Fingerprint Liveliness Detection using Stacked

Ensemble and Transfer Learning Technique

2024 IEEE International Conference on Computer Vision and Machine Intelligence (CVMI) | 979-8-3503-7687-6/24/$31.00 ©2024 IEEE | DOI: 10.1109/CVMI61877.2024.10781920

Vidya Kumari B H Shekar

Department of Computer Science Department of Computer Science
Mangalore University Mangalore University
Mangaluru, India Mangaluru, India
[email protected] [email protected]
0000-0002-9251-162X 0000-0003-4379-2960

Abstract—The effectiveness and dependability of fingerprint details and effectively differentiate between genuine finger-
recognition systems make them popular for biometric authen- prints and spoofing attempts [1] .
tication. This research addresses the challenge of distinguishing Another significant aspect of their study is the incorporation
between real and spoof fingerprint images using a comprehensive
approach that integrates data preprocessing, data augmentation, of deep learning models, specifically VGG-19 and ResNet-
feature extraction with pretrained models and stacked classifiers. 50 pretrained on ImageNet, to extract more abstract and
The study begins with rigorous data preprocessing techniques to discriminative features from fingerprint images. This approach
enhance the quality and consistency of the fingerprint images. leverages the power of deep learning to learn hierarchical
Subsequently, data augmentation is employed to increase the representations that enhance the detection accuracy under
diversity of the training dataset, enabling better generalization
and robustness of the classification model. Feature extraction is challenging conditions.
then performed using pretrained convolutional neural network Furthermore, the paper evaluates these approaches compre-
models, namely ResNet50, VGG16, and DenseNet201, to capture hensively across various scenarios, including within-dataset,
discriminative features from the fingerprint images .Stacked cross-sensor, and cross-dataset evaluations. This thorough
classifiers leverage the complementary strengths of multiple base assessment demonstrates the robustness of their proposed
classifiers, such as Random Forest, Extra Trees, and Support
Vector Classifier, to improve classification performance. The methods against different acquisition settings and types of pre-
proposed ensemble learning approach achieves an impressive sentation attacks and the exploration of fingerprint liveliness
accuracy of 99.30% in distinguishing between real and spoof detection.
fingerprint images, demonstrating its effectiveness in enhancing The research highlights the efficacy of Support Vector
biometric security systems. Machine (SVM) classifiers, which demonstrate robust perfor-
Index Terms—Fingerprint Recognition, Deep Learning, En-
semble Learning, Biometric Security mance across diverse grayscale arrays, effectively distinguish-
ing between real and spoof fingerprint images.This technique
I. I NTRODUCTION achieves great accuracy in liveness detection by using SVMs
Biometric authentication, particularly fingerprint recogni- to categorize over 50,000 fingerprint photos.
tion, has emerged as a ubiquitous method for verifying in- A significant aspect of the proposed approach involves
dividual identities in various applications ranging from access an ensemble classifier model, which outperforms traditional
control systems to mobile devices. The effectiveness of finger- image recognition methods employing RT, achieving a notable
print recognition systems lies in their ability to accurately dif- 90.34% accuracy rate. This improvement reflects a 2.5%
ferentiate between genuine fingerprint images and counterfeit reduction in error rate compared to existing methodologies,
representations. However, with the increasing sophistication bolstered by dataset augmentation techniques to enhance de-
of spoofing techniques, ensuring the security and reliability tection accuracy.
of fingerprint authentication systems has become a critical The paper’s methodology begins with the extraction of
concern. features using RT, followed by classification using SVM to
Creating a strong and dependable fingerprint identification identify liveliness in fingerprint images. The integration of
system that can successfully lessen the threat provided by Ridgelet decomposition with Fast Fourier Transform facili-
spoofing assaults is the main goal of this research. This section tates dimensionality reduction, crucial for optimizing feature
discusses some of the connected works. representation and enhancing classifier performance [5].
Using handcrafted features like Binary Statistical Image Furthermore, comprehensive evaluations across multiple
Features (BSIF), Local Binary Patterns (LBP), and Local performance metrics validate the effectiveness of the proposed
Phase Quantization (LPQ) is one noteworthy method. These ensemble classifier approach, demonstrating superior accuracy
features are chosen for their ability to capture intricate textural and precision in distinguishing between real and spoof fin-
gerprint images. These findings underscore the significance
of integrating advanced machine learning techniques with

