Computer Modeling in

Engineering & Sciences

Tech Science Press

DOI: 10.32604/cmes.2023.023674

ARTICLE

Iris Liveness Detection Using Fragmental Energy of Haar Transformed Iris

Images Using Ensemble of Machine Learning Classifiers

Smita Khade1 , Shilpa Gite1,2, *, Sudeep D. Thepade3 , Biswajeet Pradhan4,5, * and Abdullah Alamri6
1
Symbiosis International (Deemed University), Symbiosis Institute of Technology, Pune, 412115, India
2
Symbiosis Centre for Applied Artificial Intelligence, Symbiosis International (Deemed University), Pune, 412115, India
3
Computer Engineering, Pimpri Chinchwad College of Engineering, Pune, 411044, India
4
Centre for Advanced Modelling and Geospatial Information Systems (CAMGIS), School for Civil and Environmental
Engineering, Faculty of Engineering and Information Technology, University of Technology Sydney,
Sydney, New South Wales, 2007, Australia
5
Earth Observation Center, Institute of Climate Change, University Kebangsaan Malaysia, UKM, Bangi, Selangor, 43600, Malaysia
6
Department of Geology & Geophysics, College of Science, King Saud University, P.O. Box 2455, Riyadh, 11451, Saudi Arabia
*Corresponding Authors: Shilpa Gite. Email: [email protected]; Biswajeet Pradhan. Email: [email protected]
Received: 09 May 2022 Accepted: 13 September 2022

ABSTRACT
Contactless verification is possible with iris biometric identification, which helps prevent infections like COVID-19
from spreading. Biometric systems have grown unsteady and dangerous as a result of spoofing assaults employing
contact lenses, replayed the video, and print attacks. The work demonstrates an iris liveness detection approach
by utilizing fragmental coefficients of Haar transformed Iris images as signatures to prevent spoofing attacks for
the very first time in the identification of iris liveness. Seven assorted feature creation ways are studied in the
presented solutions, and these created features are explored for the training of eight distinct machine learning
classifiers and ensembles. The predicted iris liveness identification variants are evaluated using recall, F-measure,
precision, accuracy, APCER, BPCER, and ACER. Three standard datasets were used in the investigation. The main
contribution of our study is achieving a good accuracy of 99.18% with a smaller feature vector. The fragmental
coefficients of Haar transformed iris image of size 8 ∗ 8 utilizing random forest algorithm showed superior
iris liveness detection with reduced featured vector size (64 features). Random forest gave 99.18% accuracy.
Additionally, conduct an extensive experiment on cross datasets for detailed analysis. The results of our experiments
show that the iris biometric template is decreased in size to make the proposed framework suitable for algorithmic
verification in real-time environments and settings.

KEYWORDS
Iris images; liveness identification; Haar transform; machine learning; biometric; feature formation; ensemble
model

This work is licensed under a Creative Commons Attribution 4.0 International License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the
original work is properly cited.
324 CMES, 2023, vol.136, no.1

1 Introduction
Automatic human access to a system has become relatively simple in the digital era. Confirmation
of the user’s identification is critical for automated system access. Biometric authentication systems
employ biometric features to confirm the identification of a user [1]. Compared to conventional
password-based traditional verification systems, the biometric system has a distinct advantage. It
minimizes the need to memorize a passcode, pin, or keep a card in hand [2]. Biometric authentication
can be thought of as an additional layer of authentication for security-critical cyber applications
and existing traditional authentication procedures. Today, there are indeed different businesses for
biometric systems. The majority of the sectors seem to be rising quickly. According to www.statista.
com (accessed on 24th June 2022), the business for contactless biometric technologies is forecasted to
increase by roughly 30.15 billion US dollars by 2027, while the overall biometric sciences industry is
forecasted to touch 19.08 billion US dollars in 2021 [3]. Iris is widely employed in the verification and
validation of people. In most applications, it uses, because of its complex textures [4] and distinctive
features, such as the UIDAI project for citizen identification in India, the Amsterdam airport, and the
Canada-US border onon-US [5].
In comparison to fingerprint and face authentication, iris authentication delivers an additional
steadfast contactless user verification. The contactless method aids in the prevention of diseases like
COVID-19 [6]. Despite the iris having a distinct textural structure, the imposter might falsify it [7].
Table 1 shows the iris presentation attacks used that are found in the literature [8].

Table 1: Iris presentation attacks [8]

Iris presentation attacks Details
Print attacks The imposter offers a printed image of validated Iris to the
Biometric sensor [9].
Contact lenses attacks The imposter wears contact lenses on which the pattern of genuine
Iris is printed [10].
Video attacks The imposter plays the video of registered identity in front of a
biometric system [11].
Cadaver attacks Imposter uses the eye of a dead person in front of a Biometric
system [12].
Synthetic attacks Embedding the iris region into the real images makes the
synthesized images more realistic [13].

Individuals frequently assault the biometric system to get admittance to another person’s creden-
tials or to conceal their accurate individuality. The iris identification system can be readily fooled by
means of alternative contact lenses (which can be transparent, textured, colored) [10], replaying the
video, or using a print attack [9]. As a result, understanding the risk and susceptibility is critical for
safeguarding the biometric system [14]. The complicated risk of biometric spoofing is minimized by
assessing the liveness of biometric features prior to authentication [15]. The main objective of this
study is to identify iris liveness detection with reduced feature vectors.
The main objective of this study is to identify iris liveness detection with reduced feature vectors.
Following are the novelty and the main contributions of this paper.
CMES, 2023, vol.136, no.1 325

• Initiatory utilization of ‘fragmental coefficients of Haar transformed iris image data’ as

signatures in iris liveness detection;
• Determining the smallest size of fragmental coefficients that might be used for feature genera-
tion without impairing iris liveness detection performance;
• To identify which classifier is optimal in iris liveness detection, the performance of machine
learning (ML) classifiers and their ensemble combinations are analyzed;
• Testing the feasibility of the developed iris liveness detection method against a variety of existing
benchmark datasets.
The paper’s organization is presented herewith. Section 2 elaborates on an outline of existing
methodologies. Section 3 portrays the proposed approach to iris liveness detection. The experimen-
tation setup is put forth in Section 4. While Section 5 elaborates on the noted outcomes and the
conclusions taken from the findings. Section 6 is where the discussion takes place. Finally, final
thoughts, limitations, and future research suggestions are presented in Section 7.

