Face/Fingerphoto Spoof Detection under Noisy Conditions by using
Deep Convolutional Neural Network
Masakazu Fujio
1
, Yosuke Kaga
1
, Takao Murakami
2
, Tetsushi Ohki
3
and Kenta Takahashi
1
1
Security Research Dept., Hitachi, Ltd., Tokyo, Japan
2
Advanced Cryptosystems Research Group,
National Institute of Advanced Industrial Science and Technology, Tokyo, Japan
3
Department of Computing, Shizuoka University, Shizuoka City, Japan
Keywords: Biometrics, Spoofing, LBP, CNN, Deep Learning, Mobile, Blurriness.
Abstract: Most of the generic camera based biometrics systems, such as face recognition systems, are vulnerable to
print/photo attacks. Spoof detection, which is to discriminate between live biometric information and attacks,
has received increasing attentions recently. However, almost all the previous studies have not concerned the
influence of the image distortion caused by the camera defocus or hand movements during image capturing. In
this research, we first investigate local texture based anti-spoofing methods including existing popular
methods (but changing some of the parameters) by using publicly available spoofed face/finger photo/video
databases. Secondly, we investigate the spoof detection under the camera defocus or hand movements during
image capturing. To simulate image distortion caused by camera defocus or hand movements, we create
blurred test images by applying image filters (Gaussian blur or motion blur filters) to the test datasets. Our
experimental results demonstrate that modifications of the existing methods (LBP, LPQ, DCNN) or the
parameter tuning can achieve less than 1/10 of HTERhalf total error ratecompared to the existing results.
Among the investigated methods, the DCNN (AlexNet) can achieve the stable accuracy under the increasing
intensity of the blurring noises.
1 INTRODUCTION
With the exponential growth of the smartphone
market, financial services are also accelerated by the
development of various services on mobile devices,
such as mobile payments, money transfer, and all
banking related transactions.
As the mobile e-commerce continues to grow, the
biometrics authentications are attracting more
attentions for the secure and easy-to-use
authentication methods, as the alternative for the
insecure and inconvenient Password/PIN
authentication methods.
Biometrics can implement the convenient user
authentication, but on the other hand, it is vulnerable
to be spoofed by the fake copies of the user biometric
features made of commonly available materials such
as clay and gelatines. For example, the gummy finger
model attacks for the fingerprint biometrics solutions
demonstrate the possibilities of the unauthorized
accesses. Especially the generic camera based
biometrics has the higher risk of spoofing by the
printed photos and videos (preparation costs are low).
For that reason, the spoof detection, which is to
discriminate between live faces/fingers and attacks,
has received increasing attentions recently
(Keyurkumar et al., 2015; Bai et al., 2010).
Typical anti-spoofing techniques can be coarsely
classified into three categories based on clues used for
spoof attack detection: (i) motion analysis based
method, (ii) texture analysis based methods, and (iii)
image qualities analysis based methods.
(i) Motion analysis based methods
These methods, which are effective to counter
printed photo attacks, capture the movement clues
such as eye blinks (Gang et al., 2007) and lip
movements (Avinash et al., 2014), which are very
important cues for vitality. But in the case of the face
biometrics, the system needs accurate detections of
facial parts such as eyes, lips and so on. Furthermore,
54
Fujio, M., Kaga, Y., Murakami, T., Ohki, T. and Takahashi, K.
Face/Fingerphoto Spoof Detection under Noisy Conditions by using Deep Convolutional Neural Network.
DOI: 10.5220/0006597500540062
In Proceedings of the 11th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2018) - Volume 4: BIOSIGNALS, pages 54-62
ISBN: 978-989-758-279-0
Copyright © 2018 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
simply capturing movement clues are not enough for
the presentation attacks by videos.
(ii) Texture analysis based methods
These methods capture the texture features
appeared on natural scenes, photo papers, and
displays under the assumption that surface properties
of real faces and prints are different (Diego and
Giovanni, 2015; Tiago, 2012, 2014; Ivana, 2012; Juho
et al., 2012).