Authorized licensed use limited to: Indian Institute of Information Technology Kottayam. Downloaded on February 02,2025 at 11:33:17 UTC from IEEE Xplore. Restrictions apply.
 rigorous dataset augmentation for robust liveliness detection dian filtering, and Gaussian smoothing. Additionally,
in biometric systems. extensive data augmentation was performed to generate
To distinguish between real and fake fingerprints, a brand- a diverse and representative dataset, further improving
new, lightweight fingerprint liveliness detection network is model generalization and robustness.
proposed. Foreground extraction, style transfer based on Cy- • Evaluation with Confusion Matrix Analysis: High true
cleGan, fingerprint image blocking, and an improved ResNet positive rates and low false positive rates were found after
with a multi-head self-attention mechanism make up the ma- a comprehensive analysis of the model’s performance
jority of the technique. [7]. Since the recommended approach using a normalized confusion matrix. This shows how
can effectively extract ROI and generate the end-to-end data dependable and successful the methodology is in practical
structure, the volume of data has expanded. The CycleGan net- biometric security applications.
work improves the model’s ability to generalize when it comes The remaining part of the document is structured as follows:
to phony fingerprints made from unknown materials. The Section II outlines the suggested technique and talks about the
inclusion of Transformer to the improved ResNet improves several approaches we employed, including stacked classifiers,
detection performance while reducing computing overhead. transfer learning, and preprocessing. The database, evaluation
The methodology for fingerprint liveliness detection em- metrics, and experimental setting are all covered in Section
ploys a sequential attention model combined with ResNet III. The confusion matrix and a thorough analysis of the
convolutional layers. Spatial attention (SA) and channel at- experimental results are shown in Section IV. The conclusion
tention (CA) mechanisms are used in this method to improve and findings from the experiment, together with suggestions
feature learning. The model includes three layers of sequen- for further research, are provided in Section V.
tial attention alongside five convolutional layers. Using the
LivDet-2021 dataset, many pooling techniques, including Max, II. P ROPOSED M ETHODOLOGY
Average, and Stochastic, are investigated. [6]. To guarantee the quality and consistency of the fingerprint
To evaluate feature extraction performance, comparisons are dataset, the research starts with careful data collecting and
made with popular Convolutional Neural Networks (CNNs) preparation according to the suggested approach. We gather
including DenseNet121, VGG19, InceptionV3, and ResNet50 a diverse collection of fingerprint images comprising both
versions like ResNet34 and ResNet50. The final prediction is authentic and spoofed samples. These images undergo prepro-
made using a Multilayer Perceptron (MLP) classifier that is cessing steps such as noise removal, normalization, and image
merged with a fully linked layer. The study also investigates enhancement to standardize their quality and facilitate effective
the impact of attention models and pooling strategies on feature extraction. By meticulously curating and enhancing the
ResNet-based feature extraction. dataset, we aim to mitigate biases and confounding factors that
In order to improve the accuracy of the fingerprint liveliness may affect the classification performance of the models.
detection when compared to the state of art techniques. The We use advanced data augmentation strategies after data
following contributions are made: preprocessing to enhance model generalization and enrich the
• Introduction of an Advanced Integrated Methodol- training dataset. Several augmentation techniques are used to
ogy:Developed a novel approach that combines trans- create artificial changes of the fingerprint pictures, such as
fer learning with pre-trained deep learning models rotation, scaling, flipping, and noise addition. Our algorithms
(RESNET50, VGG16, and DENSENET201) and ensem- are better equipped to capture the underlying patterns and
ble learning techniques. This integrated methodology variances found in real-world fingerprint data owing to this
leverages the strengths of both deep feature extraction and augmentation procedure, which also makes the dataset more
ensemble classification to improve fingerprint liveliness robust and diverse. Our goal is to improve the model’s resis-
detection. tance to overfitting and its capacity to generalize to previously
• Implementation of Stacked Ensemble Models: created unseen samples by enriching the dataset.
a strong stacked ensemble model by combining the Subsequently, we leverage the power of pretrained convo-
Support Vector Classifier (SVC), Random Forest (RF), lutional neural network (CNN) models, specifically ResNet,
and Extra Tree Classifier (ET). This ensemble approach VGG, and DenseNet, for feature extraction from the aug-
enhances the overall classification accuracy by aggregat- mented fingerprint images. These deep learning models have
ing the predictive power of multiple classifiers. been pretrained on large-scale datasets and have learned to
• Achievement of High Detection Accuracy: Achieved extract high-level features that are discriminative and invariant
a remarkable fingerprint liveliness detection accuracy of to various transformations. By utilizing the learned repre-
99%. This performance demonstrates the efficacy of the sentations from these models, we aim to capture meaning-
proposed methodology in accurately distinguishing real ful features from the fingerprint images that can effectively
and fake fingerprints, surpassing many state-of-the-art discriminate between genuine and spoofed fingerprints. These
methods. extracted features serve as the input to our ensemble learning
• Comprehensive Data Preprocessing and Augmenta- approach, which further enhances the classification perfor-
tion: Employed rigorous data preprocessing techniques mance through the combination of multiple base classifiers.
including intensity adjustment, contrast adjustment, me- Fig. 1 depicts a flow diagram of the suggested methodology.

