Sie sind auf Seite 1von 6

Handwritten signature authentication using statistical

measures of basic on-line signature characteristics



Miroslav Baa, Tomislav Fotak, Petra Grd
Faculty of Organization and Informatics
University of Zagreb
Pavlinska 2, 42000 Varadin, Croatia
{miroslav.baca, tomislav.fotak, petra.grd}@foi.hr



Abstract: Personal signature has become an important part of human identity
verification. Therefore, it is important to protect ones signature and be able to detect
signature forgery. We present a statistical based approach to signature verification with
user oriented signature feature weights. Only basic on-line signature features are used
in the verification process. Depending on the authentication threshold, our method
achieves satisfying FAR for unskilled and skilled forgery, but there is still some research
to be done to improve user experience (reduce FRR).

Key Words: handwritten signature authentication, signature verification, on-line
signature, statistical measures, feature weights

1 Introduction
Handwritten signature is defined as the persons first and last name written in its own
handwriting [1]. It is used almost everywhere, on daily basis, and has become common
authentication method. This is not surprising, since its usage, as a mean of giving
consent to something that needs to be done, dates from ancient times.
Personal signature has become important part of the human identity verification.
Therefore, it is important to protect ones signature and be able to detect signature
forgery. This goal can be achieved through signature biometrics. Whether off-line or on-
line signature features are used, the applied biometric method must provide valid
information about signature correctness in regard to claimed identity. Handwritten
signature is behavioral biometric characteristic and almost everything can have an effect
on one signature, so the task of proving its correctness is not always as easy as it seems
to be.
To avoid static image analysis and be able to operate on live handwritten signatures,
researchers in this field have started analyzing dynamic signature features, such as pen
pressure, signature velocity, etc. Those features are usually obtained through special
electronic device used to retrieve signature. In our previous work [2] we have already
extracted the basic set of mainly global signature features that are being used to register
user, i.e. store their signature vector in the local database and use it in the verification
process.
In the rest of the paper we will describe the basics of our simple authentication
method that uses stored vector to calculate person-specific weights of each signature
feature and use it in the signature correctness decision process.
47
2 Previous work
We have already provided the set of signature features that are only the beginning in the
process of determining the ideal feature subset to be used in personal authentication.
Those features are [2]: number of strokes in signature, number of pen-ups, signature
aspect ratio, signature length, signing time, pen time-down ratio, pen time-up ratio,
signature speed, velocity along the x-axis, velocity along the y-axis, average pen
pressure and strongest pressure moment. One could notice that these features are mainly
global handwritten signature features. It does not mean that we disregard local features;
we rather give the basic set of features that can be used to compute some others and can
also be used on local level, e.g. we can determine all these features for each stroke.
Extracted features are similar to those presented in [4] and [5]. To analyze them we
have chosen five statistical measures: mean, standard deviation, median, minimum
value and maximum value; calculated from the set of all signatures retrieved during the
registration process. These measures are calculated for each feature separately and
stored as a part of the signature vector. Note that there is only one vector per person,
containing all statistical measures of all extracted features. Up till now, we have all
prerequisites for authentication process.
Before we proceed further with the authentication process one must be aware that
on-line signature authentication is well known research problem and that there are
several very important authentication methods in the literature. The newest state of the
art defines approaches to feature extraction and on-line signature verification [3]:
Dynamic Time Warping (Elastic Matching), Hidden Markov Model (HMM), Gaussian
Mixture Models (GMM), structural approaches, statistical-based approaches, Support
Vector Machine, neural networks, transform domain approaches, intelligent approaches,
protected approaches and semi-online approaches.
3 Personal signature authentication
Our authentication model uses feature based statistical approach to achieve successful
personal signature verification. The process starts with the signature acquisition via
digitizing tablet, continues with the dynamic feature extraction and feature weights
calculation, ending in the verification decision making module. This process, including
user registration, is shown in Fig 1.
Each registered user has its own signature vector stored in the database. The vector
contains five statistical measures of all features, meaning that it contains 60 statistical
measures overall (12 features, 5 measures for each). Using these measures we are able
to obtain the most and the least constant feature, i.e. calculate feature weights for
authentication process.

48


Figure 1. Personal signature registration and verification process

3.1 Feature weights calculation
Every user had to provide from 10 to 15 handwritten signatures during the registration
process. The goal of this process is to find which of twelve features was the most
constant, i.e. which feature has the same or similar value in every signature, and which
one was the least constant. According to the mentioned, features will gain their weights.
The sum of all weights must be 1.
Calculating feature weight is basically a four-step process. In the first step we
calculate coefficient of variation from the mean value, i.e. coefficient of variation for
each feature.
Let, AD
i
be the coefficient of variation of the i-th feature, STDEV
i
be standard deviation
of the i-th feature and MEAN
i
be the mean of the i-th feature, then we can calculate AD
i

using following equation:


(1)
49

The less the coefficient of variation is the more constant the feature is. Special case is
when it is zero. In the authentication process for this special case, if the users signature
does not contain specified value, user will be rejected instantly, but if it contains
specified value, the coefficient of variation will be calculated by dividing mean by
itself, leading to small weight because it is expected of that feature to have specified
value.
Second step of the process is the calculation of the relative range of values for each
signature feature. Since coefficient of variation cannot be the best measure for the
features with great standard deviation, we have to introduce another measure that will
deal properly with these features. It can be calculated as:


(2)
Where,
RRV
i
Relative range of values of i-th feature,
MAX
i
Maximum value of i-th feature,
MIN
i
Minimum value of i-th feature.
This is also the measurement of the feature stability. The more stable feature is the more
weight it should gain.
Next step in the process is normalizing previously derived measures to interval [0,
1]. This is achieved by calculating values reciprocal, converting all values to strive to
bigger number instead of zero and finally calculating the Euclidean constants for each
derived measure. The normalization is done by dividing each value from the measures
set with corresponding Euclidean constant followed by dividing each value with the
sum of all values for given measure. Since derived measures for each signature feature
are now in the interval [0, 1], we are able to compare each feature measures and
calculate final feature weight.
The last step of the process is calculating final weights of signature features. This is
done by expressing the mean value of the normalized coefficient of variation and
relative range of values. Eq. 3 describes the process mathematically:


(3)
Where,
w(k) Weight of k-th signature feature,
N_AD(k) Normalized coefficient of variation of k-th signature feature,
N_RRV(k) Normalized relative range of values of k-th signature feature.
The obtained weights are used in the authentication process.

3.2 Authentication process
Each signature feature can obtain certain amount of points, depending on its weight.
The maximum points it can obtain is the value of its weight and this is possible only if
the given signature has feature value in defined maximum deviation from mean or
median. If the value is closer to maximum or minimum value from the vector, the
feature can obtain 80% or 90% of maximum points. These margins are determined
empirically, leaving the user possibility to gain some points even if the presented
signatures feature is closer to registered extreme values. How much points can each
statistical measure of one signature feature obtain is described in the following
algorithm:
50

if ((AVG < MEAN and MEAN < MEDIAN) or (AVG > MEAN and MEAN > MEDIAN))
feature100percent = MEDIAN
else
feature100percent = MEAN
if (ABS(feature100percent MIN) < ABS(feature100percent MAX))
feature90percent = MIN
feature80percent = MAX
else
feature90percent = MAX
feature80percent = MIN

The AVG in the algorithm is the mean value of minimum value, median and maximum
value. The ABS denotes the absolute value.
Every feature of the signature acquired for verification can fall in the scope of
defined deviation. The allowed deviations from statistical measures are calculated from
AD for each feature and depend whether feature is closer to minimum or maximum
value. They are shown in Table 1 for mean and median, i.e. for the potential 100%-
feature.

Feature AD Feature
closer to
Allowed deviation from
statistical measure (-)
Allowed deviation from
statistical measure (+)
<0.15
Minimum 1 * STDEV
i
0.5 * STDEV
i

Maximum 0.5 * STDEV
i
1 * STDEV
i

0.15- 0.3
Minimum 0.5 * STDEV
i
0.25 * STDEV
i

Maximum 0.25 * STDEV
i
0.5 * STDEV
i

0.3 - 1
Minimum 0.25 * STDEV
i
0.125 * STDEV
i

Maximum 0.125 * STDEV
i
0.25 * STDEV
i


Table 1. Allowed deviations from mean and median

Similar principle is applied to calculating the allowed deviations for 80%- and 90%-
features. The amount of one half or one quarter standard deviation is allowed.
Authentication is done by comparing acquired signature features with statistical
measures and allowed deviations, giving them 100%, 90%, 80% or 0% of feature
weight. Points for all features are summed and if they pass predefined threshold the user
is authenticated, otherwise it is rejected.
4 Results and discussion
To authenticate user, a predefined threshold must be defined. We tested our system with
several thresholds to obtain system performances. The system was tested with three
measurements: false acceptance rate with unskilled forgeries (FAR-N), false acceptance
rate with skilled forgeries (FAR-S) and false reject rate. The results are presented in
Table 2.

Threshold FAR-N FAR-S FRR
0,57 0,42% 1,47% 42,4%
51
0,5 1,33% 4,98% 27,82%
0,45 4,83% 9,42% 20,45%
0,42 6,57% 11,32% 17,32%
0,38 9,66% 12,24% 12,31%

Table 2. System performance for proposed authentication method

At the first sight, the results of this system might seem unusable in commercial systems
but one must be aware that this authentication method is actually very rigorous and by
higher thresholds achieves reliable false acceptance rate. This authentication method is
very simple and uses only the basic on-line handwritten signature features. To achieve
better results, probably a larger and more descriptive set of features should be extracted.
5 Conclusion
Handwritten signature and its biometric application is one of the most common used
biometric characteristic among researchers in the academic community. We presented a
statistical based approach to signature verification with user oriented signature feature
weights. Our approach uses only the basic on-line signature features and, depending on
defined threshold, achieves satisfying false acceptance rate, but there is still more
research to be done regarding feature extraction and making authentication method
better and more robust to environmental factors.
6 References
[1] Ani, V. et al. Hrvatski enciklopedijski rjenik. In Croatian. Novi Liber, Zagreb,
Croatia, 2002.
[2] Baa M., Koruga P., Fotak T. Basic on-line handwritten signature features for
personal biometric authentication. 2011 Proceedings of the 34th International
Convention MIPRO, pages 116-121, Opatija, Croatia, 2011.
[3] El-Henawy, I.M., Rashad, M. Z., Nomir, O., Ahmed, K. Online Signature
Verification: State of the art. International Journal of Computers & Technology,
4(2):664-678, 2013.
[4] Gupta, G. K., Joyce, R. C. A Study of Some Global Features in On-Line Handwritten
Signature. The International Journal of Automated Identification Technology
(IJAIT), 1(2), 2009.
[5] Lee, L. L., Berger, T., Aviczer, E. Reliable On-Line Human Signature Verification
Systems. IEEE Transactions On Pattern Analysis and Machine Intelligence,
18(6):643-647, 1996.
52

Das könnte Ihnen auch gefallen