Texture based methods such as Local Binary
Patterns (LBP) (Matti et al., 2011) have achieved
significant success on the Idiap and CASIA databases
(Ivana et al., 2012; Zhiwei et al., 2012). For example,
The Half Total Error Rate (HTER) on the Idiap
database was reduced from 13.87% in (Ivana et al.,
2012) to 6.62% in (Samarth, 2013). Unlike motion
based methods, texture based methods need only a
single image to detect a spoofing.
Other types of texture analysis methods adopt the
frequency domain features (Jiangwei et al., 2004). For
example, low resolution printed images have a high-
frequency spectral magnitude in the frequency
domain, caused by periodic dot printing (Xiaofu et al.,
2009). Jiangwei et al. (2004) described a method for
detecting print-attack face spoofing by exploiting
differences in the 2D Fourier spectra of live and
spoofed images. The method assumes that
photographs are normally smaller in size and contain
less high-frequency components compared to real
faces. Then their method likely fails for higher-
quality samples.
(iii) Image qualities analysis based methods
These methods capture the degradation of the
image qualities caused by presenting photographs or
videos to the generic cameras (Javier et al., 2014;
Diogo and Ricardo, 2015; Di et al., 2015). For
example, printed/displayed images usually have
lower resolution, narrow dynamic range, specular
reflection, reduced image contrast and defocused
blurriness. But those image quality degradations also
appear in both genuine and spoofed face images, it is
not simple to distinguish that the image distortions are
caused by spoofing or camera operations.
The problems of the almost all the existing studies
are that they have not concerned the influence of the
image distortion caused by the camera defocus or
hand movements during an image capturing.
In this research, we first investigate local texture
based anti-spoofing methods including existing
popular methods (but changing some of the
parameters) by using publicly available spoofed
face/fingerphoto/video databases (Replay-Attack
Database and Spoofed Fingerphoto Database).
Secondly, we investigate the spoof detection under
the camera defocus or hand movements during image
capturing. To simulate image distortion caused by
camera defocus or hand movements, we create
blurred images by applying image filters (Gaussian
blur or motion blur filters) to the test datasets. For the
training images, we do not apply image filters.
The remaining of the paper is organized as follows.
Section 2 contains the examined schemes of the anti-
spoof technique and provides an explanation of the
countermeasures to photo/display attacks in
fingerphoto or face recognition. Section 3 details
experimental protocols, the dataset statistics,
parameters used in the algorithm and the results
obtained. Section 4 concludes the paper.
2 PROPOSED METHODS
To design the anti-spoofing countermeasures, we
investigated fingerphoto spoofed image database
(Archit et al., 2016). Fig. 1 shows magnified images
(The left side is a genuine image, and the right side is
a spoofed image.). From the Fig. 1, we can see block
noises in the spoofed images.
(a) genuine (b) spoofed
Figure 1: Magnified fingerphoto images.
To highlights the noises in the images, we
performed image enhancement of the above images
based on wavelet transform (Fig. 2). From the Fig. 2,
we can see repeated block noise artifacts in the
spoofed images (The square frame in the left image
in the fig. 2 shows a block of 8x8 pixels).
(a) genuine (b) spoofed
Figure 2: Wavelet transformed images (leftgenuine, right
spoofed).
To capture the noise features found in the
preliminary analysis, we focus on the LBP (Matti et
al., 2011) and the block noise indicator (Zhou et al.,
Face/Fingerphoto Spoof Detection under Noisy Conditions by using Deep Convolutional Neural Network
55
2000, 2002). The block noise indicator is used to
quantify the magnitude of block artifacts caused by
the application of lossy compression algorithms such
as JPEG compression. And used for the judgment of
the image compression algorithms (Zhou et al., 2000,
2002). In the rest of this section, we will explain anti-
spoofing techniques based on handcrafted features or
the automatic feature extraction based on CNN
(convolutional neural network).
2.1 SVM with Handcrafted Features
We used the following 3 handcrafted feature vectors
to train Support Vector Machine (SVM) classifiers.
