I.J.

Image, Graphics and Signal Processing, 2019, 2, 15-20

Published Online February 2019 in MECS (http://www.mecs-press.org/)
DOI: 10.5815/ijigsp.2019.02.02

Score-Level-based Face Anti-Spoofing System

Using Handcrafted and Deep Learned
Characteristics
Omid. Sharifi
Department of Computer and Software Engineering, Toros University, Mersin, Turkey
Email: omid.sharifi@toros.edu.tr

Received: 04 January 2019; Accepted: 20 January 2019; Published: 08 February 2019

Abstract—Recognition performance of biometric systems attacks is on biometric sensors in order to submit a fake
is affected through spoofing attacks made by fake template with the aim of impersonating a real user.
identities. The focus of this paper is on presenting a new Generally, to apply direct attacks, attackers don’t need to
scheme based on score level and decision level fusion to have any knowledge about different parts of biometric
monitor individuals in term of real and fake. The system such as feature extractors and matching
proposed fake detection scheme involve consideration of techniques. Liveness, texture and motion detection
both handcrafted and deep learned techniques on face techniques are considered as primary methods to counter
images to differentiate real and fake individuals. In this this kind of attack. On the other hand, to present indirect
approach, convolutional neural network (CNN) and attacks, attacker needs to be aware of specific information
overlapped histograms of local binary patterns (OVLBP) about the system such as template format and
methods is used to extract facial features of images. The communication protocol. Furthermore, access to internal
produced matching scores provided by CNN and OVLBP parts of the system physically or logically is needed for
then combined to form a fused score vector. Finally, the attackers. Countermeasures in this kind of attack include
last decision on real and attack images is done by physical or logical security aspects.
combining decisions of hybrid scheme using majority The main interest of this study is on direct or spoofing
vote of CNN, OVLBP and their fused vector. attacks specifically print and video attack for face
Experimental results on public spoof databases such as biometric. Print attack aims to spoof biometric systems
Print-Attack and Replay-Attack face databases by printing modality images of individuals, while video
demonstrate the strength of the proposed anti-spoofing attack concentrates on spoofing by submitting video
method for fake detection. sequences of live individuals on a screen to biometric
systems in term of fixed or hand-held to avoid liveness
Index Terms—Spoof detection, handcrafted texture detection. Spoof detection in biometrics is quite
extraction, convolutional neural network, decision level challenging and therefore has encouraged the biometric
fusion, score level fusion. society to investigate the effect of this kind of counterfeit
events on different modalities such as face, iris,
fingerprint, multimodal biometric systems, etc. [11-22].
I. INTRODUCTION As abovementioned, texture, motion and liveness
analyses can be considered in biometric systems to
Currently, the use of biometric recognition systems in
counter spoofing attacks [14, 23-25]. In general, the focus
term of identification and/or verification of individuals
of texture analysis is on detecting texture patterns such as
according to their physical or behavioral characteristics is
print failures and overall image blur to detect attacks. On
extensively studied in situations with high security
the other hand, motion detection concentrates on motion
demands [1-10]. In fact, the ability of biometrics to
features of patterns such as optical flow to overcome the
improve recognition performance and security of
problem of certain texture patterns dependency. Spoof
applications considered as increasing interest of
detection using liveness analysis attempts to solve the
researchers compared to conventional techniques such as
problem by focusing on vitality signs of biometric
token-based and knowledge-based methods. On the other
characteristics and analyzing spontaneous movements
hand, technology development causes vulnerability of
such as eye blinking and lip movements in 2D
biometric systems to fake samples specially through
recognition systems. However, taking into account a
image acquisition step [1,11-13]. Previously, the aim of
global solution to all kinds of attacks is not possible
biometric systems was only to recognize the individuals
because of nature of attacks, biometric characteristics and
without considering the spoof attacks. In general, the
spoof detection approaches.
attacks on biometric systems are categorized into direct
This paper proposes a novel solution based on score
and indirect attacks [14]. The concentration of direct
level and decision level fusion against print and video

