International Journal for Multidisciplinary Research (IJFMR)

E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

Advancing Security: Machine Learning-Based

Signature Forgery Detection in Document
Authentication Systems
Sandesh Kandel

Data Analyst, VIT University

Abstract
Handwritten signatures play a vital role in our lives. From banks to institutions to organizations,
signatures are a way of identifying a person. However, signings come with a lot challenges because any
two signatures can look very similar with slight differences written the same person. Therefore, the
identification of real and fake signatures is very difficult. To avoid similar identity related crimes
committed in banks and many others companies, the counterfeit detection system is the solution to this
problem along with the help concepts of machine learning and CNN. For better performance and time
efficiency, Parallelization concepts are used in software implementation. This software can be used to
verify signatures on many platforms such as loans, signing legal documents, applications signing,
applications and much more.

Keywords: Crucial, Banks, Organization, Forgery, CNN, Forgery, Signature, Frauds.

Introduction
The aim of this project is to improve the detection of offline signature forgeries. In this project we
analyzed and used machine learning concepts to classify, identify and also differentiate between the fake
and the original signature. In this project implemented on Jupyter, we ensured the implementation of
convolution neural network model using TensorFlow and created the model rather than traditional ways
building a CNN model. We have uniquely implemented feature extraction based on various geometric
factors of the signature image originating from the image dataset. Nowadays, a handwritten signature is
one of the most widespread personal attributes proofs of identity, whether from the banking or business
sector. People from lower society prefer to write their signatures in free handwriting due to lack of
education and knowledge. Therefore, these types of signatures can be easily forged under certain
circumstances. In this In this case, four types of fakes are possible.

Simulation Forgery
In which the forger has a sample of the signature to be forged. The quality of a simulation depends on
how much the forger practices before attempting the actual forgery, the ability of the forger, and the
forger’s attention to detail in simulating the signature. Based on a forger’s experience, known forgeries
are classified as unskilled and skilled forgeries

IJFMR23069039 Volume 5, Issue 6, November-December 2023 1

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

Unknown/Random/Blind Forgery:
This is when the forger has no idea what the signature to be forged looks like. This is the easiest type of
forgery to detect because it usually does not have the appearance of a genuine signature. This type of
forgery will sometimes allow an examiner to identify who made the forgery based on the handwriting
habits that are present in the forged signature.

Tracing:
The third type of forgery is tracing. Tracing can be done by holding the model document and the
questioned document up to light and using a pen to trace the lines of the model signature onto the
questioned document. A tracing can also be done by using a blunt stylus on the questioned document to
create an impression of the model signature on the paper. This impression is then filled in with a pen to
create the appearance of the model's signature. If the model signature used by the forger is not found, this
type of forgery is sometimes difficult to detect from a photocopy.

Optical Transfer:
It is one in which a genuine signature is transferred onto a document by the use of a photocopier,
scanner, facsimile machine, or photography. With this type of forgery, an examiner cannot positively
identify a signature as genuine without having the Original for comparison.

Objective
The objective of the software is:
● to verify if a signature is forged or original.
● to ensure the authorized use of confidential information.
● to detect any impostor trying to access any important information.

Background:
For this system, the key concept will be the convolutional neural network (CNN). The CNN will be
trained against a dataset containing many signatures such that it will have the skill to predict certain
features and find out whether a forgery has been committed. or not. We aim to bring up software that
verifies signatures and makes sure the software is more reliable, efficient, and 5% accurate than existing
systems. Signatures vary with time when a person becomes old. There are certain factors which lead to
changes in the signature, and these changes cannot be identified by ordinary people. One major concern
that must be kept in mind is that the system should not be given to people randomly. Only those who
have permission and are authorized to use this system should be taken care of since it is very
confidential. This software can be used to validate signatures across many platforms, like loans, legal
document signing, application signing, applying and a lot more. There are many organizations that have
lost tremendous amounts of money due to a single forgery, and being able to detect even a single forgery
can save money, time, and the reputation of the organization. The system will be run through a web page
since it is more efficient and easier to use. Keeping in mind the people who have low technical skills or
knowledge, the system should be easy to use and not require a lot of tasks to be done since it is time-
consuming.

IJFMR23069039 Volume 5, Issue 6, November-December 2023 2

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

Problem Description:
Online (dynamic) signature verification uses signatures that are captured by pressure-sensitive tablets
that extract dynamic properties of a signature in addition to its shape. Dynamic features include the
number and order of the strokes, Typical Signature Verification System, the overall speed of the
signature, the pen pressure at each point, etc. and make the signature more unique and more difficult to
forge. In an online signature verification system (Figure),

Fig 1: Typical signature Verification System

The users are first enrolled by providing signature samples (reference signatures). When a user presents
a test signature claiming to be an 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 must 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
decide. The single dissimilarity value can be obtained from the 6 minimum, maximum, or average of all
the distance values. Typically, a verification system chooses one of these and discards the others. In
evaluating the performance of a signature verification system, there are two important factors: the FRR
of genuine signatures and the FAR of forgery signatures. As these two errors are inversely related, the
EER where FAR equals FRR is often reported.

Section to describe or introduce new terms to the readers:

convolution neural network (CNN): A convolutional neural network (CNN) is a type of artificial neural
network used in image recognition and processing that is specifically designed to process pixel data.
equal error rate (EER): The EER is the location on a ROC or DET curve where the false acceptance rate
and false rejection rate are equal. Tensorflow: TensorFlow is a free and open-source software library for
machine learning and artificial intelligence. It can be used across a range of tasks but has a particular
focus on training and inference of deep neural networks.

IJFMR23069039 Volume 5, Issue 6, November-December 2023 3

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

LITERARY REVIEW:
According to recent studies, check fraud alone costs banks about $900 million annually, with 22% of all
check fraud attributed to signature fraud. Clearly, with more than 27.5 billion checks written each year in
the United States, visually comparing signatures with manual effort on the hundreds of millions of
checks processed daily proves [Link]: Authentic signatures of the same person will be exactly
alike in all transactions
Reality: The physical act of signing requires brain, eye, arm, finger, muscle and nerve coordination. With
all the factors at play, it's no wonder people don't sign exactly the same every time: some elements can be
left out or altered. Personality, emotional state, health, age, conditions under which individual
characters, space for signature and many other factors all affect the variation between signatures.
Types of signature forgery:
In real life, signature forgery is an event in which the forger focuses primarily on accuracy rather than
fluency. The range of signature forgeries falls into the following three categories:
1. Random/Blind Forgery — Usually bears little or no resemblance to genuine signatures. This type of
forgery occurs when the forger does not have access to an authentic signature.
2. Unqualified (tracing) forgery: The signature is traced and appears as a faint indentation on the sheet of
paper below. This indentation can then be used as a guide for the signature.
3. Skilled forgery — Made by an offender who has access to one or more specimens of a genuine
signature and can imitate it after much practice. A qualified forgery is the most difficult to verify of all
forgeries.
The goal of an accurate verification system is to minimize both types of errors. Characteristic features:
Let's understand the signature features for a human examiner to distinguish fraud from genuine. The
following is a non-exhaustive list of static and dynamic characteristics used for signature verification:
• Shaky handwriting (static)
• Pen lift (dynamic)
• Retouch marks (static and dynamic)
• Letter proportions (static)
• Signature shape/size (static)
• Slope/Angle (static)
• Very close similarity between two or more signatures (static)
• Speed (dynamic)
• Pen pressure (dynamic)
• Patterns of pressure change (dynamic)
• Acceleration pattern (dynamic)
• Smoothness of curves (static)
Based on the verification environment and sampling conditions, not all features are available for
analysis. After reading these articles, we were clear about the following topics and how to implement an
optimal signature forgery detection model.

