Classical user authentication systems have been based in something that you
have (e.g., a key, an identification card, etc.) and/or something that you know
(like a password, or a PIN). With biometrics, a new user authentication paradigm
is added: something that you are (e.g., fingerprint, iris or face) or something
that you do or produce (e.g., handwritten signature or voice). The convenience
for paper and pen in the electronic era is the reason why people still use handwriting
as a mean to convey, retain and facilitate communication. Together with this
kind of information, handwriting is also a skill that individualizes people.
Moreover, devices like PDAs, pocket PCS, tablet PCS, or 3G mobile phones might
offer handwriting capabilities, because handwriting is considered as being more
natural for humans and equally important to the possibility of size reduction
by eliminating the keyboard. From this point of view, signature is a social
and legal acceptable biometrics personal authentication method (Wu
et al., 2005). A signature is a special case of handwriting, which
includes special characters and flourishes. Many signatures can be unreadable.
They are a kind of artistic handwriting (Cemil, 2005).
Handwritten signature verification is the process of confirming the identity
of a user based on the handwritten signature of the user as a form of behavioral
biometrics (Nalwa, 1997; Jain et
al., 1999, 2002).
Biometrics technology has a great potential for automatic personal verification
and differently from other means for personal identification and verification
(Pirlo, 1994). Of the various biometrics, signature-based
verification has the advantage that signature analysis requires no invasive
measurement and is widely accepted since signature has long been established
as the most diffuse mean for personal verification in our daily life, including
commerce applications, banking transactions, automatic fund transfers and so
on (Ammar and Aqel, 2002; Plamondon
and Srihari, 2000; Plamondon, 1994a).
There are two categories of verification systems are usually distinguished: static or off-line system for which the signature is captured once the writing processing is over and thus only a static image is available and dynamic or online system for which the signature signal is captured during the writing process, thus making the dynamic information available. This study deals with the online signature verification.
In online signature verification system, users write their signature in digitizing tablet, smart pens or hand gloves. The design of a online signature verification system initially involves the following four aspects: (1) data acquisition and preprocessing (input device), (2) feature extraction, (3) matching (classification), (4) decision making as shown in Fig. 1.
|| Online signature verification system
Input device (input signature): The ordinary input device for on-line signature verification system is digitizing tablet, smart pen, or pen tablet.
Feature extraction: Some features will exhibit more discriminatory capability than others. Thus, once features are extracted, some features selection should be done. Two classes of features can be extracted in dynamic systems:
Static features: These features are extracted from the whole process
of signing, such as maximum, minimum and average of writing speed, curvature
measurements, etc. For this case, major complicatedness is related to the feature
extraction step itself. The selection of stable, pertinent and efficient feature
is not straight forward (Plamondon and Lorette, 1989).
Dynamic features: These features are the evolution of a given parameter
as a function of time f(t), such as position x(t), y(t), velocity v(t), acceleration
a(t), pressure p(t), etc. For dynamic feature methods, major difficulties are
encountered in the matching step and the feature extraction step is almost non-existent
(Plamondon and Lorette, 1989).
Matching: Matching consists of measuring the similarity between the
claimed identity model and the input features. According to Jain
et al. (2000) the four best-known approaches for pattern recognition
are: (1) template matching, (2) statistical classification, (3) structural matching;
and (4) neural networks.
Decision: Once a similarity measure is obtained, the decision implies the computation of a decision threshold. If the matching of similarity is greater than a threshold, the decision is ACCEPT, otherwise it is REJECT.
In an online signature verification system, the users are first enrolled by providing signature samples (reference signatures). Then, when a user present a signature (test signature) claiming to be a particular individual, this test signature is compared with the reference signatures for that individual. If the dissimilarity is above a certain threshold, the user is rejected.
During verification, the test signature is compared to all the signatures in the reference set, resulting in several distance values. One then has to choose a method to combine these distance values into a single value representing the dissimilarity of the test signature to the reference set and compare it to a threshold to make a decision.
There are two types of forgeries: a skilled forgery is signed by a person who has access to a genuine signature for practice and the random forgery is signed without having any information about the signature of the person whose signature is forged.
PERFORMANCE EVALUATION OF BIOMETRIC TECHNOLOGIES
Signature verification can be thought of as a two-class pattern recognition problem, one class consisting of genuine and the other consisting of forgeries. A great deal of variability can be observed in signatures from the same individual according to country, age, time, habits, psychological or mental state and physical and practical conditions. The only certainty in this domain is that when two signatures are identical, one of them is a forgery.
