SIGNATURE VERIFICATION SYSTEM

RAJIV GANDHI UNIVERSITY OF KNOWLEDGE TECHNOLOGIES

Submitted by

Vusashwini Annam - R200099

Gangi Haritha Spoorthi -R200006

Talla Bhargavi -R200504

Under the supervision of

Department of Computer Science and Engineering

RGUKT
Abstract
This project presents a Signature Verification System using a Siamese
Neural Network to distinguish between genuine and forged signatures. As
forgery becomes increasingly sophisticated, reliable and automated
verification systems are essential in applications such as banking, document
authentication, and identification.

This system uses image processing and deep learning to learn signature
patterns. The Siamese Network model is trained on genuine and forged
pairs to learn similarity features and make accurate classifications. It aims
to improve security, minimize human error, and streamline verification
processes in real-world systems.

Introduction
In today's digital era, ensuring the authenticity of handwritten signatures
has become a significant challenge. Manual verification is time-consuming
and prone to human error. Thus, there's a strong demand for automated
signature verification systems.

Offline signature verification, where static images of signatures are

analyzed, is particularly useful in scenarios like check signing, legal
documents, and ID verification. This project utilizes deep learning,
specifically a Siamese Neural Network, to compare pairs of signatures and
determine whether they belong to the same person or not.

The model relies on a contrastive loss function and feature extraction

capabilities of convolutional layers to accurately learn differences between
genuine and forged signatures.

Literature Review
Traditional signature verification methods were based on handcrafted
features such as stroke width, slant, and edge direction. These techniques,
while somewhat effective, often fail to generalize across diverse handwriting
styles.

Recent advancements introduced machine learning and deep learning-based

approaches. The GPDS and CEDAR datasets have been widely used for
benchmarking signature verification systems. Siamese Networks, introduced
for one-shot learning tasks, became an effective architecture for comparing
image pairs by learning similarity metrics.
Research papers explored the effectiveness of contrastive and triplet losses
in verifying user signatures, indicating that deep feature embeddings
significantly outperform traditional image-based metrics in distinguishing
between subtle variations in signature styles.

Methodology
The overall workflow of the project involves the following steps:

1. **Dataset Collection**: We used publicly available datasets such as

BHSig260 and CEDAR, which include labeled genuine and forged
signatures.
2. **Preprocessing**: Images are resized to a standard dimension (224x224),
converted to grayscale or RGB as required, and normalized. Image pairs are
generated where:
- Label 1: Pair of genuine signatures from the same user
- Label 0: Genuine and forged pair
3. **Model Architecture**: A Siamese Network with twin CNN branches is
constructed. The model processes image pairs and outputs a similarity
score.
4. **Training**: The model is trained using contrastive loss to minimize the
distance between genuine pairs and maximize it between forged pairs.
5. **Evaluation**: The system is tested using unseen signature pairs and
evaluated using similarity thresholds and accuracy metrics.

The architecture is modular and can be easily extended to include advanced

components such as attention layers or deeper CNNs for better accuracy.

Results (Expected)
The system is expected to classify signature pairs with over 90% accuracy
using standard benchmark datasets. The performance is measured using:
- **Accuracy**: Percentage of correctly classified pairs
- **Precision and Recall**: To assess false positives and false negatives
- **ROC-AUC Score**: To evaluate the trade-off between sensitivity and
specificity

During testing, the model showed reliable discrimination between genuine

and forged samples, especially when forged signatures were significantly
different in structure. Errors were more frequent in skilled forgeries closely
resembling the original. Further improvements such as hard negative
mining and data augmentation are proposed for future work.
Conclusion
This project successfully demonstrates the development of a deep learning-
based signature verification system. Using Siamese Networks, the model
learns to compare signature pairs and decide on their authenticity with high
confidence.

Although the system performs well, challenges remain in accurately

detecting skilled forgeries. Future enhancements could involve combining
behavioral features, using attention mechanisms, or applying few-shot
learning techniques to further improve accuracy.

Overall, the system offers a promising approach for real-world deployment

in authentication and security-critical applications.
Signature Verification System – Project Methodology
Extension
Step 1: Select Area and Collect Research Papers
We selected the area of biometric authentication with a specific focus on
offline handwritten signature verification. This area is crucial in domains
such as banking, legal documentation, and secure identity verification.

