IRJET- Face Spoof Detection using Machine Learning with Colour Features

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 05 Issue: 03 | Mar-2018 www.irjet.net p-ISSN: 2395-0072
© 2018, IRJET | Impact Factor value: 6.171 | ISO 9001:2008 Certified Journal | Page 1558
Face Spoof Detection Using Machine Learning with Colour Features
Mahitha.M.H1
1PG Scholar, Department of Computer Science and Engineering, Nehru College of Engineering and
Research Centre, Thiruvilamala, Kerala, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Now a day the authentication techniques use
biometric information as the credentialof users. Forproviding
more secure, use biometric information as authentication
method but biometric information sufferedfromsomeattacks.
Face Spoofing is one kind of biometric-based attack. In this
paper propose a face spoof detection protocol which is based
on colour texture analysis. This paper mainly focusesonphoto
and video face spoof attacks. Here use colour spaces for
extracting local texture features and distortion features. Then
collected all feature values for SVM training. SVM is a
supervised machine learning algorithm which used here to
detect genuine faces and spoofed faces.
Key Words: Face SpoofDetection,SVM,TextureFeatures,
Distortion Features, Colour Space.
1. INTRODUCTION
Now a day’s increase the cyber crimes, for providing more
security we applied biometric information in authentication
techniques. The biometric information such as finger print,
Iris, palm print and face are commonly used credential
information of user. But attackers can break the biometric
based secure system by providing fake sample of biometric
information of valid user. Face spoofing is one kind of that
attack which occurs when a fake sample of valid users face
present to the acquisition sensor. These types of attacks
mainly in three types 1) video attack 2) photo attack and 3)
mask attack. The video based face spoofing occurs when
recorded video of valid user’s face used as a fake sample to
break the security. In photo attack, attackerpresentphoto of
valid user to the acquisition sensor. In mask attack, the
attacker wears the mask which similar to the victims face.
Here propose a technique to detect spoofed faces based on
colour features. The colour features are very helpful for
discriminating fake faces from original faces. There are so
many approaches are available to detect spoofed faces but
that techniques mainly focus on the light intensity andavoid
the chroma components. Chroma components are effective
factor for discriminating the fake faces. Some existing face
spoof detection techniques are introduced in this paper.
2. PRIOR WORK
The prior works of face spoof detection mainly focus on
frequency, texture, quality and motion parameters to detect
livness of face.
2.1 Image Quality Assessment
J. Galbally [1] proposed an approach to detect genuine face
based on image quality assessment. In this method the
inputted image read as a gray scale for feature extraction.
The gray scale image I is filtered with a low-pass Gaussian
kernel for generating the distorted version I- .Then compute
quality between two images I and I- by using IQA metric.IQA
metric consider the following measures: Pixel Difference
measures, compute distortion between two images on the
basis of their pixel wise differences. Correlation-based
measures, compute the angles between the pixel vectors of
the original and distortedimages.Edge-basedmeasurestake
Total Edge Difference (TED) and Total Corner Difference
(TCD). Here simple Linear DiscriminantAnalysis(LDA)used
as a classifier for classifying genuine and fake faces.
2.2 Image Distortion Analysis
D.Wen [2] proposed a face spoofing detection algorithm
based on Image Distortion Analysis (IDA). IDA is the set of
features such as specular reflection, blurriness, chromatic
moment, and colour diversity. Here combine multiple SVM
classifier output for getting final decision and the multiple
SVM classifier is represented as ensemble classifier.
2.3 Countermeasure for Detect Face Spoofing Attack
A. Anjos [3] proposed fusion of motion and texture based
countermeasures. Motion based correlation analysis is used
to measure the correlations between the users head
movements and the background scene. Texture quality
analyzing by usinglocal binarypatterns(LBP).Herecombine
the motion and micro-texture analysis basedtechniques,the
inputted video sequences divided into N frames window.
Here the LBP face description is computed only for the last
frame but use whole time windowformotionbasedanalysis.
The fusion of these two methods is performed at score level
using linear logistic regression (LLR). For classificationhere
use linear discriminant analysis (LDA).
2.4 Context Based Face Spoofing Detection
J. Komulainen[4] proposed this work to detect the close up
fake faces by using HOG descriptors. Here useanupper body
detector for analyzing the alignment of the face and the
upper half of the torso. A specific detector is used to
determine the presence of the display medium. In this
approach if the upper body of an face image is not found it
concluded that the inputted image is fake otherwise the
inputted face image give as an input to the spoofing medium

detector for finding the medium then if the medium is found
then the inputted one is spoofed otherwise genuine.