Authorized licensed use limited to: Indian Institute of Information Technology Kottayam. Downloaded on February 02,2025 at 11:33:17 UTC from IEEE Xplore. Restrictions apply.
 P REPROCESSING M ETHODS where G(i, j) is the Gaussian kernel:
In fingerprint liveliness detection, several preprocessing 1 − i2 +j22
G(i, j) =e 2σ
methods are used to enhance image quality and prepare data 2πσ 2
for analysis. These methods include intensity adjustment, and σ is the standard deviation, controlling the amount of
contrast adjustment, median filtering, and Gaussian smoothing. smoothing.
Intensity Adjustment T RANSFER L EARNING WITH P RETRAINED M ODELS
Intensity adjustment modifies pixel values to enhance image (R ES N ET 50, VGG16, D ENSE N ET 201)
features, making the data more suitable for analysis. A simple Transfer Learning:
linear intensity adjustment can be stated as: In the context of fingerprint liveliness detection, transfer
′
I (m, n) = α · I(m, n) + β learning with pretrained models like ResNet50, VGG16, and
DenseNet201 involves initializing the models without their
where I(m, n) is the original intensity, I ′ (m, n) is the adjusted final classification layers, then customizing them by adding
intensity, α is the scaling factor, and β is the offset. and training new layers tailored for the specific detection task.
Fine-tuning follows, where the pretrained model’s weights
Contrast Adjustment are adjusted to better align with the characteristics of the
Contrast adjustment enhances the differences between vari- new fingerprint dataset. By leveraging the general features
ous regions of an image by stretching the intensity range. This learned from extensive prior training on datasets like Ima-
can be achieved through histogram equalization: geNet, these models efficiently distinguish between real and
(I(m, n) − Imin ) fake fingerprints, even when trained with limited labeled data.
I ′ (m, n) = × (L − 1) This approach accelerates model development and enhances
(Imax − Imin )
performance, particularly in tasks requiring intricate feature
where Imin and Imax are the minimum and maximum inten- extraction and robust classification.
sities, respectively, and L is the number of intensity levels.
ResNet50:
Median Filtering A deep residual network with 50 layers called ResNet50
Median filtering reduces noise by replacing each pixel’s was created to solve the degradation issue that arises when
value with the median value of its neighborhood: training extremely deep neural networks. In order to avoid
vanishing gradients when training deeper models, it employs
I ′ (m, n) = median{I(m + i, n + j)} residual blocks with skip connections. It is pretrained on a big
for (i, j) ∈ neighborhood. This is particularly effective for dataset such as ImageNet in transfer learning with ResNet50.
reducing ”salt-and-pepper” noise. The previous layers, which capture basic properties like edges
and textures, are kept and frequently adjusted to the current
Gaussian Smoothing dataset. The top classification layer is changed with a new
Gaussian smoothing reduces noise and detail using a Gaus- layer tailored to the particular task.
sian function to assign weights to pixels in the neighborhood: VGG16:
k
X k
X The convolutional neural network design VGG16, which
I ′ (m, n) = G(i, j) · I(m + i, n + j) consists of 16 layers with max-pooling layers and 3x3 con-
i=−k j=−k volutional filters, is renowned for its uniform structure and
simplicity. The pretrained model is used in transfer learning
with VGG16 in a manner akin to ResNet50. The convolutional
base, which functions as a feature extractor, can be frozen or
adjusted, and the top layers are swapped out for new ones that
are more suited to the intended purpose. This method works
particularly well for situations where spatial feature hierarchies
are significant.
DenseNet201:
With 201 layers, DenseNet201 is a densely connected
convolutional network distinguished by patterns of dense con-
nectivity among its levels. In order to promote feature reuse
and improved gradient flow during training, each layer receives
feature maps from all previous layers. DenseNet201 performs
well in transfer learning because it can identify complex
Fig. 1. The Proposed Methodology. patterns and details in data. [21].