(1) WLBP (Wavelet transformed Local Binary
Pattern)
LBP (Local Binary Patterns) is the 8-bit encoding
based on the comparisons of the magnitudes of the
luminance between the focused pixel and neighboring
8 pixels. LBP is widely used for the image
classification tasks in the literature.
Figure 3: LBP features of the each pixel.
In addition to the LBP of original image size, we
extracted LBP features from compressed images and
contrast enhanced images. All features are
concatenated and used for the training of the linear
SVM.
(a) Calculation of LBP from the original images:
(b) Image enhancement (wavelet transformation):
To enhance the noise patterns in the images,
perform wavelet transformations and calculate LBP
of the transformed images
(c) Image compression:
Base on the preliminary study about block noise
pattern, we set the compression ratio as 3/8 (“3” is the
kernel size of LBP, and “8” is the size of the observed
blockiness).
(d) Feature fusion:
Concatenate the three LBP (original, wavelet
transformed, compressed), and use for the training of
the linear SVM
(2) NRPQA (No-Reference Perceptual Quality
Assessment)
This measure applies the block artifact indicator
(2002) for the ant-spoof detections. Zhou et al. (2002)
describes perceptual quality assessment of JPEG
compressed images by calculating three measures,
inter block differential, intra block differential and
zero-crossing rate.
Figure 4: 3 types of LBP features.
We denote the test image signal as x (m; n) for
Mm ,1
and
Nn ,1
, and calculate a differential
signal along each horizontal line:
(1)
The features are calculated horizontally and then
vertically. The amount of blockiness is estimated as
the average differences across boundaries.
 
M
i
N
j
hh
jid
NM
B
1
18/
1
)8,(
)18/(
1
(2)
Second, we estimate the activity of the image
signal. The second activity measure, is the average
absolute difference between in-block image pixels.
 
h
M
i
N
j
hh
Bjid
NM
A
1
1
1
),(
)1(
1
7
1
(3)
The third activity measure , is the zero-crossing
(ZC) rate. Horizontal Zero-Crossing (ZC) means that
there is a change of the sign of the value
),( nmd
h
between n and n+1 (
 2,1 Nn
):
  
          
    
otherwise
nmdatZChorizontal
nmz
h
h
0
),(1
),(
(4)
ZC is then estimated by using
),( nmZ
h
as follows:
 
M
i
N
j
hh
nmz
NM
Z
1
2
1
),(
)2(
1
(5)
For the detail of these measures, please refer to
(Zhou et al., 2002). In this study, we trained linear-
SVM with these 6 features {
hhhhhh
ZABZAB ,,,
}.
(3) LPQ (Local Phase Quantization)
The local phase quantization (LPQ) (Timo, 2008) is a
method based on the blur invariance property of the
Fourier phase spectrum and uses the local phase
information extracted using 2-D discrete Fourier
transform or short term Fourier transform (STFT),
computed over a rectangular region. The STFT over
 
h
A
 
h
Z
BIOSIGNALS 2018 - 11th International Conference on Bio-inspired Systems and Signal Processing
56
a region of the N by N neighborhood
x
N
of image
g (x) with each position of the pixel x is defined by
x
T
u
Ny
uTyj
hweyxhxuH
x
  
 
2
)( ),(
(6)
where is the basis vector of the 2D discrete
Fourier transform at a frequency
u
while
x
h
stands
for the vector containing all
2
N
pixels.
In the task of spoofing detection, we expect that
LPQ is tolerant for the distorted images caused by
defocused images or motion blurred images, which
are commonly seen in both printed photos, replayed
video attacks and real faces/fingers.
2.2 Deep Convolutional Neural
Network (DCNN)
Convolutional Neural Networks (Alex et al., 2012)
have demonstrated state-of the-art performances in a
variety of image recognition benchmarks, such as
MNIST (Yann and Corinna., 2010), CIFAR-10,
CIFAR-100 (Alex and Geoffrey, 2009), and ImageNet
(Alex et al., 2012).
In this study, we used AlexNet (Alex et al., 2012).
This model won both classification and localization
tasks in the ILSVRC-2012 competition. This model
also exhibited good results on the spoof detection on
the Replay-Attack Database (HTER<0.5%) (Koichi
et al., 2017).