Our studies included:

Techniques such as four different CNN feature extractors have been created. Each CNN consisted of 2
convolution layers and 2 maximum pooling layers. Convolutional Neural Network (CNN) for

IJFMR23069039 Volume 5, Issue 6, November-December 2023 4

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

implementing offline authentication methods, CEDAR Signature dataset for neural network training
Convolutional Siamese Network (CSN), pooling layer, triple loss classification. I personally liked the
CSN algorithm because it is the most optimal and efficient. Another important aspect was the assessment
methods The number of hidden layers was 3 for the first paper and the number of nodes was 30, 20 and
30 for each layer. The training epoch was 500. The equal error rate (EER) is used as an evaluation metric
for signature verification. All machine learning models were implemented using Theano library, a well-
known open-source machine learning library. MLP revealed a weakness in the qualified forgery test in
the case of the second paper due to the fact that it is a 2-class model. AE and CNN_B-AE did not work
well. This result shows that the S-vector generated by the trained CNN is more efficient than the raw
signature data. CNN_D-AE showed the best result and CNN_AC- AE showed the second best accuracy.
However, note that CNN_D-AE is not a practical model. In the third document, the Dataset contains 30
users, all of which have 15 signatures, the test starts with learning 0.00000 infrequently, and the margin
is set to 0.2 0.2, the best test result is low loss and high accuracy, we end the analysis at step 180 ,
because we perceive the target result. So the accuracy is 84. There are two types of pool: maximum pool
and average pool. Max Pooling returns the maximum value from the image split reported by the kernel.
But average returns the average of all values from the part of the image that the kernel covers. To
improve the accuracy of signature verification, some studies [8,9,10] use machine learning techniques,
which are one of the most notable technologies. Buriro et al. [3] used a multilayer perceptron (MLP), a
two-class classifier, to verify a finger-drawn signature with dynamic features involving finger and phone
movements. Their method showed a verification accuracy of 94.8% for subject and other object
classification. However, this technique does not represent the ability to distinguish forged signatures.
Two-class classifiers such as MLP are at greater risk of misclassifying a forged signature as a subject
because forged signatures closely resemble the subject [10,14]. Although the signature marked on the
smartphone screen disappears immediately after verification, it is easy for an adversary to imitate the
signature through shoulder surfing and smearing, which involves tracking the smudge left on the screen
[11]. Therefore, it is necessary to improve the authentication for distinguishing forged signatures to
provide secure services. An important issue is also the difference in time between the registration of a
signature and the verification of a new signature. Unlike biometrics such as fingerprints and irises,
behavioral patterns can change over time. Although [15] focused on the problem of this time difference,
they worked on PIN verification and not dynamic finger-drawn signatures. In the first paper, we propose
a new dynamic approach to fingerprint signature verification that provides better accuracy against forged
signatures and time-delayed signatures. The proposed method uses two deep learning algorithms: a
convolutional neural network (CNN) [16,17] for feature extraction and an autoencoder (AE) [18] as a
classifier. CNNs are trained to distinguish forged signatures from genuine signatures, and the trained
CNNs are used for feature extraction, not as a classifier. Specifically, the output of the CNN
intermediate layer, which we call the S-vector, is used as input to the AE to create the subject model.
Since CNN is known to be able to extract features for classification by itself [17], we hypothesize that a
CNN trained for a specific purpose could extract features effective for that purpose. For example, a CNN
trained on forged signatures can extract features that are common in forgery, such as hesitation and
delay before drawing a complicated part of the signature. The proposed method uses AE as a classifier
due to its high accuracy in solving single-class discrimination problems such as user authentication. Our
previous work [19] and M. Fayyza et al. [10] showed that the one-class AE model is better at

IJFMR23069039 Volume 5, Issue 6, November-December 2023 5

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

distinguishing simulated signatures than the two-class model. However, AE is also highly dependent on
the accuracy of the input data [14]. The S-vector is valuable as an AE input and could subsequently lead
to an increase in the accuracy of dynamic signature verification. Although the main problem is forgery
signature discrimination, the proposed method should consider other problems, i.e. time difference
problem and classification of subject/others. They are also important issues for a signature verification
service, and a single-purpose model could reduce the performance of other issues. The proposed method
achieves better performance for time-varying signatures and subject/other classification using an
experimental approach.
The main contributions of this post are as follows.
• To the best of our knowledge, the proposed scheme is the first CNN-AE model for hand-drawn
signature verification in mobile environments.
• Experiments using real user signatures show that S-vector achieves better accuracy for dynamic
signature verification. The proposed scheme reduces the same error rate (EER) by 1.1% (subject/other),
3.2% (time difference) and 13.7% (qualified forgery) compared to previous work.
This document is organized as follows. Section 2 introduced the proposed method. In Section 3, the
proposed model is evaluated based on experimental results, where the CNN was trained with four
different classification datasets to determine the most efficient S-vector. Section 4 concludes the paper.
this includes methods for acquiring, processing, analyzing and understanding images and, in general,
high- dimensional data from the real world to produce numerical or symbolic information, e.g. in the
form of decisions [5]. Computer vision covers the basic technology of automated image analysis, which
is used in many fields [6]. Computer vision as a scientific discipline deals with the theory behind
artificial systems that extract information from images. Image data can take many forms, such as video
sequences, views from multiple cameras, or multidimensional data from a medical scanner. As a
technological discipline, computer vision tries to apply its theories and models to the construction of
computer vision systems [5].
signatures can be used to identify and authenticate the subscriber. An automated verification process
would allow banks and other financial institutions to significantly reduce check and money order
forgeries, which represent large monetary losses every year. Reliable signature verification can be of
great help in many other application areas such as law enforcement, industry, security control, and so on.
Handwritten signatures appear on many types of documents, such as bank checks and credit slips, etc.
[7][12]. A large number of such documents require automatic signature verification. A signature
verification system requires high reliability.
Images are collected for training and stored in a database. Images are collected by scanning from a
physical paper source. The database used is a self-created database that contains the signatures of three
different people. The database consists of fifteen signatures belonging to each person, for a total of forty-
five signatures. More signatures can be easily added to the database and the number of signatures per
person can be increased or decreased.
After the image has been pre-processed, various features are extracted from the image. The extracted
features from each image are then stored in a MATLAB file. The following unique features are extracted
from each image: [1] Height to Width Ratio: After cropping the image, the height to width ratio of the
signature is calculated. [2] Centroid of Signature: The center of gravity or barycenter of the image is
calculated. The centroid indicates the central point of the signature, which is a unique characteristic of

IJFMR23069039 Volume 5, Issue 6, November-December 2023 6

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

the signature. The signature is divided vertically into two halves and the centroid of each half is
calculated. [3] First derivative: The first derivative of the image matrix is calculated rowwise and
columnwise. [4] Second derivatives: After the calculation of the first derivatives, the second derivatives
of the image matrix are calculated by rows and columns. [5] Quadrant areas: The image is divided into
four quadrants and then the area of the signature pixels in each quadrant is calculated. This area is the
area of the signature strokes in that quadrant and does not include the background area. [6] COM Matrix:
COM Matrix or Co-Occurrence Matrix refers to the distribution of co- occurring values at a given offset.
It is used to measure texture in an image. Since our image is black and white after preprocessing, this
means that the image matrix has values of either 0 or 1. It looks for a pattern distribution of these values
and looks for where the patterns 00, 01, 11, and 10 occur. The co-occurrence matrix is also calculated
for the signature. [7] Calculation of edge points: The number of edge points in the signature is
calculated, which gives a distinctive characteristic of the signature. [8] Horizontal and vertical
histogram: Each row and each column of the signature is traversed and the number of black pixels is
calculated. The row and column with the maximum number of black pixels are recorded and used as a
function. All of these properties provide unique signature characteristics and are used for signature
classification.
A convolutional neural network (CNN) is a multi-layer neural network with a deep supervised learning
architecture known to extract features for classification by itself [17]. A CNN consists of two parts: an
automatic feature extractor and a trainable classifier. The feature extractor extracts features from the
input data using two operations: convolutional filtering and downsampling. Based on these features, a
trainable classifier is trained using a fully connected layer backpropagation algorithm to produce
classification results.
The proposed method uses CNN only as a feature extractor, not as a classifier.

