2013
IEEE Conference on Open Systems (ICOS), December 2

4, 2013,
Sarawak, Malaysia
Online Signature Verification Using Neural Network and Pearson Correlation Features ' ' ' Vahab Iranmanesh* , Sharifah Mumtazah Syed Ahmad** , Wan Azizun Wan Adnan , Fahad Layth Malallah , 2
"
Salman Yussof 2
Faculty of Engineering, Universiti of Putra (UPM), Selangor, Malaysia
College of Information Technology, Universiti Tenaga Nasional (UNITEN), Selangor, Malaysia *
[email protected], **
[email protected]
AbstractIn this paper, we proposed a method for
known. On the other hand, forging the signature is complex in the online approach because of some features that combine psychology, physics, and neuroscience [4].
feature extraction in online signature verification. We first used signature coordinate points and pen pressure of all signatures, which are available in the SIGMA database. Then, Pearson correlation coefficients were selected for feature extraction. The obtained features
Nevertheless, in the enrollment step, each person will submit his signatures as a template for verification in the future. Nobody could produce signatures that all are identical because of intrauser variability. Intrauser variability measures the difference between the signatures of an individual. This difference happened due to environmental causes, personal health and emotional while drawing [5]. The main challenge is to model the signature as it contains intrauser variability so that it would have less effect on the system accuracy [69].
were used in backpropagation neural network for verification. The results indicate an accuracy of 82.42%. KeywordsOnline Signature Verification; Neural Network; Pattern Recognition; Feature Extraction; Pearson Correlation Coefficients
I.
INTRODUCTION
Handwritten signatures are commonly used as a fonn of biometrics to approbate the contents of a document or to authenticate online business, especially financial, transactions [1, 2]. Signature verification is usually done by visual inspection. A person compares the appearance of two signatures and accepts the given signature if it is sufficiently similar to the stored signature, for example, on a credit card. In most situations where a signature is needed, no verification takes place at all because of the amount of time and effort that would be needed to manually verify signatures. Automating the signature verification process will improve the current situation and eliminate fraud.
To overcome the problem above, the feature extraction techniques in data mining are used to extract features representing the signature to collect a unique template in the database. Then, the collected template is stored in the knowledgebased database. In the verification process, the individual's new signature template is compared with the previous template in the knowledgebased database for authorization [lO]. Numerous studies have been done on online and offline signature verification, and a lot of techniques have been proposed and tested [1115]. The hidden Markov model (HMM) [1617], dynamic time warping (DTW) [1819], artificial neural network (ANN) [2023], and support vector machine (SVM) [2425] are among the main techniques used to verify an individual's signature.
The signature verification system is divided into two approaches: online and offline [3]. In the off line approach, the image of the signature is captured by the scanner. In the online method, several signature features such as velocity, speed, pressure, and the coordinates are recorded using a touch screen. The signature made offline can be easily mimicked as long as the shape of the signature is
18
2013
IEEE Conference on Open Systems (ICOS), December 2
•
In this study, the proposed featureextraction method on an individual's online signature is applied based on Pearson correlation coefficients. Later, the derived features are used in feedforward neural network to distinguish between the forged and genuine signature. Section 2 describes the material and database. Section 3 introduces the implementation containing Pearson correlation coefficients and a neural network. Section 4 provides the experimental results. Section 5 concludes the study. II.
•
Corr

Sarawak, Malaysia
4, 2013,
x,
Y: coordinates of each sample point. P : pressure at each coordinate.
(X, Y)
=
Lr1(xix)(yiy) 2 2 Lr=1(xiX) Lr=1(yi_y)
J
(1)
DATABASE AND EXPERIMENTAL PROTOCOL
In this study, the SIGMA signature database, [26] which contains more than 200 Malaysian users, was used. A random subset with 200 users involving 20 genuine, 10 skillforged, and 10 non skillforged signatures for each user was selected. In the training phase, 10 genuine, 5 skillforged, and 5 non skill forged were selected to represent each user's signature sample in the training phase. Similarly, the same nwnber of samples was used during the testing phase. A genuine signature was labeled 1, and a forged signature was labeled O. III.
