Intelligent Systems with Applications 16 (2022) 200107

Contents lists available at ScienceDirect

Intelligent Systems with Applications

journal homepage: www.journals.elsevier.com/intelligent-systems-with-applications

Face presentation attack identification optimization with adjusting

convolution blocks in VGG networks
Sudeep D. Thepade *, Mayuresh Dindorkar , Piyush Chaudhari , Shalakha Bang
Computer Engineering Department, Pimpri Chinchwad College of Engineering, SPPU, Pune, India

A R T I C L E I N F O A B S T R A C T

Keywords: Advancement in deep learning is mapping to every field of life and applying it to almost all research problems.
Face liveness detection Numerous Deep Convolutional Neural Network (DCNN) architectures are being proposed, giving different results
Transfer learning based on the depth and value of hyperparameters. The entire development of such DCNN architectures from
VGG16
scratch needs a lot of effort, and such architectures may not be used for other applications than the one they are
VGG19
DensNet121
structured for. Transfer learning is a way to modify these pre-trained networks to make them suitable for newer
Xception diverse applications. This paper attempts to empirically assess the performance and suitability of existing pre-
MobileNet trained DCNN architectures for human face liveness detection. Due to the advent of ambient computing for
InceptionV3 contactless identification of humans using their biometric traits, human face liveness detection proves to be an
important research area. Six pre-trained DCNN models, alias VGG16, VGG19, DensNet121, Xception, MobileNet,
and InceptionV3, are considered for empirical assessment in human face liveness detection. The method is
explored using two face liveness detection datasets - NUAA and Replay-Attack. Face Liveness Accuracy, and Half
Total Error Rate (HTER) are considered prime performance evaluation metrics. At a learning rate of 10− 4, the
VGG19 network with scenario “Original VGG” gives the highest face liveness detection accuracy, which is the
outcome of current research.

1. Introduction Large annotated datasets are the prime requirement of a highly ac­
curate model. To get such large datasets for every domain is not easy. To
In the modern era of pervasive computing, an individual’s security is minimize this requirement, an approach called transfer learning came
reliant mainly on their biometric traits such as the face (Kekre et al., into existence. It is a method that reuses an existing pre-trained model to
2010a; Kekre et al., 2010b), iris (Kekre et al., 2011a; Thepade & Bidwai, solve a different but related problem. Generally, the method which
2013; Thepade & Mandal, 2014), fingerprints (Khade et al., 2018; modifies the previously trained model developed for a particular task for
Thepade & Khade, 2018), etc. Face recognition has become the most performing some other task is called Transfer Learning. The transfer
vital physiological biometric modality used in data security owing to learning approach increases the speed and improves the performance of
rich features exhibited by the face and ease of handling the facial new models. Different transfer learning techniques are used based on the
recognition systems (Kekre et al. 2010b,Kekre et al. 2011b). Attackers research area and available data.
mislead these systems by spoofing attacks like photos, videos, cut-photo, The key contributions of the work presented in the paper are as
mask attacks, etc. To mitigate these attacks existing face recognition followse.
systems should be improvised. Existing anti-spoofing methods make use
of motion-based, shape-based, depth-based, color-based, texture-based, • The paper has attempted the empirical performance assessment of
and deep learning-based features for spoof identification. A subset of six pre-trained deep CNN architectures (VGG16, VGG19, Incep­
deep neural networks (NN) called Convolutional Neural Network (CNN) tionV3, DenseNet121, MobileNet, and Xception) for the human face
is widely explored in various image processing areas. Several pre-trained liveness detection application.
deep CNN models like VGG16, VGG19, InceptionV3, etc., are available
that are trained on large datasets for image-classification jobs.

* Corresponding author.
E-mail address: [email protected] (S.D. Thepade).

https://doi.org/10.1016/j.iswa.2022.200107
Received 11 October 2021; Received in revised form 20 July 2022; Accepted 6 August 2022
Available online 7 August 2022
2667-3053/© 2022 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
S.D. Thepade et al. Intelligent Systems with Applications 16 (2022) 200107

Fig. 1. Flowchart illustrating modifications done in the existing pre-trained deep CNN models for the task of face liveness detection.