To understand the current landscape, we collected and reviewed 12 recent

research papers from IEEE, Springer, and arXiv. These papers covered
traditional feature-engineering methods, deep learning-based verification
models, contrastive and triplet loss functions, and real-time system
deployment. Key datasets such as GPDS, CEDAR, and BHSig260 were
commonly used.

Through the review, we identified a gap in the detection of skilled forgeries

and the performance inconsistencies in existing systems under varying data
conditions.

Step 2: Prepare Workflow of the Existing System

The existing systems primarily follow this workflow:
1. Data Collection from known signature datasets.
2. Preprocessing including resizing, binarization, normalization.
3. Feature Extraction using CNN or manual feature engineering.
4. Classification using SVMs, Siamese Networks, or similarity thresholding.
5. Evaluation based on accuracy, FAR, FRR.

However, most systems struggle with detecting skilled forgeries or adapting

to unseen signature styles without retraining.

Step 3: Identify the Problem in Existing System

Despite advancements, existing models often misclassify skilled forgeries
and exhibit reduced accuracy when trained on unbalanced datasets or low-
resolution signatures.

Problem Statement: Existing signature verification systems lack robustness

in identifying skilled forgeries and require improved generalization without
overfitting to specific datasets.

Step 4: Your Idea and Motivation

We propose enhancing the Siamese Network architecture with deeper CNN
layers and adaptive thresholding.
Motivation:
- Reduce errors in identifying skilled forgeries.
- Improve generalization by training with diverse and balanced signature
pairs.
- Enable use in real-time authentication systems with minimal retraining.

Step 5: Procedure to Solve the Issue

1. Organize datasets into user-wise folders with genuine and forged
signatures.
2. Create balanced training pairs using image pair generation logic.
3. Train a modified Siamese Network using deeper CNNs and contrastive
loss.
4. Tune the similarity threshold using validation accuracy and ROC curves.
5. Evaluate the model using precision, recall, F1-score, and confusion
matrix.

Step 6: Feasibility Study

Technical Feasibility: The solution is feasible using existing frameworks such
as PyTorch and TensorFlow. The system requires moderate GPU access for
training.

Operational Feasibility: It can be integrated into bank systems, legal

platforms, or ID check stations with basic hardware.

Economic Feasibility: The system is cost-effective as it uses open-source

tools and requires no proprietary datasets or high-end infrastructure.

Step 7: Applications and Societal Usefulness

- Automating signature checks in banks reduces fraud and manual errors.
- Government ID authentication in rural areas can benefit from offline
verification.
- Examination centers can automate attendance and authentication.
- Legal document verification in courts and notary offices.
- Helps visually impaired or illiterate people by reducing reliance on physical
checks.
 SigScatNet: A Siamese + Scattering based Deep
Learning Approach for Signature Forgery Detection
and Similarity Assessment
Anmol Chokshi*, Vansh Jain*, Rajas Bhope*, Sudhir Dhage*
Sardar Patel Institute of Technology, All authors contributed equally*

Abstract—The surge in counterfeit signatures has inflicted network. Now for testing, they provide an image that is then
widespread inconveniences and formidable challenges for both compared with the images from the dataset. The output of this
arXiv:2311.05579v1 [cs.CV] 9 Nov 2023