The performance of a biometric verification system is evaluated according to
the error representation of a two-class pattern recognition problem, that is,
with Type I and II error rates. The Type I error rate (False Rejection Rate
(FRR)), measures the number of genuine signatures classified as forgeries as
a function of the classification threshold. The Type II error rate (False Acceptance
Rate (FAR)), evaluates the number of false signatures classified as genuine
ones as a function of the classification threshold.
||Curves of FRR and FAR as a function of the classification
threshold and the corresponding error trade-off curve
To evaluate the performance of our signature verification system, we adopt
the Equal Error Rate (EER) at which the percentage of FAR equal the percentage
of FRR. This EER provides an estimation of the statistical performance of the
algorithm. It can be adopted as a unique measure for characterizing the security
level of a biometric system.
Munich and Perona (1998) described that it is obvious
that it can trade-off one type of error for the other type of error. If every
signature accepted, there would have a 0% FRR and a 100% FAR and if every signature
rejected, there would have a 100% FRR and a 0% FAR. The curve of FAR as a function
of FRR, using the categorization threshold as a parameter, is called the error
trade-off curve. It provides the behavior of the algorithm for any operating
regime and it is the best descriptor of the performance of the system. From
the practical point analysis, this curve is often simplified into a scalar,
the Equal Error Rate (EER). The error rate at which the percent of false accepts
equal the percentage of false rejects. This equal error rate provides an estimate
of the statistical performance of the algorithm, it means that EER provides
as estimation of its generalization error. Figure 2a and b
show the curves of FRR and FAR as a function of the classification threshold
and the corresponding trade-off curve.
Depending on the testing conditions and on the availability of data, a signature verification system can be validated with different types of forgeries.
THE STATE OF THE ART IN ONLINE SIGNATURE VERIFICATION
According to Gupta (2006), who cited Herbst
and Liu (1977) and illustrated that the researchers used an experimental
pen which was mounted with two orthogonal accelerometers and collected the sample
signatures at the rate of 200 times per second. They observed that most signatures
were taken time from 2 to 10 sec with an average time of about 5 seco.
In addition, the researchers further reported that each signature was partitioned by heuristically into segments and after the segments were aligned on the duration of the time interval matching segments were cross correlated and inconsistency between the reference and the test signature. It revealed that with the range of 1 to 2 sec of segments showed the best performance. Seventy users evaluated the method where the first 5 sample signatures were collected from each of the users. The number of reference signature(s) either one or two was selected such that the selected signatures and the remaining signatures were at least equal to a pre-specified value in terms of distance measurement. These were considered to be the best reference signatures. An additional 695 signatures as genuine test and 287 as forged signatures were used for testing purpose. From the experimental results, it revealed that more than 20% False Rejection Rate (FRR) and around 1% False Acceptance Rate (FAR) was obtained.
Gupta (2006) cited Liu et al.
(1979) investigated and reported that the proposed two acceleration measurements
used by Herbst and Liu (1977) along with the additional
writing pressure during the signature process. They observed that the correlation
involving waveforms of pressure demonstrated slight inequity since the correlation
values dominated the gross form of the pressure waveform, but they found that
it appeared more effective when low frequency paper contact components of the
pressure waveform was removed. The researchers conducted some experiments using
the acceleration and pressure correlations and separately where they used signatures
from 24 subjects and obtained results less than 1% of FAR and near about 16%
of FRR . It demonstrated the better results than earlier of Herbst
and Liu (1977).
In the field of on-line signature verification, a number of studies have investigated.
Gupta (2006) cited Crane and Ostrem
(1983) presented a method in which testing consisted of a registration phase.
In the registration phase, the mean and Standard Deviation (SD) of each feature
was calculated from 10 or 12 sample signatures and a reference vector of feature.
Furthermore, the reference signature vector was then compared to the test signature
vector and calculated the Euclidean norm of the distance. Based on the distance,
if small enough the signature accepted as genuine, otherwise rejected it as
forgery. The system allowed up to 3 trials and a false rejection occurred only
if all three signatures failed the verification test. The experimental result
reported that the FAR (False Acceptance Rate) and FRR (False Rejection Rate)
varying from 0.5% to about 3%.
