Non-Functional Requirements for Offline Signature

Verification with Convolutional Neural Networks

Md. Amzad Hossen – 1620396642, Evan Alam – 1632230642

Abstract— A signature of a person is can use the software without any

simply a handwritten sign or marks that difficulty. For measuring the performance
resembles to his/her name often stylized of system, the neural network was trained
and unique and indicates person’s identity, with dataset of 27 person with each having
intent, and consent. Mostly used for at least 100 signature and training result
different purpose like authenticating was 97 % accuracy. Similarly, to
checks, drafts, certificates, approvals, determine the performance parameter, the
letters, and other legal documents. Since classifier was tested for 100 times and
signature are used in such critical False Acceptance Rate was calculated
activities, confirmation of their 0.1044 and False Rejection Rate was 0.363
authenticity is significantly important. and achieved the accuracy of 89 %. The
Such verification is essential in preventing end result of project was very optimistic
forgery and falsification of documents in and it encourages us for further research
numerous financial, legal, and other and development in this field.
commercial settings. Traditionally
signature was manually compared with
copies of genuine signatures for
verification. This simple method may not I. INTRODUCTION
be sufficient as the technology is becoming The Signature verification task is very
more and more advance and with these critical and often presents difficulties like
advancing techniques of forgeries and high variability i.e. a person’s signature
falsification of signature. So, in order to may vary each time and may change
tackle such problem new efficient tool is completely with age, behavior and
needed and this project proposes such environment, similarities between
signature verification tool which can assist signatures of different person and
human in correct decision making in similarity in duplication or forgery of one’s
authentication of handwritten signature.
signature. Such hinders in authentication of
For such authentication of signature this
handwritten signature can be tackled by
project presents an application software
which facilitates the feature of offline two common ways: online and offline
signature verification using the verification. Offline signature verification
convolution neural network approach. This consists of use of static features of two-
software is able to train the network with dimensional image pixel data acquired
new dataset of signature and validate the from scanning signed documents. Whereas
authenticity of new signature of trained online signature verification consists of
class. User can also perform experiment electronic signing system which results
and analyze the training and verification of dynamic data features such as the speed,
model with features result analysis feature. pressure, pen’s position, azimuth/altitude
This software consists very efficient user angle etc. There have been numerous
interface so that any non-technical person research and studies on different
approaches to verification of handwritten II. STATE-OF-THE-ART
signature, both online and offline, each
approach having its pros and cons [1]. This A. NFRs in Requirements Engineering
project adopts offline verification approach Requirements Engineering (RE) analysis
and for that various approaches can be has long created the argument that eliciting
used like image processing techniques [2], and considering NFRs is essential for the
Fuzzy logic-based approach [3], Statistical success of systems. Such systems can be
approach, Hidden Markov Model [4], technically sound, however fail because of
Support Vector Machine [5] etc. This problems in quality. Such AN argument is
project uses the offline verification method particularly relevant for cc solutions, whose
using convolution neural network for effectiveness lies mainly within the quality
training is performed on the network with of the outcomes they supply.
training dataset of signature and using the What is an NFR? merely, an NFR is any
error difference and accuracies of test quality or attribute which is non-functional.
results of the signature whose authenticity This broad definition, defining something
is needed to be determined. essential in terms of what it's not, isn't ideal,
as has been mentioned by many authors. Our
purpose here isn't to outline NFRs in a very
satisfactory method, however to explore
their application to cc. The idea of quality
has had higher luck in terms of a definite
Figure 1. Superimposed examples of multiple definition, being coated by many prominent
genuine signatures from the same ID, indicating
high intraclass variability (from [10])
ontologies, e.g., DOLCE. more modern add
RE uses ideas from DOLCE to treat NFRs as
qualities over an entity, sometimes a
A. Objectives practical demand, the system, or a system
The objective of this project is to part, e.g., “send mail (entity) quickly
develop a software application which (quality)” or “the system (entity) ought to be
can assists and improves the secure (quality)”. Although qualities of cc
verification of authenticity of signature solutions and NFRs for cc solutions are
by implementing convolution neural similar, technically one will consider NFRs
network machine learning technique. as the requirements over the standard, e.g.,