probabilities. A class-probability vector is generated by combining these

Table 1
feature class probabilities. Further, this class-probability vector is given
The number of training and testing face images taken from both datasets for the
to SVM for performing face anti-spoofing tasks.
proposed empirical performance assessment of DCNN models.
In Tu & Fang, (2017), pertinent hidden features from the input face
Face Class Replay-Attack NUAA
image are located by ResNet50. To learn temporal features, LSTM is
Train Test Train Test
applied on the top of ResNet50.
Live Face 10,800 13,656 1743 3362 These features are utilized to determine whether the face is spoofed
Spoofed Face 32,280 42,809 1748 5761
or live.
Spoofed with four types Spoofed using only one type
A simple face anti-spoofing technique is proposed in Das et al.,
(2019). Authors have extracted global deep features and handcrafted
features from VGG16 and LBP descriptors, respectively. The proposed
Table 2
method has experimented on SSIJRI, Replay-Attack, Replay-Mobile and,
The optimizer parameters set during the performance
3DMAD datasets.
assessment of existing DCNN models used in transfer
learning for human face liveness detection. Two face anti-spoofing techniques are explored (Elloumi et al.,
2020). The first technique states LBP histogram calculation & VGG16
Parameter Value
finetuning. The second technique utilizes Image Quality Measures
Optimizer Adam (IQM).
Learning rate 10− 3, 10− 4, 10− 5

Batch-size 102
Beta-1 0.9 3. Proposed method
Beta-2 0.999
Epsilon 1 * 10− 8 The pre-trained model is the saved previously trained network on a
massive dataset for broad-scale image classification. In the transfer
learning process, the knowledge gained by the machine learning (ML)
• Through the finetuning of the last two convolution blocks of VGG16
model from one dataset is used on another. Two approaches of the
and VGG19 at the learning rate of 10–5, the improvised test accuracy
transfer-learning are: (a) Using the pre-trained model as the ready-made
is attempted.
feature extractor for a specific image-classification job. (b) Finetuning
• The complexity of VGG networks is reduced by decreasing the
the pre-trained model, a whole or a portion of the model is retrained on
number of convolution blocks in explored scenarios (approaches)
new data. Transfer learning is usually employed to evade overfitting
• Experimental explorations are done on two standard datasets: the
when the amount of data is less. Saving time and computational re­
NUAA dataset and the Replay-Attack dataset.
sources required during the training phase is a major advantage of using
transfer learning.
2. Literature survey
In the proposed work here (Fig. 1), for detection of a human face, the
input face image is first converted to Grayscale and then passed through
Some of the recent attempts implementing pre-trained CNN models
the Haarcascade classifier to obtain Region-Of-Interest (ROI). ROI of
for face liveness detection have achieved prominent results. A brief re­
input RGB image is resized to the window of dimension 96 * 96 pixels.
view of these attempts is summarized here.
After this preprocessing, the proposed methodology uses finetuning
“FASNet," a CNN architecture that is an altered version of VGG16, is
approach on different pre-trained deep learning models for face liveness
explored (Lucena et al., 2017). Using back-propagation, authors have
detection. For every pre-trained model, the last fully connected layers
finetuned the weights of the FASNet – from the fourth convolution block
(FC) are substituted by a classification head composed of two FC layers
to the topmost layers. 3DMAD & Replay-Attack datasets are used to
of size 256 and 1, respectively. After the first FC layer, a dropout layer is
evaluate the performance of the proposed method.
added to shun overfitting. Most models overfit for the learning rate ’1 *
A feature fusion approach that combines ResNet18 and Rotation
10− 3′ ; hence, results for the learning rate of 1 * 10− 4 and 1 * 10− 5 are
Invariant Local Binary Pattern (RILBP) features is explored by Chen
depicted in this paper. Adam optimizer is accustomed to learning rates
et al. (2019). Further, Support Vector Machine (SVM) is trained on these
and weight decay 1 * 10− 6 (refer to Table 1). Further, sigmoid is used as
features to perform binary classification. Improved performance is
a decision function suited for binary classification instead of softmax.
observed in both intra and cross-database testing.
Finetuning the model with an arbitrarily initialized classification
ResNet-18 is fed with three different types of features, namely color,
head may cause the pre-trained base model to forget its learning due to
temporal, and patch-based (Tang et al., 2018), to obtain class
large gradient updates. Thus, we set the pre-trained model as non-

2
S.D. Thepade et al. Intelligent Systems with Applications 16 (2022) 200107

Fig. 2. Modified VGG16 architecture showing the finetuned convolution blocks by red region.

