Fingerprint Recognition System

Synopsis Report submitted in partial fulfillment

of the requirement for the degree of

B. E. (Computer Engineering)

Submitted By

SARVESH GAONKADKAR

OM GHAG

TEJAS KUMBHAR

ANURAG KUMBHARE

Under the Guidance of

Prof. SNEHA ANNAPPANAVAR

Department of Computer Engineering

(An Autonomous Institute A iliated to University of Mumbai)

Vidyalankar Institute of Technology

Wadala(E), Mumbai 400 037

University of Mumbai

2024-25
 CERTIFICATE OF APPROVAL

For

Project Synopsis

This is to Certify that

SARVESH GAONKADKAR

OM GHAG

TEJAS KUMBHAR

ANURAG KUMBHARE

Have successfully carried out Project Synopsis work entitled

Fingerprint Recognition System

in partial fulfillment of degree course in

Computer Engineering

As laid down by University of Mumbai during the academic year

2024-25

Under the Guidance of

Prof. SNEHA ANNAPPANAVAR

Signature of Guide Head of Department

Examiner 1 Examiner 2 Principal

 i

Declaration

We declare that this written submission represents our ideas in our own words and where others'
ideas or words have been included, we have adequately cited and referenced the original sources. We
also declare that we have adhered to all principles of academic honesty and integrity and have not
misrepresented or fabricated or falsified any idea/data/fact/source in our submission. We understand that
any violation of the above will be cause for disciplinary action by the Institute and can also evoke penal
action from the sources which have thus not been properly cited or from whom proper permission has
not been taken when needed.

Name of student Roll No. Signature

1. Sarvesh Gaonkadkar 21102B0001

2. Om Ghag 21102B0002

3. Tejas Kumbhar 21102B0004

4. Anurag Kumbhare 21102B0005

Date:
 ii

Acknowledgements
This Project wouldn’t have been possible without the support, assistance, and guidance of a
number of people whom we would like to express our gratitude to. First, we would like to convey our
gratitude and regards to our mentor Prof. Sneha Annappanavar for guiding us with her constructive and
valuable feedback and for her time and efforts. It was a great privilege to work and study under her
guidance.

We would like to extend our heartfelt thanks to our Head of Department, Dr. Sachin Bojewar, for
overseeing this initiative which will in turn provide every Vidyalankar student a distinctive competitive
edge over others.

We appreciate everyone who spared time from their busy schedules and participated in the
survey. Lastly, we are extremely grateful to all those who have contributed and shared their useful insights
throughout the entire process and helped us acquire the right direction during this research project.
 iii

Table of Contents

1. Abstract

2. Introduction

3. Aim and Objectives

4. Literature Survey

5. Problem Statement

6. Scope

7. Proposed System

8. Methodology

9. Analysis

10. Design

11. Conclusion

12. Hardware and Software Requirements

13. References
 iv

1. Abstract
The advancements in the deep learning and computer vision, Convolutional Neural Networks (CNNs),
has become a key framework for image related tasks including fingerprint recognition system. On one
hand, fingerprints have been matched with traditional techniques such as Scale Invariant Feature
Transform (SIFT) and Oriented FAST and Rotated BRIEF (ORB) along with CNNs. However, previous
studies have demonstrated that CNNs are not well suited to working with large and complex datasets.
SIFT, although accurate, is computationally expensive whereas ORB is fast, but precise.

For fingerprint recognition tasks of fingerprint matching, Siamese Neural Networks (SNNs) have been
employed to overcome these challenges. SNNs are unlike traditional methods, which learn the similarity
among input fingerprints, and hence are very good for verification and matching tasks. Fingerprint pattern
recognition is shown to be improved with SNNs, and SNNs are especially capable of distinguishing
genuine from false fingerprints, when there is noise or partial prints. For mobile security systems with
limited computational resources, high efficiency is a must, and this is important.

Fingerprint recognition systems based on SNNs can provide the advantages of the pairwise comparison
rather than global classification, and the more accurate and reliable fingerprint matching. While SNNs
have been designed for local feature extraction and matching, their lightweight architecture enables them
to be deployed on mobile security applications, which demand resource efficiency.

In this paper the evaluation of application of SNNs in mobile fingerprint recognition systems, from a
comprehensive point of view, in order to understand the mobile security context in which SNNs can be
applied. Results indicate that for real time biometric authentication on mobile devices, SNNs offer an
excellent balance between accuracy and computational efficiency and are thus a good solution. Through
this study the validation of the potential of SNNs to improve the performance and security of mobile based
fingerprint recognition systems, at a time when there is growing demand for secure and efficient mobile
biometric solutions.
 1

2. Introduction
Fingerprint recognition systems have largely relied on Convolutional Neural Networks (CNNs) for feature
extraction as well as pattern recognition tasks. Strong performance across multiple biometric applications
has been made possible due to their ability to capture local spatial hierarchies within images. CNNs,
however, suffer in computational efficiency and the handling of global contextual information — both of
which are important in mobile security environments with limited resources.

In order to alleviate these problems fingerprint recognition employs the techniques of Siamese Neural
Networks (SNNs) especially in mobile security applications. Using a twin network architecture, which
learns similarity metrics between two inputs, SNNs are particularly well suited to tasks like fingerprint
matching and verification. SNNs can compare pairs of fingerprint images to check if the produced images
are from the same individual, with the additional advantage of having greater accuracy and robustness
in a biometric environment that poses a challenge.

As SNNs are compute efficient and are naturally focused on pairwise comparisons, they are especially
advantageous for mobile security, where fingerprint verification requires such comparisons. Unlike
traditional CNNs, SNNs are specifically designed to compare individual image pairs, and results indicate
that they are robust to fingerprints that are noisy or partially captured. SNNs also have the appropriate
resource constrained nature of mobile platforms, where computational demands can be balanced with
accuracy.