In order to improve the efficiency of signature verification, De
Bruyne (1985) proposed 18 global features sets which included six dynamic
features and other static features. Dynamic features such as number of pen lifts,
pen-up time and writing time, total time along with the maximum writing time
and the velocity. On the other hand, the static features such as area, proportion,
Standard Deviations (SD) of x and y values and total displacements ratio in
the direction of x and y were considered. Ten sample signatures were used to
compute the reference signature. The comparison has been done using test signature
with the reference signature and with the forgeries as well. A maximum likelihood
test was applied for the comparison tests. Eleven persons signatures were
used for testing purpose. The experimental results obtained 3% of FRR and 2%
of FAR where 10 sample signatures were used.
Gupta (2006) cited Hastie et al.
(1991) and reported that a model where a test signature was assumed to consist
of a reference signature which was changed from time to time. The researchers
described the following five-step signature verification method:
Step 1:Smoothing-a cubic spline approximation was used to average out the measurement errors
Step 2:Speed-speed was computed after smoothing
Step 3:Time warping-a time warp function was computed so that correspondence was found between the reference signature and the test signature
Step 4:Segmentation-the signature was segmented using low speed regions (e.g., low speed considered as 15% of mean speed) into a sequence of segments called letters
Step 5:Averaging-estimated the reference signature based on letters
The authors further reported that the results of using the method described
above were presented in the study published by Nelson and
Kishon (1991). The authors reported that the ten genuine signatures samples
and four forgeries from each of 20 subjects were used for testing purpose. The
experimental results obtained FRR of 0% and FAR of 18%. Nelson
and Kishon (1991) also argued the point that a signature might play important
roles in hand signature verification based on the shape and dynamics of a signature.
A hand signature verification system yielding good performance for point-of-sale
applications designed by Lee (1992), who cited by Gupta
(2006). The authors described that a database where 105 human subjects contributed
total 5603 genuine signatures and 4762 forgeries in their research work. Three
types of forgeries were used in their research work such as skilled, random
and timing forgeries. To forge the each genuine signature, two forgers were
used in which each forger contributed to all three types of forgeries. From
each forger six samples of each type of forgery were collected. This process
produced 3744 forgeries where 105x2x18 = 3780 samples were used for experimental
process. Furthermore, forgeries were collected randomly from the 105 individuals
for 22 subjects. Eight dissimilar individuals contributed six skilled forgeries
each of the 22 subjects. These provided total 792 forgeries where as 22x8x6
= 1056 were used for verification purpose. It therefore appeared that a number
of the forgeries were rejected. A subset of the database were used from total
of 11 genuine signatures each for 22 individuals in which 6 were used for the
reference signature and 5 for testing purposes. On the contrary, for the verification
purpose, the researchers used 704 forgeries taken from 8 forgers for every one
in which each contributed 4 forgeries for each individual. From their experimental
result, it was reported that an EER of 3.8% was obtained.
A technique based on Bayesian neural networks for online hand signature verification
of Chinese signatures presented by Chang et al. (1993)
was reported in the paper published by Gupta (2006). The
authors investigated and reported that a set of 16 features was used in their
research which included the features such as total time, number of segments,
average velocity, width/height ratio, average distance in the eight signature
directions, upper-part/lower-part density ratio as well as left-part/right-part
density ratio. The researchers used a database from 80 individuals who contributed
total 800 genuine signatures and 200 simples. The experiment conducted for verification
using 200 skilled forgeries by 10 forgers and obtained 2% of FRR and 2.5% of
A multilevel hand signature verification system that used global features as
well as point-to-point comparison using personalised thresholds was presented
by Plamondon (1994b). The author described that the
system used a set of global features which included the percentage of pen-up
time, total pen-down time and the percentage of time while the angular velocity
is positive. Those features were used for the initial stage of verification.
The author further added that the signature was normalised by rotation as well
as scaling and local correlations were calculated between the part of test signature
velocity values and the values of the corresponding reference signature through
segments alignment based on the elastic matching. On the other hand, the second
stage was carried out by a third stage implying computation of variations amongst
the normalised values of the coordinate of the test and the reference signatures
by local elastic pattern matching.