Fig. 3. Modified VGG19 architecture showing the finetuned convolution blocks by red region.

Fig. 4. Modified DenseNet121 architecture showing the finetuned convolution blocks by red region.

trainable while enabling the classification head to be trainable during made to reduce the complexity of these VGG networks, which can be
training for the first five epochs. Later on, the last two convolution inferred from Fig. 8. Following are the reasons to select only VGG net­
blocks of the pre-trained base model, which are immediately before the works for further explorations:
classification head, are set trainable (unfrozen). Then both the base
model and classification head are trained jointly for the next 30 epochs. • Compact architecture
For VGG16, VGG19 & DenseNet121 last two convolution blocks • Potential to give higher face liveness detection accuracy.
enclosed in the red boundary are finetuned as shown in Figs. 2, 3, and 4,
respectively. To reduce the complexity of the VGG networks, we have considered
Incase of InceptionV3, the last two inception blocks are finetuned, as five different scenarios, alias Original-VGG, Scenario-D, Scenario-C,
shown in Fig. 5. The last two separable convolution blocks are finetuned Scenario-B, and Scenario-A (refer to Fig. 8). The complexity is reduced
in the case of Xception (refer to Fig. 6). Each separable convolution in each scenario by removing the last convolution block from the VGG
block includes depthwise and pointwise convolutions. For MobileNet, network; further, the remnant network is trained for 30 epochs, and
the last two depthwise separable convolution blocks are finetuned, as obtained results are justified in the “Results and Discussion” section.
shown in Fig. 7.
From Tables 3 and 4, it is observed that VGG networks (VGG16 and 3.1. Existing popular DCNN models
VGG19) are performing better compared to other considered finetuned
DCNNs hence these two networks are used for further explorations at the Here the empirical performance-appraise of six popular DCNN
learning rate of 10− 4 and 10− 5. In further exploration, an attempt is models, alias VGG16, VGG19, InceptionV3, DenseNet121, Xception, and

3
S.D. Thepade et al. Intelligent Systems with Applications 16 (2022) 200107

Fig. 5. Modified InceptionV3 architecture showing the finetuned convolution blocks by red region.

Fig. 6. Modified Xception architecture showing the finetuned convolution blocks by red region.

Fig. 7. Modified MobileNet architecture showing the finetuned convolution blocks by red region.

MobileNet, is done. These admired pre-trained DCNN models are fine­ increasing the depth of layers from 16 to 19.
tuned using a transfer learning approach for face liveness detection. With 42 layers of deep architecture (Szegedy et al., 2016), a low error
A CNN model referred to as VGG16 is proposed by Simonyan & rate is achieved, and hence it was awarded 1st runner-up ILSVRC
Zisserman (2015). It is trained on the ImageNet dataset composed of (ImageNet Large Scale Visual Recognition Competition) 2015. The
nearly 14 million images of 1000 classes using NVIDIA’s Titan Black computation cost of InceptionV3 is just 2.5 times higher than GoogleNet
GPU. Its architecture consists of 16 layers: 13 convolution layers and (InceptionV1), but it is still efficient compared to VGGNet.
three fully connected layers. DenseNet has 4 versions namely DenseNet121, DenseNet161, Den­
VGG19 (Simonyan & Zisserman, 2015) comprises 19 layers, seNet169 and DenseNet201. In DenseNet, each layer is connected in a
including 16 convolution layers and three fully connected layers. A feed-forward fashion (Huang et al., 2017). Reducing vanishing gradient
remarkable improvement in the existing VGG network has been made by problems, encouraging feature reuse, etc., are the prime advantages of

4
S.D. Thepade et al. Intelligent Systems with Applications 16 (2022) 200107

Table 3 these networks. DenseNet121 is 121 layers of deep architecture.