Figure
IMPLEMENTAnON
B.
1. Feature extraction flowchart.
Multilayer Perceptron
A multilayer perceptron (MLP) is a fully connected network with some layers between the input and output layers. This model is also called a feedforward neural network, which maps inputs to outputs. Classification, clustering, regression, and prediction are four applications of MLP. In Fig.2, the architecture of a twolayer perceptron is illustrated. Each layer contains one or more neurons to receive the inputs. Then the activation function is applied on neurons as a threshold to generate the output. The sigmoid activation function is commonly used in the range between  1 and +1.
A. Pearson Correlation Coefficient Feature Extraction One of the important steps in pattern recognition and classification is feature extraction. In this process, the raw data will be transformed and reduced into a vector (pattern) with the purpose of reducing the computing time and improving efficiency. The accuracy of verification relies on the extracted features. The Pearson correlation coefficient is a measure to investigate the relationship between two variables. The achieved correlation coefficients (r) are used to represent the correlation between variables. It's range could be between  1 and +1, where a negative value means an inverse relation, a positive value denotes a correlated relation, and 0 represents no relation.Fig.1 shows the flow chart of feature extraction in order to achieve the sufficient features for verification.
Learning in MLP is an iterative procedure to set network parameters such as weight. Back propagation is a common supervised learning technique in MLP to train the network. In this technique, the output error goes back to the input layer to calculate the errors in hidden layers. The gradient method is used to optimize the cost function to achieve minimum error, which should be close to O. The error obtained from learning is called the mean square error (MSE), which is computed from the target and output values. MSE is calculated using equation (2). The weights will be set in each iteration. This adjustment will stop when the cost function value reaches the minimum value.
In this study, the similarity of signature variables (x, y, p) is measured using Pearson correlation coefficient in equation (1) to distinguish each person's signature. The resulting nine coefficients showing the relationships between variables are selected as features representing the signature.
19
2013
IEEE Conference on Open Systems (ICOS), December 2
Inl1ut L'lyer
Hidden L"lyer
Outl)lIt L'1)'er

4, 2013,
FAR(%)
=
100
*
[FAj2000]
(3)
FRR(%)
=
100
*
[FRj2000]
(4)
Accuracy(%) Inputs
IV.
2.
Multilayer neural network architecture.
EXPERIMENTAL RESULTS

[FRR
+
FARj2]
(5)
TABLE 1.
FAR (%)
21.35
13.81
In addition, signature normalization is not needed because these nine features are extracted from whatever length of signature sample. Therefore, we have used these features to represent the signature. However, the online signature samples in SIGMA database are different from those of the other approaches. Experimental results indicate the number of false acceptance is more than false rejection.
Result
FRR (%)
DISCUSSION AND CONCLUSION
Subsequently, the MLP recognition is trained and tested on SIGMA database to classify the signature as being forged or genuine. From the output of feature extraction, nine features were used to represent the signature. According to the verification result, we could see two phenomenon. First, the only small number of features is needed to verify online signature compared with the other verification approaches, which are implemented by using more features than our technique. Second, dimension reduction is achieved by using only nine features that could reduce training and testing time to achieve a relatively accuracy rate.
As we can see from Table 1 and after the analysis, the model achieved an FAR of 21.25% and an FRR of 13.81%. The number of misclassified signature samples in the testing phase was 427 for FAR and 276 for FRR out of 4,000 signatures. Equations (3) and (4) were used to calculate the FAR and FRR, where the FA is the total number of false acceptances and FR is the total number of false rejections. The accuracy rate is calculated as 82.47% using equation (5).
82.42
100
While signature verification is one of the biometric techniques for person identification, it could also be used in document verification in the financial sector, including banking. The main weakness of the signature verification is an unstable pattern. In this work, we have presented Pearson correlation coefficients as a feature extraction on SIGMA database, which contains 8,000 online signatures taken from 200 users.
