=Paper=
{{Paper
|id=None
|storemode=property
|title=Classification of Features into Strong and Weak Features for an Intelligent Online Signature Verification System
|pdfUrl=https://ceur-ws.org/Vol-768/Paper3.pdf
|volume=Vol-768
|dblpUrl=https://dblp.org/rec/conf/icdar/TariqSH11
}}
==Classification of Features into Strong and Weak Features for an Intelligent Online Signature Verification System==
https://ceur-ws.org/Vol-768/Paper3.pdf
Proceedings of the 1st International Workshop on Automated Forensic Handwriting Analysis (AFHA) 2011
Classification of Features into Strong and Weak Features for an Intelligent Online
Signature Verification System
Saad Tariq, Saqib Sarwar & Waqar Hussain
Department of Electrical Engineering
Air University
Islamabad, Pakistan
tariq.saad@live.com
Abstract—This paper presents an efficient algorithm for the additional features such as pen pressure, pen speed and pen
classification of features into strong and weak features for tilt angle have made the process of forging online signatures
every distinct subject to create an intelligent online signature more difficult. Equal error rate of available online
verification system. Whereas Euclidean distance classifier is signature verification systems lies between 1 to 10%.
used for validation processes and low error rates obtained
Still a lot of work is needed to be done to reduce Equal
illustrate the feasibility of the algorithm for an online signature
verification system. error rate (EER) to make online signature verification the
most secure way of personal identification.
Keywords-Signature Biometrics; intellegient signature
verification ; online signature verification; classification of
features; strong features; weak features; dynamic signature
verification; euclidean distance classifer II. FEATURE EXTRACTION
Feature extraction phase is one of the crucial phases of
I. INTRODUCTION an on-line signature verification system. The
Today, with the astonishing growth of the Internet and discriminative power of the features and their flexibility
Intranet, E-commerce and E-finance become the hottest to the variation within the reference signatures of a
topics on this planet. Doing business through the public writer, play one of the major roles in the whole
network makes personal identification data security more verification process. While features related to the
and more crucial as well. How to protect the private signature shape are not dependent on the data acquisition
identification from being pirated is the key issue that the device, presence of dynamic features, such as pressure at
Internet and intranet clients would be concerned with the pen-tip or pen-tilt, depends on the hardware used.
before such E-business could be widely accepted since Features may be classified as global or local, where
authentication has become an essential part of highly global features identify signature’s properties as a whole
computerized services and/or security-sensitive and local ones correspond to properties specific to a
installations in modern society. sampling point. For example, signature bounding box,
Signature verification fulfills all the above described average signing speed, trajectory length or are global
circumstances and can play a vital role in protection and features, and Local features include curvature change
personal identification as it is a popular means of between consecutive points on the signature trajectory or
endorsement historically. Although such signatures are distance are local features. Features may also be
never the same for the same person at diverse times, classified as temporal (related to the dynamics) and
there appears to be no practical problem for human spatial (related to the shape).
beings to discriminate visually the real signature from the These features can be referred as human traits, as they
forged one. It will be extremely useful when an can vary from person to person and can be classified as
electronic device can display at least the same virtuosity. strong or weak for every distinct individual. If we make a
Signature verification systems are usually built following list of these features, more than 100 features are present
either on-line or off-line approaches, depending on the kind and even new features can be derived depending on their
of data and application involved. On-line systems generally discriminative power.
present a better performance than the off-line system but
require the necessary presence of the author during both the
acquisition of the reference data and the verification process
limiting its use. In online signature verification systems,
11
Proceedings of the 1st International Workshop on Automated Forensic Handwriting Analysis (AFHA) 2011
III. DATABASE & COMPILATION sample point of y-coordinate, zeros of x-velocity,
standard deviation of x-coordinates, standard deviation
A. System of y-coordinates, total number of samples, time taken,
For the purpose of signature verification we made an length, zero crossings of x-velocity, zero crossings of y-
experimental setup in which a person is enrolled in the velocity, zero crossings of x-acceleration, zero crossings
database by taking some of his/her signatures and a template of y-acceleration, zeros in x-acceleration, zeros in y-
is created and stored against the name and ID of the specified acceleration.
person. A new signature from that person can then be A pressure sensitive tablet was used that records pressure at
checked against the enrolled template to validate the person. every sample taken, providing with a very strong local
Furthermore we will discuss about the technique used in our feature of pressure.
system, database and how we optimized features as strong C. Optimization & Experimental Setup
and weak features.