Testing accuracies obtained for considered finetuned DCNN models over 30 Xception (Chollet, 2017) has 71 layers of deep architecture. Xcep­
epochs. tion’s architecture is motivated by Inception, where inception modules
Finetuned DCNN Learning Rate got substituted with depth-wise detachable convolutions. Xception has
(10− 4) (10− 5) the same number of parameters as InceptionV3.
Test accuracy (%) MobileNet has an organization of detachable convolution modules,
Replay-Attack NUAA Replay-Attack NUAA
formed by depthwise and pointwise convolution being done in the form
VGG16 99.99 67.14 100 73.89 of piles in the MobileNet network (Howard et al., 2017). It has 28
VGG19 99.95 68.43 100 68.30
layer-deep architecture considering depthwise and pointwise convolu­
Mobilenet 93.18 68.87 94.70 56.80
DenseNet121 91.79 65.37 98.29 62.00 tions as an individual layer.
Xception 99.86 69.71 98.96 65.79
InceptionV3 96.82 63.35 94.10 62.03
3.2. Experimentation environment

Table 4 A platform provided by Kaggle is used for experimental assessment

HTER obtained for considered finetuned DCNN models over 30 epochs. here. It is an online data science community for data scientists and
machine-learning practitioners. The Python code scribbled in the Kaggle
Finetuned DCNN Learning Rate
(10− 4) (10− 5)
notebook was executed using GPU as an accelerator. Pretrained DCNN
HTER (%) models provided by the Keras library are used for experimentation.
Replay-Attack NUAA Replay-Attack NUAA Two standard and publically available datasets used in face liveness
VGG16 0.015 37.54 0 33.14 detection, alias ’Replay-Attack dataset’ and ’NUAA dataset,’ are
VGG19 0.1 28.66 0 41.345 explored here to validate the performance appraise of existing DCNN
Mobilenet 11.08 39.895 7.345 49.61 models after transfer learning. The testing accuracy and HTER values for
DenseNet121 16.33 41.545 2.025 43.65
each modified DCNN model are used as performance measures.
Xception 0.12 26.605 1.51 35.335
InceptionV3 5.57 45.855 8.72 41
IDIAP Research Institute has constituted the Replay-Attack dataset
(Chingovska et al., 2012), having 1300 videos of 50 persons with two
diverse luminance conditions set up as adverse controlled during the
video acquisition refer to Fig. 2. All videos in the dataset are filmed at a
rate of 25 Hertz. We have extracted half of the frames per second (FPS)

Fig. 8. Reducing the complexity of VGG16 and VGG19.

5
S.D. Thepade et al. Intelligent Systems with Applications 16 (2022) 200107

Fig. 9. Average percentage test accuracy of VGG16 and VGG19 across both NUAA & Replay-Attack datasets for learning rate of 10− 4.

Table 5
Performance metrics observed for VGG16 over 30 epochs on NUAA dataset for learning rate of 10− 4.
Scenario Test Accuracy (%) Precision (%) Recall (%) F1-Score (%) FAR (%) FRR (%) HTER (%)

A 60.05 47.26 72.64 57.26 47.3 27.36 37.33

B 76.13 62.26 89.44 73.42 31.64 10.56 21.1
C 75.12 71.03 54.85 61.9 13.05 45.15 29.1
D 75.39 80.21 44.11 56.92 6.35 55.89 31.12
Original VGG16 77.65 63.85 90.72 74.95 29.98 9.28 19.63

Table 6
Performance metrics observed for VGG16 over 30 epochs on Replay-Attack dataset for learning rate of 10− 4.
Scenario Test Accuracy (%) Precision (%) Recall (%) F1-Score (%) FAR (%) FRR (%) HTER (%)

A 99.34 98.87 98.4 98.63 0.36 1.6 0.98

B 100 100 100 100 0 0 0
C 100 100 100 100 0 0 0
D 100 100 100 100 0 0 0
Original VGG16 100 100 100 100 0 0 0

Table 7
Performance metrics observed for VGG19 over 30 epochs on NUAA dataset for learning rate of 10− 4.
Scenario Test Accuracy (%) Precision (%) Recall (%) F1-Score (%) FAR (%) FRR (%) HTER (%)

A 72.308 59.33 79.42 67.92 31.77 20.58 26.175

B 76.626 62.08 94.38 74.9 33.64 5.62 19.63
C 75.681 60.49 98.51 74.95 37.55 1.49 19.52
D 73.154 63.99 62.37 63.17 20.48 37.63 29.055
Original VGG19 78.56 64.85 91.58 75.93 28.97 8.42 18.695

Table 8
Performance metrics observed for VGG19 over 30 epochs on Replay-Attack dataset for learning rate of 10− 4.
Scenario Test Accuracy (%) Precision (%) Recall (%) F1-Score (%) FAR (%) FRR (%) HTER (%)