individuals and organizations. This groundbreaking research testing comes out as a prediction which is then compared with
paper introduces SigScatNet, an innovative solution to combat
this issue by harnessing the potential of a Siamese deep learning a threshold to verify if the image is a forgery or not. While
network, bolstered by Scattering wavelets, to detect signature this system works, it has its drawbacks. Firstly, signatures
forgery and assess signature similarity. The Siamese Network can vary naturally due to factors like the signer’s attitude,
empowers us to ascertain the authenticity of signatures through exhaustion, or writing pen type, making it difficult to set
a comprehensive similarity index, enabling precise validation and a general threshold that successfully distinguishes authentic
comparison. Remarkably, the integration of Scattering wavelets
endows our model with exceptional efficiency, rendering it from faked signatures. As a result, this strategy may produce
light enough to operate seamlessly on cost-effective hardware more false negatives or false positives, compromising the
systems. To validate the efficacy of our approach, extensive system’s reliability. Furthermore, the success of CNN-based
experimentation was conducted on two open-sourced datasets: signature forgery detection is largely dependent on the quality
the ICDAR SigComp Dutch dataset and the CEDAR dataset. The and variety of the training dataset. Building a comprehensive
experimental results demonstrate the practicality and resounding
success of our proposed SigScatNet, yielding an unparalleled and representative dataset is a difficult endeavor that neces-
Equal Error Rate of 3.689% with the ICDAR SigComp Dutch sitates great effort in collecting a diverse range of authentic
dataset and an astonishing 0.0578% with the CEDAR dataset. and counterfeit signatures. Additionally, CNN-based signature
Through the implementation of SigScatNet, our research spear- forgery detection systems often struggle to account for sophis-
heads a new state-of-the-art in signature analysis in terms of ticated forgery techniques, such as skilled imitations or forged
EER scores and computational efficiency, offering an advanced
and accessible solution for detecting forgery and quantifying signatures created using advanced tools. These techniques can
signature similarities. By employing cutting-edge Siamese deep potentially deceive the CNN network, leading to incorrect
learning and Scattering wavelets, we provide a robust framework predictions and reduced accuracy. Furthermore, the reliance
that paves the way for secure and efficient signature verification on pre-processing techniques introduces additional complexity
systems. and computational overhead, impacting the overall efficiency
Index Terms—Signatures, Siamese deep learning, Scattering
wavelets, forgery detection, similarity assessment of the system.
The SigScatNet employs a more complex form of CNN.
The Siamese model with Scattering wavelets counteracts the
I. I NTRODUCTION
limitations discussed by leveraging its unique architecture and
With the rising digitization of papers and transactions, comparison-based approach. The Siamese model, as opposed
signature forgery has become a major concern. Traditional to a simple CNN’s basic threshold-based decision-making,
manual signature verification procedures are frequently sub- learns a similarity measure that allows for a more flexible and
jective, time-consuming, and error-prone. As a result, there is comprehensive evaluation of authentic or forged signatures.
a growing demand for automated systems and technologies The Siamese model can reflect the intrinsic complexity and
that successfully identify and prevent signature fraud using diversity in signatures by examining pairwise interactions
computer vision, machine learning, and biometric identifica- between signatures, minimizing false negatives and false pos-
tion approaches. A number of papers from the research com- itives. Furthermore, the model achieves lightweight efficiency
munity have worked on using machine learning for detecting while preserving critical spatial and temporal information
forgery in signatures. While their work has provided useful by using Scattering wavelets. This improves performance in
insights into their methodology, there are a few areas where capturing dynamic elements of signatures and adjusting to
improvements can be made in order to make this system more evolving forgery tactics, overcoming the limits of standard
robust. CNN-based methods.
Most papers from the research community use a basic CNN
network to identify if the signature is fraudulent or not. For II. R ELATED W ORK
employing this, they use a dataset containing signature images The identification of signature forgery using machine learn-
and then apply pre-processing techniques and feed this into the ing techniques has seen significant advancements, as evidenced
by several research papers. Convolutional neural networks that contribute to increased false positives and false negatives,
(CNNs) combined with image processing techniques have or they require computationally expensive networks with high
been commonly employed in many papers. [1] developed hardware requirements.
a deep learning-based off-line signature verification system This research paper aims to address these limitations by
using a Convolutional Neural Network (CNN) and a novel striking a balance between the robustness of the system and the
method for extracting local features to distinguish between canonical form in signature forgery detection. The objective
genuine and forged signatures, with potential applications is to develop an approach that effectively captures natural