Fig 1.2: CNN Classifier

The S-vector extracted by the CNN feature extractor is used as the input of an AE. An AE is a type of
deep neural network that has the same dimensions for input and output [14]. In the training phase, the
AE is trained by using the same data (sample signatures) as input and output. In the test phase, the

IJFMR23069039 Volume 5, Issue 6, November-December 2023 7

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

trained AE generates an output corresponding to an input (test signature). An AE can generate highly
similar output for trained data patterns, whereas it does not for unfamiliar data. Therefore, a test
signature is verified by a similarity comparison between the test signature and the output of the AE.
Consequently, an AE is used in modeling a subject for authentication.

Paper-4 Abstract:
Hand signatures are very important in our social and legal life for verification and authentication. A
signature can only be accepted if it is from the intended person. The probability that two signatures of
the same person are the same is very less. Many characteristics of a signature can differ even if two
signatures are created by the same person. Detecting a forgery thus becomes a challenging task.[4] In
this paper, a convolutional neural network (CNN)-based solution is presented where a model is trained
using a dataset of signatures and predictions are made whether a given signature is genuine or forged.

Method Proposed:
Keras library with Tensorflow backend is used to implement CNN. The directory of processed images is
loaded and then we train the model with different training and testing ratios to evaluate the
performance.[4]

Evaluation Methods:
The first raw RGB images are converted to grayscale and binarized. File management operations are
then performed to split the images into batches based on the split ratio. Training and validation
accuracies are plotted for all split ratios, and the split that gives the best results is considered.[4]

Paper-5
Abstract:
The signature capture and recognition system takes a signature image as input and trains the image by
extracting various features and stores it in a database, then it will be compared with the original source
signature using convolutional neural networks and recognize whether it is the original signature.
Algorithms such as grayscale and binarization are used for feature extraction. Once the image is
captured, it will be converted to a black and white image and then processed. This system needs to be
trained very well in order to have better results. Sample signatures will be fed into the system for
identification tests to maintain high accuracy in the system.[5]

Method Proposed:
The signature image is passed to image processing, where image enhancement, geometric
transformation, etc. are applied. During feature extraction, local and global features are extracted. After
extracting the feature, the CNN is applied to the image for comparison and then to the image with the
image in the database. The image passes through a convolutional layer, a pooling layer and a fully
connected layer. In this system, given a set of genuine signatures, the goal is to learn a model that can
distinguish between genuine signatures and forgeries. The most common classification of forgeries in the
literature is accidental forgery, where a person uses his signature to impersonate another person. Data
acquisition Preprocessing Feature extraction Comparison using the CNN algorithm File management[5]

IJFMR23069039 Volume 5, Issue 6, November-December 2023 8

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

Evaluation Methods:
First, a test signature is recognized with a given input training set using both CNN and Crest-Trough
methods. Then, forgery detection algorithms (Harris Algorithmic followed by Surf Algorithm) are
enforced on this classified image. The results from each algorithm are then compared. The signature
capture and recognition system will take the signature image as input and train the image by extracting
various features and store it in the database, then it will be compared with the original source signature
using convolutional neural networks and recognize whether it is the original signature. Algorithms such
as grayscale and binarization are used for feature extraction. Once the image is captured, it will be
converted to a black and white image and then processed. This system needs to be trained very well in
order to have better results. Sample signatures will be fed into the system for identification tests to
maintain high accuracy in the system. Feature extraction is an important stage where the features of each
signature are captured using a CNN algorithm. The idea behind this step is to identify every little detail
of the signature. Subsequent identification of features and their proper extraction will lead to better or
more accurate verification. A centralized database of correct customer signatures will be available. This
particular database may be used by many systems that require customer information and signatures. This
proposed system is focused on bank check signature verification system using artificial neural network.
Signatures are verified based on parameters extracted from the signature using various image processing
techniques. In detecting the exact person and providing more accuracy of signature verification for the
implementation above, this paper uses convolutional neural networks to recognize and verify signatures
of individuals.[5]

Paper-6
Abstract:
This paper presents an innovative approach for signature verification and forgery detection based on
fuzzy modeling. The signature images are binarized and resized to a fixed size window and are then
thinned. The thinned image is then partitioned into a fixed number of eight sub-images called boxes.
Signature verification and forgery detection relate to the process of verifying signatures automatically
and instantly to determine whether the signature is genuine or forged. There are two main types of
signature verification: static and dynamic. Static, or off-line verification is the process of verifying an
electronic or paper signature after it has been made, while dynamic or on-line verification takes place as
a subject creates his signature on a digital tablet or a similar device.[2]

Method Proposed:
After the binarization and thinning of images, the thinned image is then partitioned into a fixed number
of eight sub-images called boxes. This partition is done using the horizontal density approximation
approach. Each sub- image is then further resized and again partitioned into twelve further sub-images
using the uniform partitioning approach. The features of consideration are normalized vector angle
(alpha) and distance (gamma) from each box. Each feature extracted from sample signatures gives rise to
fuzzy sets. Since the choice of a proper fuzzification function is crucial for verification, the authors have
devised a new fuzzification function with structural parameters, which is able to adapt to the variations
in fuzzy sets. This function is employed to develop a complete forgery detection and verification system.

IJFMR23069039 Volume 5, Issue 6, November-December 2023 9

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

Since the main thrust here is to establish the genuineness of the signature thereby detecting the forgeries,
we go in for fuzzy modeling of angle features. For the purpose of signature verification and detection of
forgeries, we have employed the Takagi-Sugeno model. In this, we are following the same concept as
outlined in for considering each feature as forming a fuzzy set over large samples. This is because the
same feature exhibits variation in different samples giving rise to a fuzzy set. So, our attempt is to model
the uncertainty through a fuzzy model such as the TS model.[2]

Evaluation Methods:
Inherent variation is used to judge the test signatures. For a particular feature, if the membership value
lies within the range of variation, which is given by the difference of minimum and maximum
thresholds, it is counted as ‘true’. The total number of ‘true’ cases for a particular signature is divided by
the total number of features to get the percentage. The skill-forged and unskilled forged signatures have
corresponding figures of 88.5% and 82.3% respectively. The minimum limit or acceptable percentage
for genuine signature is set at 91% referring to the output result of signature of signer. Signatures that
have percentage less than 91% are treated as forged signatures. modified SPTA thinning algorithm- a
proposed algorithm for thinning binary patterns, Pseudo- Bacterial Genetic Algorithm (PBGA) - The
PBGA was proposed by the authors as a new approach combining a genetic algorithm (GA) with a local
improvement mechanism inspired by a process in bacterial genetics. The proposed fuzzy modeling based
on TS model discussed above has been applied on a signature database, developed in the Graphics
Visualization & Games Development (GVGD) lab at the Multimedia University, Cybejaya, Malaysia.
The efficacy of this system has been tested on a large database of signatures. The verification system is
able to detect all types of forgeries: random, unskilled and skilled with utmost precision.[5]
From our work, we have an operating convolutional network that can successfully recognize signatures
from 40 different individuals in 54% of the cases. Although the accuracy is not satisfactory for a real
application, we attained a result that is significantly better than a random draw (which would have an
accuracy of 2.5%). Based on these results, we consider our model to have learned a fair amount from the
data provided. The low accuracy obtained in the training set relative to that of the test set is an evidence
of overfitting. This result motivated us to try a few alternatives to reduce the gap between accuracies of
training and test sets. These measures include: Optimizing the regularization term in the loss function,
which did help to slightly improve the accuracy of the algorithm. We reached a point, however, where
any increase of the hyperparameter lambda led to lower training and test set accuracies, and a decrease of
it led to an increase in the gap between both accuracies[18]. Modifying the size of the network.
Increasing the size of the network by expanding the number of neurons in the hidden layers only
increased overfitting. On the other hand, decreasing the number of neurons in hidden layers did reduce
overfitting, but at the cost of reducing the accuracy of the whole algorithm. We then chose to maintain
the network structure. Stopping early. We also thought of stopping the training earlier (by epoch ~25, for
example), to try to prevent the model from overfitting.[19] However, this did not have the desired effect
and the accuracy of both the training and test sets decreased when attempting thi9s algorithm. Although
the testing accuracy is not satisfactory for real signature recognition application, we consider our model
to have learned a fair amount from the data provided and performs significantly better than a random
draw. As a future work, a further exploration of the overfitting challenge identified would be key to
improving the training accuracy of the CNN network. Additionally, additional modifications to the
algorithm such as adding Batch Normalization or additional pre-processing steps could be explored, to