the standard is usability of system X,
B. Specific Objectives whereas the NFR is “System
 To collect training data and X should be usable”, which ideally ought to
preprocess it for training the be outlined in a very measurable method,
machine learning model. e.g., “90% of take a look at users would rate
 To implement the convolution the system as AN 8/10 in terms of usability”.
neural network and train it though there's a distinction, for simplicity, in
with collected training data. this work we have a tendency to treat NFRs
 To evaluate the performance and qualities as synonyms. [8]
measures of implemented NFR Catalogues. To facilitate a thought of
CNN. NFRs, catalogues of software package
 To develop application’s GUI qualities were created. maybe, the ISO/IEC
and other features and 25010 customary divides system/software
integrate with neural network. product quality into eight classes, as well as
performance potency, compatibility,
usability, and security. every quality is any
decomposed; e.g., compatibility is refined Machine learning is a subsidiary of
into co-existence and interoperability. Such artificial intelligence that facilitates a
catalogues offer unvaried refinement of technique where machine can make decision
NFRs into subqualities, probably sub-sub- based on its experience and improve and
qualities, sometimes all the way down to learn with time and use without explicitly
measurable indicators, once doable. programmed. Machine learning focuses on
Languages and Reasoning. Abundant work the development of computer programs that
focuses on capturing NFRs in visual can access the data and use it to learn for
modeling languages, generally with an themselves. Machine learning algorithms are
underlying metamodel and linguistics, often categorized as supervised and
facilitating (semi-) automatic qualitative and unsupervised machine learning algorithms
quantitative strategies to support call
making. Usually, approaches enable users to
1. Supervised Learning
use NFRs to pick out among doable various
Supervised machine learning is the
practical requirements, given FRs and NFRs,
search for algorithms that reason from
several of that are in conflict, that needs
externally supplied instances to produce
ought to we have a tendency to implement?
general hypotheses, which then make
Runtime, Adaptation, and Evolution. NFR
predictions about future instances [9]. It
approaches were extended to think about a
can apply what has been learned in the past
requirements-based read of runtime system
to new data using labeled examples to
operation, wherever practical and quality
predict future events. Classification
needs could be monitored at runtime,
algorithms like decision tree, Naive Bayes,
supported knowledge from the running
Neural Networks, k-Nearest Neighbors
system. Add this space went further to think
and Support Vector Machines (SVM) are
about requirements-based runtime
some examples of supervised learning
adaptation, a particular quality facet isn't
algorithms.
sufficiently happy at runtime, so the system
can evolve and fits try and gain better
2. Unsupervised Learning
performance or quality, all whereas
Unsupervised learning is the type of
considering quality trade-offs.
machine learning algorithm where there is
Linking knowledge to Quality. A
connected line of labor uses an adaption of no any defined or labeled class and it itself
common needs notations to link business draws the inferences from datasets [10].
data to structure goals, as well as qualities, Unsupervised learning studies how systems
allowing for continuous goal-based business can infer a function to describe a hidden
intelligence. more modern work focuses on structure from unlabeled data. Clustering
the look of knowledge analytic systems for algorithms like K-means, mixture models,
business, which can embrace cc algorithms. hierarchical clustering, anomaly detection,
This work focuses on finding styles that Hebbian Learning Network, Generative
match domain-specific analytic queries, Adversarial Networks etc. are some
considering aspects of quality performance examples of unsupervised learning
for numerous cc choices. during this case, algorithms.
the authors adapt existing RE languages to
think about knowledge analytics at the C. Artificial Neural Network
syntax level, and that they use existing Artificial Neural Networks are composed
analysis procedures while not modification. of multiple nodes, which imitate biological
neurons of human brain. The neurons are
B. Machine Learning connected by links called axons and they
interact with each other. The nodes accept layers (often with a subsampling step) and
input data and perform operations on the then followed by one or more fully
data. The result of these operations is passed connected layers as in a standard MLP.
to other neurons. The output at each node is The architecture of a CNN is designed to
called its activation or node value. Each link take advantage of the 2D structure of an
is associated with weight. ANNs are capable input image or other 2D input such as a
of learning, which takes place by altering speech signal. This is achieved with local
weight values. connections and tied weights followed by
some form of pooling which results in
translation invariant features. Another
Hidde benefit of CNNs is that they are easier to
Inpu Outpu
n train and have many fewer parameters than
t t fully connected networks with the same
number of hidden units.