In this research, the demonstration of how the architecture of SNNs addresses the unique challenges of
mobile security and apply them to mobile fingerprint recognition systems. The combination of high
accuracy and speed makes SNNs ideal for mobile device biometric authentication systems, allowing for
fingerprint recognition systems to take advantage of SNNs to achieve both.
 2

3. Aim and Objectives

Aim:

This project aims to design and implement an efficient scalable and secure Siamese Neural Network
(SNN) based fingerprint recognition system designed for mobile devices. The application involves real
time fingerprint matching for user authentication in order to minimize computational overhead and
enhance data security. With personal and financial transactions increasingly moving to mobile devices,
and with fingerprint recognition becoming a robust biometric authentication system, it’s essential.

Key system aspects include:

- Efficiency: For low latency and resource constrained mobile environments.

- Accuracy: Fingerprinting that is reliable to distinguish between different fingerprints.
- Security: Fingerprint protection during profile registration.

An SNN is used in the project that can compute similarity score between fingerprints images to do correct
identification. Stored biometric data will be protected by AES encryption, and TFLite will optimize the
model for mobile performance.

Objectives:

- Develop an SNN model: For fingerprint recognition, to create and train a Siamese Neural
Network addressing challenges like lighting, rotation, and partial fingerprints.
- Ensure data security through AES encryption: Store fingerprint data based on AES encryption to
control the unauthorized access.
- Optimize the model for mobile platforms using TensorFlow Lite: Then, to convert and optimize
the SNN model for efficient mobile deployment using model quantization and GPU delegation.
- Integrate augmentation techniques: It is the focus to augment training with the aim of improving
the model’s robustness to variations in the images of fingerprints.
- Enable real-time fingerprint matching: Create the system for quick comparisons on the platform
using mobile hardware resources in order to authenticate users seamlessly.
- Deploy the system on mobile devices: Install the fingerprint recognition system on as a fingerprint
app on a mobile application where the users can enrol and identify fingerprints securely.
 3

4. Literature Survey
Fingerprint recognition systems have evolved from conventional Convolutional Neural Networks (CNNs)
to more complex Siamese Neural Networks (SNNs). The purpose of these advances is to drive biometric
systems to higher accuracy, greater efficiency, and increased robustness. In this review, an overview of
key SNNs and CNNs developments is provided, as applied to fingerprint recognition in mobile
applications.

Fingerprints: Fingerprint patterns are probably the most relied upon and widely used biometric identifiers
for security purposes due to their uniquely unalterable patterns and ease of acquisition. The fingerprints
of a person are unique compared to any other identical twins, as these patterns develop in the fetal stages
of human development and remain invariant across a lifetime. Due to the inherent distinctiveness and
permanence, fingerprints prove to be a good metric for identity verification and security applications.

Why Fingerprint Matching as a Security Measure: Fingerprint security is a reliable biometric

identification method, owing to the uniqueness and permanence of fingerprints, as well as their ease
of capture. Some of the main reasons why they suit secure systems are:

 Uniqueness: Different ridge patterns along with the different bifurcations of human beings
create a unique fingerprint for no two human beings; therefore, it automatically makes
fingerprints very reliable to identification systems.
 Immutable: Fingerprints are immutable during one's lifetime unlike facial features or voice
patterns. Once printed, they do not need refreshing like other biometrics. This makes them
extremely reliable over time in security systems; enrolled fingerprints cannot be deleted and
remain valid for identity verification over time.
 Universality: Fingerprint identification is quite extensively used and very easy to implement
for any population. It is also widely available for use in all fields, be it financial or law
enforcement agencies as well as on the handsets of mobile devices.
 User-Friendly: Fingerprint collection is quick and not intrusive, using commonly available
devices like optical sensors or scanners. This authentication method is practical and secure
to users and organizations since risks of unauthorized access exist as fingerprints cannot be
guessed, hacked, or shared like passwords.
 4

Fingerprint Matching Key Features: Appropriately chosen features from fingerprints are extracted
and form a biometric template to enable comparison of and matching for authentication or
identification purposes in fingerprint security systems.

 Ridge Patterns: Fingerprints contain ridge patterns in loops, whorls, and arches, mainly
due to their usage for classification and basic differentiation.
 Minutiae Points: The minutiae points, including ridge endings and bifurcations, are what
primarily help systems identify and match partially captured fingerprints.

 Ridge count and orientation: In this feature, the ridges are counted at the minutiae points
for accurate matching in fingerprints. The orientation of ridges is compared for better
results. The core is the center of a fingerprint, where the ridges form loops or spirals,
while the delta is formed by ridges, which split in a triangular fashion. Such reference
points will guide the alignment of the fingerprint images to be compared, especially if the
images have been rotated or distorted.

 Pores and Sweat Glands: Other fingerprint systems account for microscopic sweat gland
pores, which strengthens security and increases accuracy in some applications.

Advantages of Fingerprint Authentication Over the Traditional Security Systems: Fingerprint

security has many key advantages over old methods like passwords, PINs, and other biometrics. It
improves security:

 Comparing It with Other Biometrics for Resistance to Attacks: Facially, it is easily

compromised using photos or videos. Voice recognition can easily be compromised using a
 5

recording, whereas fingerprint systems are quite stronger since there is a mechanism for
liveness detection.
 Multi-dimensional Security Features: Fingerprint authentication utilizes more details like
minutiae points and ridge patterns, compared to passwords, which rely entirely on a one-
dimensional input. Therefore, it provides higher reliability on multiple dimensions as
compared to a one-dimensional password or PIN.
 Higher Entropy Compared to Passwords: Entropy calculates the complexity of security
elements. Low entropy corresponds to weak passwords and is easy prey to brute-forcing
attacks. In contrast, fingerprints correspond to much higher entropy because of the complex
and unique patterns they represent. Thus, predicting or guessing a fingerprint is almost
impossible.
 Low Effort-High Compliance: People opt for easy passwords. The reason is convenience.
Moreover, people forget to use advanced security, such as two-factor authentication.
Fingerprint Authentication is safe and easy to handle. Only a quick finger scan can do the
trick to increase compliance and ensure minimum weak security issues.
 Local matching for better privacy: New devices such as smartphones process fingerprint
data internally that prevents exposing associated biometric information to any other device.
Encryption, illustrated in systems such as Apple's Touch ID and Google's mechanism, makes
up for this deficiency.