2 Prevailing Iris Liveness Detection Techniques

Various strategies have been adopted to determine whether the acquired biometric traits are alive
prior to the authentication. Several of the most well-known techniques are addressed in this section.
For liveness identification, Agarwal et al. [16] explored fingerprints with iris. To create a fingerprint
vector function, the essential Haralick statistical characteristics use GLCM and NGTDM. The iris
texture feature is utilized to improve the device’s performance. To evaluate if this model is more
efficient than the current one, Agarwal used a standard dataset. GLCM has an extensive feature
vector size in the current system. Iris spoofing attacks are detected using rotation-invariant features
of Polar harmonic transformations and Zernike moments [5]. Spoofing assaults on numerous sensors
significantly impact the system’s overall competence. The system detects attacks like iris print and
contact lenses.
Thavalengal et al. created a that uses smartphones to take RGB along with NIR images of the
iris and eye [17]. Identification is made using pupil localization techniques and distance measures.
4096-dimensional features are examined for feature vector generation, which is a considerable number.
Although the author claims a high rate of liveness recognition, he does not work with standard
datasets. Authors Fathy et al. have not examined the segmentation or normalization processes
commonly utilized in Iris liveness discovery systems [13]. The original image is broken down into
wavelets using Wavelet Packets (WPs). Although the author claims it is 100 percent accurate, it is not
working with all genres of assaults and only covers a few spoof attacks.
Iris liveness detection will be made utilizing regional features by author Hu et al. [18]. The
interaction of the properties of nearby regions is used to create regional traits. The author utilized
one hundred forty-four relational measurements based on regional attributes during the experiment.
326 CMES, 2023, vol.136, no.1

Using pupil dynamics, the author Czajka [19] created the liveness identification system. The pupil
reaction is tested in this system using rapid changes in the intensity of light. In the case of the eye
reacting to the variations in light intensity, it is alive; otherwise, presentation attack image. In [19],
non-linear and linear SVM is employed to categorize natural reactions and impulsive oscillations. The
system’s shortcoming is that it measures a variety of functions that take time. There are inaccuracies
in the observation because the data utilized in this research does not include any details from elderly
adults.
Author Fang et al. [20] apply many data augmentation methods to generate variability. The
strategy-level and the score-level combination of fusion methods are used for Iris PAD. Bassi et al. [21]
detected PAD using NIR, visible domain, cross-datasets and cross-spectrum datasets. Authors con-
cluded that Cross-PA and cross-datasets are still challenging, as EER values above 20% in most of
cases.
A technique to detect Accurate Ocular Regions was created by Naqvi et al. [22]. This solution
uses deep neural network variants. The system’s evaluation considers publicly available databases.
Kimura et al. [23] developed a CNN-based liveness detection System that improves model accu-
racy by modifying hyperparameters. “Attack Presentation Classification Error Rate [APCER]” and
“Bonafede Presentation Classification Error Rate [BPCER]” metrics are taken to assess the system’s
performance. The hyperparameters are all studied in this work. This method is solely effective against
prints and contact lens attacks. Author [24] studied multiple transfer learning models to detect the iris
liveness and concluded that EfficeintnetB7 gives highest classification accuracy.
Only a few studies were found to be robust against all sorts of spoofing assaults [3,25]. Most of
the studies used a higher size feature vector. Based on these findings, it can be believed that there is a
necessity for a classifier or ensembles for the detection of every sort of spoofing assault.

3 Proposed Iris Liveness Detection Utilizing Fragmental Energy of Haar Transformed Iris Images
The iris recognition system is prone to a range of security threats. Because of these flaws, the
system is less trustworthy for robust authentication applications. The study employs fragmental
energy of Haar modified iris images to attempt iris liveness detection. These fragmental energies were
employed as features to detect whether the iris was real or fake. Because of these characteristics, the
suggested methodology does not require any pre-processing, such as segmentation, normalization, or
localization, which are commonly employed by methods presented in the literature. These fragmental
energies were employed as features to detect whether the iris was real or fake. Because of these
characteristics, the suggested methodology does not require any pre-processing, such as segmentation,
normalization, or localization, which are commonly employed by methods presented in the literature.
These fragmental energies make the suggested technique faster and more accessible [26]. Resizing
the iris image to 256 ∗ 256 is the sole pre-processing performed in the proposed framework. The iris
liveness detection process depicted in Fig. 1 is a block diagram. There are three phases in the proposed
system. Resizing [pre-processing] of iris images, feature formation, and classification with iris liveness
identification.
CMES, 2023, vol.136, no.1 327

Figure 1: Block diagram of the projected iris liveness detection employing fragmental energy of Haar
transformed iris images

3.1 Pre-Processing
The importance of iris pre-processing in iris liveness detection cannot be overstated. Two iris pre-
processing techniques are used in the suggested algorithm. Because images are obtained using three
standard datasets, each dataset stores images of different sizes. We normalized the original 256 ∗ 256
images in pre-processing to ensure they remained intact throughout the experiment. At the same time,
photographing various datasets with various sensors, some (LG, Content, Vista) acquired images in
RGB format, while others [LG, Dalsa] acquired grayscale images. The images were then converted to
grayscale to keep their originality.

3.2 Feature Formation with the Fragmental Energy of Transformed Iris

A scaled iris image is subjected to the Haar transform. The Haar transform allows content
with high energy to congregate in the transform domain’s lower frequency section [27]. The Haar
coefficients are described as below:
In case s = 0, the function of Haar is presented as Eq. (1).
√
ho (t) = 1/ N (1)

In case s > 0, the function of Haar is presented by Eq. (2).

⎧ p ⎫
⎪2 ⎪
√ ⎨ 2p⎬
hs (t) = 1/ N −2 2 (2)
⎪
⎩ ⎪
⎭
0
The nonzero part of the function’s amplitude and width are determined by p, whereas the nonzero
part of the function’s placement is determined by q.
328 CMES, 2023, vol.136, no.1

In the Haar transform iris image, the left highest corner has the higher energy and crucial
information, as shown in Fig. 2. This results in considerable energy compression in a limited count
of high energy coefficients. As a result, these are the preferred feature vector elements. To construct
feature vectors for proposed iris liveness detection, 256 ∗ 256, 128 ∗ 128, 64 ∗ 64, 32 ∗ 32, 16 ∗ 16, 8 ∗ 8, and
4 ∗ 4. Pixels are used to capture the high-energy portion of Haar transformed iris image coefficients.

Figure 2: Proposed fragmental energy-based feature creation approach for liveness identification from
cosine transformed iris images

These feature vectors support the reduction of the size of feature vectors. As a result, iris-liveness
detection is speedier. The compacted high energy improves iris liveness detection accuracy in these
low-frequency coefficients. These high-energy features are then employed for training the ML models
working to detect iris liveness.