IJFMR23069039 Volume 5, Issue 6, November-December 2023 10

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

investigate the impact on the model performance. Finally, memory limitations should be addressed to
incorporate more training data to the model.[12][19]
Inherent variation is used to judge the test signatures. For a particular feature, if the membership value
lies within the range of variation, which is given by the difference of minimum and maximum
thresholds, it is counted as ‘true’. [4][5]The total number of ‘true’ cases for a particular signature is
divided by the total number of features to get the percentage. The skill-forged and unskilled forged
signatures have corresponding figures of 88.5% and 82.3% respectively. The minimum limit or
acceptable percentage for genuine signature is set at 91% referring to the output result of signature of
signer. Signatures that have percentage less than 91% are treated as forged signatures. [2][4]
Each signature will have a rule so we have as many rules as the number of features. The fuzzy set Ak is
represented by the above exponential membership function where x, is the mean sigma is the variance of
kth fuzzy set. The inclusion of parameters will help track the variations in the handwriting.

When sk = 1 and tk = -1, the membership function is devoid of structural parameters

PROPOSED MODEL:-
Architecture:

Fig 2: Architecture

In this work, signature images are preprocessed batch by batch and split into training and test sets based

IJFMR23069039 Volume 5, Issue 6, November-December 2023 11

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

on the split ratio (which is chosen) as shown in Fig 2. This is done in Jupyter using functions from the
image processing toolkit. After these signatures are preprocessed, they are stored in a file directory
structure that the Keras Python library can work with. Then the CNN is implemented in Python using
Keras with a TensorFlow backend to learn the patterns associated with the signatures. Then, the derived
model was validated using accuracy and loss metrics to see how well the model fit the data. Finally, the
model was tested using the signature from the challenge set to see if the predictions were correct. The
figure shows a detailed architectural scheme of the implementation.

EVALUATION/EXECUTION:-
DATASET:-
The dataset which has been used in this project work is a collection of 120 signatures, with 60 genuine
and 60 forged signatures per subject. This dataset was carefully prepared by us, with one person making
the originals and two others making the forgeries. All the images are in RGB format.

Forged:

Fig 3: Forged Signatures

Real:

Fig 3.1: Real Signatures

IJFMR23069039 Volume 5, Issue 6, November-December 2023 12

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

METHODOLOGY WITH THE DATASET:-

The handwritten signature is a behavioural biometric that is not based on any physiology characteristics
of the individual signature but on the behaviour that changes over time. Since an individual's signature
alters over time, the verification and authentication of the signature may take a long period of time,
which includes the potential for the errors to be higher in some cases. An inconsistent signature leads to
higher false rejection rates for an individual who did not sign in a consistent way.

Evaluation Measures:
First raw RGB images are converted to grayscale and binarized. Then file management operations are
are carried out to split the images into batches based on the split ratio. Training and Validation accuracies
are plotted for all the split ratios and the the split which gives the best results is considered.
The above accuracy and loss formulae are used to plot the accuracy and loss for the test splits. Accuracy
is the total number of correct predictions divided by the total number of predictions. In loss, p is the true
distribution and q is the coding distribution.

Evaluation methods section:

One of the most common evaluation technique used in signature forgery detection is plotting accuracy
and validation graphs of various test split ratios and finding the best possible ratio. The dataset is divided
into batches using several split ratio and accuracy and validation graphs are plotted and the split ratio
with the best results is taken. The accuracy% and loss% are plotted against epoch. One unique way that
we found was taking, a test signature is recognized with a given input training set using both CNN and
Crest-Trough methods.
Custom variation is used to assess test signatures. For a particular function, if the membership value lies
within the range of variation given by the difference of the minimum and maximum thresholds, it counts
as "true". The total number of "real" occurrences for a particular signature is divided by the total number
of features to get a percentage. Skilled forged and unskilled forged signatures have corresponding
figures of 88.5% and 82.3%. The minimum limit or acceptable percentage for a genuine signature is set
to 91% with respect to the output result of the signer's signature. Signatures that have a percentage of less
than 91% are considered forged signatures[11].

Performance Measures:
In this work, the signature images are preprocessed in a batch manner and divided into training and test
sets based on the split ratio (which is chosen). This is done in MatLab with functions from the Image
Processing Toolkit. After these signatures are preprocessed, they are stored in a file directory structure
that the keras python library can work with. Then the CNN was implemented in python using Keras with
a TensorFlow backend to learn the patterns associated with the signatures. Then, the derived model was
validated using accuracy and loss metrics to see how well the model fit the data. Finally, the model was
tested using the signature from the challenge set to see if the ns predictions were correct.
In our implementation, the image goes through 3 convolutional and max pooling layers that alternate.
When an image goes through the convolution process, a predefined number of feature maps are created,
which are fed into a maximum pooling layer, which creates pooled feature maps from the feature maps

IJFMR23069039 Volume 5, Issue 6, November-December 2023 13

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

received from the convolutional layer that precedes it. This pooled feature map is sent to the next
convolutional layer and this process continues until we reach the third maximum pooling layer. The
pooled feature map from the last maximum pooling layer is merged and sent to the fully connected
layers. After several rounds of forward and back propagation, the model is trained and now a prediction
can be made.
One of the most common evaluation technique used in signature forgery detection is plotting accuracy
and validation graphs of various test split ratios and finding the best possible ratio. The dataset is divided
into batches using several split ratio and accuracy and validation graphs are plotted and the split ratio
with the best results is taken. The accuracy% and loss% are plotted against epoch. One unique way that
we found was taking, a test signature is recognized with a given input training set using both CNN and
Crest-Trough methods.
For a particular function, if the membership value lies within the range of variation given by the
difference of the minimum and maximum thresholds, it counts as "true". The total number of "real"
occurrences for a particular signature is divided by the total number of features to get a percentage.
Skilled forged and unskilled forged signatures have corresponding figures of 88.5% and 82.3%. The
minimum limit or acceptable percentage for a genuine signature is set to 91% with respect to the output
result of the signer's signature. Signatures that have a percentage of less than 91% are considered forged
signatures[11].
In our work, we will first convert raw RGB images to grayscale. We then add salt and pepper noise with
a density of 0.01 and remove all the noise using the mean and median filters. We then binarize the
images and store them appropriately. We then perform the required processing and file management
operations to split the image batches based on the desired split ratio. After the models are built, accuracy
and loss graphs are created.
We created models for vaious splits of data and plotted the training and validation accuracies to get an
idea of the presence of any overfitting or underfitting

Comparison of base papers section:Table-1

Paper-1 Paper-2 Paper-3

1. [Link] Used their The

signature own dataset signature
Database/Corpa recovery
2. method
[Link] follows
nature-verification-dataset comparable
levels for
most
common
data sets.
The user
takes a
template

IJFMR23069039 Volume 5, Issue 6, November-December 2023 14

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

with many
cells and
presents an
example of
his / her
signature in
each cell.

MLP exposed a weakness in the skilled forgery The equal The model
test due to being a 2-class model. AE and error rate accepted
Performance/Res CNN_B-AE did not perform well. This result (EER) is 100% of the
u lts shows that the S-vector generated by the trained used as an signatures
CNN is more effective than raw signature data. evaluation when only
The CNN_D-AE showed the best result and the metric for genuine
CNN_AC-AE showed the second-best accuracy. signature signatures
However, note that the CNN_D-AE is not a verification. were given.
practical model.. All machine- It gave a 3%
learning false
models were positives in
implemented Skilled
using the forgery and
Theano was
library, successful in
which is a detecting
well-known unskilful
open source and random
machine- forgeries.
learning
library.