Convolutional Neural Networks (CNNs): CNNs have become central to advances in image recognition,
from fingerprint recognition to image recognition, because they are capable of recognizing spatial patterns
and hierarchies of images. Major contributions in this area are:

- AlexNet (2012): The AlexNet was introduced in the world of image classification by Krizhevsky
et al, that brought a lot of improvements using deep layers, ReLU and drop out. In large scale
image datasets, it proved successful, and it has established itself as an important milestone in
CNN development, but beyond biometric tasks such as fingerprint recognition.

Despite the success of CNNs in feature extraction, they remain constrained in the pairwise comparison
tasks that are key to fingerprint matching and are consequently employed for fingerprint matching using
Siamese Neural Networks (SNNs).
 6

Siamese Neural Networks (SNNs): The usefulness of SNNs lies in the fact that they are specially made
for tasks where there is a need to measure the similarity between two inputs, as in biometric matching
systems such as fingerprint recognition.

- Fingerprint Recognition with SNNs (2017): SNNs have been used on fingerprint recognition tasks
where the model learns to compare two fingerprint images and learn a similarity function. A
robust matching of fingerprints is provided by this method, in the case where the prints are noisy
or incomplete.

The selection of SNNs’ architecture through which pairwise comparison is performed is particularly well
suited for mobile fingerprint recognition systems due to the fact that new fingerprints can be compared
against stored biometric templates in an efficient manner.

Application of SNNs in Mobile Fingerprint Matching: As compared to conventional SNNs, adsorption

of SNNs to fingerprint recognition in mobile devices presents several advantages, such as low
computational overhead and high accuracy.

- Pairwise Matching: The twin architecture of SNNs allows the system to achieve high accuracy in
matching tasks by comparing pairs of fingerprints, by attenuating the differences between
fingerprint images.
- Efficiency: A variety of SNNs are optimized for mobile platforms, which provides efficient
performance without sacrificing accuracy, an important property for mobile biometric systems.
- Robustness: The advantage of SNNs is that they are able to handle noisy or partial fingerprint
data, increasing system robustness in real world mobile applications where fingerprint scans may
not be the best, and making it easier to design for such situations.
 7

5. Problem Statement
Problem Statement:

With mobile devices heavily playing a role in one’s personal and work life, secure and facile authentication
systems have been an absolute necessity. Traditional password-based authentication methods are
becoming increasingly insufficient, as password breaches, phishing attacks, and negligent password
management increasingly threaten them. So, in this context, fingerprint recognition becomes more secure
and user friendly. But making fingerprint recognition systems effective on mobile devices is not a trivial
problem.

Challenges of Current Fingerprint Systems:

1. Computational Inefficiency: As conventional fingerprint recognition algorithms often need a lot of

computational resources; it takes a long time in processing. However this inefficiency makes
them impractical for real time applications on mobile devices, which have very limited processing
power and memory.
2. Accuracy and Variability: However, fingerprint recognition systems today face a lot of challenges
in terms of accuracy, especially in different lighting or partial fingerprint conditions. This in turn
can result in false positives (allowing unauthorized users to access), or false negatives
(forbidding authorized users from having access), deteriorating authenticity.
3. Data Security and Privacy: The security challenge of storing and handling sensitive biometric
data, including fingerprints, is great. If in a data breach biometric data is compromised, users
can’t easily change their fingerprint as they are doing with a password, with severe privacy
implications and potential misuse of the biometric information.
4. Integration with Mobile Hardware: Fingerprint recognition systems that run efficiently across
different platforms are difficult to develop because mobile devices come in diverse hardware
configurations. The problem of real time performance is further complicated with limited
processing power and memory which limits the execution of complex algorithms needed for
accurate fingerprint recognition.
 8

6. Scope
The focus of this work is building and deploying a Siamese Neural Network (SNN) based fingerprint
recognition system that is secure, efficient, and accurate for mobile devices. It covers several important
areas like:

- Development of the SNN Model for Fingerprint Recognition: It will be the focus of the project to
build a strong Siamese Neural Network that can accurately compare fingerprint images. Different
augmentation techniques are used to train the model to handle variations in fingerprint
presentation like difference in orientation, pressure as well as partial captures. It is to make sure
the system is still able to separate normal users from those allowed.
- Security of Fingerprint Data through AES Encryption: The system will ensure that encrypted
biometric data is used and that it will be implemented during the profile registration process using
AES. This is this critical measure that protects user’s fingerprint data from unauthorized access
and if data storage is compromised then also biometric information does not come out from the
safe unless proper decryption key is not present.
- Optimization for Mobile Devices using TensorFlow Lite: Since the resource constraint of mobile
platforms, the project will make use of TensorFlow Lite to convert and optimize the SNN model
to reduce the time consumption in mobile devices. The focus will be on reducing the
computational overhead and making the system work smoothly within the limited resources such
as processing power and memory provided by such devices as smartphones and tablets.
- Real-Time Fingerprint Matching for User Authentication: A real time fingerprint matching system
will be designed for real time fingerprint matching with a requirement of low latency to provide a
quick and smooth user authentication. The model provides a means to use mobile hardware
resources effectively and perform fast matching between the stored fingerprint and the one
during login attempts for usability.
- Integration with Real-World Mobile Applications: The last objective is to deploy the fingerprint
recognition system on Android devices. Then the system will be integrated with mobile app,
where users can register their fingerprints securely, store them safely and use them for
authentication during actions like unlocking their device or authorising transactions.
- Augmentation and Robustness: During the training phase, multiple image augmentation
techniques will be used to increase robustness to the system. The model generalizes better to
real world conditions, and it can effectively handle variations in the fingerprint quality and
presentation.
 9