3.3 Iris Liveness Detection Using Meachine Learning Classifiers

The suggested method employs a combination of machine learning (ML) classifiers and ensem-
bles. Naive Bayes (NB), Support Vector Machine (SVM), Random Forest (RF), and J48 are the ML
classifiers [15] used here, along with SVM+NB+RF SVM+RT+RF, SVM+MLP+RF ensembles of
classifiers.
Ensemble method—Using multiple models concurrently on a single set for categorization is always
preferable to just one model. Ensemble learning is the name for this technique [23]. Different classifiers
are used to train a model, and the end output is an ensemble of the classifiers. The suggested method
employs majority voting logic for an ensemble of ML classifiers.
These classifiers are trained using a tenfold cross-validation approach. The most effective method
to train ML classifiers is tenfold cross-validation. Tenfold cross-validation allows all data in the dataset
to be considered as either test or training data, giving a more unbiased classifier. The ensembles of ML
classifiers are created using the majority voting mechanism.
Any deep learning architecture to perform well needs a considerable amount of data. Due to
this, the time complexity increase. This disadvantage was overcome with the help of ML classifiers
with handcrafted feature extraction. In the case of ensembles of classifiers, multiple classifiers help
to classify correctly with majority vote logic. So, with the help of combined ML classifiers, the
performance is superior compared to deep learning architectures/CNN.
CMES, 2023, vol.136, no.1 329

4 Experimentation Setup
The investigative results of the proposed method are discussed in this section. The experiments
were performed using an Intel (R) Core (TM) i3-6006U CPU @ 2.0 GHz, 12 GB RAM, and 6a
464-bit operating system with MATLAB R2015a as a programming platform. Clarkson LiveDet2013
(Clarkson 2013), LiveDet2015 (Clarkson 2015), and IITD Combined Spoofing datasets (IIITD CSD)
were used to explore the suggested approach to iris liveness detection.

4.1 Description of Datasets

Three publicly available benchmark datasets are taken in this investigation. The dataset’s detailed
description is as follows:

4.1.1 Clarkson LivDet2013

Around 1356 iris images are included in the Clarkson 2013 dataset [28]. There are two sets of
data in this dataset: testing and training. The Dalsa sensor is utilized to acquire the iris. The images
from training data (as given by the data creator) are used in this study for training purpose, and testing
images (as given by the data creator) are used for testing purposes purpose. The dataset, sensors utilized
in image acquisition, and the Count of images taken for training and testing a model during this
exploration, with some examples, are all listed in Table 2.

Table 2: Sample of images used for exploration from Clarkson 2013 dataset
Sensor Image category Sample images Count of images taken Count of images
for exploration training taken for testing

Off (Bonafide) 270 246

Dalsa

Pattern (Contact) 400 440

4.1.2 Clarkson LivDet2015

LivDet2015—Images used in LivDet2015 dataset are captured using Dalsa and LG sensors [29].
Images are divided into three categories: live, printed, and pattern. In total, 25 subjects are used for live,
images and patterns are printed; 15 subjects are used. The whole dataset is partitioned into training
and testing. The images from training data (as given by the data creator) are used in this study for
training purposes, and testing images (as given by the data creator) are used for testing purposes. The
dataset, sensors utilized in image acquisition, and the count of images taken for training and testing a
model during this exploration, with some examples, are all listed in Table 3.
330 CMES, 2023, vol.136, no.1

Table 3: Sample of images used for an experiment from Clarkson LiveDet2015 dataset
Sensor Image Sample images Count of images Count of images
category taken for training taken for testing

Live 178 197

Dalsa Printed 275 532

Patterns 349 82

Live 166 NA

LG Printed 270 NA

(Continued)
CMES, 2023, vol.136, no.1 331

Table 3 (continued)
Sensor Image Sample images Count of images Count of images
category taken for training taken for testing

Patterns 303 NA

4.1.3 IIITD Combined Spoofing Database [IIITD CSD]

Two iris detectors, a Cogent sensor, and a Vista iris sensor, were utilized to create the images used
in this collection [30,31].
Normal, Print-Capture attack, and Print-Scan attack are the three types of images available in the
dataset [32]. A 60:40 ratio is used for training testing split. The sensors utilized in image acquisition and
the Count of images taken during this exploration with some sample images are all listed in Table 4.

Table 4: Sample of images used for an experiment from IIITD combined spoofing dataset
Sensor Image category Sample Images Count of images Count of images
taken for training taken for testing

Normal 1215 809

Vista Print-scan 718 478

Print-capture 655 437

Normal 1215 809

Content Print-scan 588 392

(Continued)
332 CMES, 2023, vol.136, no.1

Table 4 (continued)
Sensor Image category Sample Images Count of images Count of images
taken for training taken for testing

Print-capture 668 445

4.2 Performance Measures

F-measure, accuracy, recall, precision, ACER (Average Classification Error Rate), APCER
(Attack Presentation Classification Error Rate), and BPCER (Bonafide Presentation Classification
Error Rate) are employed as performance metrics utilized here. Let the true positive, true negative,
false positive, and false negative of the iris liveness detection be TP, TN, FP, and FN, respectively. The
TP designates projected authentic data instances, which are truly what they are. The TN returns data
examples that have been identified as spoofed and are also spoofed examples [2]. FP denotes that the
examples were detected as bonafide but were spoofed. The data examples were detected as presentation
attacks imaged, but bonafide iris examples are shown in FN. The formulas for the performance metric
utilized are given by Eqs. (3) to (9).
(TP + TN)
Accuracy = (3)
TP + TN + FP + FN
TP
Precision = (4)
TP + FP
TP
Recall = (5)
TP + TN
2 ∗ [Precision ∗ Recall]
F − Measures = (6)
[Precision + Recall]
FP
APCER = (7)
TN + FP
FN
BPCER = (8)
TP + FN
(APCER + NPCER)
ACER = (9)
2

5 Experimentation Setup
The benchmark datasets for all feature size variants are taken to test the proposed iris liveness
detection method. Performance measurements such as accuracy, F-measure, precision, and recall are
considered for testing versions of the proposed iris liveness detection technique.
CMES, 2023, vol.136, no.1 333

5.1 Clarkson LivDet2013 Results

Fig. 3 compares the performance of the investigated fragmental coefficients for a particular ML
classifier in the proposed technique of iris liveness detection, which was evaluated on the Clarkson
2013 dataset. In Fig. 3, it can be seen that fragmental coefficients 8 ∗ 8 outperformed other fragmental
coefficient combinations for all classifiers. The highest noted iris liveness detection accuracy comes
around 98.10%, with 8 ∗ 8 fragmental coefficients using a RF classifier.
100

95 256 X 256
Accuarcy In %

128 x 128
90
64 x 64
85
32 x 32
80
16 x 16
75
8x8

4x4

Classifiers

Figure 3: Performance assessment of considered fragmental coefficients for specific ML classifiers in

iris liveness detection for Clarkson 2013 dataset

From Fig. 4, it has been noted that the performance improves as the size of the feature vector
is reduced from 256 ∗ 256 to 8 ∗ 8 and then begins to deteriorate with feature vector size 4 ∗ 4.
This demonstrates that the fragmental coefficients of Haar transformed iris images provide more
outstanding iris liveness recognition capabilities while maintaining a small feature vector size, proving
the importance of the suggested method. The highest average accuracy, 95.94% achieved by 16 ∗ 16
fragmental coefficients.