Copyright © 2019 MECS I.J. Image, Graphics and Signal Processing, 2019, 2, 15-20
16 Score-Level-based Face Anti-Spoofing System Using Handcrafted and Deep Learned Characteristics

attacks. The facial features are extracted using of liveness detection for evaluation of trajectory of
convolutional neural network (CNN) [26, 27] and different parts of face has been presented in [24] using
overlapped histograms of local binary patterns (OVLBP) the optical flow of lines successfully. In fact, they used a
[28] methods. The strength of handcrafted and deep learnt model-based local Gabor decomposition and Support
features is then combined through score level fusion. Vector Machine (SVM) experts to detect different parts
Weighted sum rule (WS) fusion strategy is used in this of face in their liveness proposed detection method. In
paper to fuse the scores. Finally, the last decision is made [25], a holistic liveness detection paradigm has been
by combining the decisions of each classifier (OVLBP, proposed for face biometric to detect spoof attacks. The
CNN and score level fusion). Majority voting is authors suggested fusion of anti-spoofing methods in
employed in this study to fuse the results of classifiers by interactive situations for obtaining reliable liveness
outputting the label with majority of the votes. Therefore, detection strategy. In [31], image quality assessment
the contribution of the proposed scheme can be reviewed strategy is used to propose a novel software-based spoof
as: proposing a robust face anti-spoofing method with detection framework. The authors of this study employed
print and video attacks, the use of score and subsequently 25 different image quality features to distinguish fake
decision level fusion strategies improves detection samples. A face anti-spoofing framework for video attack
performance of proposed methods with consideration of has been proposed in [32] using motion magnification.
lower computational burden compared to schemes The proposed method developed performance of LBP
involving feature level fusion. The experiments feature extractor to detect attacks. The authors of this
performed on publicly available Idiap Print-Attack [29] study used a motion estimated based method using
and Replay- Attack [30] face spoofing databases Histogram of Oriented Optical Flow (HOOF) descriptor
demonstrate the superiority of proposed anti-spoofing for Print-Attack and Replay-Attack databases effectively.
method in term of detection performance, computational The researchers of [33] employed multi-scale LBP
complexity and reduction in detection alteration. method to transform the micro-textures into an improved
On the other hand, the key issue and core technologies feature histogram for face fake detection. The
of this study can be summarized as: applying OVLBP classification of fake and real samples has been done in
feature extractor as an strong and popular handcrafted their study using SVM. In [34], the structures and
method, and therefore considering more local primitive dynamics of facial micro-textures have been applied to
textures for better video and print spoofing attack propose an anti-spoofing technique on Replay-Attack and
detection, considering deep learning feature extraction CASIA Face Anti-Spoofing databases. The concentration
method as a powerful anti-spoofing extractor, combining of authors of [35] for spoof detection is on image
the methods using two level of fusions in order to distortion analysis (IDA) method to exploit specular
improve the detection rate, and then comparing the reflection, blurriness, chromic moment, and color
proposed method with state-of-the-art handcrafted and diversity feature of 2D face images. On the other hand,
deep-learnt methods in field of print and video attacks. the authors of [12] applied two different deep learning
The rest of paper is ordered as follows. Section 2 methods to detect attacks in several biometric recognition
involves previous studies of spoofing attacks and systems such as iris, face, and fingerprint. Although, their
protection techniques in field of biometrics. The experimental results demonstrated the high detection
concentration of sections 3 and 4 is on handcrafted and performance of deep learning technique, but it was not
deep learnt techniques applied in this study for spoof able to improve always the detection rate specifically for
detection. In section 5, the overall architecture of face biometric. In [36], a new method based on feature
proposed scheme is described. The demonstration of level fusion strategy is used to combine handcrafted and
experimental results and databases is presented in section deep learnt facial features in order to improve the spoof
6. Finally, Section 7 provides conclusion of this study. detection. They applied SVM classifier to differentiate
real and fake identities.
The concentration of current study therefore is to
II. RELATED WORKS combine both handcrafted and deep learnt methods with
proposing a novel scheme. The proposed scheme
The problem of spoof attack detection for face
considers a robust face anti-spoofing method for both
biometric has been studied recently using different
print and video attacks. Employing combination strategy
handcrafted and deep learnt techniques [12, 16, 23-25,31-
using score and decision level fusion leads to detection
38]. In [16], a novel double anti-spoofing pipeline has
performance improvement.
been proposed for face biometrics based on selection of
optimized textures and image quality assessment
techniques for print and video attacks. The paper applied
different texture and image quality algorithms to compare III. HANDCRAFTED FACIAL TEXTURE EXTRACTION USING
the ability of their proposed framework. Effectiveness of OVLBP
applying multiple techniques to detect print attack for The feature extraction step of this study considers an
face biometric has been studied in [23]. The authors of extension of LBP texture extractor called OVLBP in
this study compared several methods based on motion order to describe the local spatial structure of face images.
analysis, texture analysis and liveness detection for In general, LBP introduced by Ojala et al. as a gray-scale
detection of 2D facial print-based spoof attacks. The use and powerful subpattern-based texture operator for

