Content uploaded by Abdul Quaiyum Ansari
Author content
All content in this area was uploaded by Abdul Quaiyum Ansari on Mar 23, 2015
Content may be subject to copyright.
2011 International Conference on Image Information Processing (ICIIP 2011)
Proceedings of the 2011 International Conference on Image Information Processing (ICIIP 2011)
978-1-61284-861-7/11/$26.00 ©2011 IEEE
Online Signature Verification using GA-SVM
Jaspreet Kour
EIE Department
Galgotias College of Engg & Tech.
Greater Noida, India
tojaspreet@gmail.com
M.Hanmandlu
EE Department
IIT
New Delhi, India
mhmandlu@gmail.com
A.Q Ansari
EE Department
Jamia Millia Islamia
New Delhi, India
aqansari62@gmail.com
Abstract— This paper presents an online signature verification
system based on Genetic Algorithm-Support Vector Machine
(GA-SVM). The raw information, obtained from SVC 2004
database, as time functions is used to derive 75 features. Six
different groups of features have been generated from 75 features
and their performance evaluated using SVM. A method is
proposed to reduce the computational complexity and the
amount of memory required without compromising on accuracy
using the sub set of features selected by Genetic Algorithm as the
input to SVM. The experimental results show that this method
provides good performance in terms of accuracy and memory
requirement.
Keywords- Biometrics, Online Signature, GA, SVM
I. INTRODUCTION
Biometrics is an emerging field of technology. It makes use
of unique but measurable physical, biological or behavioral
characteristics to perform the identity verification of a person.
Physiological biometrics is based on direct measurements
of physical parts (such as fingerprint, face, iris, hand geometry
etc.) of human body. Behavioral biometrics is based on the
measurement of an action performed (such as signature, gait,
speech, gesture etc.) by the individual [7]. The main advantage
that signature verification has over other forms of biometric
technologies is that signature is a well accepted biometric for
identity verification in our society for years. The long history
of trust of signature verification means that people are willing
to accept a signature based biometric authentication system.
But the drawback is that some people exhibit a lot of variability
between different manifestations of their signature. Also
signatures evolve with time and are influenced by physical and
emotional condition of the signatories.
Signature analysis is categorized into two modes: offline
and online. In the offline signature verification, signatures are
captured with a scanner or camera, saved and stored in
digitized form for further processing whereas the online
signature verification uses an electronic tablet and a stylus
connected to a computer to extract information about a
signature. It provides dynamic information like pressure,
velocity, acceleration, number of strokes etc.
Online signature verification is more robust, reliable and
accurate than offline signature verification, as its dynamic
properties make forging of an online signature extremely
difficult. Therefore online signature has become an attractive
biometric method in the protection of small personal devices
(PDA, small phone, laptop), in accessing sensitive data and
building, and for the authentication of internet transactions
such as e-commerce, e-banking, e-business, e-contract, etc.
In this paper, a memory efficient online signature
verification algorithm that uses the features chosen by Genetic
Algorithm (GA) as input vector to Support Vector Machine
(SVM) is presented. The online signature verification system
uses only discriminating features selected by GA, which are
unique and thus the memory requirement as well as the
computational complexity are very less.
The outline of this paper is as follows: Background for
SVM and GA is given in Section II. Section III describes the
proposed method to construct online signature verification
system. In Section IV, the results obtained from different
experiments are discussed. Section V gives conclusions and
future scope.
II. BACKGROUND
A. Support Vector Machine
Support Vector Machines (SVMs) create a margin between
two classes, thereby facilitating an effective classification.
Given the data containing two classes, SVM finds a hyper
plane, which maximizes the distance from either class to the
hyper plane and separates largest possible number of points
belonging to same class on the same side of the hyper plane.
Therefore misclassification error between the training set and
the test set is minimized.
Although in their basic form, SVMs learn linear threshold
functions, but in nonlinear case, they can be used to learn
polynomial classifiers, radial basis function (RBF) nets and
multilayer perceptron by applying appropriate kernel functions.
The dimensionality of the feature space has no direct relation to
the learnability of SVM. In other words, SVMs judge the
complexity of hypotheses underlying the classes according to
the margin that separates them. Thus, even with large number
of features, SVM pose no problem for their classification,
provided they are separable using the functions from the
hypothesis space [8].
2011 International Conference on Image Information Processing (ICIIP 2011)
Proceedings of the 2011 International Conference on Image Information Processing (ICIIP 2011)
Fea tu re
Et ti
Feature
selection
usin
g
GA
SVM
classifier
Decision
Tra in ing
classifier
B. Genetic Algorithm
Genetic algorithm (GA) is an adaptive and robust
computational procedure, modeled on the pattern of natural
genetic systems [5]. GA typically generates a constant
population of candidate solutions to the optimization problems.
Individuals are typically represented by n-bit binary vectors
and the search space corresponds to n-dimensional Boolean
space.