7. Proposed System
For fingerprint recognition in mobile security, the proposed system implements Siamese Neural Networks
(SNNs). In contrast to standard biometric recognition techniques, the SNN architecture is specially
developed to perform pairwise comparisons, thus being a good candidate for fingerprint matching. The
goal of this system is to produce a highly efficient and accurate fingerprint recognition model that can be
deployed on mobile devices and run in real-time with secure data processing.

System Overview:

An SNN is used as the core of the system to compare two fingerprint images to identify their similarity.
With architecture optimized for small/mobile footprint, great computational overhead is minimized, and
maximum battery is attained. The design consists of a chain of convolutional layers which extract salient
features from input fingerprint images, and then dense layers that map each image to a compact feature
embedding. These embeddings are used for comparing fingerprint during authentication. The user takes
their fingerprint at the time of initial profile registration, then securely stores It on their mobile device and
uses it to generate the user’s biometric profile. In this process, the background adds fingerprint
augmentation to the fingerprint image (rotations, translations, and flips etc.) to train the model and make
it generalize. In this phase, the fingerprint data are stored secured by AES encryption so that it is secure
from sensitive information while being stored. During login attempts, the user’s input fingerprint is then
compared to the stored fingerprint data. After this initial phase the system processes fingerprint inputs in
real time without storing further fingerprint data. This is done to increase security because encryption is
handled by the mobile device’s operating system (OS) and sensitive data is stored.

Key Components:

- Siamese Neural Network Architecture: The twin network architecture is an architecture to

compare two inputs and decide their similarity. The network is split into two identical branches,
and each of them processes one of the input fingerprints. These branches are shared so that
each fingerprint goes through the same transformation, simplifying learning a similarity metric.
- Fingerprint Matching: The network is fed with two fingerprint images as inputs, outputs feature
embeddings for each, and then the system uses these features to generate features for each
fingerprint image. By taking these embeddings and comparing them using a similarity metric
 10

(such as cosine similarity), to decide if the fingerprints match or not, and then set a threshold that
verify if fingerprints belong to the same individual.
- Fingerprint Storage and AES Encryption: On initial registration, fingerprints are securely stored
on the device, protected with AES encryption, to protect biometric data. The fingerprint images
are augmented in the background using data augmentation techniques (e.g. flipping, rotation,
and translation) to improve training of model, and to make the model robust to variations in
fingerprint input.
- Real-time Processing: After initial registration, the system operates for real time processing
without storing new fingerprint data. The SNN model is placed into TensorFlow Lite (TFLite)
format, fastening the model for mobile devices. By the use of frequency and power scaling,
fingerprint matching can be achieved in milliseconds, and the system can be made responsive
and practicable for everyday use.

Advantages of the Proposed System:

- Lightweight Architecture: Traditional CNN based models require much more computational
power, and especially the SNN architecture, when optimized for mobile, demands far less. Due
to the efficiency required for use on smartphones, it is highly suitable for use in such a form.
- Secure Fingerprint Storage: During the initial registration, the user’s fingerprints are stored in the
system using AES encrypted, so that sensitive biometric data is protected. This feature enables
system to customize the fingerprint recognition process for each user. During this phase
augmentation is used to add an extra accuracy of the model.
- Scalability: The SNN model is lightweight and once registered does not require the permanent
storage of new fingerprint templates. It is easy to deploy across different mobile platforms,
irrespective of the hardware specifications. It is designed to scale, with modifications to the
underlying architecture needed only for a small amount.
- Battery Efficiency: During inference, power consumption is minimized for mobile application,
which is critical, once converted to TFLite, the SNN model. The system prevents fingerprint
recognition from using up the device’s battery with excessive energy.
 11

8. Methodology
A structured approach to developing the Siamese Neural Network (SNN) for mobile fingerprint recognition
is presented, beginning with real time data collection, preprocessing, model training, encryption, mobile
deployment and evaluation of performance.

Data Collection and Preprocessing:

Fingerprint data is collected in real time when the user first registers their identity and credentials on the
mobile device in this system. When this registration phase takes place, the fingerprint image is captured
and safely stored by AES encryption. The foundation of the user’s biometric profile is based on this
fingerprint data.

The preprocessing steps are essential to maintain consistency and quality in the input data before training
the model:

- Image Resizing: All fingerprint images are resized into a fixed size of 96x96 pixels. The
standardization that comes from this size also reduces computational resources and speeds up
processing time because this size is small enough. Smaller images, as the system is more
efficient on mobile devices, most of them have limited processing power.
- Normalization: Once resized, pixel values are scaled between 0 and 1. This last step makes the
model converge faster in training and adds the model’s ability to generalize under a variety of
lighting conditions, or fingerprint quality.
- Data Augmentation: During training, the use data augmentation techniques increase the
robustness of the model. Rotation, flipping, and translation are techniques such as these which
aid a SNN to be able to recognize fingerprints in different orientations or in different conditions.
It is important in dealing with the variability inherent in real world applications.
 12

Model Architecture and Training:

A convolutional layer, max pooling layer and dense layer are all used in building the SNN model. The
architecture is designed to extract distinctive features from fingerprint images and generate embeddings
for comparison.

