SIGNATURE VERIFICATION USING ART-2 NEURAL NETWORK

Pavel Mautner1 , Ondrej Rohlik1 , Vaclav Matousek1 , Juergen Kempf2

University of West Bohemia, Faculty of Applied Sciences, Univerzitni 8,

1

CZ – 306 14 Plzen (Pilsen), Czech Republic

e-mail: mautner | rohlik | matousek@kiv.zcu.cz
2
University of Technology (Fachhochschule) Regensburg, Pruefeninger Straße 58
D – 93049 Regensburg, Germany
e-mail: juergen.kempf@mikro.fh-regensburg.de

ABSTRACT

The ART neural network models have been developed for

the clustering of input vectors and have been commonly
used as unsupervised learned classifiers. In this paper we
describe the use of the ART-2 neural network model for
signature verification. The biometric data of all signatures
were acquired by a special digital data acquisition pen and
fast wavelet transformation was used for feature extraction. Figure 1: Digital Data Acquisition Pen
The part of authentic signature data was used for training the
ART verifier. The architecture of the verifier and achieved
results are discussed here and ideas for future research are
also suggested.

1. INTRODUCTION

There are many commercial systems designed for person

identification worldwide. Among the most popular ones are
those based on fingerprints, ID cards and signature recog-
nition using optical character recognition (OCR) methods.
The current identification systems are based on input de-
vices that consist of at least two parts (pen and tablet, pen
with an infrared transmitter and one or two receivers, etc.)
The obvious problem of such an approach is the limited
mobility of a system composed of several parts. For the
purpose of verification a special digital data acquisition pen
was developed at the University of Technology Regensburg.
The pen produces three signals in which information about
the pressure applied to ballpoint and about the side pressure
in x and in y directions is involved. Detailed information
about an acquisition device is given in Section 2. Section
3 deal with the pen output signal feature extraction method
and Section 4 describes the neural network verifier based
on the unsupervised learned neural network model of ART-
2. Results of verification experiments, and possible future
works are discussed in Section 5 and Section 6, respectively.
Figure 2: Transformed output signal waveforms
2. DATA ACQUISITION DEVICE
As was mentioned above signature verification is performed
by the special digital data acquisition pen. The first exper-
imental pen was built at University of Technology Regens-
burg during the spring of 2000. Our experiences with this
pen were discussed in [1]. The pen consists of two sensors
integrated in the pen. These are a 2-axis acceleration sensor
(based on the principle of moving mass between two capac-
itors) and a pressure sensor (based on the piezoelectric ef-
fect). As the acceleration sensor was not very sensitive and
the output signal was noisy, it was later replaced by sensors
working on a principle similar to that of the piezoelectric
pressure sensor. These new sensors scan the side pressure
applied to the ball-pen in x and y directions. The third sen-
sor scanning the pressure to the ball-pen nib is integrated
into the pen and is the same as in the previous version of
the pen. The prototype of the pen is illustrated at Figure 1.
To obtain the right signal waveform, the pen should be held
so that the x side sensors are parallel to the writing axis and
the y sensors are perpendicular to it. Failing that, the signals
corresponding to the x and y side pressure are distorted. In
order to reduce this effect, these signals are transformed to
polar coordinates mag and phi. The transformed output
signal waveforms are illustrated in Figure 2.
3. FEATURE EXTRACTION
4. NEURAL NETWORK SIGNATURE VERIFIER
The neural network models are commonly used for pro-
cessing classification problems. But signature verification
differs from the general classification problem. The goal
of the general classification problem is to choose one class
from several classes, whereas the training data contain data
from all classes. For our application all the training data
are authentic signatures and we have no data for the sec-
ond class fake signatures. This is the reason why the fre-
quently used supervised learned neural network model such
as multi-layer perceptron cannot be applied to the signature
verification task.