Authorized licensed use limited to: Indian Institute of Information Technology Kottayam. Downloaded on February 02,2025 at 11:33:17 UTC from IEEE Xplore. Restrictions apply.
 S TACKED C LASSIFIER Unlike Random Forests, Extra Trees choose random thresh-
A stacked classifier is an ensemble learning approach olds for each feature instead of searching for the optimal
designed to enhance prediction accuracy by combining the ones. This approach reduces computational complexity while
outputs of multiple individual classifiers. The main idea be- increasing the level of randomness in the model.
hind stacking is that by integrating predictions from diverse
classifiers, the ensemble can capture different aspects of the III. E XPERIMENTAL S ETUP
data, potentially leading to improved performance compared
The experimental results, evaluation results using various
to any single classifier. This technique is particularly effective
metrics, and comparison with state-of-the-art techniques are
when the base classifiers possess complementary strengths or
presented in this part.All experiments were executed in Light-
when the dataset is complex and varied.
ning AI, an advanced online platform designed for stream-
Support Vector Classifier (SVC) lined and efficient AI model development and training. Light-
The Support Vector Classifier (SVC) is a supervised learn- ning AI provides a robust environment with capabilities that
ing algorithm used for classification tasks. It functions by enhance productivity and performance, including optimized
identifying a hyperplane in the feature space that best separates model training workflows, scalable computing resources, and
the classes. comprehensive monitoring and management tools.
Given a training dataset {(mi , ni )}ni=1 , where mi ∈ Rd
represents the feature vectors and ni ∈ {−1, +1} indicates the A. Datasets
class labels, SVC seeks to determine the optimal hyperplane The LivDet Dataset comprises approximately 19,000 images
defined by w · m + b = 0 that maximizes the margin between of genuine and spoofed fingerprints collected from various
the closest data points from different classes. sensors, including the Biometrika FX2000 and Cross-match
The decision function for SVC is expressed as: Verifier 300 LC. Fake fingerprints were created using materials
f (m) = sign(w · m + b) such as silicone, play dough, and gelatin. In this study, the
Alive Dataset provided authentic data, while silicone images
Here, w is the weight vector perpendicular to the hyper- represented fake data. More than 3,000 images were used for
plane, b is the bias term, and sign(·) yields +1 or −1 based training, and over 1,500 for testing. Fig 2 showcases examples
on the sign of the expression. of live and silicone fingerprints from the dataset.
The objective of SVC is to maximize the margin while
ensuring that all training examples are correctly classified:
1
min ∥w∥2
w,b 2