CNN Model AlexNet without pre-trained
weights.
As the preliminary experiments, we tried various
sizes of the image compressions and cropping sizes
for the training of the DCNN (AlexNet) models.
Based on the results of the preliminary experiments,
we choose the combination of image resizing:256
pixels and image cropping size:227 pixels, which
exhibited the highest accuracy.
We set the parameters for the training of CNN as
follows:
Output layer number: 2, Resized image size: 256,
random cropping size: 227, batch size: 100, epoch
size: 300,000, learning rate: 0.01, weight decay:
0.004.
In the above settings, we do not use image
augmentation (except for the random image
cropping). We consider that not only foreground
regions but also of images (such as faces and finger)
but also both background images and foreground
images have discriminative cues for the spoof
detection.
3 EXPERIMENTS
In this section, we first provide an overview of the
datasets in our experiments, and present our initial
results for the proposed methods in previous section
(one is linear-SVM with the handcrafted feature
vectors, and the other is Deep Convolutional Neural
Network (DCNN)).
3.1 Database Descriptions
To evaluate the spoofing detection accuracies, we
used two databases, one is Replay Attack Database
(face) and the other is Spoofed Fingerphoto
Database (finger).
A) Spoofed Fingerphoto DB
Table 1: Summary of the Spoofed Fingerphoto DB.
#
Condition
Description
1
Lighting
condition
2types (Indoor/Outdoor
2
Background
2typesWhite/Natural
3
Image size
Genuine image3264x2448
4
Spoofed image2332x1132
5
Number of
images
Genuine image: 409664x2x8x4
6
Spoofed image: 819264x2x2x8x4
Table 2: Attack Protocols for Fingerphoto Spoofing.
#
Capture
Display
1
OnePlus One phone
iPad
2
OnePlus One phone
Laptop
3
OnePlus One phone
Nexus
4
OnePlus One phone
Printout
5
Nokia
iPad
6
Nokia
Laptop
7
Nokia
Nexus
8
Nokia
Printout
Following the setup of Tanja (2016), genuine
images (4096) were split into the gallery and probe
data (each has 2048 images). Then the images used
for generating spoof images (512) were excluded
from genuine images (1536), and spoofed images
were added (4096) to the training of the each model
(total 5632).
From probe data, genuine images (2048) +
imposter images (4096) were used for the evaluation
data set.
B) Replay-Attack DB
The 2D face spoofing attack database consists of
1,300 video clips of photo and video attack attempts
of 50 clients, under different lighting conditions. The
size of the image is 320 x240.
u
w
Face/Fingerphoto Spoof Detection under Noisy Conditions by using Deep Convolutional Neural Network
57
Table 3: Attack protocols for face spoofing.
#
Setting
Description
Capture
condition
Lighting
condition
2types (controlled/adverse
Attack
protocols
Display
devices
5types
(mobile photo/mobile video/high-
resolution photo/high-resolution
video/high-resolution print)
Attack modes
2types (hand/fixed)
Training data size:
Attack Video: 300 clips (15 (clients) × 2
(lighting) × 5 (devices) × 2 (modes) )
Real Video: 60 clips (15 (clients) × 2 (lighting)
× 2 (shots))
Test data size:
Attack Video: 400 clips (20 (clients) × 2
(lighting) × 5 (devices) × 2 (modes))
Real Video: 80 clips (20 (clients) × 2 (lighting)
× 2 (shots))
Following the setup of Ito (2017), we used the
training set for the training, and evaluated the
spoofing detection accuracies with the test set. To
train the face spoof detection model, we extracted the
all frame images from the video clips, and both
training and evaluations are performed for those
images.
3.2 Spoof Detection Accuracy
In this experiment, we compare the spoof detection
accuracy of the investigated methods with the
baseline methods and state of the art methods on two
databases: Replay-Attack Database and Spoofed
Fingerphoto Database. Table 4 and Table 5 show the
HTER (half total error rate) of the spoof detection
accuracies for the each attack type, such as printing,
photo, movies. For the Replay-Attack DB, we used
all frames in the movie clips both for the training and
testing.