Scope It can be improved further by improving the They plan In all our
dataset quantity and quality to train the model. verify investigatio
signatures ns we train
drawn by a CNN with
hand gripping one × one
a kernel and

IJFMR23069039 Volume 5, Issue 6, November-December 2023 15

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

smartphone, TL as a
i.e., a user classifier.
draws a The margin α
signature in is set to
the air by a 0.2. We
hand gripping used the
a Kaggle data
smartphone, set, and the
and not on a result is
screen by a shown in the
finger. They tables above.
expect that
the large
signature
drawn by
hand could
have positive
effects of
verification
accuracy and
user
convenience
.

Applications This can be used to verify signatures on Verify Can be used

important documents to verify the owner. signatures in banks,
drawn on the government
phone. offices and
other
sensitive
fields
because the
accuracy is
high.

IJFMR23069039 Volume 5, Issue 6, November-December 2023 16

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

Algorithms/appr Convolutional Neural Network (CNN) to Convolutio nal Convolution

o ach implement the offline authentication methods, neural al Siamese
network Network
CEDAR Signature dataset to train the neural (CNN for (CSN).
network feature
extraction, Pooling
and an Layer,
autoencod er Triplet Loss
(AE) as a Classificati
classifier on

Table-2:
Paper-1 Paper-2 Paper-3

A collection of 6000 The signatures collected A signature database is

Database/Corpa signatures is used which were manually made. Few used which was developed
contains 1000 genuine and features are extracted in Graphics Visualization
1000 forged signatures per from the signatures to and Games Development
subject. The images taken crease a knowledge base lab at the Multimedia
were in RGB format. for every individual. The University, Cybejaya,
dataset contains 750 Malaysia.
signatures which has 5
signatures were taken
from each person from
150 people in total.
The highest training and The speed of execution of The model accepted 100%
Performance/Result validation accuracy is found the method is found out to of the signatures when only
s to be 98.11% and 98.23% be pretty fast. The genuine signatures were
respectively. accuracy is found out to given. It gave a 3% false
be around 80%. positives in Skilled forgery
and was successful in
detecting unskilful and
random forgeries.
Scope A custom loss function can be An optimized algorithm Global learning techniques
created in the future which can be created which can can be used which will
can predict which user the differentiate the errors improve accuracy and
signature belongs to. made by the real user to speed of the prediction.
remove false negatives.
Applications This can be used to verify Can be used in bank Can be used in banks,
signatures on important letters cheque signature government offices and
etc to verify the writer. verification systems. other sensitive fields

IJFMR23069039 Volume 5, Issue 6, November-December 2023 17

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

because the
accuracy is high.

Algorithms/approac The dataset is divided into First, a signature image is Signature images are
h batches using several split passed after the image binarized and window size
ration and accuracy and preprocessing where with a fixed size and then
validation graphs are plotted normalization, image they are thinned out. The
and the split ratio with the enhancements, geometric the thinned image is then
best results is taken. CNN is transformations, etc divided into a fixed
used. applied to the image so number eight sub-images
that the image perfectly called boxes. This section
processed and can be is complete using a
passed on feature horizontal density
extraction. When approximation
extracting features of approach. Each sub-image
local a global features are is then further resized and
extracted from the figure resized split into twelve
for comparing uploaded more sub-images using
image to image stored in unified distribution
the database. approach.
Generally for low level Properties the normalized
applications such as vector angle (a) and are
object detection and taken into account
classification, global distance (y) from each
functions are used and for box. Each function
higher-level application, extracted from sample
such as an object signatures give rise to
recognition, local features fuzzy sets. With regard to
are used. A it regarding to it,
combination of those will Choosing the right
provide higher accuracy. fuzzification function is
key

IJFMR23069039 Volume 5, Issue 6, November-December 2023 18

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

verification, we invented a
new fuzzification function
with structural parameters
that is able to adapt to
changes in hzzy sets. This
feature is used to develop
complete counterfeit
detection and
authentication system.
Evaluation metrics The dataset is divided into First, a test signature is Custom variation is used
batches using several split recognized with a given to assess test signatures.
ration and accuracy and input training set using For a particular function, if
validation graphs are plotted both CNN and Crest- the membership value lies
and the split ratio with the Trough methods. Then, within the range of
best results is taken. The forgery detection variation given by the
accuracy% and loss% are algorithms Harris difference of the minimum
plotted against epoch. Algorithmic and Surf and maximum thresholds,
Algorithm are carried out it counts as "true". The
on this classified image. total number of "real"
The results from each occurrences for a
algorithm are then particular signature is
compared. divided by the total
number of features to get a
percentage. Skilled forged
and unskilled forged
signatures
havecorresponding figures
of 88.5% and 82.3%. The
minimum limit or
acceptable percentage for
a genuine signature is set
to 91% with respect to the
output result of the signer's
signature. Signatures that
have a percentage of less
than
91% are considered forged
signatures.

IJFMR23069039 Volume 5, Issue 6, November-December 2023 19

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

Table-3:

Paper-1 Paper-2 Paper-3

In this Paper group of These handwritten The signature

5000 signatures were signatures are from a database consists of a
Database/Corpa used with 2500 forged dataset containing 750 total of 510
and 2500 genuine sigs. signatures from 150 handwritten signature
All the Images are in individuals providing 5 images. Out of these,
RGB format. signatures each. Second, 255 were authentic
these signatures are stored signatures and others
in the form of a matrix, were
then converted from the forged ones. These
RBG image to a black and signatures were
white image with the help obtained Erom 17
of a grayscale algorithm volunteers with each
person contributing
15 signatures

IJFMR23069039 Volume 5, Issue 6, November-December 2023 20

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

The highest training The result of this system is An off-line signature

and validation that it has 94% and 85-89% verification and
Performance/Resul accuracy is found to accuracy for signature forgery detection
ts be 98.11% and 98.23% recognition and forgery system is modeled by
respectively. detection, respectively. We TS model, which
can conclude that this involves structural
system may have some parameters in ’ its
flaws, but it is a useful one. exponential
membership function.
The features
consisting of angles
are extracted using
box approach. Each
feature yields a fuzzy
set when its values are
gathered from all
samples because of
the variations in
handwritten
signatures

Scope A custom made This algorithm can be Machine learning

technology can be created which can techniques can be
created in the future differentiate the errors used which will
which can predict the made by the real user to improve accuracy and
identity of the users. remove false signatures. speed of the
prediction with the
minimal time
complexity.

Applications This can be used to Can be used in paper bank Can be used in banks,
verify signatures on cheque signature government offices
entry pass, letters, verification and and other sensitive
authentication to authentication. fields because the
verify the owner. accuracy is extremely
sensitive.