in organizations with a limited number of individuals for factors influencing signatures while maintaining computational
training and future forgery detection. [2] proposes a system efficiency. By employing a Siamese network and incorporating
for signature verification and forgery detection wherein this novel techniques such as Scattering wavelets, the SigScatNet
paper uses CNN and deep learning to extract unique features aims to achieve accurate and lightweight signature forgery
from pre-processed signatures, which are then compared to detection. This would contribute to enhancing the reliability
the signatures stored in the system to determine whether the and effectiveness of signature verification systems, improving
signature is real or fake. The system was tested on a dataset overall security and trust in various domains where signatures
containing 750 signatures from 150 individuals, and the results play a crucial role.
showcased how the proposed system works and then stated that
III. P ROPOSED M ETHODOLOGY
their method is more accurate and faster in the detection of
forged signatures than a trained professional. [3] developed A. SigScatNet Architecture
a solution using Convolutional Neural Network (CNN) to Fig. 1 illustrates the SigScatNet architecture, showcasing the
detect forgery in handwritten signatures by training the model integration of Scattering wavelets and the Siamese network.
on a dataset of signatures and predicting the authenticity This architecture forms the core of SigScatNet, enabling the
of provided signatures. [4] developed an interactive software robust detection of signature forgery and precise assessment
implementation of signature verification that includes both of signature similarity.
general learning from a population of signatures and special The SigScatNet architecture can be divided into two essen-
learning from multiple samples of an individual’s signature, tial parts: the Scattering wavelet component and the Siamese
resulting in improved accuracy. [5] proposes a signature network component.
verification system that uses pre-processing techniques such 1) Scattering Wavelets: Scattering wavelets are an effective
as RGB to Gray Scale Conversion, image binarization, and technique for extracting useful characteristics from signals or
Gaussian blurring. The system uses contour and SIFT features images in signal processing and image analysis. Both local
for feature extraction and SVM and K-Means algorithms for and global information is captured in a multiscale, translation-
model training. invariant representation provided by this method. The wavelet
Siamese networks have gained attention in recent research. transform, which breaks down a signal into several frequency
In [6], the methodology used involves the CEDAR dataset. components at various scales, serves as the foundation for the
The signatures are processed using various techniques such scattering transform. The scattering transform is better suited
as median filter, Otsu method, and morphological operations. for tasks like pattern identification and signal classification
The Siamese neural network is then trained and tested on this since it keeps the information about phase and magnitude
dataset, resulting in high accuracy rates. [7] conducted offline in contrast to conventional wavelet transforms. The scatter-
signature verification using a Siamese neural network based ing transform works via cascading wavelet transformations
on one-shot learning, which achieved successful classification and modulus non-linearities. A series of wavelet filters are
with fewer signature images. By utilizing the one-shot learning convolved with the input signal or image at various scales.
method, the model distinguished between genuine and forged Insight into local frequency is captured by the resulting
signatures without the need for a large amount of labeled data. wavelet coefficients. The magnitude data is then captured by
Experimental results on signature datasets demonstrated across applying the modulus non-linearity to the wavelet coefficients.
4NSigComp2012, SigComp2011, 4NSigComp2010, and BH- A depiction of multiscale scattering is created by repeating this
sig260 datasets, showcasing the effectiveness of their proposed procedure at several scales.
approach. [8] uses a Siamese Neural Network (SNN) with The scattering transform is incredibly helpful when it comes
a contrastive loss function and robust embedding to verify to signature analysis. The level of depth in a signature can
genuine and forged offline signatures. The experiments were range from broad strokes to finer details. We can record
conducted on challenging datasets like MCYT-75, GPDS, and these various informational scales by applying the scattering
CEDAR, and the proposed method outperformed other state- transform on signature images. The signatures’ global and
of-the-art methods in terms of accuracy. Their results show- low-frequency components, as represented by the coefficients
cased that the use of inter-quartile range and median absolute determined from lower scales, are what give the signatures
deviation measures in the embedding vector increases the their overall shape and structure. On the other hand, finer
accuracy of the algorithm for GPDS and MCYT-75 databases details and high-frequency components, such as minor texture
relating to the signature forgery domain. However, existing patterns and variation in stroke, are captured by the coefficients
approaches often face limitations in handling natural variations from higher scales. We consider a handwritten signature as
 Fig. 1. SigScatNet Architecture