Table 4 shows that the proposed DCNN method
outperforms the other methods. For the Fingerphoto
Spoof DB (Archit et al., 2016), spoof detection
accuracies of the proposed methods (NRPQA+SVM
and DCNN) exhibited less than 1/10 error rates
compared to the Archit et al. (2016) s LBP based
method.
Table 5 shows that the proposed DCNN method
outperforms the other methods, even better than the
other state-of-the-art DCNN method (Koichi et al.,
2017; Jianwei et al., 2014). The differences of the our
DCNN model and the Koichis DCNN model are the
preprocessing of input images and the cropping size
of the images. Our model compresses the input
images (256x256) before cropping, but Itos model
does not. The cropping size of our model is (227x227),
a little bit larger than Itos model (240x180). As is the
case with WLBP features, the image compression
process may contribute to capturing the differences
between the real /spoofed images. In the case of the
Replay-Attack Database, the method NRPQA+SVM
shows the low detection accuracies, especially for the
Highdef samples, which means high-resolution
photos and videos images. This may mean that the
examined NRPAQ features may be specific features
that is prominent to the Spoofed Fingerphoto
Database. But only our DCNN model shows high
accuracy (HTE < 2/10
-6
, by using rule of three) for
the Highdef images.
Table 4: Summary of the evaluation results (finger).
Half Total Error Rate%)
Attack
Scenario
Proposed Methods
Baseline
Display
Capture
DCNN
WLBP+SVM
NRPQA+SVM
LBP+SVM
LPQ+SVM
LBP+SVM
[Tanja2016]
Print
Nokia
0.024
0.05
0.0
0.66
4.71
6.05
OPO
0.024
0.02
0.0
0.42
1.83
4.85
iPad
Nokia
0.2
0.12
0.0
2.52
3.15
3.12
OPO
0.024
0.83
0.0
0.39
0.71
5.27
Nexus
Nokia
0.024
0.78
0.0
0.85
3.79
1.39
OPO
0.024
0.32
0.24
0.56
3.13
0.24
Laptop
Nokia
0.024
0.20
0.24
5.76
20.8
4.48
OPO
0
0.17
0.39
0.42
1.8
2.31
Total
0.04
0.9
0.06
3.56
4.8
3.71
Table 5: Summary of the evaluation results (face).
Half Total Error Rate%)
Attack
Scenario
Proposed Methods
Baselines
State-of-
the-art
DCNN(Ours)
WLBP+SVM
NRPQA+SVM
LBP+SVM
LPQ+SVM
DCNN
[Itoh17]
Print
0.0
0.01
1.84
12.4
7.9
N/A
Mobile
0.0
0.4
1.14
2.5
25.0
N/A
Highdef
0.0
7.8
17.2
12.7
32.2
N/A
Total
0.01
4.98
11.1
16.5
30.2
0.52
To investigate that which parts of the image areas
contribute to detect spoofing, we adopted Grad-CAM
visualization method (Ramprasaath et al., 2016) to
our DCNN models.
Grad-CAM highlights importance of each
BIOSIGNALS 2018 - 11th International Conference on Bio-inspired Systems and Signal Processing
58
neurons for an each prediction. To obtain the class-
discriminative localization map (Fig. 5 (c)(d), Fig. 6
(c)(d)), Grad-CAM calculate the gradient of the score
for the class of interest (in this study, “real” or
“spoofed”), with respect to CNN feature maps (for
example, relu5 layer of the DCNN(AlexNet)).
Fig. 5 (a) is the original fingerphoto image, and
(c) is the Grad-CAM visualization of the image (a).
Fig. 5 (b) is the spoofed fingerphoto image, and (d) is
the Grad-CAM visualization of the image (b).
From Fig. 5 (a)(c), we can see that our DCNN
model detects spot areas of the genuine image, mainly
inside of the finger areas. On the other hand, for the
spoofed image, the model detects border areas
between background areas and finger areas.
(a) genuine (b) spoofed
(c) Grad-CAM of (a) (d) Grad-CAM of (b)
Figure 5: (a-b) Original finger image and the generated
spoofed image. (c-d) Grad-CAM maps for the original
image and the spoofed image.