subject to ni (w · mi + b) ≥ 1 for all i

Random Forest Classifier (RF)
Random Forest is an ensemble learning technique that
constructs a multitude of decision trees during training. It
predicts the class by taking the average (for regression) or the
mode (for classification) of the individual trees’ predictions.
Let T denote the total number of trees in the forest. Each
tree Ti is built using a subset of the training data and a random
selection of features.
For classification, the final prediction from the random
forest ŷ(m) is the majority vote among the predictions of the
individual trees:
Fig. 2. Alive and Silicone images of LivDet Dataset.
ŷ(m) = mode(T1 (m), T2 (m), . . . , TT (m))
where Ti (m) is the prediction of tree Ti for the input m. The photos’ sizes range from 240x320 pixels to 700x800
Extra Trees Classifier (ET) pixels, depending on the sensor; nevertheless, they were all
scaled to 224x224X3 pixels in accordance with the pretrained
Extra Trees, or Extremely Randomized Trees, is another models’ input size.
ensemble method similar to Random Forests but introduces
additional randomness in both the selection of splitting thresh-
B. Evaluation Metrics
olds and the subset of data for each tree.
The prediction process for Extra Trees mirrors that of Ran- The formulas used for the performance metrics used in
dom Forests, with the final prediction ŷ(m) being determined evaluating the classifiers are as follows: Accuracy, Precision,
by aggregating the outputs from all the trees. Recall, and F1-Score.

Authorized licensed use limited to: Indian Institute of Information Technology Kottayam. Downloaded on February 02,2025 at 11:33:17 UTC from IEEE Xplore. Restrictions apply.
 on accuracy, precision, recall, and F1-score. The results are
tabulated in TABLE II and TABLE III below:

TABLE II
DENSNET201:

Classifier used Accuracy Precision Recall F1-Score

Random Forest Classifier 0.9723 0.9725 0.9723 0.9723
Support Vector Classifier 0.8748 0.8466 0.8428 0.8446
Extra Tree Classifier 0.9723 0.9725 0.9723 0.9723
Stacked Classifier 0.9786 0.9792 0.9786 0.9789