A 99.56 99.5 98.72 99.11 0.17 1.28 0.725

B 100 100 100 100 0 0 0
C 100 100 100 100 0 0 0
D 100 100 100 100 0 0 0
Original VGG19 100 100 100 100 0 0 0

(i.e., 25/2 = 12) from each video. possesses appearance variations like gender, with-or-without glasses,
Nanjing University of Aeronautics and Astronautics has invented the and light. Few sample faces from the NUAA dataset are given in Fig. 3.
NUAA dataset (Tan et al., 2010). Real and Spoofed face images of 15 The number of training and testing samples taken across live and
persons are acquired by a web camera. The dataset consists of 7509 fake spoofed face images for both ‘Replay-Attack’ and ‘NUAA’ datasets are
and 5105 real face images, segregated into training and testing sets. shown in Table 1.
Every face image has dimensions of 640 * 480 pixels. The database in­ The performance metrics used in the paper are as follows:
cludes only one type of spoofing attack, a ’photo-attack.’ The database

6
S.D. Thepade et al. Intelligent Systems with Applications 16 (2022) 200107

Table 9 FP
FAR (False Acceptance Rate) = (1)
Average percentage test accuracy of VGG networks across both the datasets for TN + FP
each scenario considering learning rate of 10− 4.
FN
Scenario VGG16 VGG19 FRR (False Rejection Rate) = (2)
TP + FN
A 79.7 85.94
B 88.07 88.32 ( )
FAR + FRR
C 87.56 87.84 HTER (Half Total Error Rate) = ∗ 100 (3)
D 87.7 86.58
2
Original VGG 88.83 89.28
Average over Scenarios 86.37 87.59 TP
Precision or Positive Predictive Value (PPV) = (4)
TP + FP

TP
Recall or Sensitivity = (5)
TP + FN

Table 14
Average test percentage accuracy of VGG networks across both the datasets for
each scenario considering learning rate of 10− 5.
Scenario VGG16 VGG19

A 82.57 81.07
B 81.94 83.95
C 87.09 87.12
D 86.26 87.59
Original VGG 87.21 84.67
Average over Scenarios 85.01 84.88

Table 10
Performance metrics observed for VGG16 over 30 epochs on NUAA dataset for learning rate of 10− 5.
Scenario Test Accuracy (%) Precision (%) Recall (%) F1-Score (%) FAR (%) FRR (%) HTER (%)

A 65.32 52.52 61.30 56.57 32.34 38.69 35.52

B 63.87 51.00 50.27 50.63 28.19 49.73 38.96
C 74.18 60.41 86.85 71.26 33.22 13.15 23.185
D 72.51 63.47 59.85 61.61 20.10 40.15 30.125
Original VGG16 74.41 76.70 43.87 55.82 7.78 56.13 31.955

Table 11
Performance metrics observed for VGG16 over 30 epochs on Replay-Attack dataset for learning rate of 10− 5.
Scenario Test Accuracy (%) Precision (%) Recall (%) F1-Score (%) FAR (%) FRR (%) HTER (%)

A 99.81 99.79 99.40 99.59 0.065 0.6 0.3325

B 100 100 100 100 0 0 0
C 100 100 100 100 0 0 0
D 100 100 100 100 0 0 0
Original VGG16 100 100 100 100 0 0 0

Table 12
Performance metrics observed for VGG19 over 30 epochs on NUAA dataset for learning rate of 10− 5.
Scenario Test Accuracy (%) Precision (%) Recall (%) F1-Score (%) FAR (%) FRR (%) HTER (%)

A 62.14 49.11 75.16 59.40 45.46 24.84 35.15

B 67.89 54.00 87.00 66.64 43.26 13.00 28.13
C 74.23 60.89 84.06 70.62 31.50 15.94 23.72
D 75.18 61.81 84.98 71.57 30.64 15.02 22.83
Original VGG19 69.34 59.3 53.54 56.27 21.44 46.46 33.95

Table 13
Performance metrics observed for VGG19 over 30 epochs on Replay-Attack dataset for learning rate of 10− 5.
Scenario Test Accuracy (%) Precision (%) Recall (%) F1-Score (%) FAR (%) FRR (%) HTER (%)

A 100 100 100 100 0 0 0

B 100 100 100 100 0 0 0
C 100 100 100 100 0 0 0
D 100 100 100 100 0 0 0
Original VGG19 100 100 100 100 0 0 0

7
S.D. Thepade et al. Intelligent Systems with Applications 16 (2022) 200107

Fig. 10. Average percentage test accuracy of VGG16 and VGG19 across both NUAA & Replay-Attack datasets for learning rate of 10− 5.