Fig. 6 (a) is the original face image, and (c) is the
Grad-CAM visualization of the image (a). Fig. 6 (b)
is the spoofed face image, and (d) is the Grad-CAM
visualization of the image (b).
From Fig. 6 (a)(c), we can see that our DCNN
model detects spot areas of the genuine image, mainly
lateral side of the face areas. On the other hand, for
the spoofed image, the model detects border areas
between background areas and face areas (but mainly
background areas), and covers more wider areas,
compared to the original (real) image.
Those visualization results suggest that our
DCNN models learns the border between background
areas and finger/face areas, and utilizes them to detect
spoof/genuine images.
(a) genuine (b) spoofed
(c) Grad-CAM of (a) (d) Grad-CAM of (b)
Figure 6: (a-b) Original face image and the generated
spoofed image. (c-d) Grad-CAM maps for the original
image and the spoofed image.
3.3 Performance Comparison
We compare the performances of the investigated
spoofing detection methods (three types of
handcrafted features with SVM, and automatic
feature extraction/classification by using DCNN
(AlexNet)).
For the SVM methods based on handcrafted
features, we examined feature extraction + SVM
classification times by using linear SVM classifier
(liblinear). For the DCNN (AlexNet) based method,
we examined input image classification times.
Table 6 shows that DCNN and NRPQA are more
than 10 times faster than WLBP or LPQ. The
spoofing detection accuracies of those two methods
are also much better than the WLBP or LPQ.
Table 6: Summary of the performance (finger).
Processing time per 1000 images (second)
AlexNet
WLBP
NRPQA
LPQ
Feature
extraction
4.33
193.0
3.76
49.0
Classification
0.22
Total
4.33
193.0
3.98
49.22
3.4 Blurriness Tolerant Analysis
In the task of the spoof detection, presentation images
are affected by camera defocus or hand movements,
which are commonly seen in both printed photo,
replayed video attacks and real faces/fingers.
To simulate image distortions caused by camera
defocus or hand movements, we create blurred
images by applying image filters (Gaussian blur or
motion blur filters) to the test datasets. For the
Face/Fingerphoto Spoof Detection under Noisy Conditions by using Deep Convolutional Neural Network
59
training images, we do not apply image filters (only
use original data sets).
For the Gaussian blurriness, we changed the
standard deviation σ parameter, from 0.1 to 2.6 by
the step size 0.5. For the motion blurriness, we
changed the length of the PSF point spread
functionfrom 1 to 10 pixels by step size 1, and the
filter angle was fixed (11°).
(a)σ=0.1 (b) σ=2.6
(c)σ=0.1 (d) σ=2.6
Figure 7: Examples of Gaussian blurred images (face).
(a) PSF size=1 (b) PSF size=10
(c) PSF size=1 (d) PSF size=10
Figure 8: Examples of motion blurred images (face).
Fig. 13 and Fig. 15 show the spoof detection
accuracies for face and spoofed fingerphoto database
for Gaussian blurred images. Fig.14 and Fig. 16 show
the spoof detection accuracies for face and spoofed
fingerphoto database for motion blurred images. In
each figure, NRPQA represents the method
NRPQA defined in the section 2.1 (2), LBPs
represents the method WLBP defined in the section
2.1 (1), LPQs represents the method LPQ defined
in the section 2.1 (3) (but combined with the LPQ
from wavelet transformed images too), and DCNN
represents the method defined in the section 2.2.
ALL represents the features concatenation of all
NRPQA, LBPs and LPQs.
(a)σ=0.1 (b) σ=2.6
Figure 9: Examples of Gaussian blurred images (finger).
(a) PSF size=1 (b) PSF size=10
Figure 10: Examples of motion blurred images (finger).
In the case of Spoofed Fingerphoto Database, Fig.
13 shows that the more standard deviation of the
Gaussian blur increases, the accuracies of the spoof
detection decrease steeply. Especially it is prominent
for the NRPQA, which exhibited the hight
performance for the normal (no additive blurring)
images. NRPQA is based on the block noise of the
images, which may disappear by the blurring noises.