92.67
4x4
95.68
95.94 8x8
AVG 93.73 16 x 16
90.73 32 x 32
89.74
64 x 64
93.33
128 x 128

86.00 87.00 88.00 89.00 90.00 91.00 92.00 93.00 94.00 95.00 96.00 97.00 256 X 256
Accuracy in %

Figure 4: The performance assessment for the Clarkson 2013 dataset by averaging the specific
fragmental coefficients in iris liveness detection

Table 5 appraises the performance of specific ML classifiers and ensembles of classifiers for
iris liveness detection tested on the Clarkson 2013 dataset. From Table 5, it can be noted that NB
classifiers give the highest average ACER, whereas the lowest average ACER, 2.63% achieved by
SVM+MLP+RF ensembles of classifiers.
334 CMES, 2023, vol.136, no.1

Table 5: ML classifier’s performance evaluation in the proposed iris liveness detection approach for
Clarkson 2013 dataset using an average of % accuracy, % APCER, % BPCER, and % ACER values
Classifiers/EOC AVG APCER BPCER ACER
Random forest 95.83 4.1 3.38 3.74
SVM 95.63 4.3 3.54 3.92
J48 91.00 8.89 8.12 8.505
NB 88.67 11.04 9.92 10.48
SVM+NB+RF 95.07 4.53 3.63 4.08
SVM+RT+RF 96.47 3.43 3.09 3.26
SVM+RF+MLP 96.93 3.05 2.21 2.63
Note: The highest performance is represented in bold.

5.2 IIITD Combined Spoofing Database Results

Fig. 5 reflects the performance comparison of considered Fragmental coefficients for particular
ML classifiers in the proposed iris liveness detection, tested on IIITD CSD. Here, it is noted that, for
all classifiers, 4 ∗ 4 fragmental coefficients outperformed other fragmental coefficient combinations for
IIITD CSD. The highest accuracy, 97.94% achieved by using an RF classifier. Because of its significant
energy compaction, Haar can produce the best results with 4 ∗ 4 fragmental coefficients.
100

98
Accuarcy In %

256 X 256
128 x 128
96
64 x 64
32 x 32
94 16 x 16
8x8
92 4x4

Classifiers

Figure 5: Performance assessment of considered fragmental coefficients for specific ML classifiers in

iris liveness detection for IIITD CSD dataset

From Fig. 6, it can be seen that the performance improves as the feature vector size is compacted
from 256 ∗ 256 to 8 ∗ 8 and then begins to deteriorate with feature vector size 4 ∗ 4. This demonstrates
that the fragmental coefficients of Haar transformed iris images provide more excellent iris liveness
recognition capabilities while maintaining a small feature vector size. The highest average accuracy,
96.80% achieved by 8 ∗ 8 fragmental coefficients.
CMES, 2023, vol.136, no.1 335

96.60
4x4
96.80
8x8
96.63
AV 16 x 16
96.00
G 32 x 32
95.62
64 x 64
94.97
128 x 128
96.40 256 X 256

94.00 94.50 95.00 95.50 96.00 96.50 97.00

Accuracy in %

Figure 6: Performance assessment by averaging the specific fragmental coefficients in iris liveness
detection for IIITD CSD

Table 6 provides the performance assessment of specific ML classifiers and ensembles of classifiers
in the projected iris liveness detection explored on the IIITD CSD dataset. It is noted from the table
that Decision Tree (J48) classifiers give the highest average ACER, whereas the lowest average ACER,
2.29%, is achieved by SVM+NB+RF ensembles of classifiers. The majority voting technique generates
ensembles of classifiers, so they provide the best classification accuracy.

Table 6: Performance evaluation of ML classifiers in the proposed iris liveness detection for IIITD
CSD dataset with an average of % accuracy, % APCER, % BPCER, and % ACER values
Classifiers/EOC AVG APCER BPCER ACER
Random forest 96.75 3.23 2.18 2.71
SVM 95.30 4.69 3.78 4.24
J48 95.49 4.42 3.14 3.78
NB 96.74 3.25 2.19 2.72
SVM+NB+RF 97.37 2.59 1.98 2.29
SVM+RT+RF 96.92 2.89 2.45 2.67
SVM+RF+MLP 96.47 3.51 2.67 3.09
Note: The highest performance is represented in bold.

5.3 Clarkson 2015

Fig. 7 reflects the performance comparison of considered Fragmental coefficients for particular
ML classifiers in proposed iris liveness detection, tested on the Clarkson 2015 dataset. Here, it is noted
that for all classifiers, 16 ∗ 16 fragmental coefficients outperformed other fragmental coefficient combi-
nations for Clarkson 2015 dataset. The highest accuracy, 99.90% achieved by using an SVM+RF+RT
ensemble classifier. Because of its significant energy compaction, Haar can produce the best results
with 8 ∗ 8 fragmental coefficients.
From Fig. 8, it has been noted that the performance improves as the feature vector size is
compacted from 256 ∗ 256 to 8 ∗ 8 and then begins to deteriorate with feature vector size 4 ∗ 4. This
demonstrates that the fragmental coefficients of Haar transformed iris images provide more excellent
336 CMES, 2023, vol.136, no.1

iris liveness recognition capabilities while maintaining a small feature vector size. The highest average
accuracy, 99.58% achieved by 8 ∗ 8 fragmental coefficients.
100

95
Accuarcy In %

256 X 256
90
128 x 128
64 x 64
85
32 x 32

80 16 x 16
8x8
75 4x4

Classifiers

Figure 7: Performance assessment of considered fragmental coefficients for specific ML classifiers in

iris liveness detection for Clarkson 2015 dataset

98.91
99.58 4x4
99.53 8x8
AVG 92.66 16 x 16
89.66 32 x 32
88.67 64 x 64
92.26 128 x 128
256 X 256
82.00 84.00 86.00 88.00 90.00 92.00 94.00 96.00 98.00 100.00 102.00
Accuracy in %

Figure 8: Performance assessment by averaging the specific fragmental coefficients in iris liveness
detection for Clarkson 2015 dataset

Table 7 gives the performance assessment of specific ML classifiers and ensembles of classifiers
in the projected iris liveness detection explored on the Clarkson 2015 dataset. It is noted from the
table that Decision Tree (J48) classifiers give the highest average ACER, whereas the lowest average
ACER, 3.2%, is achieved by SVM+RF+MLP ensembles of classifiers. The majority voting technique
generates ensembles of classifiers, so they provide the best classification accuracy.