- Training Objective: A contrastive loss function is used to train the model, minimizing the distance
between embeddings of matching fingerprints and maximizing the distance between
nonmatching pairs. For the pairwise comparison tasks, this is an ideal approach, and the SNN
also accurately distinguishes genuine and impostor fingerprints during the authentication
process.
- Hyperparameter Tuning: These key hyperparameters, learning rate, batch size and number of
epochs are fine-tuned in order to find the optimal model performance. In order to avoid overfitting,
early stopping is employed, allowing the model to generalize well among new unseen data.
- Encryption and Secure Storage: Biometric data is stored securely with the fingerprint data stored
using AES encryption at the time of registration. During a login attempt, the encrypted fingerprint
data is used for future comparison. Once the fingerprint has been matched against the user’s
input fingerprint, the system executes according to the mobile device’s security protocols of data
handling validating both data integrity and privacy.

Model Architecture and Training:

The SNN model is built with convolutional layers, max pooling layers and dense layers. The architecture
is designed to extract distinctive features from fingerprint images and generate embeddings for
comparison.
 13

- Training Objective: A contrastive loss function is used to train the model, minimizing the distance
between embeddings of matching fingerprints and maximizing the distance between
nonmatching pairs. For the pairwise comparison tasks, this is an ideal approach, and the SNN
also accurately distinguishes genuine and impostor fingerprints during the authentication
process.
- Hyperparameter Tuning: The model performance is fine-tuned on key hyper parameters such as
learning rate, batch size and the number of epochs. To prevent overfitting, early stopping is
performed, so that once the model sees new, unseen data, it generalizes well.
- Encryption and Secure Storage: In order to guarantee sensitive biometric information’s security,
the user’s fingerprint data is stored at the time of registration, using AES encryption. During a
login attempt, the encrypted fingerprint data is used for future comparison. A comparison
between the pre-stored fingerprint and the user input fingerprint is done, and after the
comparison, the data is processed in line with the mobile device’s security protocols to match
data integrity and privacy.
 14

9. Analysis
Figure A: The relationship between time per matching and similarity score:

The relationship between two important variables in the matching process is illustrated in this scatter plot.

Key Observations:

- X-axis (Time per Matching in seconds): Usually indicates the time needed to do a matching
operation, e.g. nearest neighbours or similarity matching algorithm.
- Y-axis (Similarity Score per Matching): It is a measure of similarity between two entities in terms
of matching process, from score of 0 to 1, where 1 is the maximal similarity.

Analysis:

- The plot most of the data points cluster in the top left area, around a similarity score of 1 and
time per matching around 0.06 seconds. This means that a large number of matching operations
are both very accurate (similarity close to 1) and very fast (small time per matching).
- Lower similarity scores (between 0.5 and 0.9) with different matching time (between 0.065 and
0.085 seconds) are shown in a few outlier points. These cases correspond to situations where
matching took longer or produced less similarity.
 15

Figure B: SNN (Subtract) Model Train and Test Accuracy Across 100 Epochs:

Here the comparison of train and test accuracy for a Spiking Neural Network (SNN) model with subtract
operation over 100 epochs is plotted in this graph.

Key Observations:

- X-axis (Epochs): Number of training iterations, from 0 to 100.

- Y-axis (Accuracy): Accuracy is measured from 0.70 to 1.0 (70% to 100%).

Analysis:

- The blue line represents the train accuracy of the SNN (subtract) model for which the accuracy
begins at approximately 70% and increases up to a little above 96% at the end of training
indicating effective learning.
- The orange line represents the test accuracy, starting at a comparable value, and increasing
more quickly to almost 98% after the 20th epoch and then staying at a stable level. Generalization
to unseen data shows strong evidence of minimal overfitting and this trend.
 16

Figure C: Using SNN(L1): Similarity Score vs Matching Time

Here similarity scores are compared against matching time of a fingerprint recognition system based on
the Siamese Neural Network (SNN) with L1 distance.

Key Observations:

- Y-axis (Similarity score per matching): It is a measure of how similar two fingerprint templates
are (0.55 to 0.90).
- X-axis (Time per matching): It shows the time taken for each matching from 0.055 to 0.090
seconds.

Analysis:

- Most matches fall into the top left (higher similarity scores, faster matching times), with some
points close to a maximum similarity score (0.90) at low times (0.055 to 0.065 seconds).
- At slightly higher matching times (above 0.065 seconds) there are some outliers with lower
similarity scores (around 0.55-0.75).
- A clear trade-off between similarity scores at different time intervals, with the highest
performance in both accuracy and time near 0.90 score mark.
 17

Figure D: Training Accuracy Comparison of SNN Models

This line graph compares the training accuracy of two SNN (Siamese Neural Network) models over 15
epochs.

Key Observations:

- Y-axis (Accuracy): Represents the accuracy of each model during training.

- X-axis (Epochs): Number of epochs during training, ranging from 1 to 15.

Analysis:

- L1 SNN (Blue Line): Shows a gradual increase in accuracy, starting at around 60% and steadily
climbing to approximately 78% over 15 epochs. This model improves consistently but at a slower
rate.
- Subtract SNN (Yellow Line): Begins with a higher initial accuracy, around 83%, and reaches a
peak of about 92% accuracy in the first few epochs. The improvement slows after 5 epochs but
maintains a higher accuracy compared to the L1 SNN.
 18

Figure E: Test Accuracy Comparison of SNN Models over 100 Epochs

This line graph compares the test accuracy of two SNN (Siamese Neural Network) models over 100
epochs.

Key Observations:

- Y-axis (Accuracy): Represents the test accuracy of each model.

- X-axis (Epochs): Number of epochs during the testing phase, ranging from 1 to 100.