Architecture of a CNN consists of a

number of convolutional and subsampling
layers optionally followed by fully
connected layers. For an example, in a CNN
where input is image, the input to the
convolutional layer be m x m x r image
where m is the height and width of the image
and r is the number of channels, e.g. an RGB
Figure 1. Three-layer representation of
image has r = 3. The convolutional layer will
ANN
have k filters (or kernels) of size n x n x q
where n is smaller than the dimension of the
Single layer perceptron, Multi layers
image and q can either be the same as the
perceptron, Backpropagation network,
number of channel r or smaller and may vary
Recurrent neural network, Self-Organizing
for each kernel. The size of the filters give
Map (SOM), Hebbian network, Kohenon
rise to the locally connected structure which
network, Hopfield Network, Convolutional
are each convolved with the image to
neural network etc. are the different types of
produce k feature maps of size m – n + 1.
neural network architectures each suitable
Each map is then subsampled typically with
for different purposes.
mean or max pooling over p x p contiguous
Neural networks in practice like sales
regions where p ranges between 2 small
forecasting, industrial process control,
images and is usually not more than 5 larger
customer research, and targeted management
inputs. Either before or after the
etc. Neural networks in medicine like
subsampling layer an additive bias and
modelling and diagnosing the cardiovascular
sigmoidal nonlinearity is applied to each
system, in business etc. are few of the many
feature map
application area of artificial neural network.

1. Convolution Neural Network

A Convolutional Neural Network (CNN)