Copyright © 2019 MECS I.J. Image, Graphics and Signal Processing, 2019, 2, 15-20
 Score-Level-based Face Anti-Spoofing System Using Handcrafted and Deep Learned Characteristics 17

analyzing face biometric [37]. By using LBP texture images are divided into 36 sub-regions of 10 × 10
extractor images are divided into several blocks and window size without and with 50% overlapping of sub-
subsequently local textures of blocks are extracted. In regions in this study.
fact, histogram extraction of features associated to the
patterns on a set of pixels in each block is performed and
then concatenating the local extracted features is done to IV. DEEP LEARNING EXTRACTION USING CNN
present a single global feature vector. Therefore, binary
patterns are calculated by comparing its value with those In order to include more representative feature set of
information in the proposed method, the paper attempts to
of its neighbor according to a central pixel in an image.
The concentration of first version of LBP for feature apply CNN learning-based strategy to improve the
extraction is done on a 3 × 3 window size which leads to detection rate of print and video attacks. In general, any
CNN structure contains two significant layers called
preventing large scale structures capturing. However,
Uniform Local Binary Pattern (ULBP) [28] as extension convolutional layers and fully connected layers. The
of original version focuses on implementation of the convolutional layer aims to extract image features and
manipulate them using the convolution operation. A
operator to circular neighborhoods with a different radius
size to solve the problem of basic LBP. The main goal of training process is needed based on the characteristics of
uniform LBP is to provide an independent output label images to achieve filter coefficients. In order to make the
final CNN structure, consideration of a cross-channel
for each uniform pattern of mapping while for non-
uniform patterns a single output label is considered. normalization layer, a rectified linear unit (ReLU) and a
In general, the use of histogram-based methods for face pooling layer is needed for each convolutional layer.
Finally, the constructed feature map is sent to fully-
spoof scenarios to detect attacks specifically print and
video attacks has been considered as sufficient factor in connected layers for further classification.
several studies [16, 18, 32-34]. This study applies idea of The construction of CNN method for the anti-spoofing
overlapped LBP histograms to obtain more significant proposed framework to extract the deep facial features is
histograms as handcrafted facial texture extractor for based on VGG-16 architecture introduced by the Oxford
Visual Geometry Groups’ model in ImageNet Large-
print and video attacks. In OVLBP method [28], images
are divided into overlapped blocks with the aim of Scale Visual Recognition Challenge (ILSVRC) [39].
achieving more sub-windows over the image. Therefore, Compared to earlier version of CNN structure VGG-16 is
more extensive and richer with including five batches of
the extracted sub-windows contain micro-textons of more
local primitive textures related to spots, flat areas, edges, convolution operations. Fig.1 illustrates the architecture
edge ends and curves [28]. The extracted textures from of VGG-16. In general, each batch includes 2–3 adjacent
convolution layers connected via max-pooling layers. All
small windows usually include more specific and precise
texture information because of involving more convolution layers consist of kernel sizes of 3 ×3 with
informative histograms of overlapped regions and same number of kernels inside each batch starting 64 in
the first group to 512 in the last one.
consequently capable to better managing of print and
video attacks in face spoofed framework. All the face

Fig.1. The architecture of VGG-16.