Analysis:

- L1 SNN (Blue Line): The test accuracy of the L1 model remains fairly consistent, fluctuating
between 55% and 60% throughout the epochs. There is no significant improvement over time.
- Subtract SNN (Yellow Line): The Subtract SNN model starts with a high accuracy of
approximately 90% and shows slight improvements, peaking close to 98% before stabilizing.
 19

Figure F: Comparison of Similarity Score of Models:

In the comparison, comparison of the similarity scores of two SNN (Siamese Neural Network) models
with different matching methods for the same matching time.

Key Observations:

- Y-axis (Similarity Score per Matching): This is the amount of how well each model can match the
data. The closer the score is to 1, the stronger the match.
- X-axis (Time per Matching): It is the time taken by each model to match.

Analysis:

- SNN (Sub) (Blue Dots): This method shows high similarity scores consistently and most of the
clusters are near 1.0, indicating strong performance. And the time per matching stays low and is
concentrated around 0.06 seconds.
- SNN (L1) (Orange Dots): The similarity scores of this method are more spread out, with values
ranging from 0.4 to 0.9 which exhibit more variance in performance. The matching times are
more widespread, beyond 0.1 seconds.
 20

Figure G: The Matching Rate Comparison of Models:

The matching rates of three models are compared in a verification or classification task and this bar graph
compares them.

Key Observations:

- Y-axis (Matching Rate %): It represents the accuracy of each model to identify or match data.
- X-axis (Models): It compares three different approaches.

Analysis:

- Traditional Minutiae-based (Gray Bar): It achieves a matching rate of about 60% which is
relatively low.
- CNN-based (Orange Bar): It demonstrates a matching rate of around 85% which is a better
accuracy compared with the traditional method.
- Siamese Neural Networks (SNNs) (Blue Bar): It achieves matching rates close to 100% and
demonstrates the best performance of the models.
 21

Figure H: Comparison of Storage Requirements of Models:

The storage requirements for each of the three models are shown in this bar graph.

Key Observations:

- Y-axis (Storage Requirement in MB): It is the memory required per model.

- X-axis (Models): The same three models are then compared.

Analysis:

- Traditional Minutiae-based (Gray Bar): It takes about 1 MB of storage, making it the least memory
demanding option.
- CNN-based (Orange Bar): This means that the storage requirements are significantly higher than
that.
- Siamese Neural Networks (SNNs) (Blue Bar): Strikes the balance between minutiae-based
model and CNN based model, using around 3 MB.
 22

Figure I: Speed Comparison of Models:

Below is a graph depicting the time (in milliseconds) taken per match by each model.

Key Observations:

- Y-axis (Time per Match in ms): Indicates how long each model takes to perform a match.
- X-axis (Models): Compared to the following models:

Analysis:

- Traditional Minutiae-based Model: The time of shows the highest at around 40 ms which means
it is one of the slowest to do among the options.
- CNN-based Model: One that is faster than the minutiae based (approximately 30 ms per match).
- Siamese Neural Network (SNN): With a time per match below 10 ms, the fastest model.
 23

Figure J: Accuracy Comparison of Models:

This graph shows the accuracy of each model, as a percentage.

Key Observations:

- Y-axis (Accuracy %): They represent the accuracy of each model.

- X-axis (Models): Compares them in the following models.

Analysis:

- Traditional Minutiae-based Model: It has an accuracy around 70%.

- CNN-based Model: Approximately 90% accuracy shows improved.
- Siamese Neural Network (SNN): It achieves the highest accuracy, approximately 95%.
 24

Figure K: Comparison of Fingerprint-Matching Models in Computational Complexity:

On a logarithmic scale, this graph compares the computational complexity of Traditional Minutiae based,
CNN based, and SNN based fingerprint matching models using the number of Floating-Point Operations
per Second (FLOPs).

Key Observations:

- Y-axis (FLOPs): Is the computational complexity represented on a logarithmic scale.

- X-axis (Models): It compares the following models.

Analysis:

- Traditional Minutiae-based Model (Gray Bar): Has the lowest FLOPs, around 106. It is found that
this method demands much less computational resources than the other models.
- CNN-based Model (Orange Bar): The FLOPs are displayed at approximately 10 11 This suggests
that CNN based models are very computationally expensive and require much higher number of
operations to do fingerprint matching.
- Siamese Neural Network (SNN)-based Model (Blue Bar): Shows a FLOPs value around 1010 Still
orders of magnitude above traditional minutiae based approaches, , which is slightly lower than
CNN based models.
25
 26

10. Design
1. Data Preprocessing Flowchart:
This flowchart describes the critical steps for the preparation of a biometric input data for
processing within the model.
- Start (Raw Biometric Input): Biometric data, specifically, fingerprint images are captured
at the raw level, and after that the process starts.
- Input Formatting (Resize to 96 x 96 Pixels): The raw input images are then resized to a
common dimension of 96x96 pixels to standardize the image size to help ease efficient
processing. Along with this, the computational load is reduced and the resizing
guarantees consistency in the data set.
- Normalization (Scale Values to Range (0,1)): All images are normalized to the range
from 0 to 1 in terms of pixel values. The scaling improves the training efficiency and
achieves higher consistency of the model input to learn from the data.
- Data Augmentation (Apply Flipping, Rotation, Translation): Different augmentation
techniques are used to change the images, such as flipping, rotating and translating in
order to enlarge the training dataset. In this approach, the model generalizes better and
is more robust to the variation of fingerprint presentation.
- Output (Pre-processed Biometric Data): To a certain extent, the biometric data is
rendered suitable for model input by following these preprocessing steps; therefore, this
biometric data can be further processed and analysed.