In the stage of the performance evaluation, 3 signatures from each of 8 human beings were used and 8 other persons contributed 3 forgeries for each of the 8 genuine signers in 64 sessions after having access to genuine signatures and with features on the dynamics of these signatures. Six other subjects were used to produce an additional set of genuine signatures, nine signatures were used for each one where three of which were used as reference signatures. Tests were carried out with the two databases for reconciling the differentiating function to minimised inaccuracies. For that reason, it appeared that test signatures as well as reference signatures were used in determining individual thresholds that minimised inaccuracies. Hence the results achieved FAR of 0.5% as well as FRR of 0.0%, which cannot be pondered as trustworthy. Besides, the testing was very restricted and the number of signatures evaluated was minimum.
Nelson et al. (1994) proposed a statistical
method for hand signature verification which was found in the article published
by Gupta (2006). The authors reported that the proposed
method used a set of 25 features which was included two time-related features,
6 features related to velocities and accelerations, four shape-related features,
eight features giving the distribution density of the path tangent angles and
four giving angle-sector densities of the angular changes and a feature relating
to the correlation between the two components of pen velocity. They also articulated
the statistical basis of hand signature verification and then used three different
methods for computing the distance between the reference signature and the test
signature, such as the Euclidean distance method, Mahalanobis distance method
and the quadratic discriminant method. A simple method of feature selection
is described which essentially consists of computing the ratios of the Standard
Deviation (SD) to the mean for each feature and rank-ordering the features according
to this ratio. It was not reported that why a feature with least normalised
standard deviation would contribute competent discrimination amongst the genuine
signatures and forgeries. A combination of schemes were used such as individual
best 8, 10, 12 or 14 of the 25 features, were evaluated. The achievement of
all these sets was alike even though the individual best 8 and 10 found to perform
the excellent result with FRR near 0%. The authors identified that using an
Euclidean distance approach along with the best 10 features out of the 25 features
and achieved as outcomes with 0.5% of FRR and 14% of FAR.
Lee et al. (1996) designed an on-line dynamic
signature verification systems with a data base of more than 10,000 signatures
in (x(t), y(t))-form was acquired using a graphics tablet. The authors further
reported that they extracted a 42-parameter feature set at first and advanced
to a set of 49 normalized features that tolerate inconsistencies in genuine
signatures while retaining the power to discriminate against forgeries. They
studied algorithms for selecting and perhaps orthogonalizing features in accordance
with the availability of training data and the level of system complexity. For
decision making the researches studied several classifiers types. A modified
version of majority classifier yielded 2.5% EER and, more significantly, an
asymptotic performance of 7% FAR at 0% FRR was reported using 15 parameter features.
Gupta and Joyce (1997) cited by Gupta
(2006) proposed an algorithm with the aspire of using a small set of global
features that are easy to compute and invariant under most two-dimensional transformations
such as rotation, slant and size. They used 6 features in the initial experiments
like total time, number of velocity sign changes in the x and y
directions, number of acceleration sign changes in the x and y directions
and total pen-up time. The authors further reported that the hand signature
verification algorithm was used based on Euclidean distance. It showed that
time by itself was the best single discriminator and pen-up time was also a
good discriminator. Furthermore, the authors also added that the included path
length in the set of attributes improves the performance of the technique and
good results were obtained when path length was included and the reference signature
built using 10 sample signatures. An FRR of about 0.5% was obtained with FAR
of little more than 10%. The more comprehensive explanation can be found in
the authors published article.
According to Gupta (2006) cited Nalwa
(1997) investigated and reported that his proposed technique was based on
applying jitter, parameterisation over normalised length, aspect normalisation,
centre of mass, sliding computation window, moments of inertia, torque, weighted
cross-correlation and warping, moving to coordinate frame and saturation. The
author further enlightenment that to test the proposed algorithm three signature
databases were used. The first comprised of 904 genuine signatures and 325 skilled
forgeries from 59 different individuals. The second set comprised of 982 signatures
from 102 individuals, collected in a solitary session. There were 401 skilled
forgeries. A number of genuine signatures and forgeries were removed from the
data set as well. The third data set comprised of 790 genuine signatures and
424 skilled forgeries from 43 signers. The outcomes from the three test databases
as well as one that included all three, applying 4, 5 and 6 reference signatures
were demonstrated. The experimental results obtained the EER within the range
from 2 to 5%.