Table 15
Comparison of the existing DCNN models of literature survey with the best performing finetuned DCNN model observed in the current empirical assessment.
Face Anti-spoofing Method Pre-trained DCNN Performance Dataset Used Test Accuracy (%) HTER/ACER EER (%)
model Measure (%)

VGG16 (Lucena et al., 2017) VGG16 HTER, Test Accuracy Replay-Attack 99.04 1.20 –
3DMAD 100 0.00 –
ResNet50 + RI-LBP (Chen et al., 2019) ResNet-50 HTER, EER Replay-Attack – 2.6 2.3
NUAA – – 0.5
CASIA-FASD – – 4.4
MSU-MFSD. – – 3.1
Multiple Deep Features (Tang et al., 2018) ResNet-18 EER, ACER CASIA-FASD – – 2.2
Replay- – 0.0 –
Mobile
OULU-NPU – 2.4 –
ResNet50 + LSTM (Tu & Fang, 2017) ResNet-50 HTER, EER CASIA-FASD – 1.22 1.0
Replay-Attack – 1.18 1.03
VGG16 + LBP (Das et al., 2019) VGG16 Test Accuracy SSIJRI 92.05 – –
Replay-Attack 75.25 – –
Replay- 90.52 – –
Mobile
3DMAD 96.97 – –
VGG16 (Elloumi et al., 2020) VGG16 HTER Replay-Attack – 2.5 –
CASIA-FASD – 0.0 –
Replay- – 0.0 –
Mobile
Proposed method VGG19 Test Accuracy, HTER Replay-Attack 100 0.00 –
VGG19 NUAA 78.56 18.7 –
(Learning rate = 10− 4, Scenario = “Original VGG”)

( )
2 ∗ Precision ∗ Recall The graph in Fig. 9 depicts that for the learning rate of 10− 4, scenario
F1 − Score = (6)
Precision + Recall “Original VGG” outputs the highest accuracy among the considered
scenarios. Thus, from Table 9 and Fig. 9, it can be jointly concluded that
( )
TN + TP a combination of VGG19 and scenario “Original VGG” results in the best
Accuracy = ∗ 100 (7)
TN + TP + FN + FP performance for face liveness detection, considering a learning rate of
10− 4.
Where, Similarly, for a learning rate of 10− 5, from Table 14 and Fig. 10, it
TP (True Positive) = Number of Real Faces predicted as Real. can be concluded that a combination of VGG16 and scenario “C” outputs
TN (True Negative) = Number of Fake Faces predicted as Fake. second-best results compared to a combination of VGG19 and scenario
FP (False Positive) = Number of Fake Faces predicted as Real. “original VGG” with a learning rate of 10− 4. By comparing graphs drawn
FN (False Negative) = Number of Real faces predicted as Fake in Figs. 9 and 10, it can be concluded that for all scenarios, higher test
accuracy is obtained at a learning rate of 10− 4 compared to the learning
4. Results and discussion rate of 10− 5.
Table 15 shows the comparison of a few of the existing DCNN models
For each scenario, we evaluated the model over 30 epochs and used in face spoofing detection methods proposed in the literature sur­
selected the model of a particular epoch having maximum test accuracy vey with the best-observed performance in the empirical assessment
for comparing with models of other corresponding scenarios. Learning done in this paper with the help of two datasets. The performance
rates of 10− 4 and 10− 5 are employed for all scenarios. Performance comparison is not easy as the experimentation environment used in each
metrics used to evaluate the scenarios are Test accuracy, Precision, existing method is different as well as the performance measures and
Recall, F1-score, FAR, FRR, and HTER. datasets also differ. But if we compare the explorations done using

8
S.D. Thepade et al. Intelligent Systems with Applications 16 (2022) 200107

Replay-Attack datasets (as shown in Table 15), the testing accuracy and Declaration of Competing Interest
HTER observed in the proposed work presented here with VGG19 (with
the scenario "Original VGG") is superior to the considered existing No conflict of interest exists.
similar attempts from the literature (Chen et al., 2019; Das et al., 2019; We wish to confirm that there are no known conflicts of interest
Elloumi et al., 2020; Lucena et al., 2017; Tang et al., 2018; Tu & Fang, associated with this publication and there has been no significant
2017). financial support for this work that could have influenced its outcome.