is comprised of one or more convolutional
 Inp Conv Pool Con2 Pool Hidde protection
Outp and information honesty. In
ut 1 1 v 2 n ut
controlled ventures, enactment or guideline
can give requirements on least principles to
basic NFPs inside the industry. These
Subsam Convolutilimitations might be ordered to give
Subamp Fully
Connect
s le Convoluti ple
on on customer assurances or to oversee
ed
fundamental dangers or negative monetary
Convoluti externalities inside the business. NFPs are
on
Figure 2: Architecture of sample additionally significant in getting
Convolutional Neural Network advancement: NFPs are quality or execution
measurements for innovation, and
mechanical advancement pushes out the
III. NFRS for Signature Verification outskirts of execution on these different
Example measurements. Requests of extent upgrades
in execution on NFP measurements open up
While determining a framework, opportunities for new markets and new plans
programming engineers frequently recognize of action utilizing that innovation
practical requirements from non-utilitarian development.
requirements. For a PC framework,
straightforward functional requirements Whenever any documents is verified on
describe the connection between noticeable the basis of signature in it, person verifying
data sources and outputs. Non-functional it is taking a great risk and he/she should be
requirements (NFRs) are needs absolutely certain of his/her decision. The
communicated for non-functional properties validation of signature in many cases are
(NFPs), which are otherwise called highly critical and any inaccuracy in the
'characteristics', or 'ilities'. These include authentication may result serious
characteristics, for example, cost, security consequences and damages. With the
(secrecy, respectability, accessibility, advancement in technology, new and
privacy, nondisavowal), execution (latency, complex forgery and fraud techniques are
throughput), modifiability, and ease of use. emerging. In order to avoid such scenario
NFRs are communicated independently from and prevent potential damage, modern robust
functional requirements since they are often approach must be adopted in verifying the
'cross-cutting worries' that range numerous genuineness of signature. Adopting such
framework capacities. For instance, a approach will assists person in making
requirement for the versatility of framework decision over authenticity of signature and
execution may compel the resources that can prevents mistakes.
be utilized to react to a given degree of
simultaneous interest in an opportune
manner, up as far as possible on that request. IV. Challenges
The interest right now typically be a blend of
a wide range of sorts of framework works in In the past decade, there have been ample
ordinary use. Distinctive use cases convey amount of research in the field of pattern
diverse NFRs. For instance, in security basic recognition and also in the field of offline
ventures, for example, clinical gadgets or signature verification. A bunch of solutions
aviation frameworks, NFRs for wellbeing has been introduced, to overcome the
are paramount. In endeavor programming limitations of off-line signature verification
frameworks, administrative prerequisites and to compensate for the loss of accuracy.
regularly oblige NFPs, for example,
Researchers come across two problems in of the curves. In the same way, angled letters
offline signature verification – (i) Most of in the original signature can become smooth
the dynamic information in the signature is curves.
lost and (ii) Low quantity of available 3. Retouching: Many times the forger
signature samples versus high number of makes correction at later stage after the
extracted features. imitation has already been. Due to this
retouching, in the forge signature, lines may
The first issue is addressed by some appear to be thicker at these points, or there
researchers but may be lines that do not follow the continual
this is still a challenging problem. flow of the pen as in the original signature.
4. Poor line quality: The pressure put on
Luana Batista et al. [1] have mentioned some
the paper by the pen is not same in case of
remedies for the second issue; they are –
forged and original signature. It is found that
1. Select the most discriminating features
the pressure used for the questioned
2. Use regularization techniques to obtain a signature is harder than that of the real
stable estimation of the covariance matrix
signature. The ink reveals variation in light
3. Generate synthetic samples and shade, pressure and speed, with either
4. Use dissimilarity representation more or less ink appearing on the page.
However, in a forged signature sometimes a
lighter pressure can be detected. But this
IV.1 Characteristics of forgeries may cause a tremor caused by trembling of
the hand, poor line quality, or writing too
In offline signature verification, some slowly [23].
general characteristics of genuine signatures 5. Hesitation: In the process of creating a
and forgeries need to be understood. forgery, the forger may pause to consult the
Knowledge of these characteristics is genuine signature and then continue
important for determining those aspects or duplicating it. This can often create blobs.
features of the signatures that are most 6. Punctuation: In forgery full stops, dots
important for automatic signature on small letter „i‟ are found to be in the
verification. In [7] Vamsi Krishna Madasu wrong place, missing or added.
and Brian C. Lovell have mentioned few 7. Differing pressure: It is hard to vary pen
such characteristics outlined by several pressure in the same way as a genuine
document examiners in the past in various signer. Forger cannot imitate identical pen
literatures: pressure profile as like as the genuine author.
1. Enlargement of characters: A forgery is The pen pressure may be too heavy or too
usually larger than the original signature. As light, depending on the style of the forger.
compared to the original author, a forger Pressure differences occur at different places
takes more time drawing each letter in the from the genuine signature.
signature. This makes a forgery larger than 8. Sudden endings: Sudden endings are a
the original both in terms of the size of characteristic feature of a forgery. It is seen
letters and the size of the entire signature. that in many cases the original signature
2. Tendency of curves to become angles: trails off, but the forgery just stops. It is very
Curved letters are often observed in the difficult to trail off in the same way as the
forgery as being more angular. The forger genuine.
takes care to obtain the correct letter shape 9. Forger’s characteristics: Everyone has
by using a slower speed to produce the curve his/her own characteristics of handwriting.
accurately. This results in more angular The forger unconsciously exposes his/her
letters as greater time elapses in the making own handwriting characteristics when doing
the forgery. It is observed that forger cannot
avoid revealing some of his/her writing Data Collection and Pre-processing
characteristics like the basic letter shapes, In order to train the classification model,
spacing and position of letters in relation to signature data which in this case was images
base line even in a forgery. of signature specimens was collected. About
10. Baseline error: The imaginary line that 30 students of Kathford international college
runs across the base of the signature is not of engineering and management provided
similar in the forged signature and the their signatures dataset in hard copy. Each
genuine signature. The baseline in a dataset consisted about 100 signature
signature is not horizontal and any notable specimens of person providing it. These data
variances in the baseline indicate forgery. in hard copy was scanned into computer and
11. Spacing: Imitating the spacing between preprocessing was done on them with
individual letters, whole words, and between cropping the specimens, cleaning and scaling
punctuation and letters is difficult. These to appropriate size, gray scaling and giving
spacing may be larger or smaller that cannot proper name and putting them in separate
be copied by tracing a signature. directory.
12. Bad line quality: A slow forgery results
hesitant or shaky pen strokes and domino Implementation of Convolutional neural
effect is bad line quality. network and training
13. Forming characters not appearing in Convolutional neural network for this
signatures: When the forgers know the name project is based on the Inception architecture
of the author of genuine signature that they of Google Net [13]. For the project,
are trying to forge, unconsciously they implemented neural network is Inception-v3
include letters in the forgery that do not trained tensor flow model. Training an
actually appear in the genuine signature. On Inception-v3 network from scratch is a
the other hand, if the forger is unsure about computationally intensive task and
the name, then incorrect letters may appear depending on computer setup may take
clearly in the forgery. several days or even weeks which is not
Even though the above-mentioned points possible with limited resources. In order to
will help detecting forgery, it is very difficult overcome this problem transfer learning
to apply most of these points to mechanism was adopted and retraining was
computerized signature verification. performed on the Inception v3 model, which
is trained for the ImageNet Large Visual
Recognition Challenge using the data from
V. RESEARCH DIRECTIONS 2012, with collected training data.