Since in spoof detection method only two type of method for detecting print and video attacks. This study
classes as real and fake is used, the number of output considers score level fusion due to ease in accessing,
neurons in the last layer of model is changed to two. On simplicity, low computational complexity and usually
the other hand, in order to reduce the overfitting affection similar and/or equivalent performance compared to
of training step of CNN, this study employs different feature level fusion. Therefore, the proposed method first
learning rate policy for different layers. Furthermore, data produces the scores of handcrafted and deep learnt facial
augmentation technique by cropping different regions of extractors using Manhattan Distance technique and then
input and their fillips is used in training part of this work. the produced scores are combined using weighted sum
rule (WS) strategy.
In general, weighted sum rule technique combines the
V. PROPOSED ANTI-SPOOFING SCHEME matching scores of individual matchers. In this study,
computation of weights is performed according to quality
The proposed anti-spoofing framework detects print of individual classifiers in term of performance
and video attacks in face databases based on score level
improvement. Weighted sum rule of different matching
and decision level fusion strategies as depicted in Fig.2. scores is shown as equation 1.
The facial features are extracted using CNN and OVLBP
methods in order to improve the ability of the proposed
𝑤𝑠 = 𝑤1 × 𝑠1 + 𝑤2 × 𝑠2 + ⋯ + 𝑤𝑛 × 𝑠𝑛 (1)

Copyright © 2019 MECS I.J. Image, Graphics and Signal Processing, 2019, 2, 15-20
18 Score-Level-based Face Anti-Spoofing System Using Handcrafted and Deep Learned Characteristics

Where 𝑤1 , 𝑤2 , … , 𝑤𝑛 are the computed weights for contains 450 real, 450 print attack and 600 video attack
different classification methods and 𝑠1 , 𝑠2 , … , 𝑠𝑛 are the images while the number of images for real, print attack
set of matching scores. and video attack in test dataset is 300, 300 and 400
Nearest neighbor classifier (NNC) is then applied on respectively. The division of datasets into train and test is
set of calculated scores to provide singular classification done three times and the averaged result of these three
of each set of score. Finally, the last decision as real or sets is reported in this study.
attack is made by combining the decisions of each In order to perform the experiments in CNN part of the
classifier (OVLBP, CNN and score level fusion) using proposed method, the images are resized to 256 ×256 size.
majority votes of the three classifiers in this study. The data augmentation part of training contains ten
Majority voting combines the results of classifiers by different cropped size of 227 ×227 and their flip.
outputting the label with majority of the votes. Therefore, train augmented database takes into
consideration of 9000 real, 9000 print attack and 12000
video attack images. In order to avoid overfitting of
training data this study also considered the regularization
as 0.1, momentum parameters as 0.9 and learning rate as
0.001 with batch size of 32. The process of training is
done for 50 epochs. Half Total Error Rate (HTER) that is
half of sum of False Genuine Rate (FGR) and False Fake
Rate (FFR) of spoof detection errors is considered as
evaluation protocol of proposed anti-spoofing method in
this study.
The first set experiments done in this study
concentrates on applying handcrafted texture extraction
method for print and video attacks and comparison with
some other handcrafted methods in field of fake detection.
In order to classify the images nearest-neighbor classifier
(NN) is applied for all the implemented methods. In
general, NN is a method of data classification to
approximate how likely a data point is to be a member of
one group. In fact, the algorithm classifies a sample
according to the category of its nearest neighbor.

Table 1. Results of applying handcrafted texture extraction methods for

print and video attacks in HTER (%)
Method Print-Attack Video-Attack
HTER (%) HTER (%)
Gabor [33] 37.45 37.30
LBP [38] 18.26 22.18
HOG [40] 25.40 27.47
Fig.2. The proposed anti-spoofing framework.
IQA [31] 34.72 33.35
OVLBP 14.35 15.75

Analyzing the results of Table 1 shows the superiority

VI. EXPERIMENTAL RESULTS AND DATABASES of OVLBP for both print and video attacks as 14.35% and
In order to evaluate the performance of the anti- 15.75%. As shown in the table, all LBP based methods
spoofing method, Idiap Print-Attack and Repla-Attack outperform other implemented handcrafted methods for
databases are employed in this study. Print-Attack video and print attacks. It should be stated that, in order
database consists of 200 short videos of printed to apply image quality assessment method, mean-squared
photograph and real-access of 50 individuals while error (MSE) measurement is used in this study. On the
Repla-Attack databases contains 1300 video clips of other hand, in order to investigate the effect of CNN on
photograph and video attack access of 50 individuals print and video attacks another set of experiment was
under different lighting conditions. In order to perform performed in Table 2, using only CNN and CNN +
the experiments, this study first detected the face images OVLBP with Score Level Fusion. Selection of OVLBP
from video sequences and then the detected face images here is due to providing better detection performance in
were aligned according to center position of left and right term of HTER for spoofing in this study.
irises. This study randomly, assigned 30 subjects for As depicted in Table 2, employing CNN in both
training and 20 persons for testing to extract the real and scenarios with and without the handcrafted feature
attack images from videos. In total, in order to perform extractor leads to detection improvement for both print
the experiments, 750 real access images, 750 print attack and video attacks. However, in case of applying only
images and 1000 video attack images were extracted CNN (without OVLBP combination) video attack
from the videos in this work. Therefore, training dataset achieved better detection in term of HTER in this study.