5. Conclusion References

Human face liveness detection is important in today’s ambient Chen, F., Wen, C., Xie, K., Wen, F., Sheng, G., & Tang, X. (2019). Face liveness detection:
Fusing colour texture feature and deep feature. IET Biometrics, 8(6), 369–377.
computing scenario, where user identity is verified through the con­ https://doi.org/10.1049/iet-bmt.2018.5235
tactless biometric traits acquired. The paper has attempted the empirical Chingovska, I., Anjos, A., & Marcel, S. (2012). On the effectiveness of local binary
performance assessment of six pre-trained deep CNN architectures patterns in face anti-spoofing. In A. Bromme, & C. Busch (Eds.), Proceedings of the
BIOSIG - international conference of biometrics special interest group (pp. 1–7). IEEE.
(VGG16, VGG19, InceptionV3, DenseNet121, MobileNet, and Xception) Chollet, F. (2017). Xception: deep learning with depthwise separable convolutions. In
for the human face liveness detection application. All pre-trained DCNN Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR)
models are modified to make them suitable for human face liveness (pp. 1800–1807). https://doi.org/10.1109/CVPR.2017.195, 2017.
Das, P. K., Hu, B., Liu, C., Cui, K., Ranjan, P., & Xiong, G. (2019). A new approach for face
detection. The testing accuracy & HTER observed over a diverse number anti-spoofing using handcrafted and deep network features. In Proceedings of the IEEE
of epochs are used as the performance parameters for assessing the pre- international conference on service operations and logistics, and informatics (SOLI) (pp.
trained DCNN models. Firstly, by finetuning the last two convolution 33–38). IEEE Press. https://doi.org/10.1109/SOLI48380.2019.8955089.
Elloumi, W., Chetouani, A., Charrada, T. B., & Fourati, E. (2020). Anti-Spoofing in face
blocks of VGG16 and VGG19 at the learning rate of 10− 5, we obtained
recognition: Deep learning and image quality assessment-based approaches (pp. 51–69).
test accuracy of 100%, which is the highest for the Replay-Attack dataset Cham: Deep Biometrics. Unsupervised and semi-supervised learning. Springer.
compared to methods mentioned in Table 15. https://doi.org/10.1007/978-3-030-32583-1_4
Secondly, the aim to reduce the complexity of VGG networks was Howard, A. G., Menglong, Z., Chen, B., Dmitry, K., Wang, W., Weyand, T., et al. (2017).
MobileNets: Efficient convolutional neural networks for mobile vision applications. ArXiv
fulfilled by decreasing the number of convolution blocks in each sce­ abs/1704.04861.
nario, as shown in Fig. 8. In this approach, from results, it is evident that Huang, G., Liu, Z., Maaten, L. V. D., & Weinberger, K. Q. (2017). Densely connected
the VGG19 network with the scenario "Original VGG" at a learning rate convolutional networks. In Proceedings of the IEEE conference on computer vision and
pattern recognition (CVPR) (pp. 2261–2269). https://doi.org/10.1109/
of 10− 4 gives remarkable overall results in terms of test accuracy of CVPR.2017.243, 2017.
78.56% on the NUAA dataset and 100% for the Replay-Attack dataset. Kekre, H. B., Thepade, S. D., & Chopra, T. (2010a). Face and gender recognition using
Closely following this result, the VGG16 network with scenario “C” at a principal component analysis. International Journal on Computer Science and
Engineering (IJCSE), 2(4), 959–964.
learning rate of 10− 5 outputs second-best results. Lastly, for all consid­ Kekre, H. B., Thepade, S. D., Jain, J., & Agrawal, N. (2011a). Iris recognition using
ered scenarios, higher test accuracies are obtained for the learning rate texture features extracted from walshlet pyramid. In B. K. Mishra (Ed.), Proceedings
of 10− 4. In the present research work, the proposed method has of the ACM international conference & workshop on emerging trends in technology
(ICWET ’11) (pp. 76–81). Association for Computing Machinery. https://doi.org/