The goodness of each candidate solution can be evaluated
using fitness function. Evolutionary algorithms, using some
form of fitness-dependent probabilistic method, select
individual solutions from the current population to produce
solutions for the next generation. Genetic operators are applied
to selected individuals that constitute next generation. Mutation
and cross-over are two common operators used in GAs.
Mutation operator is applied on a single string to change its bits
randomly. Crossover, on other hand operates on two parent
strings to produce two offspring. The process of fitness
depends on selection of input features and application of
genetic operators to generate successive population of solutions
until a satisfactory solution is found.
III. PROPOSED METHOD
Figure 1 shows the block diagram of the proposed online
signature verification system. The whole process involves three
steps: Feature Extraction, Features Selection and Verification.
Figure 1. Block diagram of proposed online signature verification system
A. Data Base
The First International Signature Verification Competition
was held in 2004, to provide a landmark on signature
verification systems referred to as SVC2004 [10]. It contains
40 sets of signatures collected from different people and each
set consists of 20 genuine signatures and 20 skilled forgery
signatures. Signatures are acquired using WACOM Intuos
tablet dynamically, when the instrumented pen moves on the
tablet. Each signature is simply represented as a discrete time
dynamic sequence.
Figure 2. Azimuth angle and inclination angles of the pen with respect to the
plane of graphic card
The raw data of each signature consists of the following
information:
• Position on the x-axis
• Position on the y-axis
• Time stamp
• Button status
• Azimuth angle of pen with respect to the tablet (Fig 2)
• Altitude angle of pen with respect to the tablet (Fig 2)
• Pressure
A signature along with its raw dynamic features are shown
in Figure 3
Figure 3. Example of signature(top) and its associated dynamic
information(bottom)
2011 International Conference on Image Information Processing (ICIIP 2011)
Proceedings of the 2011 International Conference on Image Information Processing (ICIIP 2011)
B. Feature Extraction
Our aim is to find the most reliable and suitable features, as
the discriminative power of features play a major role in the
whole verification process. Using a set of raw dynamic data, 75
features have been derived as shown in Table 1. They represent
a collection of some of the features that have been used, studied
and reported in literature [2].
Six different groups of features generated include: shape,
dynamics, time, velocity, geometry, and miscellaneous
features [3].
TABLE I. LIST OF FEATURES
Shape related features
1 Height to width ratio ;
2 Length to width ratio ; Lw =
3 Horizontal mean-min diff; Ycn = Ym -min(Y(k))
4 Vertical mean-min diff ; Ycn = Ym -min(Y(k))
5 Left to right side ratio ; LR =
6 Upper to lower side ratio ; UL =
7-
10 Direction histogram S =
k=2.-----K ; l= 1,2,3,4
Dynamics related features
11 Total Signature time ; T=t
k
-t1
12 Pen down time ; Td =
13 Pen down time Ratio ; Tdr = T
d
/ T
14 RMS speed ;
15 RMS – min- difference of speed ; Vm-min = V-
min(V(K))
16 Max-RMS- difference of speed`; Vma
x
-
m
= max(V(K)) – V
17 Num of positive x velocity ; NV xp = card {Vx(K):
V
x
(K)>0}
18 Num of negative x velocity ; NV xn = card {Vx(K):
V
x
(K)<0}
19 Num of positive y velocity ; NV yp = card {Vy(K):
V
y
(K)>0}
20 Num of negative y velocity ; NV yn = card {Vy(K):
V
y
(K)<0}
Time related features
21 T(2nd pen down/Ts
22 (1st T(xma
x
))/Tw )
23 (1st T(xmin))/Tw )
24 (1st T(yma
x
))/Tw )
25 (1st T(ymin))/Tw )
26 (T 2nd pen up)/Tw
27 T (V
x
<0 | penup)/ Tw
28 T (V
x
>0 | penup)/ Tw
29 T (V
y
<0 | penup)/ Tw
30 T (V
y
>0 | penup)/ Tw
31 T (V
x
<0)/ Tw
32 T (V
x
>0)/ Tw
33 T (V
y
<0)/ Tw
34 T (V
y
>0)/ Tw
35 1st t (V
y,
Min))/ Tw
36 1st t (V
y,
Ma
x
))/ Tw
37 1st t (Vx
,
Mi
n
))/ Tw
38 1st t (Vx
,
Ma
x
))/ Tw
Velocity related features
39 Std. dev. Of V
y
40 Std. dev. Of V
y
41 (Average vel. )/ Vx, max
42 / V
y
, max
43 / Vmax
44 Velocity correlation Vx,y )/
Geometry related features
45 No. of penups ; N (pen-ups)
46 No. of sign changes of dx/dt ; N(sign changes Dx/dt)
47 No. of sign changes of dy/dt ; N(sign changes Dy/dt)
48 (Standard deviation of y) /
49 (Standard deviation of x) /
50 No. of local maxima of x ; N (local maxima in x)
51 No. of local maxima of y ; N (local maxima in y)
52 (xlast
p
e
n
-u
p
– xma
x
) /
53 (x1st
p
e
n
-dow
n
– xmin) /
54 (ylast
p
e
n
-u
p
– ymin) /
55 (x1st
p
e
n
-down – ymin) /
56 (ylast
p
e
n
-u
p
– yma
x
) /
57 (x1st
p
e
n
-down – yma
x
) /
58 (xlast
p
e
n
-u
p
– xmin) /
59 (x1st
p
e
n
-dow
n
– xma
x
) /
60 (Tw)/ (xma
x
– x min )
61 (Tw)/ (yma
x
– y min )
62 Std (x)/ (xma
x
– x min )
63 Std (y) / (yma
x
– y min )
Miscellaneous features
64 Average azimuth
65 Maximu m azimuth
66 Minimum azimuth
67 RMS azimuth
68 Average Altitude
69 Maximum Altitude
70 Minimum Altitude
71 RMS Altitude
72 Average Pressure
73 Maximum Pressure
74 Minimum Pressure
75 RMS Pressure
2011 International Conference on Image Information Processing (ICIIP 2011)
Proceedings of the 2011 International Conference on Image Information Processing (ICIIP 2011)
C. Feature Selection using GA
The task of selecting the most discriminative features for a
particular classification problem in a high dimensional space is
known as feature selection. Given a d-dimensional problem,
there exist, 2d possible subset of features. Even for reasonable
values of d, an exhaustive search is usually not feasible. Many
different algorithms have been presented in the literature to
cope up with problem of feature selection, GA being the most
popular out of them. The optimization criterion is taken as the
classification accuracy of SVM for convergence of GA.