IJFMR23069039 Volume 5, Issue 6, November-December 2023 21

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

Algorithms/approa The handwritten We intend to use a Since the main thrust

ch signature is a Convolutional Neural here is to establish the
behavioural biometric Network (CNN) to genuineness of the
which is not based on implement the offline signature thereby
any physiolo authentication methods, detecting the
characteristics of the since all the offline forgeries, we go in for
individual signature methods exploit the fuzzy modeling of
but on the behaviour content-based features and angle features. For the
that change over time. the visual information of purpose of signature
Since an individual's the signature, it is better to verification and
signa alters over time use a CNN since a CNN detection of forgeries,
the verification and can classify the extracted we have employed the
authentication for the features from the signature. Takagi- Sugeno
signature may take a We will use the CEDAR model. In this, we are
long period which Signature dataset to train following the same
include the for the the neural network. concept as outlined in
errors to be higher in The CEDAR Signature [ 131 for considering
some cases. dataset is a signature each feature as
Inconsistent signature verification database. 55 forming a fuzzy set
leads to higher false individuals contributed 24 over large samples.
rejection rates for an signatures each and hence This is because the
indivi who did not the dataset consists of 1320 same feature exhibits
sign in a consistent genuine signatures. People variation in different
way were asked to forge the samples giving rise to
three other writers a fuzzy set. So, our
Data Acquisition signatures, eight times attempt is to model
every subject creating a the uncertainty
total of 1320 forged through a fuzzy
signatures. The obtained model such as the TS
Handwritten dataset consists of 24 model.
signatures are genuine and 24 forged
collected and some signatures for each writer.
unique features are
extracted to create
knowledgebase each
and every individual.
A standard database of
signatures for every
individual is needed
for evaluating
performance of the

IJFMR23069039 Volume 5, Issue 6, November-December 2023 22

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

signature verification
system and also for
comparing the result
obtained using other
techni on the same
database

Pre-processing

RGB to Grayscale

In layman's terms, any

RGB image is
represented as matrix
of X, Y dimensions
and

IJFMR23069039 Volume 5, Issue 6, November-December 2023 23

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

depth of 3 pla where

each plane comprises
of Red, Green and
Blue values ranging
from 0 to
255. Whereas, grays
image is represented
as matrix of X. Y
dimensions and depth
of only I plane. Each
cell value ninges fre to
255 Any RGB (Red,
Green, Blue) image
which has to undergo
Digital Image
Processing (DIP) neec
be converted to
grayscale image. By
doing this the
computational
complexity of DIP
decreases drastic and
helps to run image
processing algorithms
in much smoother
way. This operation is
seen in Fig.

Evaluation metrics We create models for For signature recognition, The signature
different splits of data forgery detection and database consists of a
and plot the training verification based on total of 510
and validation CNN(Convolutional Neural handwritten signature
accuracies to get an Network), Crest-Through images. Out of these,
idea of the presence of Method, SURF algorithm 255 were authentic
any overfitting or and Harris corner detection signatures
underfitting.
algorithm. In this system, and others were
CNN and Crest-Through forged ones. These

IJFMR23069039 Volume 5, Issue 6, November-December 2023 24

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

On splitting the Method are used for signatures were

dataset in the ratio of recognition and obtained Erom 17
8:2 we obtained a verification. Whereas, volunteers with each
maximum accuracyof- SURF and Harris corner person contributing
98 percent on the detection algorithms are 15 signatures as
validation set as used forgery detection. This shown in Table 1. The
shown in Fig 9. There paper has literature surveys signatures were
is very little overfitting on other papers also. These collected over a
as the training and signatures undergo pre- period of a few weeks
testing
accuracies are almost
equal to each other.
processing using CNN and to account for
On splitting our Crest-Through Method. variations in the
dataset in the ratio 73, signature with time.
we get a maximum ?he forgeries of these
accuracy of 96 percent signatures were
on the validation set collected over a
as shown in Fig 10 at similar time frame.
the third epoch. There The random forgeries
is some underfitting were obtained by
here, and further supplying only the
epochs show over
fitting. The validation

IJFMR23069039 Volume 5, Issue 6, November-December 2023 25

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

accuracy at the fifth names of the

epoch is-94 percent individuals to the
casual forgers who
did not have any
access to the actual
genuine signatures.
The unskilled
forgeries in turn,
were obtained by
providing sample
genuine signatures to
the forgers who were
then allowed to
practice for a while
before imitating them
to create their
forgeries. Each
volunteer had to
provide five
imitations of any one
of the genuine
signatures, apart
from his or her own
signatures. These
samples constituted
the set of unskilled
forged signatures for
the set of genuine
signatures. We then
requisitioned the
services of a few
expert forgers who
provided five
forgeries of each
genuine signature in
the test set to create
the skilled forged
samples of all the
persons.

IJFMR23069039 Volume 5, Issue 6, November-December 2023 26

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

Data Source List of 5000

signatured were up

Conclusion and future directions:

Handwritten signatures are very important in our social and legal life for verification and authentication.
A signature can only be accepted if it is from the intended person. Probability two signatures of the same
person are very small. Many properties the signature may differ even if the two signatures are of the
same person. So the detection of a forgery becomes a challenging task. In the organization without
affecting the current system. In parallel with the new technology that opens up with new possibilities,
there is a clear demand for new and improved methodologies and algorithms. The proposed approach is
suitable for use as an efficient signature verification system. The the proposed technique effectively
performed offline signature verification with increased efficiency and accuracy, as well as the ability to
detect expert forgeries. We have successfully detected the signature cheats using Python and its modules
combined with a convolutional neuron based solution network (CNN). In the future, the model will be
improved by reducing the error rejection rate.
Another interesting project is the merging of offline and online signature verification technologies,
which it will strengthen the system by requiring both speed of execution and authenticity. Another
promising project would be to merge offline and online signature verification systems, which would
make the system more robust due to execution speed and authentic appearance signature, making it
difficult to produce signatures. This project was only implemented for one language. When uploading
digital signatures to apps or websites, many additional languages can be incorporated using a GUI based
on the Flask platform for better users participation. The planned system is highly economical in detecting
and tracking counterfeits on the fly, and therefore the system's responsiveness can be increased by
training the extracted features on artificial neural networks by storing the extracted features. Negligible
misclassification or however, a bug is required in such sensitive applications This comes at the cost of a
high recognition rate (HRR). Another goal is that the probability of a
a forged signature as if it were real is zero. As future work, we can also focus on increasing the resulting
accuracy of the system by trying new and better parameter coefficients that increase the difference
between real and forged signatures.

References:
1. Hilton O, Scientific Examination of Questioned Documents:[Link] [Link]
n/books/about/Scientific_Examination_of Questioned _Doc.himl?id=nQlIw2_gE-cC&redir_ese=y
Specin Forensics LLC, Handwriting and Forgery Examination http:/4n6,com/handwriting-and-
forgery- examination
2. Zaidi S.F.A., Mohammed S, Biometric Handwritten Signature Recognition, TDDD17: Information
Security Course, Linkopingsuniversitet,Sweden
3. Srinivasan H., Srihari S.N., Beal M.J. (2019) Machine Learning for Signature Verification. In: Kalra
P.K., Peleg S. (eds) Computer Vision,Graphics and Image Processing. Lecture Notes in Computer
Science, vol 4338. Springer, Berlin, Heidelberg,
4. Bhattacharya L, GhoshP.,Biswas S. (2018) Offline Signature Verification Using Pixel Matching
Technique [6]Drott B. and Hassan-Reza T. On-line Handwritten Signature Verification using

IJFMR23069039 Volume 5, Issue 6, November-December 2023 27

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

Machine Learning Techniques with a Deep Learning Approach(2019), In Master's Theses in