In order to evaluate the perfonnance of obtaining features, a twolayer MLP classifier was proposed. The LevenbergMarquardt algorithm was used for training the neural network. The number of neurons in the input layer was equal to the number of features, which was nine. The neuron in the output layer was selected to hold a value of either 1 or O. The number chosen for hidden layers was 20, and the sigmoid activation function was selected. To evaluate the performance of the classifier, the false acceptance rate (FAR) for forged signatures and false rejection rate (FRR) for genuine signatures were applied. The MSE as a cost function for error calculation was considered to apply as well.
Accuracy (%)
=
Outputs
V.
Figure
Sarawak, Malaysia
Number of
Number of
FA
FR
427
276
Further study will concentrate more on accuracy through more features that could be retrieved from available signature features in the database. In addition, a new training algorithm, such as the genetic algorithm for adjusting weights, could be established. ACKNOWLEDGMENT We would like to acknowledge Malaysian Ministry of Higher Education for the provision of
20
2013
IEEE Conference on Open Systems (ICOS), December 2

CybernetiCS,
Exploratory Research Grant Schemes through which this research was made possible.
[14]
Sarawak, Malaysia
4, 2013,
Part
C:
Transactions on, vol.
Applications
and
38,pp. 609635,2008.
Reviews,
iEEE
T. Wilkin and Y. Ooi Shih, "State of the art: Signature verification system," in information Assurance and Security (iAS), 2011 7th international Conference on,
REFERENCES
[15]
115.
2011, pp. 110
Y. M. AIOmari, S. N. H. S. Abdullah, and K. Omar, "State oftheart in offline signature verification system," in Pattern
[I]
Analysis S. Nanavati, M. Thieme,
and R. Nanavati, "In Biometrics:
[16]
Identity verification in a networked world". New York: John
[2]
Wiley & Sons,
2002.
"Online Handwritten Signature Verification for Electronic
[3]
Conference
2001.
S.
A.
Rashidi,
[17]
on Web
Intelligence:
Research
and
and
F.
Towhidkhah,
"Feature
[18]
19,pp. 18101819,121/2012.
Scientiairanica, vol.
[6]
17881792.
and
Z.
Lin,
"Signature
Encyclopedia of BiometriCS, pp. A.
Sample
12051210,2009.
2009, pp.
[19]
Synthesis,"
K. Jain, K. Nandakumar, and A. Nagar, "Biometric
[20]
2008,pp. 117,2008.
2nd international Congress on,
2009,pp. 15.
S. M. S. Ahmad, A. Shakil, and M. A. M. Balbed, "Study on and
online
Symposium on,
signature
verification
systems,"
in
2008. iTSim 2008. international
2008,pp. 16.
G. Taherzadeh, R. Karimi, A. Ghobadi, and H. M. Beh,
international Conference on,
20II,pp. 772777.
E. Zhan, J. Guo, 1. Zheng, C. Ma, and L. Wang, "Online Back Propagation Neural Network," in intelligent Ubiquitous
Neri, "Cancelable Templates for SequenceBased Biometrics
Computing and Education, 2009 international Symposium
with Application to Online Signature Recognition," Systems,
on,
Transactions on, vol.
[21]
40,pp. 525538,2010.
2009,pp. 202205.
S. K. Ahmed, A. K. Ramasamy, A. Khairuddin, and J. Omar, "Automatic online signature verification: A prototype using
E. ArgonesRua, E. Maiorana, J. L. Alba Castro, and P.
neural networks," in TENCON 2009  2009 iEEE Region 10
Campisi, "Biometric Template Protection Using Universal
Conference,
vol.
[22]
verification
E. Grosso, L. Pulina, and M. Tistarelli, "Modeling biometric update
with
Ant
Colony
Optimization,"
on,
2009, pp. 14.