Table 7: Performance evaluation of ML classifiers in the proposed iris liveness detection for Clark-
sonclarkson 2015 dataset with an average of % accuracy, % APCER, % BPCER, and % ACER values
Classifiers/EOC AVG APCER BPCER ACER
Random Forest 93.49 6.06 5.98 6.02
(Continued)
CMES, 2023, vol.136, no.1 337

Table 7 (continued)
Classifiers/EOC AVG APCER BPCER ACER
SVM 96.02 3.07 3.54 3.305
J48 93.32 5.78 6.64 6.21
NB 91.34 8.29 8.61 8.45
SVM+NB+RF 95.21 3.97 4.03 4
SVM+RT+RF 95.28 4.15 4.71 4.43
SVM+RF+MLP 96.70 3.31 3.09 3.2

5.4 Cross Datasets Evaluation Results

In this section, the results of cross datasets performances are explained in detail [33]. The first
scenario, where model train on Clarkson 2015 datasets was evaluated on the Clarkson 2013 and IIITD
test datasets. The evaluation results are presented in Table 8, where bold digits indicate the highest
accuracy. From Table 8, we observed that Clarkson’s 2015 datasets give a lower ACER of nearly zero
percent. However, Clarkson’s 2013 datasets do not perform well and give a high ACER of around
74%.

Table 8: Cross datasets evaluation

Train datasets Clarkson 2015
Test datasets Clarkson 2013 IIITD
Metric Accuracy APCER BPCER ACER Accuracy APCER BPCER ACER
Random forest 89.02 10.94 10.09 10.52 74.89 25.07 24.89 24.98
SVM 83.33 16.63 15.67 16.15 48.08 51.88 50.92 51.40
J48 50.00 49.96 47.43 48.70 52.34 47.62 47.43 47.53
NB 98.78 1.18 1.09 1.14 82.97 16.99 16.78 16.89
SVM+NB+RF 91.86 8.10 8.54 8.32 72.55 27.41 27.09 27.25
SVM+RT+RF 86.58 13.38 12.90 13.14 65.95 34.01 34.15 34.08
SVM+RF+MLP 89.43 10.53 10.05 10.29 53.61 46.35 45.9 46.13
AVG 84.14 15.82 15.11 15.47 64.34 35.62 35.30 35.46
Train datasets Clarkson 2013
Test datasets Clarkson 2015 IIITD
Metric Accuracy APCER BPCER ACER Accuracy APCER BPCER ACER
Random Forest 37.19 62.77 62.09 62.43 27.65 72.31 72.45 72.38
SVM 35.92 64.04 63.29 63.67 35.95 64.01 63.29 63.65
J48 41.00 58.96 57.76 58.36 31.27 68.69 67.9 68.30
NB 30.42 69.54 69.27 69.41 66.38 33.58 33.56 33.57
SVM+NB+RF 26.71 73.25 73.06 73.16 41.91 58.05 58.01 58.03
SVM+RT+RF 33.22 66.74 66.47 66.61 28.51 71.45 72.05 71.75
SVM+RF+MLP 25.39 74.57 74.52 74.55 37.44 62.52 62.56 62.54
AVG 25.69 67.12 66.63 66.88 38.44 61.52 61.40 61.46
(Continued)
338 CMES, 2023, vol.136, no.1

Table 8 (continued)
Train datasets IIITD
Test datasets Clarkson 2013 Clarkson 2015
Metric Accuracy APCER BPCER ACER Accuracy APCER BPCER ACER
Random Forest 94.71 5.25 5.12 5.19 100.00 0.00 0.00 0.00
SVM 100.00 0.02 0.01 0.02 100.00 0.00 0.00 0.00
J48 90.65 9.31 9.45 9.38 99.73 0.27 0.27 0.27
NB 94.30 5.66 5.76 5.71 100.00 0.00 0.00 0.00
SVM+NB+RF 99.19 0.77 0.57 0.67 100.00 0.00 0.00 0.00
SVM+RT+RF 96.74 3.22 3.26 3.24 100.00 0.00 0.00 0.00
SVM+RF+MLP 100.00 0.03 0.04 0.04 100.00 0.00 0.00 0.00
AVG 96.51 3.49 3.45 3.47 99.96 0.04 0.04 0.04
Note: The highest performance is represented in bold.

In this section, the results of cross datasets performances are explained. The first scenario, where
model train on Clarkson 2015 datasets was evaluated on the Clarkson 2013 and IIITD test datasets.
The evaluation results are presented in Table 8, where bold digits indicate the highest accuracy. From
Table 8, we observed that Clarkson 2015 gives a lower ACER of nearly zero percentage. However,
Clarkson 2013 datasets do not perform well and give a high ACER of around 74%.
Fig. 9 shows performance evaluation on cross datasets. It can be seen that our model outperforms
in IIITD and Clarkson 2015 datasets, however, shows low performances for Clarkson 2013 dataset.
One possible reason for this is Clarkson 2013 dataset has a smaller number of images compared to the
other two datasets.

Figure 9: Performance assessment on cross datasets evaluation

CMES, 2023, vol.136, no.1 339

Table 9 represents the performance comparison of fragmental coefficients across all datasets used
for implementation with an average of percent accuracy, percent precision, percent recall, and percent
F-ratio values. The highest performance is represented in bold and underlined. From Table 8, it can
be seen that reducing the number of higher energy coefficients from 128 ∗ 128 to 8 ∗ 8 improves
performance since the common part is reduced and discriminative is emphasized more.

Table 9: Performance comparison of fragmental coefficients with an average of percent accuracy,