Accuracy is the ratio of correctly predicted instances to the

TABLE III
total instances. VGG16:
Precision, sometimes referred to as Positive Predictive
Value, gauges how accurate positive forecasts are. It is cal- Classifier used Accuracy Precision Recall F1-Score
culated as the proportion of actual positive forecasts to all Random Forest Classifier 0.9725 0.9723 0.9723 0.9723
positive predictions. Support Vector Classifier 0.8489 0.8166 0.8428 0.8294
Recall, sometimes referred to as True Positive Rate or Extra Tree Classifier 0.9725 0.9723 0.9723 0.9723
Stacked Classifier 0.9774 0.9790 0.9766 0.9777
Sensitivity, measures how well the classifier can identify every
positive example. It is the proportion of actual positives to true
positive predictions. The following diagram Fig 3 shows one of the confusion
The F1-Score is the harmonic mean of Precision and Recall. matrix using stacked classifier.
It provides a single metric that balances both Precision and
Recall.
IV. R ESULTS
Different classifiers, such as Random Forest (RF), Extra
Trees Classifier (ET), and Support Vector Classifier (SVC),
were assessed in the first stage of research utilizing ResNet-
50 for fingerprint liveliness detection based on metrics like
accuracy, precision, recall, and F1-score. Results showed that
RF and ET classifiers achieved high accuracy of 98.16%.
SVC, while slightly lower in performance with an accuracy
of 84.88%, still demonstrated respectable metrics in precision,
recall, and F1-score . Notably, the stacked classifier combining
ResNet-50 predictions with RF, ET, and SVC outputs showed
the most promising results, achieving an accuracy of 98.26%
and significantly higher precision, recall, and F1-score of
98.32%, 98.26%, and 98.29%, respectively. This ensemble Fig. 3. Confusion matrix for Resnet50 with Stacked Classifier.
approach suggests that integrating multiple classifiers enhances
overall performance metrics for fingerprint liveliness detection
Upon observing the strong performance of the stacked clas-
tasks. The results are tabulated in TABLE I as shown below:
sifier across ResNet-50, VGG16, and DenseNet models during
TABLE I
initial experiments for fingerprint liveliness detection, our
RESNET50: investigation progressed to explore transfer learning strategies.
These strategies involved combining various pretrained models
Classifier used Accuracy Precision Recall F1-Score
such as VGG16 with ResNet-50, DenseNet with ResNet-50,
Random Forest Classifier 0.9816 0.9826 0.9816 0.9821 and VGG16 with DenseNet. The objective was to harness the
Support Vector Classifier 0.8488 0.8166 0.8428 0.8292 collective strengths of these models through transfer learning
Extra Tree Classifier 0.9816 0.9826 0.9816 0.9821 and enhance their classification capabilities.
Stacked Classifier 0.9826 0.9832 0.9826 0.9829
The results of these experiments were tabulated in TABLE
IV:
Similarly, experiments were conducted using VGG16 and VGG16 + ResNet-50: This combination aimed to blend
DenseNet models for fingerprint liveliness detection, employ- the feature extraction capabilities of VGG16 with the deep
ing various classifiers and evaluating their performance based learning prowess of ResNet-50.

Authorized licensed use limited to: Indian Institute of Information Technology Kottayam. Downloaded on February 02,2025 at 11:33:17 UTC from IEEE Xplore. Restrictions apply.
 TABLE IV
C OMPARISION OF P RETRAINED M ODEL USED WITH S TACKED
C LASSIFIERS :

Pretrained Model used Accuracy Precision Recall F1-Score

Resnet50+Densenet201 0.9886 0.9872 0.9884 0.9878
Resnet50+Vgg16 0.9786 0.9792 0.9786 0.9789
Vgg16+Densenet201 0.9826 0.9832 0.9826 0.9829
Resnet+Densenet+Vgg 0.9930 0.9931 0.9930 0.9930