It is also the case for the motion blurring, as you can
see in the Fig. 14.
In the case of Replay-Attack Database, Fig. 15
and Fig. 16 also show that the more adding blurring
noise, the accuracies of the spoof detection decrease
steeply. In this case, the decreases of the accuracies
are almost all the same for the features NRPQA,
LBPs, and LPQs.
Among the examined methods, DCNN shows not
only the highest accuracy but also the highest
robustness for the blurring noises. To investigate the
stabilities of the examined DCNN model for the
increasing intensity of the blurring noise, we adopted
Grad-CAM visualization for blurred images. Fig. 11
shows the results of Grad-CAM visualizations for the
Gaussian blurred images. Fig. 12 shows the results of
Grad-CAM visualizations for the motion blurred
images. We can see that highlighted areas are not
affected by the increasing intensity of the Gaussian
blur, or motion blur. Those results support the results
of spoof detection accuracies of the DCNN models.
BIOSIGNALS 2018 - 11th International Conference on Bio-inspired Systems and Signal Processing
60
(a) σ=0.1 (b) σ=2.6
Figure 11: (a) Grad-CAM map for the Gaussian blurred
image ( σ =0.1). (b) Grad-CAM maps for the motion
blurred same image (σ=2.6).
(a) PSF size=1 (b) PSF size=10
Figure 12: (a) Grad-CAM map for the motion blurred
image (PSF size=1). (b) Grad-CAM maps for the motion
blurred same image (PSF size=10).
Figure 13: Spoof detection accuracy on the Gaussian blur
(finger).
Figure 14: Spoof detection accuracy on the motion blur
(finger).
Figure 15: Spoof detection accuracy on the Gaussian blur
(face).
Figure 16: Spoof detection accuracy on the motion blur
(face).
4 CONCLUSIONS
In this paper, we address the problem of face/finger
spoof detection for the generic camera based
biometrics, particularly under noisy conditions. We
first propose the anti-spoofing methods based on the
local texture features, and achieved less than 1/10 of
HTER (half total error rate) compared to the previous
methods, for the two different modality databases,
Replay-Attack Database (face) and Spoofed
Fingerphoto Database (finger).
Furthermore, to simulate the real-life scenarios,
we investigate spoof detection under additive noise,
such as defocused blurriness and motion blurriness.
Our experiments show that using the model trained
from clean data, most of the system performance
degrades significantly for blurred images. Among the
proposed methods, only the DCNN based method
shows not only the highest accuracy but also the
highest robustness for the blurred images.
For future work on spoof detection of the generic
camera based biometrics, we intend to include i)
evaluation under the ambient illumination in the use
case scenario of interest, ii) investigation for the
Face/Fingerphoto Spoof Detection under Noisy Conditions by using Deep Convolutional Neural Network
61
cross-modal scenario, iii) develop compact models
that can be used on mobile devices.
REFERENCES
Alex, K., Geoffrey, H., 2009. Learning multiple layers of
features from tiny images. University of Toronto.
Alex, K., Ilya, Sutskever., Geoffrey.H., 2012. Imagenet
classification with deep convolutional neural networks.
In NIPS’12, Advances in Neural Information
ProcessingSystems. NIPS FOUNDATION , Inc.
Archit, T., et al., 2016. Fingerphoto Spoofing in Mobile
Devices: A Preliminary Study. In BTAS’16,
International Conference on Biometrics: Theory,
Applications, and Systems. IEEE.
Avinash, K.S et al, Face recognition with liveness detection
using eye and mouth movement. In ICSPCT’14,
International Conference on Signal Propagation and
Computer Technology. IEEE.
Bai, J., Ng, T.T., Gao, X., Shi, Y.Q., 2010. Is physics-based
liveness detection truly possible with a single image?.
In ISCAS’10. IEEE.
Di, W., Hu, H., Anil, K.J., 2015. Face spoof detection with
image distortion analysis. IEEE Transactions on Image
Process., vol.10, no. 4. IEEE.