Copyright © 2019 MECS I.J. Image, Graphics and Signal Processing, 2019, 2, 15-20
 Score-Level-based Face Anti-Spoofing System Using Handcrafted and Deep Learned Characteristics 19

Generally, combining CNN and OVLBP using score using handcrafted texture analysis for print attack while
level fusion achieved 4.05% and 3.20% spoof detection video attack over handcrafted technique obtained 7.1%
rate enhancement over applying only CNN against print improvement in this study.
and video attacks respectively in this work. Moreover, in
order to examine the effect of proposed scheme against REFERENCES
video and print attack the paper compare the result of
[1] A.K Jain,.; A.Ross,; Prabhakar, S. An introduction to
applying only handcrafted, CNN and CNN + OVLBP + biometric recognition. IEEE Trans. Circuits Syst. Video
Score Level Fusion with the proposed scheme as the last Technol. 2004, 14, 4–20.
set of experiment in Table 3. [2] A. K., Jain, S. Z Li,. Handbook of face recognition. New
York: springer, 2011.
Table 2. Results of applying CNN and combination of CNN and [3] M. Ç. Yildiz, O.Sharifi, and M. Eskandari, Log-Gabor
handcrafted texture extraction methods using score level fusion for print Transforms and Score Fusion to Overcome Variations in
and video attacks in HTER (%) Appearance for Face Recognition, International
Method Print-Attack Video-Attack Conference on Computer Vision and Graphics, –
HTER (%) HTER (%) Proceedings of International Conference on Computer
CNN 14.45 14.20 Vision and Graphics, ICCVG 2016, Warsaw, Poland,
CNN + OVLBP 10.40 11.00 September 19-21, 2016. Springer 2016 Lecture Notes in
using Score Level Fusion Computer Science.
[4] H. S. Bhatt., S. Bharadwaj, R. Singh,& M, Vatsa.
Table 3. Results of applying proposed method for print and video Recognizing surgically altered face images using
attacks in HTER (%) multiobjective evolutionary algorithm. IEEE Transactions
Method Print-Attack Video-Attack on Information Forensics and Security, 2013, 8(1), 89-100.
HTER (%) HTER (%) [5] O. Sharifi, M. Eskandari, and M. Ç Yildiz., Scheming an
OVLBP 14.35 15.75 Efficient Facial Recognition System using Global and
CNN 14.45 14.20 Random Local Feature Extraction Methods, 2nd
CNN + OVLBP + Score Level 10.40 11.00 International Conference on Computer Science and
Fusion Engineering UBMK’17, October 5-8, 2017, Antalya,
Proposed Method 8.50 8.65 Turkey, DOI: 10.1109/UBMK.2017.8093508.
[6] D. Zhang, Z. Guo, G. Lu,.; L. Zhang,; Y. Liu,; W. Zuo,
The experiments performed in Table 3 demonstrate the Online joint palmprint and palmvein verification. Expert
effectiveness of proposed anti spoofing method based on Syst. Appl. 2011, 38, 2621–2631.
score and decision level fusion strategy for both print and [7] M. Eskandari, O. Sharifi. Optimum scheme selection for
face–iris biometric. IET Biometrics, 2016, 6(5), 334-341.
video attacks. The general analysis of table however
[8] K. Nguyen, C. Fookes,; R. Jillela,; S. Sridharan,; A. Ross,.