an illustration in Fig. 2. The scattering transform produces The Euclidean distance D between the feature embeddings
a scattering representation with several scales of coefficients of input image 1 and input image 2 is computed as seen in
when applied to the signature image. Lower scale coefficients equation (5), where n denotes the number of features (128 in
encode the signature’s overall shape and structure, encapsulat- our case):
ing its general properties. The factors pick up finer features v
at larger scales, including differences in the thickness of the u n
uX
signature’s strokes, their curvature, and their texture patterns. D = t (xi − xi )2 (5)
In this process, the selection of scales and scattering factors i=1
is critical. We may regulate the level of detail recorded in the
scattering representation by choosing the right scales. Finer This Euclidean distance is normalized between 0 and 1,
features can be captured at higher scales, however doing so where a lower D denotes a higher degree of similarity between
may also result in an increase in noise or artifacts. It’s crucial the signatures.
to find a balance and pick scales that accurately depict the The Siamese model is trained using a triplet loss function
pertinent distinctive traits without obscuring the representation. during the training phase [9]. The triplet loss function is a
The scattering transform is further influenced by the choice of distance-based loss function designed to learn embeddings
wavelet filters, their number per scale, and their sizes. These by minimizing the distance between the anchor and positive
variables affect how the transform behaves and how many samples while maximizing the distance between the anchor
features are extracted. These parameters need to be carefully and negative samples. The triplet loss function plays a crucial
thought out in order to guarantee ideal feature extraction and role in learning discriminative embeddings by encouraging the
discriminative power for signature analysis jobs. model to bring genuine signatures (positive samples) closer
to the anchor signature and push forged signatures (negative
2) Siamese Model: The Siamese model is a potent method
samples) farther away from the anchor. This enables effective
for comparing two signatures to assess their similarity. In our
differentiation between genuine and forged signatures based
scenario, the Siamese model is used to compare a signature
on their feature representations.
under scrutiny with a base or reference signature, allowing
We utilize the trained Siamese model to determine a signa-
us to determine whether the signature is genuine or not. The
ture’s similarity score to the reference signature for validation.
Siamese model comprises of two identical neural networks
If the similarity score exceeds a predefined threshold, the
with the same architecture and weights. Each network receives
questioned signature is considered authentic; otherwise, it is
an input signature, processes it through a series of convolu-
labeled as a forgery.
tional and fully connected layers to extract relevant features.
The formula for the Triplet loss is shown in equation (6):
The Conv2D layers in the model use filters with the following
configurations: 16 filters in the first two layers and 32 filters
in the last two layers. The kernel size for all Conv2D blocks L = max(d(a, p) − d(a, n) + margin, 0) (6)
is set to 3 × 3, and the activation function used is ReLU. The
features from the two networks are concatenated and passed Where L represents the Triplet loss, d(a, p) represents the
into a similarity function. distance between the anchor sample a and the positive sample
The model takes an input image of size 180×300×3. Before p, d(a, n) represents the distance between the anchor sample
the scattering network’s operation, the image is converted a and the negative sample n. The margin is a hyperparameter
to grayscale. It then goes through the convolutional blocks, that specifies a margin or threshold value, max(x, 0) ensures
resulting in an output array of size 128, which represents that the loss is non-negative, as the loss is only calculated
the signature’s features. A similarity score is generated by when the positive distance plus the margin is greater than the
comparing the feature embeddings of the two signatures using negative distance. The model was trained with a learning rate
these 128-sized arrays. One commonly used similarity metric of 0.0005, a batch size of 32, and it underwent 100 epochs of
is the Euclidean distance, also known as the L2 distance. training.
 Fig. 2. Signatures at J=2 on performing Scattering Wavelet Operation.