Kashi et al. (1998) described a method for the
automatic verification of on-line handwritten signatures using both global and
local features. The global and local features ware captured for various aspects
of signature shape and dynamics of signature production. The researchers demonstrated
that adding a local feature based on the signature likelihood obtained from
Hidden Markov Models (HMM), to the global features of a signature, considerably
improved the performance of verification. The authors further added that the
performance of signature verification methods tested on the Murray Hill database.
The test database 542 genuine signatures and 325 forgeries were used. Each reference
set used the first 6 signatures of every one of the 59 subjects. There were
32 volunteers, who provided a total of 325 forgeries. The best result obtained
from their research method with an EER of 2.5%.
Jain et al. (2002) cited by Kholmatov
and Yanikoglu (2004) and Gupta (2006) reported that
the researchers were used a method in which specific critical points like start
and end points of a stroke as well as changes of trajectory points, were extracted
for every signature. Moreover, the authors enlightened that the number of strokes
was used as a global feature. Two types of local features, spatial and temporal,
were extracted from the x and y coordinates. The proposed technique was tested
using two datasets. The first dataset contained 520 signatures, ten signatures
each from 52 writers, collected in one session. The second dataset was a superset
of this dataset and contained a total of 1,232 signatures collected from 102
writers, seventeen of which contributed more than ten signatures in multiple
sessions over a period of up to one year. Twenty writers provided three skilled
forgeries each (a total of only 60) after viewing an original signature. The
best error rates using a common threshold were 3.3% FRR and 2.7% FAR and the
best error rates using writer-dependent thresholds were 2.8% FRR and 1.6% FAR.
The FAR rates appeared to be based on random forgeries. No FAR for skilled forgeries
A new Dynamic Time Warping (DTW) technique for the signature verification was
proposed by Feng and Wah (2003). The authors argued
that the technique was originally used in speech recognition and has been applied
in the field of signature verification with some success since few decades ago.
The new warping technique proposed the authors named as Extreme Points Warping
(EPW). The authors further reported that the techniques proved to be more adaptive
in the field of signature verification than DTW, given the presence of the forgeries.
EPW and DTW were compared on a database of 1000 signatures of 25 users. With
the use of EPW, the equal error rate was improved by a factor of 1.3 and the
computation time was reduced by a factor of 11. From the experimental observation
the authors further pointed out that EPW was much faster than DTW and considerably
Ortega-Garcia et al. (2003) cited by
Gupta and Joyce (2007) investigated and reported the signature verification
results using the five time sequences, x and y coordinates, pressure, inclination
and attitude as well as three derived sequences, path tangent angle, path velocity
and log curvature radius. Furthermore, the authors described that if the first
and second derivative of each of these sequences computed, the total time sequences
were found 24. Hence, a signature sample that has say 1000 samples would generate
24,000 values. The functional values were then normalised to obtain zero mean
and unit standard deviation. Signatures were modelled using Hidden Markov Models
(HMM) based on the sequences. The performance was tested using a signature database
of 15 genuine signatures and 15 forgeries each from 50 people. The tests were
conducted using the same threshold for all, resulted in 4.83% EER which reduced
to 0.98% by using user-specific thresholds.
A new stroke-based algorithm for dynamic signature verification was presented
by Qu et al. (2004). The algorithm was developed
to convert sample signatures to a template by considering their spatial and
time domain characteristics and by extracting features in terms of individual
strokes. Individual strokes were identified by finding the points where there
found a (1) decrease in pen tip pressure, (2) decrease in pen velocity and (3)
rapid change in pen angle. Experimental results were achieved for signatures
from 10 volunteers over a 4 months period. All the collected genuine signatures
were classified into training and verification classes. An experiment was performed
to evaluate the performance of the verification system. A total of 110 signatures,
split into 50 reference and 60 test signatures, from 10 volunteers were used
in this experiment. Each volunteer performed 5 signatures to train their signature
template and performed another 3 genuine signatures as test signatures. In addition,
for each template, 3 skilled forgery signatures were performed by other volunteers.
First, found the best 4 non-stroke features for the system (total time during
the signing process, average writing speed, variance of pressure signal in 10
sliding windows and mean of the x displacement signal in 10 sliding windows).
When the threshold was set to be 75%, the system achieved a FRR of 30% and FAR
of 46.67%. In order to evaluate the performance of the stroke based features,
the proposed system added one or two stroke based features to the 4 non-stroke
feature system. Based on the previous non stroke based feature system, if adding
time duration for velocity significant stroke as the 1st stroke based feature
and correlation coefficient for the pressure significant stroke as the 2nd feature,
both of them improved the system with 6.67% of FRR and 13.33% of FAR.