shows detection of print attack is more successful using Long range iris recognition: A survey. Pattern Recognit.
combination of CNN and handcrafted methods with score 2017, 72, 123–143.
level fusion and decision level fusion according to the [9] O. Sharifi, M. Eskandari, Optimal Face-Iris Multimodal
experiments done in this study. The proposed scheme Fusion Scheme. Symmetry, 2016, 8(6), 48.
obtained 5.85% and 7.10% improvement over using only [10] D.T Pham, Y.H. Park, D.T Nguyen, S.Y. Kwon,; K.R
OVLBP for print and video attacks. In addition, Park,. Nonintrusive finger-vein recognition system using
improvement of 5.95% and 5.55% detection rate in term NIR image sensor and accuracy analyses according to
of print and video attack detection over employing just various factors. Sensors 2015, 15, 16866–16894.
[11] D.T. Nguyen, H.S. Yoon, D.T. Pham,; K.R Park,. Spoof
CNN extractor.
detection for finger-vein recognition system using NIR
camera. Sensors 2017, 17, 2261.
[12] D. Menotti, G. Chiachia,; A. Pinto, W.R Schwartz, H.
VII. CONCLUSION Pedrini,; A.X. Falcao,; A. Rocha, Deep representation for
iris, face and fingerprint spoofing detection. IEEE Trans.
This paper presented a new anti-spoofing method
Inf. Forensic Secur. 2015, 10, 864–879.
based on combination of CNN and handcrafted [13] K.R. Nalini, H.C. Jonathan, M.B. Ruud,: An analysis of
techniques in two level of fusion against print and video minutiae matching strength. In: Audio- and Video-Based
attacks. The proposed method first applied handcrafted Biometric Person Authentication, Proceedings of 3rd
and CNN methods to extract the facial features and then AVBPA ed.,2001, vol. 2091, pp. 223–228.
the computed scores of each method along with [14] M. JMarcos, F. Julian, , et al.: An evaluation of indirect
combination of them using weighted sum rule strategy attacks and countermeasures in fingerprint verification
are sent to a decision level fusion for further spoof systems. Pattern Recognit. Lett. 2011, 32(12), 1643–1651.
detection. The paper improved the detection rate of attack [15] T. Santosh, P. Norman, et al.: Detection of face spoofing
using visual dynamics. IEEE Trans. Inf. Forensics Secur.
by combining the decisions obtained from CNN, OVLBP
2015, 10(4), 762–777.
and fused scores of these two methods. Performing [16] M., Eskandari, & O.Sharifi, Designing Efficient Spoof
different set of experiments in this study showed the Detection Scheme for Face Biometric. In International
effectiveness of print attack over video attacks detection Conference on Image and Signal Processing, 2018, July;
through proposed scheme. The concentration of this pp. 427-434. Springer, Cham.
paper for experiments was on Idiap Print-Attack and [17] A. Anjos, M.M. Chakka, S. Marcel,: Countermeasures to
Repla-Attack face spoofing databases with 50 individuals. photo attacks in face recognition. Biom. IET, 2014, 3(3),
The proposed method has 5.85% improvement over only 147–158.