Mathematical Sciences FMAS20 20151, Mathematics. (Faculty of Engineering)
5. Hafemann L, G., Sabourin R., Oliveira L.S., Learning features for offline handwritten signature
verification using deep convolutional neuralnetworks, Pattern. Recognition, Volume —70,2017,
Pages «163-176,
6. ISSN 0031-32037302017)
7. Jerome Gideon S, Anurag Kandulna, Aron Abhishek Kujur, Diana A Kumudha Raimod
“Handwritten Signature Forgery Detection Using Convolutional Neural Networks”, International
Conference on Advances in Computing and Communication, P – 978-987 (ICACC-2018)
[Link] S877050918320301
8. Bhattacharya L., Goshp., Biswas S, (2013) “Offline Signature Verification Using Pixel Matching
Technique”, International Conference on Computational Intelligence : Modeling Technique and
Applications(CIMTA), Proceida Technology, P – 970-977, 2013.
9. Ashish A Dongare, Prof R. D. Ghongade(2016) “Artificial Intelligence Based Bank Cheque
Signature Verification System”, Vol – 03, P-ISSN 2395-0072(IRJET)
10. Walid Hussein, Mostafa A. Salama and Osman Ibrahim (2016),“Image Processing Based Signature
Verification Technique to Reduce Fraud in Financial Institution”, MATEC Web of Conferences
2016.
11. Saroj Ramdas, Geethu P. C. (2015) “Comparative Study on Offline Handwritten Signature
Verification Scheme”, International Journal of Advanced Research Trends in Engineering and
Technology (IJARTET), Vol - 02, March 2015.
12. [ 13] R. Plamondon and F. Leclerc, “Automatic signature verification: the state of the art 1989-
1993”, lnternationai Journal of Pattern Recognition and Artificial Intelligence, Vol. 8, No. 3, pp.
643-660, 1994.
13. R. Plamondon and G. Lorette, “Automatic signature verification and writer Identification: the state of
the art”, Pattern Recognition, Vol. 22, No. 2, pp. 107-131, 1989.
14. R. Sabourin, R. Plamondon and G. Lorette, Offline identification with handwritten signature images:
Survey and Perspectives, Structured Image Anaiysis, Springer-Verlag, New York, 1992, pp. 2 19-
234.
15. R. Plamondon and S.N. Srihari, “On-line and offline Handwriting Recognition: A Comprehensive
Survey”, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 22, No. I, pp. 63-
x4,2000.
16. M.A. Ismail and Samia Gad, “Off-line Arabic signature recognition and verification”, Pattern
Recognition, Vol. 33, No. 10, pp. 1727-1740, 2000.
17. A. El-Yacoubi, E.J.R. Justino, R. Sabourin and F. Bortolozzi, “Off-line signature verification using
HMMS and cross-validation”, Proceedings of the IEEE Workshop on Neural Networks for Signal
Processing, USA, 2000, pp. 859-868.
18. C. Quek and R.W. Zhou, “Antiforgery: a novel pseudo-outer product based fuzzy neural network
driven signature Verification system”, Pattern Recognition Letters, Vol. 23, pp. 1795-18i6, 2002.
19. M. Hanmandlu, K.R. Murali Mohan, S. Chakraborty, S. Goel and D. Roy Choudhury,
“Unconstrained handwritten character recognition based on fuzzy logic”, Pattern Recognition, Vol.
20. [21]M. Hanmandlu, K.R. Murali Mohan, S. Chakraborty and G. Garg, “Fuzzy modeling based
signature verification system”, Proceedings of the sixth International Conference on Document

IJFMR23069039 Volume 5, Issue 6, November-December 2023 28

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

Analysis and Recognition, USA, 2001, pp. 110-1 14.

21. M. Ammar, Y. Yoshida and T. Fukumura, “A new effective approach for off-line verification of
signatures by using pressure features”, Proceedings of the International Conference on Pattern
recognition, 1986, pp, 566- 569.
22. M. Ammar, “Progress in verification of skillfully simulated handwritten signatures”, International
Journal of Pattern Recognition and Artificial Intelligence, Vol. 5, pp. 337-351, 1991.
23. Jinhong K. Guo, D. Doermann and A. RosenfeId, “Off-line skilled forgery detection using stroke
and sub- stroke properties”, Proceedings of the International Conference on Pattern Recognition,
24. M. Hanmandlu, K.R. Murali Mohan and Vivek Gupta, “Fuzzy logic based character recognition”,
Proceedings of the International Conference on Image Processing, Santa Barbara, USA, pp.7 14-
717.
25. [26]Y. Xuhua, T. Furuhasbi, K. Obata, Y. Uchikawa, Study on signature verification using a new
approach to genetic based machine learning. Proceedings of the IEEE International Conference on
Systems, Man and Cybernetics, USA, 1995, pp. 36, NO. 3, pp. 603423,2003. 2000, pp. 355-358.
4383-4386.
26. Roh, J.H.; Lee, S.H.; Kim, S. Keystroke Dynamics for Authentication in Smartphone. In Proceedings
of the International Conference Information and Communication Technology (ICTC), Jeju Island,
Korea, 19–21 October 2016; pp. 1155–1159.
27. Liu, D.; Dong, B.; Gao, X.; Wang, H. Exploiting Eye Tracking for Smartphone Authentication. In
Proceedings of the International Conference on Applied Cryptography and Network Security
(ACNS), New York, NY, USA, 2–5 June 2015; pp 457–477.
28. Buriro, A.; Crispo, B.; Delfrari, F.; Wrona, K. Hold and Sign: A Novel Behavioral Biometrics for
Smartphone User Authentication. In Proceedings of the IEEE Symposium on Security and Privacy
Workshops (SPW), San Jose, CA, USA, 22–26 May 2016; pp. 276–285.
29. Feng, H.; Wah, C.C. Online signature verification using a new extreme points warping technique.
Pattern Recognit. Lett. 2003, 24, 2943–2951.
30. Fierrez-Aguilar, J.; Nanni, L.; Lopez-Peñalba, J.; Ortega-Garcia, J.; Maltoni, D. An on-line signature
verification system based on fusion of local and global information. In Proceedings of Audio- and
Video-Based Biometric Person Authentication (AVBPA), Rye Brook, NY, USA, 20–22 July 2005;
pp. 523–532.
31. Nanni, L. An advanced multi-matcher method for on-line signature verification featuring global
features and tokenised random numbers. Neurocomputing 2006, 69, 2402–2406.
32. Guru, D.S.; Prakash, H.N. Online signature verification and recognition: An approach based on
symbolic representation. IEEE Trans. Pattern Anal. Mach. Intell. 2009, 31, 1059–1073.
33. Gruber, C.; Gruber, T.; Krinninger, S.; Sick, B. Online signature verification with support vector
machines based on LCSS kernel functions. IEEE Trans. Syst. Man Cybern. Part B (Cybern.) 2010,
40, 1088–1100.
34. Iranmanesh, V.; Ahmad, S.M.S.; Adnan, W.A.W.; Malallah, F.L.; Yussof, S. Online signature
verification using neural network and Pearson correlation features. In Proceedings of the IEEE
Conference on Open Systems (ICOS), Subang, Selangor, Malaysia, 26–28 October 2014; pp. 18–21.
35. Fayyaz, M.; Saffar, M.H.; Sabokrou, M.; Hoseini, M.; Fathy, M. Online signature verification based
on feature representation. In Proceedings of the International Symposium on Artificial Intelligence
and Signal Processing (AISP), Mashhad, Iran, 3–5 March 2015; pp. 211–216.

IJFMR23069039 Volume 5, Issue 6, November-December 2023 29

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

36. Sae-Bae, N.; Memon, N. Online signature verification on mobile devices. IEEE Trans. Inf. Forensics
Secur. 2014, 9, 933–949.
37. Antal, M.; Szabó, L.Z. On-line verification of finger drawn signatures. In Proceedings of the
International Symposium on Applied Computational Intelligence and Informatics (SACI),
Timisoara, Romania, 12–14 May 2016; pp. 419–424.
38. Paudel, N.; Querini, M.; Italiano, G.F. Handwritten Signature Verification for Mobile Phones. In
Proceedings of the International Conference on Information Systems Security and Privacy (ICISSP),
Rome, Italy, 19–21 February 2016; pp. 46–52.
39. R. Plamondon and F. Leclerc, “Automatic signature verification: the state of the art 1989-1993”,
lnternationai Journal of Pattern Recognition and Artificial Intelligence, Vol. 8, No. 3, pp. 643-660,
1994.
40. R. Plamondon and G. Lorette, “Automatic signature verification and writer Identification: the state of
the art”, Pattern Recognition, Vol. 22, No. 2, pp. 107-131, 1989.
41. R. Sabourin, R. Plamondon and G. Lorette, Offline identification with handwritten signature images:
Survey and Perspectives, Structured Image Anaiysis, Springer-Verlag, New York, 1992, pp. 2 19-
234.
42. R. Plamondon and S.N. Srihari, “On-line and offline Handwriting Recognition: A Comprehensive
Survey”, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 22, No. I, pp. 63-
x4,2000.
43. M.A. Ismail and Samia Gad, “Off-line Arabic signature recognition and verification”, Pattern
Recognition, Vol. 33, No. 10, pp. 1727-1740, 2000.
44. A. El-Yacoubi, E.J.R. Justino, R. Sabourin and F. Bortolozzi, “Off-line signature verification using
HMMS and cross-validation”, Proceedings of the IEEE Workshop on Neural Networks for Signal
Processing, USA, 2000, pp. 859-868.