Before the feature vector is evaluated from the output sig- A F

nals, only the active part of the signature has to be deter-
mined. This is done from the first difference of output sig-
nal z . To determine of the beginning (or the end) of the F
signature, the z signal is scanned from left to right (or con- 2
versely) and the first difference is evaluated. The beginning
(or the end) of the signature is determined if the value of
the first difference of signal z exceeds the threshold value ρ
Θ at the first occurrence and value of the signal is greater ri
than the reference value σ The threshold values Θ and σ are G
determined according to the type of the piezoelectric sensor F1 R
used.
For the extraction of features from signals, the fast wave-
let transform (FWT) was used. At first, each signal of the
signature was filtered by an average filter, afterwards it was
decomposed by FWT and coefficients of AS5 and Dm S
, for
m = 1, 2, 5, were determined [2]. For decomposition the Fv
Daubechies and Coiflet wavelet families were tested, the 5-
th order Daubechies wavelet gave the best result. Using of Figure 3: ART-2 signature verifier
this wavelet, the following features were used to describe
the signature:
• rl = sln , where sl is the number of samples of the 4.1. Architecture of the neural network verifier
signature and n is the total number of samples in one
scan (n = 500 for all tested signatures). The adaptive resonance theory (ART), developed by Car-
penter and Grossberg, was designed for clustering binary
X
Nm
input vectors (ART-1) or continuous-valued input vectors
S
• Dm = d2mj , m = 1, 2, ..., 5; s = x, y, z,
j=1
(ART-2). With regards to the features what we used for de-
where Nm is the number of coefficients in scale m. scription of signals, the ART-2 model is suitable for signa-
ture verification. The general architecture and description
X
N5
of the ART-2 network is not discussed here, for details see
• AS5 = a25j , m = 1, 2, ..., 5; s = x, y, z, [3],[4].
j=1
The basic structure of the network verifier is illustrated
where N5 is the number of coefficients in scale 5.
in Figure 3. The network consists of two layers of process-
Features AS5 and Dm s
represent the energy of the m-th signal ing elements labelled F1 (input and interface units) and
decomposition level, rl represents the relative length of the F2 (cluster units), each fully interconnected with the oth-
active signature signal. These features are components of ers, and supplemental unit G and R (called gain control
the feature vector F vi of the i-th signature. The feature unit and reset unit), which are used to control the process-
vector is presented to the input of the signature verifier. ing of the input data vector and creating of the clusters.
 a)
b)
Figure 4: a) Authentic signatures with quality mark
b) corresponding fake signatures
The input and interface layer F1 consists of six sub-
layers (these are not illustrated in Figure 3); each sub-layer
has the same number of processing units as is the size of
the feature vector. The purpose of these sub-layers is to
allow the ART-2 network to process continuously varying
inputs. Moreover, they normalize the components of the
feature vector and suppress the noise. The size of the F1
layer (and hidden sub-layers) was 19 in our application.
The clustering layer F2 consists of two processing units
only, the former (labelled A) is active only if the feature vec-
tor corresponding to the authentic signature appears at the
input of the network, the latter (labelled F) is active in other
cases. More clusters are not enabled in our application.
4.2. Training and verification
As was mentioned above, only the data for the authentic
signature are known. Moreover, the number of template
signatures cannot be too high because the acquisition of a
large training set, e.g. at a bank counter could be boring
and unpleasant for the customer. Hence only 5 signatures
were used for the training of the ART neural network in
our application. For these signatures corresponding feature
vectors were evaluated and repeatedly presented to the in-
put layer of the network (the slow learning mode was used
for ART-2 network training). The parameters of the hidden
sub-layers of F1 and vigilance parameter ρ were set so that
only the unit labelled A of layer F2 was active during the
a)
b)
Figure 5: Example of the output signal waveforms for au-
thentic and fake signatures in Figure 4
whole training procedure (the network places the template
signatures only in one cluster and adapts the correspond-
ing weights between F1 and F2 layers). When the train-
ing is completed, the network is prepared for verification.
The parameters of F1 sub-layers are not changed during the
verification, only the vigilance parameter ρ have to be set
properly to the authentic and the fake signatures were set
to right clusters. The three methods of setting the vigilance
parameter were tested in our work:
• manual setting M : vigilance parameter ρ is set to the
fixed value (ρ = 0.98) manually, for all verifications,
• automatic setting A1 : ρA1 = min {ri } i = 1 · · · Nt ,
i
1 X
Nt
• automatic setting A2 : ρA2 = ri i = 1 · · · Nt .
Nt i
In the equations above, Nt is a number of training vectors
and ri is activation level of unit R (see Figure 3 and [3],[4]
for detailed description and evaluation of ri ). The results
achieved by corresponding setting of parameter ρ are pre-
sented in Table 1.
5. EXPERIMENTAL RESULTS
To test the verifier, signatures by 10 authors were taken.
For each author, 20 authentic signatures and 15 fakes were
recorded. The fake signatures were written by three differ- Authentic signatures Overall
ent authors (5 fakes for each person). Sometimes the author No. classified as Accuracy
is not satisfied with his/her own signature. The quality of authentic fakes Ratio [%]
signature depends on his/her physical and mental condition. M A 1 A2 M A 1 A2 M A1 A2
1 17 19 16 3 1 4 85.7 91.4 85.7
In such a case the signatures can be classified as fakes. For
2 14 15 9 6 5 11 80.0 80.0 68.6
the evaluation of such cases, the authors marked their au-
3 19 17 12 1 3 8 88.6 91.4 77.1
thentic signatures by a mark from the scale 1 - 4 (1 means
4 19 18 12 1 2 8 94.3 91.4 77.1
a best form of the signature). For the verifier training, only
5 20 20 11 0 0 9 100 100 74.3
the five signatures labelled by mark 1 or 2 were chosen. Ex- 6 16 18 8 4 2 12 88.6 94.3 65.7
amples of the authentic signatures of one author and their 7 16 18 8 4 2 12 85.7 91.4 62.9
fakes are presented in Figure 4, corresponding output signal 8 19 18 9 1 1 11 88.6 91.4 68.6
waveforms are presented in Figure 5. The summary of test 9 19 19 14 1 1 6 91.4 94.3 82.8
results for 10 authors is presented in Table 1. The overall 10 14 13 12 6 7 8 77.1 77.1 74.3
accuracy ratio presented in Table 1a) was evaluated without
the respect of signature quality mark mentioned above. But a)
in most cases the authentic signatures classified as fakes are Fake signatures Overall
the ones labelled by their authors as poor signatures (with No. classified as Accuracy
a quality mark worse than 2). It means that the overall ac- fakes authentic Ratio [%]
curacy ratio will be a slightly higher if these signatures will M A 1 A2 M A 1 A2 M A1 A2
be recognized as a fakes. The Table 1b) show this situa- 1 13 13 14 2 2 1 91.4 94.3 94.3
tion. Accuracy ratio presented here was evaluated with the 2 14 13 15 1 2 0 91.4 91.4 88.6
respect of signature quality mark assigned to each signa- 3 12 15 15 3 0 0 88.6 91.4 82.9
ture by its author. We can see the accuracy ratio is getting 4 14 14 15 1 1 0 97.1 94.2 82.9
higher for all kind of vigilance parameter setting. The best 5 15 15 15 0 0 0 100 100 85.7
result was achieved if the vigilance threshold ρ was set au- 6 15 15 15 0 0 0 88.6 94.3 71.4
tomatically during training process as the minimum value 7 14 14 14 1 1 1 91.4 94.3 74.3
of activation level of unit R (automatic setting A1 ). 8 12 14 15 3 1 0 91.4 94.3 82.9
9 13 14 15 2 1 0 94.3 97.1 88.6
10 13 14 14 2 1 1 85.7 85.7 82.9
6. CONCLUSION AND FUTURE WORK b)