DenseNet + ResNet-50: Here, DenseNet’s dense connectiv-

ity pattern was integrated with ResNet-50’s residual learning
Fig. 4. Confusion matrix for Transfer learning and Ensemble model
framework to exploit their complementary advantages.
we also present the comparison of accuracy obtained in
VGG16 + DenseNet: Combining VGG16 and DenseNet various state of art methods tabulated in TABLE VI.
sought to capitalize on their respective architectural designs
and learning mechanisms. TABLE VI
C OMPARISION OF VARIOUS STATE OF ART METHODS :
All Three Models (VGG16, DenseNet, ResNet-50) with Method Accuracy
Stacked Classifier: Finally, integrating all three models in LFLDNet using Resnet [7] 95.27%
conjunction with the stacked classifier aimed to leverage en- Attention based Resnet50 [6] 97.78%
Ensemble Classifier [5] 90.34%
semble learning, aggregating predictions from multiple sources EaZy learning [4] 95.4%
to achieve robust performance. A-Stacking [1] 92.15%
Proposed Methodology 99.30%
These experiments were designed to demonstrate how com-
bining transfer learning with ensemble techniques can poten-
tially optimize the accuracy, precision, recall, and F1-score V. C ONCLUSION
metrics crucial for effective fingerprint liveliness detection In this study, we have developed a robust methodology
systems. for distinguishing between real and spoof fingerprint images
by integrating data preprocessing, data augmentation, and
In order to assess the method’s resilience and efficacy in feature extraction using pretrained deep learning models such
a variety of settings, it has also been evaluated using noisier as ResNet, VGG, and DenseNet. Leveraging these high-level
data. The results are as tabulated in TABLE V features, we employed an ensemble learning approach with
stacked classifiers, combining Random Forest, Extra Trees,
and Support Vector Classifier to achieve a superior classifi-
cation performance. Our experimental results demonstrated an
impressive accuracy of 99.30%, showcasing the effectiveness
TABLE V of our approach in enhancing the security and reliability of
C OMPARISION OF P RETRAINED M ODEL USED WITH S TACKED fingerprint recognition systems. This work not only highlights
C LASSIFIERS FOR DIFFERENT DATASETS :
the potential of combining advanced deep learning techniques
Dataset Accuracy Precision Recall F1-Score with ensemble methods but also provides a comprehensive
Green Bit 0.9520 0.9488 0.9520 0.9488 framework for future research in biometric authentication. The
DigitalPersona 0.9930 0.9931 0.9930 0.9930
CrossMatch 0.9780 0.9764 0.9780 0.9764
success of our approach underscores its applicability in real-
Biometrica 0.9930 0.9931 0.9930 0.9930 world scenarios, offering a significant improvement in the
detection of spoof fingerprints and contributing to more secure
and dependable fingerprint recognition systems.
ACKNOWLEDGEMENT
In particular, Biometrika, Crossmatch, and Identix from
Fig. 4 shows the confusion matrix obtained for all the three LivDet 2009, Biometrika, Digital, Italdata, and Sagem from
pretrained models and all the three classifiers, that is a result LivDet 2011, Biometrika, CrossMatch, Italdata, and Swipe
obtained by transfer learning and ensemble classifiers. from LivDet 2013, and Crossmatch, Digital person, Hi Scan,

Authorized licensed use limited to: Indian Institute of Information Technology Kottayam. Downloaded on February 02,2025 at 11:33:17 UTC from IEEE Xplore. Restrictions apply.
 and GreenBit from LivDet 2015, we acknowledge the Stanford [18] Kantipudi J, Dubey SR, Chakraborty S. Color channel perturbation
University-provided LivDet dataset that contained the images attacks for fooling convolutional neural networks and a defense against
such attacks. IEEE Transactions on Artificial Intelligence. 2020 Dec
used for the Liveliness Detection competitions in 2009, 2011, 21;1(2):181-91.
2013, and 2015. [8] [9] [12]. [19] Nagpal C, Dubey SR. A performance evaluation of convolutional neural
We acknowledge the use of OpenAI’s ChatGPT for assis- networks for face anti-spoofing. In 2019 International Joint Conference
on Neural Networks (IJCNN) 2019 Jul 14 (pp. 1-8). IEEE.
tance in refining the grammar, improving sentence structure, [20] Rajesh B, Dusa N, Javed M, Dubey SR, Nagabhushan P. T2CI-GAN:
and enhancing the clarity of the text throughout the paper. The Text to Compressed Image generation using Generative Adversarial
AI system was utilized solely for language enhancement, and Network. arXiv preprint arXiv:2210.03734. 2022.
[21] Huang G, Liu Z, Van Der Maaten L, Weinberger KQ. Densely con-
the content of the research, ideas, and conclusions presented nected convolutional networks. InProceedings of the IEEE conference
remain entirely the authors’ original work. on computer vision and pattern recognition 2017 (pp. 4700-4708).