D. R. Shashikumar, K. B. Raja, R. K. Chhotaray, and S. Pattanaik, "Biometric security system based on signature
7, pp. 269282,2012.
using
neural
in
[23]
2012,pp. 506511.
international Conference on,
international Symposium on,
intelligent
Robotics
(lCPAiR),
international Conference on,
2011,pp. 5964.
D.
Pirlo,
and
G.
"Automatic
Verification:
The State of the Art," Systems,
Cybernetics,
Part
C:
Transactions on, vol.
Applications
and
38,pp. 609635,2008.
[24]
2011
verification,"
in
Uncertainty
Knowledge Engineering (URKE), Conference on, D.
lmpedovo
2012,pp. 5457. and
2010,pp. 16.
G.
Pirlo,
Systems
and
Applications
(iNiSTA),
2011,pp. 3438.
2011
S. Fauziyah, O. Azlina, B. Mardiana, A. M. Zahariah, and H. Haroon, "Signature verification system using Support Vector
Signature Man,
Reviews,
Machine," in Mechatronics and its Applications, 2009. iSMA
and
[25]
iEEE
'09. 6th international Symposium on,
2009,pp. 14.
M. Saeidi, R. Amirfattahi, A. Amini, and M. Sajadi, "Online signature verification using combination of two classifiers,"
A. Sanmorino and S. Yazid, "A survey for handwritten signature
Computational
and recognition: Neural network approach," in innovations in
oftheart in offline signature verification system," in Pattern
lmpedovo
in
S. M. Odeh and M. Khalil, "Offline signature verification intelligent
and
networks,"
intelligence and Computing Research (iCCiC), 2010 iEEE
Y. M. AIOmari, S. N. H. S. Abdullah, and K. Omar, "State Analysis
[13]
2008,pp. 173179.
and
iEEEliFlP
Handwritten Signature Verification Based on Two Levels
E. Maiorana, P. Campisi, 1. Fierrez, J. OrtegaGarcia, and A.
template
[12]
'08.
"Evaluation of online signature verification features," in
Biometrics (ICB), 2012 5th iAPR international Conference
[II]
EUC
ModelBased
Embedded
Advanced Communication Technology (iCACT), 2011 13th
information Forensics and Security, iEEE Transactions on,
[10]
2008.
in
2011
F.j. Luan, L. Lin, and H. Cheng, "The Algorithm of OnLine
information Technology,
Wan,
"Markov
Verification,"
international Conference on,
based on correlation image," in Machine Learning and
Background Models: An Application to Online Signature,"
[9]
Trevathan,
Computing,
(iCPAiR),
the effect of number of training samples on HMM based
Man and Cybernetics, Part A: Systems and Humans, iEEE
[8]
J.
offline
template security, " EURASiP J. Adv. Signal Process, vol.
[7]
and
Signature
D. HaoRan and W. YiDing, "Online signature verification
L.
Robotics
2011,pp. 5964.
Model, " in image and Signal Processing, 2009.ClSP '09.
Fallah,
Cybernetics, 2009 international Conference on,
[5]
intelligent
Handwritten Signature Verification Based on Hide Markov
extraction based DCT on dynamic signature verification, "
[4]
McCabe
Ubiquitous
Commerce over the Internet, "Proceedings of the First Asia Development,
A.
Handwritten
W. S. Wijesoma, K. W. Yue, K. L. Chien, and T. K. Chow,
Pacific
and
international Conference on,
Reasoning
in Machine Vision and image Processing (MViP), 2010 6th
and
[26]
20i2 2nd international
iranian,
2010,pp. 14.
S. M. S. Ahmad, A. Shakil, A. R. Ahmad, M. Agil, M. Balbed, and R. M. Anwar, "SIGMA  A Malaysian signatures
"Automatic
Signature
database," in Computer Systems and Applications,
Verification: The State of the Art, " Systems, Man, and
AiCCSA 2008. iEEEIACS international Conference on,
pp. 919920.
21
2008.
2008,