2. Model Architecture:
The architecture of the Siamese Neural Network used for fingerprint recognition is detailed in this
flowchart which showcases its multistage design.
- Two Inputs (Fingerprint 1 and Fingerprint 2): An architecture is presented, which starts
with two different inputs, each corresponding to a fingerprint. First, the user’s profile
(previously registered fingerprint) provides one input, and the second input is captured
when the user attempts to login. The setup allows to compare the biometric data.
 27

- Feature Model: Each fingerprint is processed by a specialized feature extraction model

that extracts salient features necessary for good comparison.
- Subtraction Layer: A subtraction layer turns into feature representations of both inputs,
and they are compared using it. That allows the model to work with the differences in
the fingerprints.
- Convolution and Pooling: The resultant differences are processed through a series of
convolutional layers followed by pooling operations. At this stage, there is a need to
extract hierarchical features whilst simultaneously keeping dimensionality low, in order
to enable the model to find relevant patterns.
- Flatten: The flattened processed features are then fed to feed to the following dense
layers.
- Dense 64, Dense 1: It is flattened vector through fully connected (dense) layers. The first
dense layer with 64 neurons is intended to capture complex feature interactions, and the
second, dense layer contains a single neuron which outputs a single prediction score.
- Sigmoid: Finally, a final dense layer feeds to an application of a sigmoid activation
function to transform raw output into a probability score indicating the probability of
match.
- Output (1: final output binary (Same, 0: Not Same): The final output is binary, a value of
1 means match (inputs belong to the same person), a value 0 means do not match
(inputs do NOT belong to the same person).

3. Security Architecture:
The secure fingerprint data processing flow using a Siamese Neural Network (SNN) for
fingerprint matching in an Android device context is shown in this diagram. It highlights the key
security items that protect biometric data from the time of capture to the time of processing, and
ultimately to the time of storage.
- Biometric Data Capture: The process of security begins with capturing biometric data
which is fingerprints via the device sensors. This initial stage is very important to make
sure that the input data is handled securely.
 28

- Data Encryption (AES-256): The biometric info is captured and after that the biometric
information goes through AES-256 encryption, which is considered one of the most
robust encryptions. This step in fact protects from exposure of sensitive biometric data
raw form transmission, storage, or via unauthorized access.
- Secure Storage (Android Keystore): And finally, in the Android Keystore is the encrypted
biometric data and the TFLite model used for inference. This component ensures that
cryptographic keys stay contained within the device, so that defense against
unauthorized tampering and access to encrypted fingerprint data and the machine
learning model used to learn who is who is strengthened.
- BiometricPrompt API Authentication: Before further processing it relies on the
BiometricPrompt API to check for user authentication. This API makes it easy for a
secure authentication flow using the biometric capabilities of the device, e.g., fingerprint
or facial recognition, to be used only with those who are allowed to activate the SNN
model for fingerprint matching.
- Access SNN Model for Inference: The system retrieves the Siamese Neural Network
(SNN) model for inference upon successful authentication. The fingerprint comparison
process with this model is highlighted with the secure TensorFlow Lite model and the
input fingerprint is run through it to compare the stored biometric data to make a similarity
determination and produce a matching score.
- Output (Encrypted Result Storage): The SNN inference produces a result, i.e. does it
match or not, which is securely stored. The matching decision is encrypted and stored
in a storage location that limits access by only authorized personnel, leaving the
confidentiality of the matching decision undisturbed. It protects the results from possible
data breach or unauthorised exploitation.

4. Mobile Deployment Flow:

The deployment of a trained Siamese Network model (SNN) for fingerprint authentication on a
mobile device is illustrated by this flowchart.
 29

- Trained SNN Model (TensorFlow): Using TensorFlow framework provides robust tools
to build and optimal deep learning models, the process starts with SNN model training.
- Model Conversion (TensorFlow Lite): The model is then trained after which the model is
converted into TensorFlow Lite (TFLite) format. It is crucial for the model to be able to
operate efficiently on mobile devices to achieve performance optimality and accuracy at
the same time.
- Integrate into Android App Using TFLite Interpreter: Finally, the converted model is used
in an Android application using the TFLite interpreter to do on device inference. It
reduces the reliance on external servers, and it improves user privacy.
- Input1 and Input2 (Biometric 1 and Biometric 2): For comparison, the model takes two
biometric inputs (fingerprints). The first input is taken from the user’s profile (previously
registered fingerprint), while the second input is captured when the user is trying to log
in. With this setup, the user’s authentication data can be directly analysed.
- TFLite Inference: The on-device inference engine is used to process the inputs and
generate embeddings, generating 64 dimensional vectors which represent the feature
sets of the input fingerprints.
- Output1 and Output2 (64-D Vector): Outputs are two 64 dimensional vectors, one for
each biometric input, which contain the essential features for comparison.
- Compute Similarity: Between the two output vectors, a similarity measure is computed,
which measures degree of similarity between the fingerprints and aid in authentication
decisions.
- Result (Similarity Score for Authentication): The system calculates the similarity score
and based on the similarity score decides if the two fingerprints come from the same
person. If the score is high, there is a match, and authentication occurs; if low, there is
a mismatch.
 30

11. Conclusion
Conclusion from fingerprint recognition of image: The fingerprint recognition image is made up of
three panels that present the fingerprint recognition process in the context of biometric security. The
different labels and scores demonstrate how the system evaluates the similarity between fingerprints:

1. Left Panel (Input Image): - Label: Input: [504 0 0 3]

It is an input fingerprint image on this panel. The fingerprint looks a bit rotated perhaps an initial
raw (or unprocessed image) that needs correction or normalization. The print is clear, ridge lines
are visible, but there are some borders or artifacts around the edges of the image.
2. Middle Panel (Matched Fingerprint): - Label: O: 0.94, [504 0 0 3]
This panel has a high match score (0.94) making this is a successful matching attempt or
recognition. The fingerprint pattern looks corrected or aligned with a well-centred clear fingerprint
pattern. This score of 0.94 is high, meaning input fingerprint and a stored template are very
similar (i.e. a success match).
3. Right Panel (Unmatched Fingerprint): - Label: X: 0.00, [182 0 1 3]
As this score is 0.00, this is a failed matching attempt. Here the fingerprint seems darker and
more distorted, and a few ridge lines seem less defined than in the middle panel. This fingerprint
does not match because the lower quality or distortion of it.