percent precision, percent recall, and percent F-ratio values
EOC/ Fragmental coefficients
Classifiers 256 ∗ 256 128 ∗ 128 64 ∗ 64 32 ∗ 32 16 ∗ 16 8 ∗ 8 4∗4 AVG
Random forest 97.08 84.34 86.88 93.14 97.667 98.1 96.79 93.43
SVM 97.23 94.02 94.75 95.77 97.667 96.79 92.41 95.52
J48 93 88.09 88.48 89.5 90.52 91.1 93.29 90.57
NB 76.09 88.34 88.92 91.1 92.27 90.81 82.07 87.09
Clarkson 2013 SVM+NB+RF 97.08 89.45 90.23 95.18 97.66 97.23 91.25 94.01
SVM+RT+RF 97.37 88.9 90.52 95.33 97.959 97.81 96.35 94.89
SVM+RF+MLP 95.45 95.06 95.33 96.06 97.81 97.95 96.5 96.31
AVG 93.33 89.74 90.73 93.73 95.94 95.68 92.67 —–
Random forest 97.16 94.2 94.51 95.31 96.57 97.71 97.94 96.20
SVM 96.89 95.67 96.68 96.11 96.45 94.62 93.6 95.72
J48 94.56 93.98 94.4 95.08 95.65 95.08 96.57 95.05
NB 95.89 95.09 95.65 95.31 97.02 98.17 96.91 96.29
IIITD_CSD SVM+NB+RF 96.78 96.01 96.78 97.6 97.71 97.82 96.68 97.05
SVM+RT+RF 96.9 95.48 96.02 96.45 96.68 97.37 97.48 96.63
SVM+RF+MLP 96.65 94.35 95.32 96.11 96.34 96.8 97.02 96.08
AVG 96.40 94.97 95.62 96.00 96.63 96.80 96.60 —–
Random forest 96.01 83.27 85.81 92.07 99.67 99.7 97.57 93.44
SVM 96.16 92.95 93.68 94.7 97.57 98.64 98.44 96.02
J48 91.93 87.02 87.41 88.43 99.50 99.7 98.73 93.32
NB 75.02 87.27 87.85 90.03 99.78 99.7 99.61 91.34
Clarkson 2015 SVM+NB+RF 96.01 88.38 89.16 94.11 99.89 99.8 99.02 95.20
SVM+RT+RF 96.3 87.83 89.45 94.26 99.9 99.8 99.41 95.28
SVM+RF+MLP 94.38 93.99 94.26 94.99 99.78 99.7 99.61 96.67
AVG 92.26 88.67 89.66 92.66 99.53 99.58 98.91 —–
Note: The highest performance is represented in bold.

6 Discussion
The proposed experiment was performed using the Haar transform. The fundamental goal of
experimenting with the transform domain is to learn more about how the image is split into low and
340 CMES, 2023, vol.136, no.1

high-energy parts, reducing the feature vector size and speeding up retrieval. The Haar statistic is used
to transform data [27].
By applying Haar transform on Iris images, high energy coefficients of transformed iris images
sized 256 ∗ 256, 128 ∗ 128, 64 ∗ 64, 32 ∗ 32, 16 ∗ 16, 8 ∗ 8, and 4 ∗ 4 do generate feature vectors for
the projected iris-liveness detection. The procedures outlined in Section 3.2 are used to create the
feature vector. Seven distinct ML and ensembles of classifiers are trained using these features. These
classifiers are trained using the tenfold cross-validation method to detect presentation attacks. Three
benchmark datasets are taken for testing: Clarkson 2013, Clarkson 2015, and the IIITD combined
spoofing database. These three datasets explain in Section 4.1. Accuracy, Precision, Recall, and F-ratio
and ISO standard metrics APCER, BPCER, and ACER are utilized to compare the performance of
all the variants of the suggested approach. Section 4.2 describes several performance measures.
As stated in Sections 5.1 and 5.2, feature extraction using Haar has shown outstanding average
classification accuracy. For the Clarkson 2013 dataset, the highest noted iris liveness detection
accuracy comes around 98.10%, with 8 ∗ 8 fragmental coefficients obtained using the classifier RF.
The highest average accuracy, 95.94% was achieved by 16 ∗ 16 fragmental coefficients, whereas the
maximum average accuracy was 96.31%, and the average ACER was around 2.63%, achieved by
SVM+MLP+RF ensembles of classifiers. For Clarkson 2015, the highest accuracy achieved was
99.90% by using an SVM+RF+RT ensemble classifier. For IIITD CSD, the uppermost accuracy of
97.94% was obtained by using the RF classifier. The uppermost average accuracy was 96.80% achieved
by 8 ∗ 8 fragmental coefficients, whereas the utmost average accuracy was 97.05%, and average ACER
was around 2.29%, achieved by SVM+NB+RF ensembles of classifiers. The findings show that our
suggested approach distinguishes between the bonafide and presentation attack images artifacts using
the Haar transform approach. Table 10 shows a comparison of the suggested strategy to recent studies
in this area.

Table 10: The comparison of the prevailed methods with the proposed approach
Author/year Feature Classifiers Performance Outcome [%] Dataset
generation metrics
Arora et al. CNN VGGNet Accuracy (ACC) Acc = 97.98 IIITD
(2021) [34] FAR
LeNet Acc = 89.38

Khade et al. TSBTC, RF Accuracy, 78.88 IITD

(2021) [8] GLCM NPCER, 95.57 Clarkson 2015
precision, recall,
APCER, ACER,
F-Measure

Comparison Omran et al. IRISNet, [KNN, SVM, Accuracy (ACC), Acc = 96.43 IIITD
with same (2020) [35] CNN NB, DT Precision,
datasets F-Measure, and
Recall
(Continued)
CMES, 2023, vol.136, no.1 341

Table 10 (continued)
Author/year Feature Classifiers Performance Outcome [%] Dataset
generation metrics
Fang et al. ResNet50, NA APCER, BPCER ACER = 10.55, IIITD
(2022) [20] VGG16 and ACER ACER = 18.53
MobileNetv3 ACER = 11.41

Bassi et al. DenseNet, NA APCER, BPCER APCER = 10.7 Clarkson 2013/2015

(2021) [21] PBS and HTER, EER APCER = 76.51 IIITD
A-PBS APCER = 7.38
Das Notre Dame SVM, MLP, APCER, ACER, APCER = 2.61 Clarkson University
et al. PAD MSU RF, and BPCER BPCER = 2.18 [CU], Warsaw
(2021) [36] PAD1 CNN. ACER = 28.96 University of
MSU PAD2 Technology [WUT],
University of Notre
Dame [ND]

Wang et al. CNN-Joint CNN, SDH Accuracy (ACC) Acc = 90.71 PolyU bi-spectra
(2019) [37] Bayesian,
CNN-SDH

Comparison Cheng et al. CNN Hadamard + Accuracy Acc = 97.41 CASIA-Iris-L

with different (2019) [38] CNN
datasets
Chatterjee DWT, ResNet Accuracy Acc = 92.57 ATVS
et al. ResNet
(2019) [39]

Proposed Haar RF Accuracy, Acc = 98.10 Clarkson 2013

Approach Transform APCER, ACER = 2.05
Precision,
BPCER, Recall,
F-Measure, and
ACER

RF Acc = 97.94 IIITD Combined

ACER = 2.29 Spoofing

Compared to similar current techniques based on fragmented energy, the Haar transformation
better discerns between real and artificial artifacts. The results reveal that the proposed method reduces
classification error and gets better accuracy when compared to earlier ways of detecting presentation
attacks using an iris liveness detection. Table 10 summarizes this information. The proposed strategy
outperforms some recent existing studies. As many recent studies used the different train and test
datasets, so we partition Table 10 into two parts, comparing with same datasets and different datasets.
Even though some studies outperform, our approach achieved this performance with reduced feature
vector size (only with 64 features).
While implementing this study, we faced a few challenges-one of them was getting access
to datasets with permissions for use in experimentation. The second challenge we faced was the
explorations of the proposed method with the number of test runs was time-consuming task.
342 CMES, 2023, vol.136, no.1

The limitations of the study are: (i) it was applied only two pre-processing techniques like resizing
and converting an image into grayscale; (ii) this experiment is limited to image size 256 by 256 and
only grayscale images were used during this study.