2.5 Fourier Spectra Analysis
J. Li [5] proposed this approach for analyzing the Fourier
Spectra of face image. This method mainly focuses on two
principles for the detection of spoofed samples. First
principle is that the high frequency componentsofthephoto
images is less than the real face images and Second principle
is that the standard deviation of the frequency components
in the sequence must be small[5]. If the median of the HFD is
smaller than threshold value then inputted face is a spoofed
sample otherwise the sample is a live face.
2.6 Visual Dynamics Based Face Spoofing Detection
S.Tirunagari[6] proposed an approach to identify the
liveliness detection of a user by using the pipeline of
DMD+LBP+SVM. Dynamic Mode Decomposition (DMD)
algorithm is used to transform video sequences into
corresponding image sequences. Extract features from each
image sequences by using Local Binary Pattern (LBP) and
LBP histogram feature values give as a input to SVM for final
classification.
2.7 Motion Based Face Antispoofing
S.Bharadwaj [7] introduced an approach for spoofing
detection in face videosusingmotionmagnification. Eulerian
motion magnification approach used here to enhance facial
expression from captured video. Two types of feature
extraction algorithms are implemented here: (i) LBP that
provides texture based analysis and (ii) HOOF descriptor
which used to extract motion based features.
2.8 Texture and Local Shape Analysis
J. Komulainen [8] proposed a method which adopted two
powerful texture features, LBPs and Gabor wavelets, for
describing not only the micro-textures but also more
macroscopic information. HOG extract local shape
characteristics by counting occurrences of gradient
orientation in localized portions of an image. Each low-level
descriptor produces itsownfacerepresentationbuthere use
homogeneous kernel map to transform the data into
compact linear representation. Each vector applied to a
linear SVM classifier and combine the individual SVM
outputs for determines whether there is a live person or a
fake image in front of the camera.
3. METHODOLOGY
This paper proposes colour texture based face spoofing
detection. Spoofing occurs when an attacker present a fake
sample to the acquisition sensor. The prior spoofing
detection approaches doesn’t work with colour spaces. But
the proposed spoofing detection protocol mainly focus on
the colour spaces to detect spoofed faces. In this approach
the inputted face image converted into different colour
spaces, Table 1 shows different colour representation of an
image.
Three different colour spaces are used here, RGB, YCbCrand
HSV. RGB is the commonly used colour space which is the
set of red, green and blue colours. These three colours are
also known as primary colour. By using these three colour
we can generate any colour. YCbCr is the combination of
luminance and chroma components such as chroma blue
(Cb) and chroma red(Cr). HSV represent the Hue, Saturation
Value.
The proposed system use different feature vectors for
extracting local texture information from the image and
some feature extraction techniques are help to avoid
specular reflection. Here use the Viola Jones algorithm for
face detection. In this approach there have a chance for false
detection based on the light intensity. For avoiding this
problem here use specular reflection, blurriness, and
chromatic moment feature vectors. Fig 1 shows the system
architecture of this proposed system.
Table -1: Different Colour Space Representation of
original, photo and video Image.
IMAGE
TYPE
COLOUR SPACE
RGB HSV YCbCr
ORIGINAL
PHOTO
VIDEO
3.1 Viola Jones Face Detection Algorithm
The main goal of this algorithm to determine whether there
are any face present or not. The face detection depend some
factors such as illumination, location, view point etc. This
algorithm consist three ideas 1) image integration, 2)
adaBoost learning 3) Cascade classifier.
Steps:
1: compute the integral image from the inputted image. The
integral image also known as summed area table. The
integral image pixel(x, y) is equal to the sum of the pixels
above and to the left of (x, y).

2: Then extract two types of Haar-like features: the vertical
feature and the horizontal feature from integral image.
3: Give feature set and a training set of positive and negative
images, any number of machinelearningapproachestolearn
a classification function. Here use AdaBoost learning
function which consist a weighted sum of many weak
classifiers, where each weak classifier is a threshold on a
single Haarlike rectangular feature.
4: A strong classifier from AdaBoost contained stages
composed into Cascade classifier. The role of each stage isto
find whether a given sub-window is not a face or a face. In
first stage strong classifier used to classify training samples
and calculate number of false positiveandfalsenegative.The
next stage will train a strong classifier using the samples
which are classified as positive by the first stage. Then use
the strong classifier to classify the remaining samples and
calculate the number of false positive and false negative for
this stage. Repeat the stages until one stage get zero false
positive and false negative. All threshold value of strong
classifiers from each stage are saved to form the final
Cascade classifier.