Diogo, C.G., Ricardo. L. de Queiroz., 2015. Face-spoofing
2d-detection based on moire-pattern analysis. IEEE
Transions on Information Forensics and Security, vol.
10, no. 4. IEEE.
Diego, G., Giovanni P., 2015. An Investigation of Local
Descriptors for Biometric Spoofing Detection. IEEE
Transions on Information Forensics and Security, vol.
10, no. 4. IEEE.
Gang, P., Lin, S., Zhaohui, W., Shihong, L., 2007.
Eyeblink-based anti-spoofing in face recognition from
a generic web camera. In ICCV’07, 11
th
international
conference on Computer Vision. IEEE.
Ivana, C., Andre, A., Sebastien, M., 2012. On the
effectiveness of local binary patterns in face anti-
spoofing. In BIOSIG’12, International Conference on
Biometrics Special Interest Group. IEEE.
Javier. G., Sebastien, M., Julian, Fierrez., 2014. Image
quality assessment for fake biometric detection:
Application to iris, fingerprint and face recognition.
IEEE Transactions on Image Process., vol. 23, no. 2.
IEEE.
Jiangwei, Li., et al., Live face detection based on the
analysis of Fourier spectra. In SPIE’04, Biometric
Technology for Human Identification.
Jianwei, Y., Zhen, L., Stan.Z. L., 2014. Learn convolutional
neural network for face anti-spoofing. In CoRR’14,
International Conference on Computer Vision and
Pattern Recognition. IEEE.
Juho, M., Abdenour, H., Matti, P., 2012. Face spoofing
detection from single images using texture and local
shape analysis. In IET Biometrics. IET Journals &
Magazines.
Keyurkumar, Patel., Hu Han, Anil K. Jain, 2015. Secure
Smartphone Unlock: Robust Face Spoof Detection on
Mobile. Technical Report MSU-CSE-15-15. MSU.
Koichi, I., Takehisa, O., Takafumi, A., 2017. A Study of
Anti-Spoofing Using Deep Learning for Face
Recognition. In SCIS 17, 2017 Symposium on
Cryptgraphy and Information Security. IEICE.
Matti, P., Abdenour, H., Guoying, Z., Timo. A., 2011. In
Computer Vision Using Local Binary Patterns.
Springer Science & Business Media.
Xiaofu. H., Yue, L., Pengfei. S., 2009. A new fake iris
detection method. In ICB’09, Advances in Biometrics.
SPRINGER.
Ramprasaath R.S et al, 2016. Grad-CAM: Visual
Explanations from Deep Networks via Gradient-based
Localization. preprint, arXiv:1610.02391
Samara, B., Tejas, I.D., Mayank, V., and Richa. S., 2013.
Computationally efficient face spoofing detection with
motion magnification. In CVPRW’13, IEEE
Conference on Computer Vision and Pattern
Recognition Workshop. IEEE.
Tiago, d.F.P. et al., 2012. LBP-TOP based counter measure
against facial spoofing attacks. In ACCV’12 Workshop
on Computer Vision with Local Binary Pattern
Variants. IDIAP.
Tiago, d.F.P. et al., 2014. Face liveness detection using
dynamic texture. In EURASIP J. Journal on Image and
Video Processing. SPRINGER.
Timo, A., et al., 2008. Recognition of blurred faces using
local phase quantization. In ICPR’08, 19th
International Conference on Pattern Recognition.
IEEE.
Yann, L., Corinna, C., 2010. MNIST handwritten digit
database. http://yann.lecun.com/exdb/mnist/.
Zhiwei, Z., et al., 2012. A face anti-spoofing database with
diverse attacks. In ICB’12, International Conference on
Biometrics. IEEE.
Zhou, W. et al., 2000. Blind Measurement of Blocking
Artifacts in Images. In ICIP2000, Intertational
Conference on Image Processing. IEEE.
Zhou, W., et al., 2002. No-Reference Perceptual Quality
Assessment of JPEG Compressed Images. In
ICIP2002, Intertational Conference on Image
Processing. IEEE.
BIOSIGNALS 2018 - 11th International Conference on Bio-inspired Systems and Signal Processing
62