7 Conclusion
The paper proposed a new method for determining iris liveness. Until now, several approaches
have relied on pre-processing, such as iris segmentation, localization, and normalization; however,
this method of iris liveness detection is computationally intensive. The suggested method employs
Haar transforms on iris images to address this issue, obtaining fragmental coefficients as feature
vectors. The Haar transformed iris image fragmental coefficients are used to train various ML and
ensemble algorithms. Seven criteria are considered to compare the performance of variants of the
suggested approach. Various metrics such as accuracy, precision, recall, f-measure, APCER, BPCER,
and ACER are used to check the performance of the models. Presentation attack images are detected
with 98.10% accuracy in the Clarkson 2013 dataset. The best accuracy for IIITD-CSD was 97.94%.
The experimental results prove the effectiveness of the projected method for detecting iris spoofing
attacks. The study’s main contribution is achieving a good accuracy of 98.10% with lesser feature
vector size by using the fragmental coefficients of the Haar transformed iris image of size 8 ∗ 8
utilizing a RF algorithm with reduced featured vector size. The reduction in considered feature vector
size of iris images with improved accuracy of liveness detection is achieved by exploiting the energy
compaction property of Haar transform in the proposed method. The method is tested on three
available benchmark datasets for validation of results in a generic form. The cross-dataset validations
are performed to prove the worth of the proposed method. The main limitations of this study are as
follows: only two pre-processing techniques were applied, such as resizing and converting an image into
grayscale. Moreover, this experiment is limited to image size 256 by 256 and only grayscale images are
used. In future work, this framework may be extended with the best performance features. Currently,
the presented work is limited to the exploration of explored Haar transform features only. However, the
hybridization of transform using Haar, DCT and Kekare transforms would be an exciting exploration
in the future. Moreover, the proposed framework may be applied for the liveness detection of other
biometric traits, like face, fingerprints, etc. The best performance features a level fusion of fragmental
coefficients of Haar may be added to this framework in future work.

Author Contributions: Data curation: Smita Khade; Writing original draft: Smita Khade; Supervision:
Shilpa Gite, Biswajeet Pradhan; Project administration: Shilpa Gite, Biswajeet Pradhan; Conceptual-
ization: Sudeep Thepade; Methodology: Sudeep Thepade, Shilpa Gite; Validation: Biswajeet Pradhan;
Visualization: Sudeep Thepade, Smita Khade, Shilpa Gite, Biswajeet Pradhan; Resources: Biswajeet
Pradhan, Abdullah Alamri; Review & Editing: Sudeep Thepade, Biswajeet Pradhan; Funding acqui-
sition: Biswajeet Pradhan, Abdullah Alamri.

Funding Statement: The Centre for Advanced Modelling and Geospatial Information Systems
(CAMGIS), Faculty of Engineering and Information Technology, the University of Technology
Sydney, Australia, has funded the research. This research is also partially supported by the Researchers
Supporting Project No. RSP-2021/14, King Saud University, Riyadh, Saudi Arabia.

Conflicts of Interest: The authors declare that they have no conflicts of interest to report regarding the
present study.
CMES, 2023, vol.136, no.1 343

References
1. Khade, S., Thepade, S. D., Ambedkar, A. (2018). Fingerprint liveness detection using directional ridge
frequency with machine learning classifiers. 2018 Fourth International Conference on Computing Commu-
nication Control and Automation (ICCUBEA), pp. 1–5. Pune, India.
2. Khade, S., Thepade, S. D. (2019). Fingerprint liveness detection with machine learning classifiers using
feature level fusion of spatial and transform domain features. 2019 5th International Conference On
Computing, Communication, Control and Automation (ICCUBEA), pp. 1–6. Pune, India.
3. Khade, S., Ahirrao, S., Phansalkar, S., Kotecha, K., Gite, S. et al. (2021). Iris liveness detection for
biometric authentication: A systematic literature review and future directions. Inventions, 6(4), 65. DOI
10.3390/inventions6040065.
4. Su, L., Shimahara, T. (2019). Advanced iris recognition using fusion techniques. NEC Technical Journal,
13(2), 74–77.
5. Kaur, B., Singh, S., Kumar, J. (2019). Cross-sensor iris spoofing detection using orthogonal features.
Computers & Electrical Engineering, 73(1), 279–288. DOI 10.1016/j.compeleceng.2018.12.002.
6. Khade, S., Ahirrao, S., Thepade, S. (2020). Bibliometric survey on biometric iris liveness detection. Library
Philosophy and Practice, 1–29.
7. Nguyen, K., Fookes, C., Jillela, R., Sridharan, S., Ross, A. (2017). Long range iris recognition: A survey.
Pattern Recognition, 72(9), 123–143. DOI 10.1016/j.patcog.2017.05.021.
8. Khade, S., Gite, S., Thepade, S. D., Pradhan, B., Alamri, A. (2021). Detection of iris presentation attacks
using feature fusion of Thepade’s sorted block truncation coding with gray-level co-occurrence matrix
features. Sensors, 21(21), 7408. DOI 10.3390/s21217408.
9. Kaur, J., Jindal, N. (2019). A secure image encryption algorithm based on fractional transforms and
scrambling in combination with multimodal biometric keys. Multimedia Tools and Applications, 78(9),
11585–11606. DOI 10.1007/s11042-018-6701-2.
10. Choudhary, M., Tiwari, V., Venkanna, U. (2019). An approach for iris contact lens detection and classifi-
cation using ensemble of customized DenseNet and SVM. Future Generation Computer Systems, 101(1),
1259–1270. DOI 10.1016/j.future.2019.07.003.
11. Chen, Y., Zhang, W. (2018). Iris liveness detection: A survey. 2018 IEEE Fourth International Conference on
Multimedia Big Data (BigMM), pp. 1–7. China.
12. Trokielewicz, M., Czajka, A., Maciejewicz, P. (2016). Human iris recognition in post-mortem subjects:
Study and database. 2016 IEEE 8th International Conference on Biometrics Theory, Applications and Systems
(BTAS), pp. 1–6. USA.
13. Fathy, W. S. A., Ali, H. S. (2018). Entropy with local binary patterns for efficient iris liveness detection.
Wireless Personal Communications, 102(3), 2331–2344. DOI 10.1007/s11277-017-5089-z.
14. Gupta, R., Sehgal, P. (2016). A survey of attacks on iris biometric systems. International Journal of
Biometrics, 8(2), 145–178. DOI 10.1504/IJBM.2016.077833.
15. Khade, S., Thepade, S. D. (2018). Novel fingerprint liveness detection with fractional energy of cosine
transformed fingerprint images and machine learning classifiers. 2018 IEEE Punecon, pp. 1–7. Pune, India.
16. Agarwal, R., Jalal, A. S., Arya, K. V. (2020). A multimodal liveness detection using statistical tex-
ture features and spatial analysis. Multimedia Tools and Applications, 79(19), 13621–13645. DOI
10.1007/s11042-019-08313-6.
17. Thavalengal, S., Nedelcu, T., Bigioi, P., Corcoran, P. (2016). Iris liveness detection for next generation
smartphones. IEEE Transactions on Consumer Electronics, 62(2), 95–102. DOI 10.1109/TCE.2016.7514667.
18. Hu, Y., Sirlantzis, K., Howells, G. (2016). Iris liveness detection using regional features. Pattern Recognition
Letters, 82, 242–250. DOI 10.1016/j.patrec.2015.10.010.
19. Czajka, A. (2015). Pupil dynamics for iris liveness detection. IEEE Transactions on Information Forensics
and Security, 10(4), 726–735. DOI 10.1109/TIFS.2015.2398815.
344 CMES, 2023, vol.136, no.1