R EFERENCES
[1] Agarwal S, Rattani A, Chowdary CR. A comparative study on hand-
crafted features v/s deep features for open-set fingerprint liveness
detection. Pattern Recognition Letters. 2021 Jul 1;147:34-40.
[2] Sabri MA, Ennouni A, Aarab A. An effective facial spoofing detection
approach based on weighted deep ensemble learning. Signal, Image and
Video Processing. 2024 Feb;18(1):935-42.
[3] Agarwal S, Chowdary CR. A-Stacking and A-Bagging: Adaptive ver-
sions of ensemble learning algorithms for spoof fingerprint detection.
Expert Systems with Applications. 2020 May 15;146:113160.
[4] Agarwal S, Chowdary CR, Sourabh V. EaZy Learning: An Adaptive
Variant of Ensemble Learning for Fingerprint Liveness Detection. arXiv
preprint arXiv:2103.02207. 2021 Mar 3.
[5] Adam EE, Sathesh A. Evaluation of fingerprint liveness detection by
machine learning approach-a systematic view. Journal of ISMAC. 2021
Mar 1;3(01):16-30.
[6] Kothadiya D, Bhatt C, Soni D, Gadhe K, Patel S, Bruno A, Mazzeo
PL. Enhancing Fingerprint Liveness Detection Accuracy Using Deep
Learning: A Comprehensive Study and Novel Approach. Journal of
Imaging. 2023 Aug 7;9(8):158.
[7] Zhang K, Huang S, Liu E, Zhao H. LFLDNet: Lightweight Fingerprint
Liveness Detection Based on ResNet and Transformer. Sensors. 2023
Aug 1;23(15):6854.
[8] Jung HY, Heo YS, Lee S. Fingerprint liveness detection by a
template-probe convolutional neural network. IEEE Access. 2019 Aug
22;7:118986-93.
[9] Yuan C, Xia Z, Jiang L, Cao Y, Wu QJ, Sun X. Fingerprint liveness
detection using an improved CNN with image scale equalization. IEEE
Access. 2019 Feb 27;7:26953-66.
[10] H. Y. Jung, Y. S. Heo and S. Lee, Fingerprint Liveness Detection by a
Template-Probe Convolutional Neural Network, in IEEE Access, vol. 7,
pp. 118986-118993, 2019, doi: 10.1109/ACCESS.2019.2936890.
[11] Nogueira RF, de Alencar Lotufo R, Machado RC. Evaluating software-
based fingerprint liveness detection using convolutional networks and lo-
cal binary patterns. In2014 IEEE workshop on biometric measurements
and systems for security and medical applications (BIOMS) Proceedings
2014 Oct 17 (pp. 22-29). IEEE.
[12] Nogueira RF, de Alencar Lotufo R, Machado RC. Fingerprint liveness
detection using convolutional neural networks. IEEE transactions on
information forensics and security. 2016 Jan 22;11(6):1206-13.
[13] Marcialis GL, Lewicke A, Tan B, Coli P, Grimberg D, Congiu A,
Tidu A, Roli F, Schuckers S. First international fingerprint liveness
detection competition—LivDet 2009. InInternational Conference on
Image Analysis and Processing 2009 Sep 8 (pp. 12-23). Springer, Berlin,
Heidelberg.
[14] Ghiani L, Yambay D, Mura V, Tocco S, Marcialis GL, Roli F, Schuckcrs
S. Livdet 2013 fingerprint liveness detection competition 2013. In2013
International Conference on Biometrics (ICB) 2013 Jun 4 (pp. 1-6).
IEEE.
[15] Ghiani L, Yambay DA, Mura V, Marcialis GL, Roli F, Schuckers SA.
Review of the fingerprint liveness detection (LivDet) competition series:
2009 to 2015. Image and Vision Computing. 2017 Feb 1;58:110-28.
[16] Pandya B, Cosma G, Alani AA, Taherkhani A, Bharadi V, McGinnity
TM. Fingerprint classification using a deep convolutional neural net-
work. In2018 4th International Conference on Information Management
(ICIM) 2018 May 25 (pp. 86-91). IEEE.
[17] Win KN, Li K, Chen J, Viger PF, Li K. Fingerprint classification and
identification algorithms for criminal investigation: A survey. Future
Generation Computer Systems. 2020 Sep 1;110:758-71.

Authorized licensed use limited to: Indian Institute of Information Technology Kottayam. Downloaded on February 02,2025 at 11:33:17 UTC from IEEE Xplore. Restrictions apply.