B. Dataset same 55 writers. Therefore, for each person, we have access to

24 original signatures and 24 forgeries. Each image is resized
In our study, we addressed the task of signature verification to 300x180 pixels before being fed into the model.
and forgery detection using two different datasets: the ICDAR Out of the 55 individuals, 40 are utilized to train the model,
2011 SigComp Dutch Signature Dataset and the CEDAR while the remaining 15 people are reserved for testing. This
Dataset. approach ensures that the trained model is tested on entirely
There are two folders in the ICDAR 2011 SigComp Dutch new signatures that it has not encountered during the training
Signature Dataset: ”train” and ”test.” The ”train” folder con- phase.
tains samples from 64 people, each of whom has at least 24
genuine and 8 forged signatures. This dataset serves as the IV. E XPERIMENTS AND R ESULTS
primary training set for our model. The ”test” folder, on the The evaluation of our SigScatNet model on the ICDAR
other hand, contains samples from 42 different individuals, SigComp Dutch dataset and the CEDAR dataset resulted
each with a minimum of 12 authentic and 12 forged signa- in impressive performance metrics. The Receiver Operating
tures. We use this dataset to evaluate the effectiveness and Characteristic (ROC) curve, which plots True Positive Rate
generalizability of our trained model. (TPR) against False Positive Rate (FPR), exhibited an excep-
The CEDAR dataset comprises 1320 forgeries, with 24 tional Area Under the Curve (AUC) of 99.6% on the ICDAR
forgeries for each of the 55 writers. Additionally, it contains SigComp Dutch dataset and 99.9% on the CEDAR dataset.
1320 genuine signatures, with 24 original signatures from the This indicates the model’s ability to achieve high true positive
rates while maintaining low false positive rates, a crucial
aspect for accurate signature forgery detection. Additionally,
the Precision-Recall (PR) curve demonstrated an outstanding
Average Precision (AUPR) of 99.4% on the ICDAR SigComp
Dutch dataset and 99.9% on the CEDAR dataset, further
reinforcing the model’s effectiveness in correctly identifying
genuine and forged signature pairs. These results validate
the discriminative power of our SigScatNet model and its
robustness in capturing signature characteristics. Fig. 3 and
Fig. 5. Distance histograms for the Testing set on the ICDAR SigComp Dutch
Fig. 4 showcases the ROC and PR curve on the two datasets. dataset