Conclusion derived from the figures: The figures are used for an in-depth comparison of different
fingerprint matching models including Siamese Neural Networks (SNNs), CNN based models and
traditional minutiae-based methods.

1. Performance and Accuracy:

 31

- The Siamese Neural Networks (SNNs) are nearly perfect with accuracy rates nearly 100%
and also display high similarity scores. For the fingerprint recognition tasks, the SNN
(Subtract) model outperforms the SNN (L1) model, and it is the most effective model.
- As with traditional methods, CNN based models have accuracy rates of 85–90% but are
dramatically slower and less efficient than SNNs.
- The matching rates of traditional minutiae-based models are lowest (around 60%-70%) and
thus the least suitable for high demanding fingerprint matching tasks.
2. Speed and Efficiency:
- The fastest matching times (below 10 milliseconds) make SNNs particularly well suited to
real time fingerprint recognition, for example unlocking mobile devices or secure access
points.
- SNNs are moderately slow, but much more efficient than CNNs.
- Slowest to perform are traditional minutiae-based methods, which take about 40
milliseconds to match, which makes them impractical for use in real time applications like
mobile security.
3. Computational and Memory Efficiency:
- The SNNs come up with its good trade-off that uses less computational power than any
CNN model and yet produces good performance. They also have moderate storage
requirements which are suitable for devices with little resources, such as mobile devices.
- CNN based models are computationally expensive in terms of memory and processor
requirement and may not be convenient for mobile devices constrained with hardware
limitations.
- On the other hand, models based on traditional minutiae can make use of the least
resources, However, their accuracy and speed are not high enough to be desirable in
secure fingerprint recognition.
4. Training and Generalization:
- The SNN models learn extremely well and there is no overfitting on both training and
unseen data. For instance, the SNN (Subtract) model obtains 98% test accuracy, which is
very accurate for mobile applications.
- SNN models generalize as well but are outperformed by the CNN models that are more
efficient.
- Training and generalization are not evaluated on traditional methods, yet they are
considered less adaptable to more complex tasks like fingerprint matching.
 32

Implications for Mobile Fingerprint Security:

- Compared to other security systems, such as those based on keystroke dynamics, this analysis
is particularly relevant for mobile fingerprint security systems that need to achieve high accuracy,
fast response times, and efficient use of computational resources.
- Mobile security applications require a quick and accurate fingerprint matching and SNNs are the
best choice for mobile security. They are able to achieve near perfect accuracy within
milliseconds so that it doesn’t impact the user experience and at the same time keep the security
strong.
- Mobile use for CNN based models may be plausible but they would face a ceiling on energy
consumption possibly because of greater computational requirements or memory, causing
slower unlock times or battery drain.
- Modern mobile security requires models that are faster and more accurate, but traditional
minutiae-based models are too slow and too inaccurate, resulting in frequent misidentifications
or slower unlocking times, decreasing the user experience.

Therefore, the conclusion is that Siamese Neural Networks (SNNs) are the most appropriate models for
mobile fingerprint security due to the excellent balance between speed, accuracy and resource
efficiency, compared to other proposed models. Thanks to their quick response time and high accuracy,
mobile devices can be unlocked quickly and securely in a modern mobile environments.
 33

12. Hardware and Software Requirements

Hardware Requirements:

- Mobile Device: Android smartphones or tablets with built in fingerprint sensor, minimum 2GB
RAM, running Android 8.0 or later.
- PC for Development: Processor: At least 2.5 GHz Intel i5 or equivalent. RAM: Training and
preprocessing tasks take up 8GB or higher.
- Storage: Datasets and models require at least 100GB available. GPU: Accelerated model
training recommended NVIDIA GPU.

Software Requirements:

- Operating Systems:
 Development: Windows 10, macOS, or Linux.
 Deployment: If you have Android, version 8.0 or later.
- Development Tools: Python: For SNN implementation and the model training.
- TensorFlow & Keras: For building, training and optimizing of the SNN model.
- Android Studio: It can be used to develop the mobile application.
- AES Encryption Tools: AES encryption is implemented using PyCryptodome.
- Libraries and Dependencies: OpenCV, NumPy, Pandas, TensorFlow/Keras, TFLite Converter.
- Version Control and Collaboration Tools:
 Git: For version control.
 GitHub/GitLab: It is for cloud-based code repositories.
 34

13. References
1. "Fingerprint Matching Algorithm for Android." International Journal of Engineering Research
and Technology (IJERT), vol. 2, no. 10, 2013
2. Chen, Y., & Wang, H. "Efficient Fingerprinting-Based Android Device Identification With
Zero-Permission Identifiers." ResearchGate, 2016
3. Gune, S., & Khedkar, S. "The Recognition of Fingerprints on Mobile Applications - An
Android Case Study." ResearchGate, 2016.
4. "Fingerprint Hardware Abstraction Layer (HAL)." Android Developer Documentation.
5. Vangala, P. "How Fingerprint Scanners Work." Android Authority, 2017.
6. B. Bakhshi and H. Veisi, "End to End Fingerprint Verification Based on Convolutional Neural
Network," 2019 27th Iranian Conference on Electrical Engineering (ICEE), Yazd, Iran, 2019,
pp. 1994-1998, doi: 10.1109/IranianCEE.2019.8786720.