45. C. Quek and R.W. Zhou, “Antiforgery: a novel pseudo-outer product based fuzzy neural network
driven signature Verification system”, Pattern Recognition Letters, Vol. 23, pp. 1795-18i6, 2002.
46. M. Hanmandlu, K.R. Murali Mohan, S. Chakraborty, S. Goel and D. Roy Choudhury,
“Unconstrained handwritten character recognition based on fuzzy logic”, Pattern Recognition, Vol.
47. M. Hanmandlu, K.R. Murali Mohan, S. Chakraborty and G. Garg, “Fuzzy modeling based signature
verification system”, Proceedings of the sixth International Conference on Document Analysis and
Recognition, USA, 2001, pp. 110-1 14.
48. M. Ammar, Y. Yoshida and T. Fukumura, “A new effective approach for off-line verification of
signatures by using pressure features”, Proceedings of the International Conference on Pattern
recognition, 1986, pp, 566- 569.
49. M. Ammar, “Progress in verification of skillfully simulated handwritten signatures”, International
Journal of Pattern Recognition and Artificial Intelligence, Vol. 5, pp. 337-351, 1991.
50. Jinhong K. Guo, D. Doermann and A. RosenfeId, “Off-line skilled forgery detection using stroke
and sub- stroke properties”, Proceedings of the International Conference on Pattern Recognition,
51. M. Hanmandlu, K.R. Murali Mohan and Vivek Gupta, “Fuzzy logic based character recognition”,
Proceedings of the International Conference on Image Processing, Santa Barbara, USA, pp.7 14-
717.
52. Y. Xuhua, T. Furuhasbi, K. Obata, Y. Uchikawa, Study on signature verification using a new

IJFMR23069039 Volume 5, Issue 6, November-December 2023 30

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

approach to genetic based machine learning. Proceedings of the IEEE International Conference on
Systems, Man and Cybernetics, USA, 1995, pp. 36, NO. 3, pp. 603423,2003. 2000, pp. 355-358.
4383-4386.
53. Kshitij Swapnil Jain, et al. “HANDWRITTEN SIGNATURES FORGERY DETECTION” (2021).
International Research Journal of Engineering and Technology (IRJET) Volume: 08 Issue: 01 | Jan
2021.
54. Kiran, Lakkoju Chandra, et al. "Digital signature Forgery Detection using CNN." (2021).
55. Raj Balsekar, et al. "OFFLINE SIGNATURE FORGERY DETECTION USING
CONVOLUTIONAL NEURAL NETWORK", International Journal of Emerging Technologies and
Innovative Research ([Link]), ISSN:2349-5162, Vol.7, Issue 5, page no.13-18, May-2020.
56. Alajrami, Eman, et al. "Handwritten signature verification using deep learning." International Journal
of Academic Multidisciplinary Research (IJAMR) 3.12 (2020).
57. Poddar, Jivesh, Vinanti Parikh, and Santosh Kumar Bharti. "Offline signature recognition and
forgery detection using deep learning." Procedia Computer Science 170 (2020): 610-617.
58. Based Off-Line Handwritten Signature Verification. Computer Vision and Image Understanding,
76(3), 173-190. DOI: 10.1006/cviu.1999.0799
59. [ 60] R. Plamondon and F. Leclerc, “Automatic signature verification: the state of the art 1989-
1993”, lnternationai Journal of Pattern Recognition and Artificial Intelligence, Vol. 8, No. 3, pp.
643-660, 1994.
60. R. Plamondon and G. Lorette, “Automatic signature verification and writer Identification: the state of
the art”, Pattern Recognition, Vol. 22, No. 2, pp. 107-131, 1989.
61. R. Sabourin, R. Plamondon and G. Lorette, Offline identification with handwritten signature images:
Survey and Perspectives, Structured Image Anaiysis, Springer-Verlag, New York, 1992, pp. 2 19-
234.
62. R. Plamondon and S.N. Srihari, “On-line and offline Handwriting Recognition: A Comprehensive
Survey”, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 22, No. I, pp. 63-
x4,2000.
63. M.A. Ismail and Samia Gad, “Off-line Arabic signature recognition and verification”, Pattern
Recognition, Vol. 33, No. 10, pp. 1727-1740, 2000.
64. A. El-Yacoubi, E.J.R. Justino, R. Sabourin and F. Bortolozzi, “Off-line signature verification using
HMMS and cross-validation”, Proceedings of the IEEE Workshop on Neural Networks for Signal
Processing, USA, 2000, pp. 859-868.
65. C. Quek and R.W. Zhou, “Antiforgery: a novel pseudo-outer product based fuzzy neural network
driven signature Verification system”, Pattern Recognition Letters, Vol. 23, pp. 1795-18i6, 2002.
66. M. Hanmandlu, K.R. Murali Mohan, S. Chakraborty, S. Goel and D. Roy Choudhury,
“Unconstrained handwritten character recognition based on fuzzy logic”, Pattern Recognition, Vol.
67. M. Hanmandlu, K.R. Murali Mohan, S. Chakraborty and G. Garg, “Fuzzy modeling based signature
verification system”, Proceedings of the sixth International Conference on Document Analysis and
Recognition, USA, 2001, pp. 110-1 14.
68. M. Ammar, Y. Yoshida and T. Fukumura, “A new effective approach for off-line verification of
signatures by using pressure features”, Proceedings of the International Conference on Pattern
recognition, 1986, pp, 566- 569.
69. M. Ammar, “Progress in verification of skillfully simulated handwritten signatures”, International

IJFMR23069039 Volume 5, Issue 6, November-December 2023 31

 International Journal for Multidisciplinary Research (IJFMR)
E-ISSN: 2582-2160 ● Website: [Link] ● Email: editor@[Link]

Journal of Pattern Recognition and Artificial Intelligence, Vol. 5, pp. 337-351, 1991.
70. Jinhong K. Guo, D. Doermann and A. RosenfeId, “Off-line skilled forgery detection using stroke
and sub- stroke properties”, Proceedings of the International Conference on Pattern Recognition,
71. M. Hanmandlu, K.R. Murali Mohan and Vivek Gupta, “Fuzzy logic based character recognition”,
Proceedings of the International Conference on Image Processing, Santa Barbara, USA, pp.7 14-
717.
72. Y. Xuhua, T. Furuhasbi, K. Obata, Y. Uchikawa, Study on signature verification using a new
approach to genetic based machine learning. Proceedings of the IEEE International Conference on
Systems, Man and Cybernetics, USA, 1995, pp. 36, NO. 3, pp. 603423,2003. 2000, pp. 355-358.
4383-4386.
73. Kshitij Swapnil Jain, et al. “HANDWRITTEN SIGNATURES FORGERY DETECTION” (2021).
International Research Journal of Engineering and Technology (IRJET) Volume: 08 Issue: 01 | Jan
2021.
74. Kiran, Lakkoju Chandra, et al. "Digital signature Forgery Detection using CNN." (2021).
75. Raj Balsekar, et al. "OFFLINE SIGNATURE FORGERY DETECTION USING
CONVOLUTIONAL NEURAL NETWORK", International Journal of Emerging Technologies and
Innovative Research ([Link]), ISSN:2349-5162, Vol.7, Issue 5, page no.13-18, May-2020.
76. Alajrami, Eman, et al. "Handwritten signature verification using deep learning." International Journal
of Academic Multidisciplinary Research (IJAMR) 3.12 (2020).
77. Poddar, Jivesh, Vinanti Parikh, and Santosh Kumar Bharti. "Offline signature recognition and
forgery detection using deep learning." Procedia Computer Science 170 (2020): 610-617.
78. Based Off-Line Handwritten Signature Verification. Computer Vision and Image Understanding,
76(3), 173-190. DOI: 10.1006/cviu.1999.0799

IJFMR23069039 Volume 5, Issue 6, November-December 2023 32