Yeung et al. (2004) SVC2004: First International
Signature Verification Competition was organized on 2004. A signature database
involving 100 sets of signature data was created, with 20 genuine signatures
and 20 skilled forgeries for each set using small pen-based input devices such
as Personal Digital Assistants (PDA). Of the 100 sets of signature data, only
the 40 sets were released. When evaluated on data with skilled forgeries, the
best team for competition gives an Equal Error Rate (EER) of 2.84%.
Quan and Ji (2005) cited by Gupta
and Joyce (2007) presented a novel approach that applied the dynamic time
warping (DTW) to match the crucial points of signatures. Firstly, the signatures
were aligned through the DTW and the crucial points of signatures were matched
according to the mapping between the signatures. Then the signatures were segmented
at these matched crucial points and the comparisons were accomplished between
these segments. The distance between the two was computed using a simplified
Mahalanobis distance. The testing procedure was somewhat unclear but it appeared
6 samples for each signer were used to find a reference signature. It involved
comparing each of the 6 samples with the other 5 and counting the number of
matching points. The signature that has the largest total matching points was
then selected as the reference for the individual. An EER of 3.8% was obtained
using random forgeries.
The authors Fierrez-Aguilar et al. (2005b) presented
an on-line signature verification system exploiting both local and global information
through decision-level fusion. Global information was extracted with a feature-based
representation and recognized by using Parzen Windows Classifiers. Local information
was extracted as time functions of various dynamic properties and recognized
by using Hidden Markov Models. Experimental results were given on the large
MCYT signature database where 330 signers and total 16500 signatures were collected
for random and skilled forgeries. Feature selection experiments based on feature
ranking were carried out. The two proposed systems were also shown to give complementary
recognition information which is successfully exploited using decision-level
score fusion. The experimental results obtained from the system was promising
where skilled forgeries EER of between 5 and 7% were obtained with 5 training
signatures as well as 1-2% with 20 training signatures. On the other hand, the
random forgeries EER were 1-1.5% for 5 training signatures and around 0.5% for
20 training signatures.
Fierrez-Aguilar et al. (2005a) used target dependent
score normalization technique using SVC2004 database which is consists of 40
sets of signatures. Each set contains 20 genuine signatures from one contributor
and 20 skilled forgeries from five other contributors. Obtained result with
7.14% of EER.
Shintaro et al. (2006) used user-generic Fusion
model with Markov Chain Monte Carlo Method. The database consists of pen position,
pen pressure, angle, altitude and azimuth based on the time sequence. From 330
individuals, 25 genuine signatures and 25 skilled forgeries were collected for
each individual and obtained the best results with 4.06% of EER.
A multivariate autoregressive (MVAR) modeling in combination with a Dynamic
Time Warping-based (DTW) segmentation technique was proposed by Osman
et al. (2007). The authors described that the database consists of
16 genuine writers, each writer provided 150 signature samples over 15 sessions
at a rate of 10 samples per session. Thus, a total of 2400 genuine. The skilled
forgeries population consists of 8 forgers. The 8 forgers were used to forge
8 genuine writers; each forger provided 30 samples for each of the 8 writers
Thus, a total of 1920 forgeries were collected. Obtained result 96.6% of accuracy
in skilled forgery test.
A new stroke-based signature verification system was proposed by Chang
and Shin (2007). According to the authors, it was crucial to find correct
points of a testing signature to be spilt according to its template signature.
In their study, they proposed a modified dynamic time warping algorithm (DTW)
for the problem. Foremost, the stroke information of both template and testing
signature was considered precious to prevent wrong splitting. Next, the number
of stroke difference in genuine signature is not limited because the proposed
method does not require much extra computational time. During the experiment,
proposed method was verified with 1359 signatures written by 17 Japanese writers.
679 authentic signatures and 680 forged signatures were used. All the signatures
were written by Japanese writers and composed Chinese characters such as Japanese
Kanji. The signatures were written by pen tablet input device. All the forgery
signatures were written by other writers watching the authentic signature. 40
authentic signatures were taken from every writer. Along with these data, 10
signatures per a writer were used for training data and rest authentic signatures
were used for verification. Nevertheless, since the number of the signatures
having the same number of stroke is smaller than 10 for one writers signature,
only 8 signatures were used for the training in the case. All the 40 forged
signatures were used as a test data without selection. The test data were all
the semi-skilled forgeries and the result obtained from the proposed was 3.85%
Signature verification has been an attractive field of research area because
of the social and legal acceptance and widespread use of written signatures.