20. Fang, M., Damer, N., Boutros, F., Kirchbuchner, F., Kuijper, A. (2022). The overlapping effect and fusion
protocols of data augmentation techniques in iris PAD. Machine Vision and Applications, 33(1), 1–21. DOI
10.1007/s00138-021-01256-9.
21. Li, Y. H., Aslam, M. S., Harfiya, L. N., Chang, C. C. (2021). Conditional wasserstein generative adversarial
networks for rebalancing iris image datasets. IEICE Transactions on Information and Systems, 104(9),
1450–1458.
22. Naqvi, R. A., Lee, S. W., Loh, W. K. (2020). Ocular-net: Lite-residual encoder decoder network for accurate
ocular regions segmentation in various sensor images. 2020 IEEE International Conference on Big Data and
Smart Computing (BigComp), pp. 121–124. Busan, Korea.
23. Kimura, G. Y., Lucio, D. R., Britto Jr, A. S., Menotti, D. (2020). CNN hyperparameter tuning applied to
iris liveness detection. arXiv preprint arXiv:2003.00833.
24. Khade, S., Gite, S., Pradhan, B. (2022). Iris liveness detection using multiple deep convolution networks.
Big Data and Cognitive Computing, 6(2), 67. DOI 10.3390/bdcc6020067.
25. Khade, S., Gite, S., Thepade, S. D., Pradhan, B., Alamri, A. (2021). Detection of iris presentation attacks
using hybridization of discrete cosine transform and haar transform with machine learning classifiers and
ensembles. IEEE Access, 9, 169231–169249. DOI 10.1109/ACCESS.2021.3138455.
26. Vyas, R., Kanumuri, T., Sheoran, G., Dubey, P. (2019). Recent trends of ROI segmentation in iris biometrics:
A survey. International Journal of Biometrics, 11(3), 274–307. DOI 10.1504/IJBM.2019.100842.
27. Thepade, S. D., Mhaske, V. (2015). New clustering algorithm for vector quantization using hybrid Haar
slant error vector. 2015 International Conference on Computing Communication Control and Automation,
pp. 634–640. USA.
28. Yambay, D., Doyle, J. S., Bowyer, K. W., Czajka, A., Schuckers, S. (2014). LivDet-iris 2013-Iris liveness
detection competition 2013. IEEE International Joint Conference on Biometrics, pp. 1–8. USA.
29. Yambay, D., Becker, B., Kohli, N., Yadav, D., Czajka, A., (2017). LivDet-iris 2015 – Iris liveness detection
competition 2017. 2017 IEEE International Joint Conference on Biometrics (IJCB), pp. 733–741.
30. Kohli, N., Yadav, D., Vatsa, M., Singh, R., Noore, A. (2016). Detecting medley of iris spoofing attacks using
DESIST. 2016 IEEE 8th International Conference on Biometrics Theory, Applications and Systems (BTAS),
pp. 1–6. USA.
31. Gupta, P., Behera, S., Vatsa, M., Singh, R. (2014). On iris spoofing using print attack. 2014 22nd
International Conference on Pattern Recognition, pp. 1681–1686. USA.
32. Yadav, D., Kohli, N., Doyle, J. S., Singh, R., Vatsa, M. et al. (2014). Unraveling the effect of textured contact
lenses on iris recognition. IEEE Transactions on Information Forensics and Security, 9(5), 851–862. DOI
10.1109/TIFS.2014.2313025.
33. Fang, M., Damer, N., Boutros, F., Kirchbuchner, F., Kuijper, A. (2021). Cross-database and cross-attack iris
presentation attack detection using micro stripes analyses. Image and Vision Computing, 105(12), 104057.
DOI 10.1016/j.imavis.2020.104057.
34. Arora, S., Bhatia, M. P. S., Kukreja, H. (2020). A multimodal biometric system for secure user identification
based on deep learning. International Congress on Information and Communication Technology, pp. 95–103.
Singapore.
35. Omran, M., AlShemmary, E. N. (2020). An iris recognition system using deep convolutional neural network.
Journal of Physics: Conference Series, 1530(1), 012159.
36. Das, P., McFiratht, J., Fang, Z., Boyd, A., Jang, G. et al. (2020). Iris liveness detection competition (livdet-
iris)-the 2020 edition. 2020 IEEE International Joint Conference on Biometrics (IJCB), pp. 1–9. Houston,
TX.
CMES, 2023, vol.136, no.1 345

37. Wang, K., Kumar, A. (2019). Cross-spectral iris recognition using CNN and supervised discrete hashing.
Pattern Recognition, 86(2), 85–98. DOI 10.1016/j.patcog.2018.08.010.
38. Cheng, Y., Liu, Y., Zhu, X., Li, S. (2019). A multiclassification method for iris data based on the
hadamard error correction output code and a convolutional network. IEEE Access, 7, 145235–145245. DOI
10.1109/ACCESS.2019.2946198.
39. Chatterjee, P., Yalchin, A., Shelton, J., Roy, K., Yuan, X. et al. (2019). Presentation attack detection using
wavelet transform and deep residual neural net. International Conference on Security, Privacy and Anonymity
in Computation, Communication and Storage, pp. 86–94. Cham.