Here is an important discussion that how we opted only 9
B. Database Completion features out of those 26 features for our system. As we know
A comprehensive database was created by obtaining the that a large number of features have been proposed by
signatures from the students. Signatures were gathered from researchers for online signature verification [2], [3], [4].
a total of hundred subjects with ten signatures from each However, a little work has been done in measuring the
subject. So a total of thousand signatures were collected to consistency and discriminative power of these features [5],
create the original signature database. WACOM INtuous4 [6]. On the basis of consistency and discriminative power
tablet with a sampling rate of 200 samples per seconds was features can be divided into strong and weak features, where
used for this purpose. presence of the strong features decreases the FRR while on
To form the forgeries database we performed a total 10 the other hand presence of some weak features also
forgeries per person, among which were five zero-effort decreases FRR but increases FAR. Thus there is a need to
forgeries and five skilled forgeries. The forgeries that are select the best features set.
performed by first training the counterfeiter to copy the The approach we used for classification of strong and
precise dynamics of the original signer are skilled forgeries. weak features is by using difference between mean to
A forger is trained by showing him plots of the original standard deviation ratio of each feature from the feature
signature being performed or by training the original signer vector and from the forgeries features vector set. Thus the
himself. mean/standard-deviation difference of each feature from the
template of 100 subjects was taken. The standard deviation
of a feature shows how large a deviation from the enrolled
IV. SIGNATURE VERIFICATION TECHNIQUE template can be tolerated (i.e. large deviated signature could
In the first phase, a signature verification technique was be classified as true for large standard deviation).
successfully put into operation for the classification of
original and forged signatures using Euclidean Classifier.
The technique is previously implemented by H. Dullink, B.
Van Daalen, J. Nijhuis, L. Spaanenburg, and H. Zuuidhof
[1].
A. No Pre-Processing
The technique we implemented did not use any
preprocessing because the tablet used had a sampling rate of In (1), Mo/STDo is the mean/standard-deviation ratio of
200 samples per second. Therefore it was not essential to the feature of original signatures and Mf/STDf is the
smooth or normalize the signature datasets, which were mean/standard-deviation ratio of the feature of forgery
required if we had used the signatures collected from a signature. The features with large value of mean/standard-
tablets with low resolution. Re-sampling and resizing was deviation difference as compared to others were taken as
also skipped considering the fact that valuable data is lost strong features and others as weak features eliminating
while pre-processing the data. which results in considerable good results.
A number of original signature’s features have a large
B. Feature Extraction mean/standard-deviation ratio and of course it will decrease
Among the list of features that can be extracted a total of FRR but contrary to it forgery signature’s features having a
26 features were extracted. The features extracted were large mean/standard-deviation will decrease FAR. So
standard deviation of x-acceleration, standard deviation therefore to obtain best results we took the difference
of y-acceleration, average pressure, standard deviation of between the original signature and forgery signature.
x-velocity, standard deviation of y-velocity, number of D. Optimization Results
pen-up samples, pen down time/total time taken,
standard deviation of y / change in y, pen down time, As computed using (1) nearly 14 features have greater C
RMS velocity / maximum velocity, average jerk, jerk than other 16 features. As researchers have discussed earlier
RMS, maximum sample point x-coordinate, maximum that too many features may decrease FRR but increase FAR
12
Proceedings of the 1st International Workshop on Automated Forensic Handwriting Analysis (AFHA) 2011
[7] therefore we have to choose between the best of them. Features such as total time, pen-down time and total
The 14 features with greater C are standard deviation in y- samples are all time dependent features so therefore for a
velocity, total samples, number of zeros in y- versatile verification engine we opted total time to be the
acceleration, number of zeros in x-acceleration, zero best among them. Moreover standard deviation of y-velocity
crossings in x-acceleration, zero crossings in y- is another feature having a greater result but on the standard
acceleration, zero crossings in x-velocity, zero crossings deviation of x-velocity has a very small difference, therefore
in y-velocity, length, average pressure, total time, this ambiguous result made us step down with these features
number of zeros in y-velocity, number of zeros in x- too.
velocity and pen-down time.
TABLE I. CALCULATIONS OF EQUATION (1)
V. INTELLIGENT ONLINE SIGNATUARE VERIFICATION
Feature Mo/STDo Mf/STDf C The experimental setup and optimization proposed above
gave very good results but still as we have discussed earlier
Std Dev y/∆y -4.8766 -1.9296 2.9
that signature and its features are personal traits and they
T(pen-down)/T(total) 23.3710 17.6752 5.4 may vary person to person. Thus to make this system
N (pen-ups) 3.8719 0.8551 2.95 efficient and intelligent we made it route person to person.
Standard Deviation vy 25.2692 13.9054 12.3
As we had a list of 9 most efficient features, we decided to
choose 5 out it but based on subject itself. These 5 features
Standard Deviation vx 2.8116 1.9122 0.9 may vary person to person. While recording a template from
N(vy=0) 5.8355 1.2074 4.6 a subject all these features were stored in the template but at
Average v/v( max.) 5.7595 3.3267 2.45 the time of verification we proposed a system in which only
5 features were compared against its template based on the
(x1-xmin)/average x 4.5197 2.8109 1.7
following results.
