During the acquisition process, we have found two

types of errors that could alter the signatures from

that database: acquisition errors and operating errors.

The acquisition errors are caused by the

improper handling of the electronic pen, for instance

by holding it at an improper angle or spin related to

its own axis. The visual feedback provided by the

acquisition application represents the first step in

order to minimize this type of error. Also, at the end

of the acquisition process, these errors were

removed by visual inspection of each set of

signatures images belonging to a subject. The

operating errors are caused by mixing the subject

IDs and signatures. This type of errors was

eliminated by defining validation rules in the

acquisition application and in the database. By using

generated forms, customized for each subject, for

handwritten signatures in the acquisition process, we

have eliminated operator’s errors that can occur by

mixing subject IDs.

After removing the errors, we used z-score or

standard score which is a common statistical method

for data standardization. We apply z-score to detect

signatures having significantly differences in length

from the rest of signatures collected from a user. The

length variation may appear due to the fact that this

type of signatures may be incorrectly acquired in

terms of the acceleration signals and graphic form.

Also, the properties of the normal distribution

were helpful in assessing and improving out

signature data set (Larsen and Marx, 2006).

We calculated z-score in order to understand

how various subsets of data signatures contribute to

the authentication system performance results. Z-

scores were used to find subsets of signatures which

could negatively affect the rates in the evaluation

process. We computed z-scores by formula (1),

where μ represents the mean of the distribution and

σ the standard deviation.

−

(1)

By using the three-sigma rule or 68-95-99.7 rule,

we have considered the interval

a 95%

confidence interval (2).

Pr( 2 2 0.09545x

−σ≤ ≤μ+σ) ≈

(2)

For every subject in the database we have

computed the z-score for each of his genuine

signature. Then, we eliminated from the dataset the

signatures with a z-score outside the interval of

confidence. For the 95% prediction interval chosen,

the corresponding z-score is 1.96 (Larsen and Marx,

2006). The score is computed in terms of the

quantile function using the formula (3), where Φ

represents the standard normal cumulative

distribution function; μ represents the mean of the

distribution and σ the standard deviation.

() ()

−−

μσ

Φ=μ+σΦ

(3)

Another procedure applied on the signatures

database was to eliminate the signatures given by the

subject in the first acquisition session, when he was

not fully accustomed to the biometric acquisition

device. We observed that by removing the signatures

acquired in the first acquisition session we obtain

better results in the evaluation phase.

4 EXPERIMENTAL RESULTS

In order to prove the effects of the post-processing

procedures presented in previous section we will

compute two performance coefficients: FAR (False

Accept Rate) and FRR (False Reject Rate) over the

collected signatures database. To compute these

performance coefficients we will use an online

authentication system. It is a distance-based system

ANAPPROACHTODATACOLLECTIONINANONLINESIGNATUREVERIFICATIONSYSTEM

253

containing a set of algorithms which compare two

strings of symbols extracted from the two signatures:

the input signature from the authentication phase and

the specimen signatures from the registration phase.

The pen signals are translated into these symbols

arrays called “invariants “which have attached a cost

and are used in several algorithms like the

„Levenshtein” algorithm (Andrei, Rusu, Diaconescu

and Dinescu, 2011).

We compute the performance coefficients FAR

and FRR for each user registered in the database,

using 5 signatures declared as being genuine, and we

consider them to be specimens. Then, for each user,

we send all the remaining genuine signatures to be

authenticated, computing the FRR coefficient. For

all subjects in the database, to compute the FAR, we

send for authentication all the signatures that were

captured as forgeries for a subject.

The performance coefficients obtained when

using the collected database before any of the post-

processing procedures presented above are: FRR

19.44% and FAR 2.29%. By removing the

acquisition errors and the operating errors, the FRR

decreased with 2.29 percentages, meaning that there

will be more with 2.29 percentages genuine

signatures accepted correctly. The FAR also

decreased with 0.45 percentages, meaning that more

forgery attempts will be rejected correctly. By

applying also z-score for data standardization over

the post-processed dataset, we obtain the same value

for FRR while the value for FAR decreased with

another 0.57 percentages. The above results prove

that, in order to build a strong data collection of

dynamic signatures, you need to make sure that the

signatures of a user are consistent; otherwise there

will be negative effects in the evaluation process.

We observed that, the additional method applied

for eliminating the set of signatures acquired in the

first session (Table 1) of the acquisition process,

further improves results: FRR decreased with 2.56

percentages. This is due to the fact that the signer

was not fully accustomed to the signing pen in the

first day of acquisition. After applying all post-

processing procedures mentioned above, over the

data collection, the performance coefficients

improved: FRR decreased with approximately 5

percentages, while FAR decreased with

approximately 1 percentage.

5 CONCLUSIONS

In this paper we presented our methods to achieve a

major step in evaluating performances of online

signature authentication systems: data collection.

We described our acquisition process, acquisition

device and acquisition application. We presented our

approach used to collect genuine signatures and

forgeries. We applied post-processing methods on

the collected database. Besides, using the raw

databases and also the post-processed database we

report the performance results of our dynamic

signature verification system. The obtained results

support the claim that signature data collecting

strategies have an impact over the performance

coefficients of an authentication system.

In future works, it is interesting to study how the

performance results will change if we use forgeries

collected from professional forgers which would be

motivated to break the system and also signatures

which are collected over a long period of time.

REFERENCES

Rusu, S., Dinescu, A., Diaconescu, S., 2011. Systems and

methods for assessing the authenticity of dynamic

handwritten signature. World Intellectual Property

Organization WO/2011/112113.

Bernadette, D., Chollet, G., Petrovska-Delacrétaz, D.,

2009. Guide to Biometric Reference Systems and

Performance Evaluation. In London: Springer-Verlag

London, pp. 125-166.

Yeung, D. Y., Chang, H., Xiong,Y., George, S., Kashi, R.,

Matsumoto, T., Rigoll, G., 2004. SVC2004: First

International Signature Verification Competition. In

Springer LNCS, Volume 3072, pp. 16-22.

Houmani, N., Mayoue, A., Garcia-Salicetti, S., Dorizzi,

B., Khalil, M. I., Moustafa, M. N., Abbas, H.,

Muramatsu, D., Yanıkoğlu, Berrin, Kholmatov,

Alisher, Martinez-Diaz, M., Fierrez, J., Ortega-Garcia,

J., Roure Alcobe, J., Fabregas, J., Faundez-Zanuy, M.,

Pascual-Gaspar, J. M., Carednoso-Payo, V.,

Vivaracho-Pascual, C., 2011. BioSecure signature

evaluation campaign (BSEC’2009): evaluating online

signature algorithms depending on the quality of

signatures. In: Pattern Recognition, Volume 45, Issue

3, pp. 993-1003.

Larsen, R. J., Marx, M. L., 2006. An Introduction to

Mathematical Statistics and Its Applications. 4th ed.

Prentice Hall, pp. 308-869.

Andrei, V., Rusu, M. S., Diaconescu S., Dinescu, A.,

2011. Securing On-line Payment using Dynamic

Signature Verification. LISS (3) 2011, pp. 58-65.

WEBIST2012-8thInternationalConferenceonWebInformationSystemsandTechnologies

254