An automatic signature verification technique was proposed by Hu
and Wang (2007). The authors stated that it is still a challenging issue
because of small sample size problem as well as large intra-class differences
and, when considering forgeries, small inter-class variations. In order to solve
these problems the researchers proposed a two-stage fusion method to get high
accuracy. At first, an Enhanced Dynamic Time Warping (EDTW) algorithm and a
normalized feature measure were used to build a classifier based on local features.
The former enhanced the separability between genuine and forgery signatures,
while the latter approaches the problem as a two-class pattern recognition problem,
which make it possible to use training signatures as many as possible. However,
local method is time and resource consuming, so they then designed another classifier
based on global features using majority voting rule. The method fused the global
and local method by two-sage serial strategy to build an on-line signature verification
system. In their experiment, Task 2 of SVC2004 was used for skilled forgeries
and achieved 3.02% of EER.
Rabasse et al. (2007) presented a method for
the generation of synthetic handwritten signatures, in the form of a series
of time-stamped pen data channels, for use in dynamic signature verification
experimentation. The technique introduced a modelled variability within the
generated data based on variation that is naturally found within genuine source
data. In order to assess the quality of the synthesized images, a commercial
dynamic signature engine was used within a verification scenario. A mode of
operation of the selected verification engine is to provide a binary decision
on whether a presented signature is genuine or forged when compared against
a reference template formed from 3 signatures. The measure of confidence associated
to this binary result is also returned a default confidence value over 80 out
of 100 is taken to indicate a genuine signature. Signatures were obtained from
the publicly available database used in the Signature Verification Competition
(SVC2004). This database consists of 1600 text files containing separate signatures
in the form of time stamped x, y and pen-on-tablet, pressure, azimuth and altitude
sequences. Forty separate signers are represented in the data set. For each
signer, 20 files represent genuine signatures and the remaining 20 represent
skilled forgeries. The skilled forgeries were not used in this experiment. In
order to examine this lower verification performance, the individual verification
rates of the synthesized signatures with variability were assessed according
to their position within the synthesis cycle. The researchers further investigated
and reported that the synthesised signatures 1 and 100 were the closest to the
seed signatures 1 and 2, respectively with the other 98 signatures being interpolations
in between the seeds, because the synthesised signature 50 represented the mid-point
interpolation. It was found that between positions 23 and 78 the average verification
rate was above the value of 87.88% achieved by the genuine signatures.
An online signature verification system based on local information and on a
one-class classifier, the Linear Programming Descriptor classifier (LPD) was
presented by Nanni and Lumini (2008). The authors investigated
and described that the information was extracted as time functions of various
dynamic properties of the signatures, then the discrete 1-D wavelet transform
(WT) was performed on these features. The Discrete Cosine Transform (DCT) was
used to reduce the approximation coefficients vector obtained by WT to a feature
vector of a given dimension. Moreover, the Linear Programming Descriptor classifier
is trained using the DCT coefficients. The experimental results using all the
5000 signatures from the 100 subjects of the SUBCORPUS-100 MCYT Bimodal Biometric
Database were presented, yielding performance improvement both with Random and
Skilled Forgeries and obtained an EER of 5.2% in the Skilled Forgeries.
An approach to online signature verification using data glove has been presented
by Kamel et al. (2008). To verify the efficiency
of the proposed technique in handwritten signature verification, the 5DT Data
Glove 14 Ultra was used. This glove uses 14 sensors to measure finger flexure
(two sensors per finger) as well as the abduction between each finger. The system
was interfaced with computer via cable to USB port. Data glove is a new dimension
in the field of virtual reality environments, initially designed to satisfy
the stringent requirements of modern motion capture and animation professionals.