experimented on the datasets that contain only the photo attack and the 10.1145/1980022.1980038.
video attack. To enrich the conclusion regarding the robustness level of Kekre, H. B., Thepade, S. D., & Maloo, A. (2010b). Eigenvectors of covariance matrix
the current finding, it can experiment on datasets containing mask at­ using row mean and column mean sequences for face recognition. International
Journal of Biometrics and Bioinformatics (IJBB), 4(2), 42–50.
tacks in the future.
Kekre, H. B., Thepade, S. D., & Maloo, A. (2011b). Face recognition using texture features
extracted from Walshlet pyramid. ACEEE International Journal on Recent Trends in
Author declaration Engineering and Technology (IJRTET), 5(1).
Khade, S., Thepade, S. D., & Ambedkar, A. (2018). Fingerprint liveness detection using
directional ridge frequency with machine learning classifiers. In Proceedings of the
[Instructions: Please check all applicable boxes and provide additional fourth international conference on computing communication control and automation
information as requested.] (ICCUBEA) (pp. 1–5). IEEE. https://doi.org/10.1109/ICCUBEA.2018.8697895.
Lucena, O., Junior, A., Moia, V., Souza, R., Valle, E.&, & Lotufo, R. (2017). Transfer
learning using convolutional neural networks for face anti-spoofing, 10317 pp. 27–34).
Funding Springer Cham. https://doi.org/10.1007/978-3-319-59876-5_4
Simonyan, K., & Zisserman, A. (2015). Very deep convolutional networks for large-scale
No funding was received for this work. image recognition. arXiv 1409.1556.
Szegedy, C., Vanhoucke, V., Loffe, S., Shlens, J., & Wojna, Z. (2016). Rethinking the
inception architecture for computer vision. In Proceedings of the IEEE conference on
a Dr. Sudeep D. Thepade computer vision and pattern recognition (CVPR) (pp. 2818–2826). IEEE. https://doi.
https://orcid.org/0000–0001–7809–4148 org/10.1109/CVPR.2016.308, 2016.
Tan, X., Li, Y., Liu, J., & Jiang, L. (2010). Face Liveness Detection from a single image
with sparse low rank bilinear discriminative model. In Computer vision – eccv 2010,
b Piyush Chaudhari 6316 pp. 504–517). Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-
https://orcid.org/0000–0001–9023–0157 642-15567-3_37
Tang, Y., Wang, X., Jia, X., & Shen, L. (2018). Fusing multiple deep features for face anti-
spoofing. In Biometric recognition. CCBR 2018, 10996 pp. 321–330). https://doi.org/
c Mayuresh Dindorkar 10.1007/978-3-319-97909-0_35
https://orcid.org/0000–0001–8089–3301 Thepade, S. D., & Bidwai, P. (2013). Iris recognition using fractional coefficients of
transforms, wavelet transforms and hybrid wavelet transforms. In Proceedings of the
IEEE international conference on control, computing, communication and materials
d Shalakha Bang (ICCCCM) (pp. 1–5). https://doi.org/10.1109/ICCCCM.2013.6648921
https://orcid.org/0000–0003–3472–9636 Thepade, S. D., & Khade, S. (2018). Novel fingerprint liveness detection with fractional
energy of cosine transformed fingerprint images and machine learning classifiers. In
Proceedings of the IEEE Punecon. IEEE,. https://doi.org/10.1109/
CRediT authorship contribution statement PUNECON.2018.8745407.
Thepade, S. D., & Mandal, P. R. (2014). Novel iris recognition technique using fractional
Sudeep D. Thepade: Conceptualization, Methodology, Writing – energies of transformed iris images using haar and kekre transforms. International
Journal of Scientific & Engineering Research, 5(4), 305–308.
original draft, Supervision. Mayuresh Dindorkar: Data curation, Tu, X., & Fang, Y. (2017). Ultra-deep neural network for face anti-spoofing. In , 10635.
Investigation, Visualization, Software, Validation. Piyush Chaudhari: Proceedings of the neural information processing. ICONIP (pp. 686–695). Springer,
Data curation, Investigation, Writing – original draft. Shalakha Bang: Cham. https://doi.org/10.1007/978-3-319-70096-0_70.
Data curation, Investigation, Visualization, Software, Validation.