D. Verification
To evaluate effectiveness of the proposed method, we have
considered two cases. First, different groups of features
separately and then all features together are input to SVM.
Second, the input vectors for SVM consist of feature subset
chosen by GA.
IV. EXPERIMENTAL RESULTS
Experiment 1:
Each group of features has been considered separately for
classification using SVM. The classification accuracy due to
shape and time features is higher than that of other groups.
Experiment 2:
In this step, two groups are considered together. A
combination of dynamic and time related features shows better
accuracy than that obtained from other combinations.
Experiment 3:
To classification accuracy is improved further by
combining more number of groups and by considering all the
features, as shown in Table II.
TABLE II. RESULTS OF VERIFIC ATION USING SVM
Feature Vector Accuracy
%
Shape Dynamics Time Velocity Geometry Misc.
√ 70.1
√ 70.4
√ √ 74.5
√ √ 78.0
√ √ 76.0
√ √ √ 78.6
√ √ √ √ 75.9
√ √ √ √ √ √ 83.4
Experiment 4:
In this experiment, GA is used for feature selection. A
reduced set of 41 features selected using GA (shown as grey in
Table I) produces 83.6% accuracy, thereby reducing the
computational complexity by 53%
V. CONCLUSION
This paper proposes a memory efficient method of online
signature verification by integrating GA into SVM.
Comparative experiments show that, shape, dynamics and time
features have better performance than that of other group of
features. Also a method to reduce the computational
complexity and memory requirement has been suggested using
genetic algorithm in the online signature authentication.
Future work aims at exploring new features and testing
them on other classifiers using a local database
REFERENCES
[1] Anil K. Jain, Friederike D. Griess, Scott D. Connell, “On-line signature
verification” Pattern Recognition ,2002, pp. 2963 – 2972.
[2] H .Lei ,V.Govinaraju, “A study on the consistency of features for online
signature verification” SSPR/SPR,2004,p.444
[3] J. Galbally, J. Fierrez, M.R. Freire, and J. Ortega-Garcia, “Feature
Selection Based on Genetic Algorithms for On-Line Signature
Verification”, In Proc. of IEEE Workshop on Automatic Identification
Advanced Technologies, AutoID, pages 198–203, 2007a.
[4] Jonas Richardi,Hamed Ketabdar,Andrzej Drygajlo, “Local and Global
feature Selection for On-line Signature Verification”,Proceedings of
Eight International Conference on Document Analysis And Recognition
(ICDAR’05),2005, pp 625-629
[5] K.Lee,Y Chung and H.Byun, “Face recognition using support vector
machine with the feature set extracted by genetic algorithm”
Proceedings of Audio and Video Based biometric person
authentication,2001,pp 32-37
[6] L. Nanni , A. Nanni, “ A novel local on-line signature verification
system”,2008, Pattern Recognition Letters, Volume 29 Issue 5
[7] Mohammad A. U. Khan, Muhammad Kh alid Khan Niazi, and
Muhammad Aurangzeb Khan,“Velocity-Image Model for Online
Signature Verification”, 2006, IEEE Transactions on Image Processing,
vol. 15, no. 11, pp. 3540-3549
[8] V.Vapnik,Statistical Learning Theory,John Wiley & Sons,New
York,1998
[9] V. S. Nalwa, “Automatic On-Line Signature Verification,” Proc. IEEE,
85 (2) 1997, pp. 215-239.
[10] http://www.cse.ust.hk/svc2004/download.html
[11] Chih-Chung Chang and Chih-Jen Lin, LIBSVM: a library for support
vector machines,2001. http://www.csie.ntu.edu.tw/~cjlin/libsvm