Copyright © 2019 MECS I.J. Image, Graphics and Signal Processing, 2019, 2, 15-20
20 Score-Level-based Face Anti-Spoofing System Using Handcrafted and Deep Learned Characteristics

[18] P. Gupta, et al. "On iris spoofing using print attack." [32] Samarth, et al. "Computationally efficient face spoofing
Pattern Recognition (ICPR), 2014 22nd International detection with motion magnification." Proceedings of the
Conference on. IEEE, 2014. IEEE Conference on Computer Vision and Pattern
[19] H. Abdenour, G. Mohammad, et al.: Can gait biometrics Recognition Workshops. 2013.
be spoofed. In: 2012 21st International Conference on [33] J. Määttä, H. Abdenour, and P. Matti. "Face spoofing
Pattern Recognition. detection from single images using micro-texture
[20] B. Biggio, Z. Akhtar, G. Fumera, , G. L Marcialis., & F. analysis." Biometrics (IJCB), 2011 international joint
Roli,. Security evaluation of biometric authentication conference on. IEEE, 2011.
systems under real spoofing attacks. IET biometrics, 2012, [34] P. de Freitas, Tiago, et al. "Face liveness detection using
1(1), 11-24. dynamic texture." EURASIP Journal on Image and Video
[21] M. Gomez-Barrero, G. Javier, and F. Julian. "Efficient Processing 2014.1 (2014): 1-15.
software attack to multimodal biometric systems and its [35] D. Wen, H. Hu, and A. K. Jain. "Face spoof detection
application to face and iris fusion." Pattern Recognition with image distortion analysis." IEEE Transactions on
Letters 36, 2014: 243-253. Information Forensics and Security 10.4 (2015): 746-761.
[22] A., Zahid, S. Kale, and N. Alfarid. "Spoof attacks on [36] D. Nguyen, Tien, et al. "Combining Deep and
multimodal biometric systems." Proc. International Handcrafted Image Features for Presentation Attack
Conference on Information and Network Technology Detection in Face Recognition Systems Using Visible-
(IPCSIT). Vol. 4. 2011. Light Camera Sensors." Sensors 18.3 (2018): 699.
[23] M.M. Chakka, A. Anjos, et al.: Competition on counter [37] T. Ojala, M. Pietikäinen, D. Harwood: A comparative
measures to 2D facial spoofing attacks. In: 2011 study of texture measure with classification based on
International Joint Conference on Biometrics. feature distributions. Pattern Recognit. 29, 51–59.
[24] K. Kollreider, H. Fronthaler, J. Bigun, Evaluating liveness [38] A. Benlamoudi; D. Samai.; A. Ouafi.; S.E Bekhouche, A.
by face images and the structure tensor. In: Automatic Taleb-Ahmed, Hadid, Face spoofing detection using local
Identification Advanced Technologies, 2005. binary patterns and Fisher score. In Proceedings of the 3rd
[25] K. Kollreider, H. Fronthaler, J. Bigun, Verifying Liveness International Conference on Control
By Multiple Experts In Face Biometrics. In: IEEE [39] K. Simonyan; A. Zisserman, Very deep convolutional
Computer Society Conference on Computer Vision and neural networks for large-scale image recognition. In
Pattern Recognition Workshops, 2008. Proceedings of the International Conference on Learning
[26] A. Krizhevsky, I. Sutskever, G.E Hinton, ImageNet Representations, Kunming, China, 25–27 September 2013.
classification with deep convolutional neural networks. In [40] A. Pinto, et al. "Using visual rhythms for detecting video-
Proceedings of the Advances in Neural Information based facial spoof attacks." IEEE Transactions on
Processing Systems, Lake Tahoe, NV, USA, 3–8 Information Forensics and Security 10.5 (2015): 1025-
December 2012. 1038G. Eason, B. Noble, and I. N. Sneddon, “On certain
[27] P.N. Druzhkov , V.D. Kustikova , A survey of deep integrals of Lipschitz-Hankel type involving products of
learning methods and soft-ware tools for image Bessel functions,” Phil. Trans. Roy. Soc. London, vol.
classification and object detection, Pattern Recognit. Im- A247, pp. 529–551, April 1955.
age Anal. 26 (1) (2016) 9–15.
[28] Z. Guo, D. Zhang, D. Zhang, A completed modeling of
local binary pattern operator for texture classification.
IEEE Transactions on Image Processing, 19(6) (June 2010)
Author’s Profile
1657-1663.
[29] Print Attack face database,
Omid. Sharifi received his Ph.D. degree
https://www.idiap.ch/dataset/printattack, Accessed
from the Department of Computer
October 2014.
Engineering, Eastern Mediterranean
[30] Replay Attack face database,
University, North Cyprus in 2014. Currently,
https://www.idiap.ch/dataset/replayattack, Accessed
he is an Assistant professor in the
October 2014.
Department of Computer and Software
[31] J. Galbally, M. Sébastien and J. Fierrez, "Image quality
Engineering, Toros University, Turkey. His
assessment for fake biometric detection: Application to
current research interests include biometrics,
iris, fingerprint, and face recognition." IEEE transactions
face recognition, iris recognition and multimodal fusion.
on image processing 23.2 (2014): 710-724.

How to cite this paper: Omid. Sharifi, "Score-Level-based Face Anti-Spoofing System Using Handcrafted and Deep
Learned Characteristics", International Journal of Image, Graphics and Signal Processing(IJIGSP), Vol.11, No.2, pp.
15-20, 2019.DOI: 10.5815/ijigsp.2019.02.02

Copyright © 2019 MECS I.J. Image, Graphics and Signal Processing, 2019, 2, 15-20