The system consists of two modules core:

implementation of convolutional neural
network and cross platform desktop
application with various functionalities like
signature verification, training features for
new signature classes and training result
analysis. At the initial phase of project,
application development and CNN
implementation were done separately and
integrated together to single system later in
project.
 Figure 3: Inception v3 Model aspect like technical, economical, and
layers social feasibility etc. were realized. After
the feasibility study it was decided that it
The top layer receives as input a 2048- was impractical to reinvent the whole
dimensional vector for each image. neural network architecture and train it
Training is done on a softmax layer on top from scratch to perform required job so
of this representation. Assuming the implementing existing neural network with
softmax layer contains N labels, model transfer learning technique was the
corresponds to learning N + 2048*N suitable option. Similarly it was realized
model parameters Corresponding to the that task of acquiring many signature
learned biases and weights. Training is specimens from individual person was
done with collected data of signatures and very risky but essential aspect of the
training result was analyzed until project and was to be accomplished
satisfactory result was achieved. trustily.
The application is based in object
Application development and system oriented analysis and design approach so
modules integration various object oriented diagrams and
Desktop application was developed with system designs like dataflow diagrams, use
python programming language and case diagrams, sequence diagrams, class
anaconda python data science platform. diagrams etc. were designed which laid the
For graphical user interface, python’s de- foundations for system implementation.
facto standard GUI package Tkinter was System designed were implemented with
used. The convolutional neural network Python programming language and
module was integrated to main application. anaconda platform consisting different
Other features such as verification of libraries numpy, tkinter, tensorflow, thread
signature, training features for new classes etc.
of signature holder, analyzing results with In testing phase, integration testing was
graph representation etc. were integrated. done for integration of two modules of
system i.e. neural network classifier
Development Model module and application module to test the
This project adopted agile software function of whole system after integration
development approach as this software and eliminate present errors. Performance
development model is very adaptable and and functionality testing was done in order
gives project members flexibility detect and eliminate bugs and errors in
throughout project duration with support system. These detected bugs and problems
for rapid and flexible response to change were solved with changes in codes and
[15]. sometimes in changes in system
configuration and even system design
when needed, exploiting the iterative and
continuous improvement nature of agile
development model.