The researchers tried to shift the implementation of data glove from motion
animation towards signature verification problem, making use of the offered
multiple degrees of freedom for each finger and for the hand as well. Their
proposed technique was based on the Singular Value Decomposition (SVD) in finding
r singular vectors sensing the maximal energy of glove data matrix A, called
principal subspace and thus account for most of the variation in the original
data, so the effective dimensionality of the data can be reduced. Having modeled
the data glove signature through its r-principal subspace, signature authentication
was performed by finding the angles between the different subspaces. A demonstration
of the data glove was presented as an effective high bandwidth data entry device
for signature verification. The SVD-based signature verification technique was
tested and its performance was shown to be able to produce Equal Error Rate
(EER) of less than 2.37%.
Furthermore, in the area of online signature verification using data glove,
a number of research have investigated and the experimental results have been
reported by the researchers (Sayeed et al., 2007,
2008, 2009; Kamel
and Sayeed, 2008).
Recently, Elahen and Mohsen (2009) was presented online
signature verification system based on global information and an Adaptive Network
Based Fuzzy Interface System (ANFIS). According to the authors, the proposed
method for signature verification was divided into two phases. The first phase
was based on the analysis performed by the method known as fractal dimension
and the second phase used Adaptive Network Based Fuzzy Interface System for
output. The system was tested with two different data sets: SUBCORPUS-100-MCYT
database and Persian signature database. SUBCORPUS-100-MCYT was captured using
a WACOM pen tablet model INUOS A6 USB with resolution 2540 lines per inch and
sampling frequency of 100 Hz. The dataset consists of 100 signature contributors
where 25 genuine signatures and 25 skilled forgeries were captured from each
of the signature contributor. On the contrary, the Persian dataset was captured
using digitizing tablet with sampling frequency of 50 Hz.
|| Experimental results in percentage (%)
A total number of 400 genuine signatures ware captured from 40 signature contributors.
In addition, 200 random and 200 skilled forgeries were used to test the system
performance. In case of producing the random forgeries, forgers tried to forge
only the signature shape whereas skilled forgeries were captured when forgers
were provided with the animation of each signing process and they could repeat
the animation several times to learn the signing process.
Furthermore, the signature databases were divided into two parts: training and test sets for the purpose of the experimental setup. In case of skilled forgeries, training set consists of 10 genuine and 10 forgery signatures in MCYT database as well as 6 genuine and 6 forgery signatures in Persian database. On the contrary, test set consists of 3000 (30x100) in MCYT and 480 (12x40) in Persian databases. In case of random forgeries the same number of training set was used as well as signatures were used of every other user to evaluate the forgery detection. The experimental result obtained from their proposed system is shown in Table 1.
Online hand signature verification is a extremely potential field of research
from both scientific and commercial points of view. In recent years, along with
the continuous development of the Internet and the increasing security protection
necessities for the growth of the e-society, the field of online signature verification
is being considered with renewed significance given that it uses a customary
individual confirmation technique that is accepted at both legal and societal
levels. In addition, recent results achieved in international competitions using
standard databases and test protocols have revealed that signature verification
systems can have an accuracy level similar to those achieved by other biometric
systems (Vielhauer, 2005). Finally, different from physiological
biometrics, handwritten signature is an active method that requires the user
to perform the unambiguous act of signing. Thus, online signature verification
is principally useful in all applications in which the confirmation of both
transaction and user is essential (Plamondon and Srihari,
2000; Vielhauer, 2005).
Therefore, the number of possible applications for online signature verification is constantly rising along with the development of various sophisticated and user-friendly input devices for online handwriting acquisition.
The ultimate result is that in the near future, along with a broad array of
prospective applications, a noteworthy yearly growth is predicted in the global
signature verification market (Ureche and Plamondon, 1999).
Obviously, this tendency has been further exaggerated by research results in
recent years, which have notably advanced the state of the art in the field.
However, in order to reinforce the commercial and societal benefits associated
with the online signature verification, extra efforts are essential.
In this article, the state of the art in online signature verification has been presented and the most important results have been addressed. Moreover, a number of most potential directions for research in this field have been highlighted. In the coming future, research need not be paying attention absolutely on accuracy excellence, since it has mostly been in the past. As an alternative, it should concentrate on a huge number of issues associated to miscellaneous circumstances of the application themselves.
Therefore, in the era of the e-society, online signature verification can no longer be pondered definitely limited to academics and research laboratories as the prospect of applying online signature verification in an array of applications is becoming a reality. Certainly, further research is essential to completely examine and interpret the potential of handwritten signatures, which remain extremely distinct signs, unambiguously representing the encouragement and convolution of human beings.