Fig. 6. Distance histograms for the Testing set on the CEDAR dataset

Fig. 3. (i)ROC and (ii)PR curve on the Testing set of ICDAR SigComp Dutch the EER at the threshold coordinate came out to be 3.22%
dataset for the SigComp Dutch dataset and 0.058% for the CEDAR
dataset. This balance between false match and false non-match
errors denotes the point at which the model performs at its
best.
Fig. 7 and Fig. 8 showcase the FMR and FNMR plot with
the EER representation for the ICDAR SigComp Dutch dataset
and the CEDAR dataset, respectively.
These assessment metrics—the ROC and PR curves, dis-
tance histograms, and the False Error Rate versus Threshold
curve—all demonstrate the effectiveness of our Siamese model
in detecting signature forgeries. The model’s capacity to distin-
guish between authentic and forged signatures is supported by
its strong AUC and AUPR values, and the distance histograms
draw attention to the distinct differences between the two
Fig. 4. (i)ROC and (ii)PR curve on the Testing set of CEDAR dataset
groups.
We examined two histograms showing the Density vs. A quantitative evaluation of the model’s performance at the
Euclidean distances and Count vs. Euclidean distances be- best operating point is also provided by the EER computation.
tween embeddings to acquire additional insights into the These results demonstrate the accuracy and dependability of
model’s performance. These histograms visually represent the SigScatNet model in addressing the challenges related to
the distribution of distances between genuine pairings and signature forgery detection.
imposter pairs, allowing us to assess the model’s capability to
V. C OMPARISON WITH OTHER M ODELS
distinguish between authentic and forged signatures based on
the estimated embeddings. An important factor in successfully To evaluate the efficacy of our SigScatNet model, it is
categorizing signatures is the optimal threshold. imperative to compare its performance against other state-of-
Fig. 5 and Fig. 6 showcase the distance histograms for the the-art verification models in the research community, pro-
Testing sets on the ICDAR SigComp Dutch dataset and the viding valuable insights into the strengths and weaknesses of
CEDAR dataset, respectively. our model and highlighting the advancements achieved by
The False Match Rate (FMR) and False Non-Match Rate our proposed method. The results of these comparisons are
(FNMR) curve, which depicts the relationship between the presented in Table I and Table II.
error rates and the threshold value, provided another tool we Another important comparison to be done was to compare
used to assess the model’s effectiveness. The Equal Error Rate the SigScatNet model in two ways: one with Scattering
(EER), a crucial measure in the detection of signature fraud, is parameters and one without Scattering parameters. Table III
represented at the junction of these two curves. In our instance, showcases the advantages of using the Scattering parameters.
 TABLE I
C OMPARISON OF PROPOSED S IG S CAT N ET WITH OTHER METHODS ON THE S IG C OMP D UTCH DATASET USED IN THE RESEARCH COMMUNITY

Method Used Feature Extraction Classifier Accuracy EER

Cozzens et al. (2018) [10] CNN CNN 84.74 -
Pham et al. (2014) [11] Global Features Distance Based 87.80 -
Khan et al. (2015) [12] Local and Global Features DTW - 5.27%
Alvarez et al. (2016) [13] CNN VGG-16 97.00 -
Chen et al. (2018) [14] Correlation Coefficients Score based Likelihood 96.17 -
Parodi and Gomez (2014) [15] - Random Forest - 5.5%
Proposed SigScatNet Scattering + Siamese CNN Distance Based 97.76 3.22%

TABLE II
C OMPARISON OF PROPOSED S IG S CAT N ET WITH OTHER METHODS ON THE CEDAR DATASET USED IN THE RESEARCH COMMUNITY

Method Used Feature Extraction Classifier Accuracy EER

Zois et al. [16] K-SVD SVM 84.74 2.78%
Hadjadji et al. [17] Curvelet Transform Bi-SVM, OC-PCA 97.99 -
Çalik et al. [18] CNN 1-NN, CNN 98.3 -
Okawa [19] VLAD with KAZE SVM - 1.0%
Rateria et al. [20] CNN CNN - SVM 99.87 -
Zois et al. [21] Poset Grid Features SVM - 3.02%
Eskander et al. [22] Feature learning SVM - 4.63%
Kumar et al. [23] Surrounding pixels RBF-SVM 91.67 -
Hafemann et al. [24] CNN (Signet-SPP) SVM - 2.33%
Proposed SigScatNet Scattering + Siamese CNN Distance Based 99.91 0.058%