Figure 4: Agile Development VII. Conclusion

Methodology
At the analysis stage, fundamental This project will experiment and
requirements were drafted and feasibility implement the signature verification task
with utilization of latest and powerful International Journal of Computer
convolutional neural network available Applications, vol. 10, 2010.
today. This project will not only experiment [5] B. Zhang, "Offline signature
the classification or verification of offline verification and identification by hybrid
signature but also will propose an original features and Support Vector Machine,"
application software for experimenting with International Journal of Artificial
new dataset of signature and training with Intelligence and Soft Computing, vol. 2, pp.
that dataset for further new verification 302-320, 2011.
problems. The final outcomes of the project [6] A. Pansare and S. Bhatia, "Off-line
Signature Verification Using Neural
will be very optimistic and it’ll genuinely
Network," International Journal of Scientific
encourage us for further research and
Engineering Research, vol. 3, no. 2, 2012.
development in this field. The performance
[7] G. Alvarez, B. Sheffer and M.
of the developed software will be Bryant, "Offline Signature Verification with
encouraging, but the absence of online Convolutional Neural Networks,"
verification technique will be realized as the Stanford.edu, Stanford, California, 2016.
inclusion of such dynamic features of online [8] M. Shirdhonkar and M. Kokare,
verification technique like pen’s speed, "Off-Line Handwritten Signature Retrieval
pressure, azimuth angle etc. would have using Curvelet Transforms," International
significantly improved the verification Journal of Computer Science and
performance and in coming future we are Information Security, vol. 8, 2010.
very eager to work on that. After the [9] S. B. Kotsiantis, "Supervised
conclusion of this project we are very Machine Learning: A Review of
optimistic that will immerge many Classification Techniques," Department of
wonderful outcomes and possibilities in this Computer Science and
field in the coming future. Technology,University of Peloponnese,
Peloponnese, Greece, 2007. 26
[10] P. Dayan, "Unsupervised Learning,"
The MIT Encyclopedia of the Cognitive
References Sciences, Massachusetts, 1999.
[11] M. T. Hagan, H. B. Demuth and M.
H. Beale, Neural Network Design,
[1] Surabhi Garhawal and N. Shukla, "A
Oklahoma, 2014. [12]
Study on Handwritten Signature Verification
[12] C. Stergiou and D. Siganos,
Approaches," International Journal of
"NEURAL NETWORKS," SURPRISE 96
Advanced Research in Computer
Journal, vol. 4, 1996.
Engineering & Technology (IJARCET), vol.
[13] L. Torrey and J. Shavlik, "Transfer
2, no. 8, 2013.
Learning," University of Wisconsin,
[2] W. Hussein, A. S. Mostafa and O.
Madison WI, USA, 2008.
Ibrahim, "Image Processing Based Signature
[14] N. Wang, L. W., J. Y., S. P., R. S., .
Verification Technique to Reduce Fraud in
A. D., . E. D., V. V. and R. A., "Going
Financial Institution," MATEC Web of
deeper with convolutions," Proceedings of
Conferences, p. 76, 2016.
the IEEE, pp. 1-9, 2015.
[3] M. Hanmandlu, "Neuro-Fuzzy
[15] K. N. Rao, G. K. Naidu and P.
Approaches to Signature Verification,"
Chakka, "A Study of the Agile Software
2006.
Development Methods, Applicability and
[4] D. S. A. Daramola and P. T. S.
Implications in Industry," International
Ibiyemi, "Offline Signature Recognition
using Hidden Markov Model (HMM),"
Journal of Software Engineering and Its
Applications, vol. 5, 2011.