Total Samples 15.9329 2.1116 13.79 X = C/ Vx - STDf (2)
(x1-xmax)/average x -7.4158 -8.2712 0.8
N(max. y) 15.9590 17.7610 1.81 Where C is the difference between the mean/standard-
deviation ratio of the feature of original signatures and the
Standard Deviation of ay 3.1448 4.0654 0.92 mean/standard-deviation ratio of the feature of forgery
Standard Deviation of ax 1.6747 2.1500 0.48 signature from (1) which is already calculated and Vx is
Number of zeros in ay 7.7817 1.0288 6.78 current value of the sample and STDf is the standard
deviation of the forgery signature already stored. So among
Number of zeros in ax 8.5880 1.2653 7.30
the 9 features, only 5 features are opted which have a greater
Zero cross. X- 9.0654 1.3230 7.68 value of X from (2).
acceleration
Zero cross. Y- 9.6669 1.2263 8.44 A. Comparison
acceleration For comparison we need a reference. So for the
Zero cross. X-velocity 12.8354 1.5204 11.31 enrollment process we selected 5 original signatures from
each signature extracted the 9 features described above to
Zero cross. Y-velocity 13.5760 1.2228 12.35
create a reference template. The template contains the mean,
Length 7.5981 1.7094 5.89 standard deviations and their difference stored in 3 vectors R,
rms jerk 2.6554 1.9491 0.71 S and C respectively. If we want to compare a signature
average jerk 2.7470 2.4410 0.26
(original or forged) with the template we will first compute
the feature vector of that signature and corresponding vector
N(max. x) 15.4440 13.6379 1.81 X using (2). Then the greater 5 features depending on the
Average Pressure 12.1289 2.2516 9.87 value of X will be stored in a vector T. To compare the
Total Time 15.9329 2.1116 13.82 signature we will simply opt out those 5 features from R and
S and a distance vector D will be computed using Euclidean
Number of zeros in vy 8.8355 1.2074 7.63 classifier.
Number of zeros in vx 8.5746 1.2781 7.30 D=R–T (3)
(y1-ymax)/average y -3.9525 -3.1871 0.77 Then the distance vector V will be normalized by dividing
each value by the corresponding standard deviation in the
(x1-xmin)/average x 8.0218 5.4126 2.62
vector S to obtain a vector Z whose mean is then computed
Pen-down Time 29.7766 2.9390 26.87 and finally the computed norm is compared to a pre-defined
Highlighted features are with greater results threshold.
13
Proceedings of the 1st International Workshop on Automated Forensic Handwriting Analysis (AFHA) 2011
this system to make it more efficient by using other
classifiers and updating signature over time with tablets
B. Results
with better sampling rates.
Results for FRR, FAR of the template of 5 signatures of 100
subjects were computed with threshold from 4 to 9 for this
intelligent online signature verification system and best REFERENCES
results were obtained.
[1] H. Dullink, B. van Daalen, J. Nijhuis, L. Spaanenburg and H.
Zuidhof, Implementing a DSP Kernel for Online Dynamic
TABLE II. CALCULATIONS OF FFR AND FAR (1) Handwritten Signature Verification Using the TMS320 DSP Family,
EFRIE, France December 1995, SPRA 304.J. Clerk Maxwell, A
Treatise on Electricity and Magnetism, 3rd ed., vol. 2. Oxford:
Threshold FRR FAR Clarendon, 1892, pp.68–73.
[2] Charles E. Pippin, Dynamic Signature Verification using Local and
4 11.57% 0.72% Global Features, Georgia Institute of Technology, July 2004.
[3] T. S. Tolba, A Virtual-Reality-Based System for Dynamic Signature
5 11.20% 3.92% Verification, Digital Signal Processing Vol. 9, pp. 241-266, 1999.
(article available online at http://www.idealibrary.com)
6 4.53% 8.02% [4] V. S. Nalwa, Automatic On-Line Signature Verification, Proceedings
of IEEE, vol. 85, pp. 215-239, 1997.M. Young, The Technical
7 2.06% 13.62% Writer’s Handbook. Mill Valley, CA: University Science, 1989.
[5] Hao Feng and Chan Choong Wah, Online signature verification using
8 1.13% 19.89% a new extreme points warping technique, PRL(24), No. 16, pp. 2943-
2951, December 2003.
[6] H. Goto, Y. Hasegawa, and M. Tanaka, “Efficient Scheduling
9 0.66% 27.02% Focusing on the Duality of MPL Representatives,” Proc. IEEE Symp.
Computational Intelligence in Scheduling (SCIS 07), IEEE Press,
Dec. 2007, pp. 57-64, doi:10.1109/SCIS.2007.357670.
[7] G. Lorette and R. Plamondon, Dynamic approaches to handwritten
Results obtained from our implementation are very better signature verification, Computer Processing of handwriting, World
Scientific, 1990, 21-47.
than a number of techniques implemented because we used
very strong features and an intelligent system to classify
them person to person. Anyways more work can be done on
14
Proceedings of the 1st International Workshop on Automated Forensic Handwriting Analysis (AFHA) 2011
15