TABLE III
C OMPARISON OF PROPOSED S IG S CAT N ET WITH AND WITHOUT S CATTERING PARAMETERS ON S IG C OMP D UTCH AND ON CEDAR DATASET

Methodology Used Number of Parameters EER (For SigComp) EER (For CEDAR)
Only Siamese Network 3,327,056 3.22% 0.069%
SigScatNet (Siamese CNN + Scattering) 257,232 3.87% 0.058%

Fig. 7. False Match Rate and False Non-Match Rate Plot with EER
representation for the ICDAR SigComp Dutch dataset

Fig. 8. False Match Rate and False Non-Match Rate Plot with EER
The number of parameters in the plain Siamese network representation for the CEDAR dataset
rose to around 3.3 million as illustrated in Table III to give
comparable results. On the other hand, while incorporating the
Scattering parameters in the network, we were able to reduce similarity assessment by combining a Siamese deep learning
the parameters down to 257,232, which is a 92.27% decrease network with Scattering wavelets. This system proves to be a
in the number of parameters, still yielding a comparable or reliable and effective model that can accurately differentiate
even better EER. This further showcases that incorporating between authentic and fake signatures while also offering a
wavelets into the network drastically reduces the complexity measure of similarity by utilizing the strength of deep learn-
and dramatically increases computational efficiency. ing and the distinctive characteristics of scattering wavelets.
Furthermore, the addition of Scattering wavelets to our model
VI. C ONCLUSION AND F UTURE W ORK helped us achieve excellent efficiency without sacrificing
This research paper presents SigScatNet: a comprehensive accuracy. In order to help the model make wise decisions,
and effective solution for signature forgery detection and the wavelet-based feature extraction procedure extracted cru-
cial structural and textural data from the signatures. The [10] Brittany Cozzens, Richard Huang, Maxwell Jay, Kyle Khembunjong,
experimental evaluation of the SigComp Dutch and CEDAR Sahan Paliskara, Felix Zhan, Mark Zhang, Shahab Tayeb, Aeop Unite,
and Unlv Stem. Signature verification using a convolutional neural
datasets proved the viability of our suggested methodology. network. 2017.
The SigScatNet model demonstrated excellent performance [11] T.-A. Pham, H.-H. Le, and N.-T. Do. Offline handwritten signature
with an Equal Error Rate of 3.689% for the SigComp Dutch verification using local and global features. volume 75, page 231–247,
2014.
Dataset and 0.05787% for the CEDAR dataset after being [12] S. Khan, M. Ali Akbar, F. Shahzad, M. Farooq, and Z. Khan. Secure bio-
trained on a wide range of real and fake signatures. This metric template generation for multi-factor authentication. volume 48,
demonstrates the model’s capacity to balance false match pages 458–472. Elsevier, 2015.
[13] G. Alvarez, B. Sheffer, and M. Bryant. Offline signature verification
and false non-match errors, resulting in reliable and accurate with convolutional neural networks: Technical report, stanford university.
detection of signature fraud. 2016.
The proposed approach presents a practical solution to [14] X. h. Chen, C. Champod, X. Yang, S. pei Shi, Y. wen Luo, N. Wang,
Y. chen Wang, and Q. meng Lu. Assessment of signature handwriting
address the challenges associated with signature forgery. Our evidence via score-based likelihood ratio based on comparative measure-
model illustrates its potential for real-world deployment, ment of relevant dynamic features. volume 282, page 101–110, 2018.
enabling individuals and organizations to verify signatures [15] Marianela Parodi and Juan Gómez. Legendre polynomials based feature
extraction for online signature verification. consistency analysis of
quickly and accurately. It does this by utilizing cutting- feature combinations. volume 47, pages 128–140, 01 2014.
edge deep-learning techniques and harnessing the special [16] Elias Zois, Ilias Theodorakopoulos, Dimitrios Tsourounis, and George
properties of scattering wavelets. By offering a strong and Economou. Parsimonious coding and verification of offline handwritten
signatures. pages 636–645, 07 2017.
effective framework that combines the strength of Siamese [17] Bilal Hadjadji, Youcef Chibani, and Hassiba Nemmour. An efficient
deep learning networks and Scattering wavelets, this research open system for offline handwritten signature identification based on
makes a contribution to the field of signature fraud detection curvelet transform and one-class principal component analysis. volume
265, pages 66–77, 2017. New Trends for Pattern Recognition: Theory
and paves the way for increased security and confidence in & Applications.
applications for signature verification. Since this model is [18] Nurullah Calik, Onur Kurban, Ali Yilmaz, Tülay Yildirim, and Lutfiye
extremely lightweight, it can in theory be implemented into Durak Ata. Large-scale offline signature recognition via deep neural
networks and feature embedding. volume 359, 05 2019.
an app that can make the signature verification task in the [19] Manabu Okawa. From bovw to vlad with kaze features: Offline signature
palm of everyone’s hands. verification considering cognitive processes of forensic experts. volume
All other authors have no conflicts of interest. SigComp 113, pages 75–82, 2018. Integrating Biometrics and Forensics.
[20] Avani Rateria and Suneeta Agarwal. Off-line signature verification
Dutch dataset analyzed during the current study is available through machine learning. In 2018 5th IEEE Uttar Pradesh Section
in the ICDAR 2011 Signature Verification Competition (Sig- International Conference on Electrical, Electronics and Computer En-
Comp2011) repository. CEDAR dataset analyzed during the gineering (UPCON), pages 1–7, 2018.
[21] Elias N. Zois, Linda Alewijnse, and George Economou. Offline
current study is available in the CEDAR from University of signature verification and quality characterization using poset-oriented
Buffalo repository. grid features. volume 54, pages 162–177, 2016.
[22] George S. Eskander, Robert Sabourin, and Eric Granger. Hybrid writer-
R EFERENCES independent–writer-dependent offline signature verification system. vol-
ume 2, pages 169–181, 2013.
[1] Gowri P, Sivapriya G, Karthick Raja K, Kamaleshwar N K J, and [23] Rajesh Kumar, J.D. Sharma, and Bhabatosh Chanda. Writer-independent
Kesavaraj N. Real time signature forgery detection using machine off-line signature verification using surroundedness feature. volume 33,
learning. In 2022 Second International Conference on Advances in pages 301–308, 2012.
Electrical, Computing, Communication and Sustainable Technologies [24] Luiz Gustavo Hafemann and Luiz Soares de Oliveira. Fixed-sized
(ICAECT), pages 1–5, 2022. representation learning from offline handwritten signatures of different
[2] Subramaniam M, Teja E, and Mathew A. Signature forgery detection sizes. volume 21, pages 219–232, 09 2018.
using machine learning. In International Research Journal of Modern-
ization in Engineering Technology and Science, volume 04, 2022.
[3] S Jerome Gideon, Anurag Kandulna, Aron Abhishek Kujur, A Diana,
and Kumudha Raimond. Handwritten signature forgery detection using
convolutional neural networks. volume 143, pages 978–987, 2018. 8th
International Conference on Advances in Computing & Communications
(ICACC-2018).
[4] Harish Srinivasan, Sargur N. Srihari, and Matthew J. Beal. Machine
learning for signature verification. In Prem K. Kalra and Shmuel Peleg,
editors, Computer Vision, Graphics and Image Processing, pages 761–
775, Berlin, Heidelberg, 2006. Springer Berlin Heidelberg.
[5] Dr. Chayadevi M L, Anagha R, and Chandan Kumar G. Signature
Recognition and Forgery Detection. 11 2022.
[6] A. Agarwal and A. BS. International journal of creative research
thoughts (ijcrt), 9(8). volume 9, pages 659–662, 2021.
[7] M. Varol Arısoy. Signature verification using siamese neural network
one-shot learning. volume 3, pages 248–260, 2021.
[8] Amruta Jagtap, Dattatray Sawat, Rajendra Hegadi, and Ravindra Hegadi.
Verification of genuine and forged offline signatures using siamese
neural network (snn). volume 79, 12 2020.
[9] Soumitri Chattopadhyay, Siladittya Manna, Saumik Bhattacharya, and
Umapada Pal. Surds: Self-supervised attention-guided reconstruction
and dual triplet loss for writer independent offline signature verification.
01 2022.