The using of unsupervised learned ART-2 network for sig-

nature verification was discussed here. The tests showed
that this network can be used as signature verifier and gives
a good result with respect to the training set size. Before the
network training and the verification, only a small number
of the network parameters had to be set manually. These
parameters have remained at most the same in verification
process too. In our future work, we plan to focus it on the
setting of these parameters automatically during the train-
ing phase. To improve the overall accuracy ratio, we plan
to include the new valuable features to the feature vector
describing the signature. Finally, we also plan to check the
possibility of the application of other unsupervised learned
neural network models (e.g. Kohonen self organizing fea-
ture map) for further improving of the signature verification
task.
Table 1: Results of verification tests for different setting of
vigilance parameter ρ
[2] S. Pitnerr and S.V. Kamarthi, Feature Extraction from
Wavelet coefficients for Pattern Recognition Tasks,
IEEE Transactions on PAMI, Vol. 21, No. 1, 1999
[3] L. Fausett, Fundamentals of Neural Networks,
Prentice-Hall, New Jersey, 1994.
[4] G.A. Carpenter and S. Grossberg, ART-2: Self-
organization of Stable Category Recognition Codes
for Analog Input Patterns, Applied Optics, No. 26, pp.
4919-4930, 1987

7. REFERENCES

[1] O.Rohlik et al., A New Approach to Signature Verifi-

cation: Digital Data Acquisition Pen , Neural Network
World, Vol. 11, No. 5, pp. 493-501,2001