3.2 Feature Extraction Techniques
1) Local Binary Pattern (LBP)
LBP is a grayscale local texture descriptor. This feature
extraction technique converts selected pixels into binary
code. The following step describes how LBP work.
Steps:
1: the inputted image divided into cells.
2: select center pixels from each cell.
3: compare the selected pixels with neighbors of selected
pixel.
4: Then replace the neighbor pixel with 0 if center pixel is
greater than neighbor pixel or 1 if center pixel is less than
neighbor pixel.
5: collect all replaced neighbor pixel value and convert that
binary value into decimal. that represent the center pixel
which selected from the cell.
2) Co-occurrence of Adjacent Local Binary Patterns
(CoALBP)
CoALBP is a local descriptor which helps to exploit spatial
information from adjacent LBP in four different directions
such as {upper left, upper right},{lowerleft,lowerright},{left
top, left bottom},{right top, right bottom}.
3) Local Phase Quantization (LPQ)
This feature descriptor deal with the blurred images. It use
Short Term Fourier Transform (STFT) to extract the local
phase information from a targeted pixel x.
4) Binarized Statistical Image Features (BSIF)
This feature extraction technique usedtofindbinarypattern
for each pixel in an image by using filter.Thenumberoffilter
depend the length of the binary pattern of pixels.
5) Specular reflection
This feature helps to avoid specularreflectionandnormalize
illumination of face. Hereuse aniterativemethodtoseparate
the components of specular reflection.
6) Blurriness Features
Attackers can conceal the spoofing medium by defocusing
the camera. The camera can’t focus when face locateinshort
distance. Spoof faces have a tend to be defocused. Due to
defocus there have a chance to occur blur. The blurriness
calculated by taking the difference between the input image
and its blurred version.
7) Chromatic Moment Features
This feature extraction technique helps to detectrecaptured
images such as photo and video image. The human eye can
detect the colour variations of fake and original samples but
system fail to detect the colour variation due to the
illumination and variations of camera. For avoiding this
problem chromatic moment features used here.
Fig -1: System architecture
3.3 Classification Algorithm
Support vector machine used for classificationwhichhelpto
detect spoofed faces and original faces. SVM is a supervised
machine learning algorithm. The two main goal of SVM is, to
maximize the distance between decision boundaries and
classify all input variables correctly.Extractedfeaturevalues
and types of input sample give to SVM for training. After

training the SVM generate a classification matrix which
consist set of similar feature vector withtypesof class.When
an image inputted for testing SVM compare the feature
values of inputted image to the feature values of different
classes. If feature vector of inputted image similar with any
class of features value then SVM determine the inputted
image belonging into that types of class.
4. RESULTS AND DISCUSSION
We evaluate different types of spoof attacks based on colour
features: LBP, CoALBP, LPQ, BSIF and Distorted features
such as specular moment, blurriness, chromatic moment
There are so many face databases are publically available
NUAA, CASIA & MSU are commonly used databasesforspoof
test experiment. The efficiency result on spoof attacks
without using any colour spaces shown in table 2 and the
efficiency result on spoof attacks using any colour spaces
shown in table 3 under 13,754 training samples. In table 4
shows the efficiency result on spoof test using both colour
texture and distorted features. As per the table results, from
table 4 give better performance result compared to table 2
and 3.
Table -2: The efficiency result on the test of different types
of faces without using colour space
Types of
face
Correct
Prediction %
Wrong
prediction %
Original face 6.96 1.08
Photo face 3.4 9.42
Video face 5.6 8.3
of faces using colour space
Types of
face
Correct
prediction %
Wrong
prediction%
Photo face 5.7 3.1
Video face 4..08 6.02
of faces using colour space and distorted features
Types of
face
Correct
prediction %
Wrong
prediction%
Photo face 8.75 2.8
Video face 8.0 1.34
5. CONCLUSION
In this paper, proposed a solution for avoiding face spoof
attackers based on colour texture analysis with distortion
features. If this system use only colour texture analysis to
detect spoofed faces, the attackers can break the system by
defocusing the camera. For avoiding these problem here
applied distorted features with colour features which give
effective result in spoof detection.
REFERENCES
[1] J. Galbally and S. Marcel, “Face anti-spoofing based on
general image quality assessment,” in Proc. IAPR/IEEE
Int.Conf. on Pattern Recognition,ICPR, 2014, pp. 1173–
1178.
[2] D. Wen, H. Han, and A. Jain, “Face spoof detection with
image distortion analysis,” TransactionsonInformation
Forensics and Security,vol. 10, no. 4, pp. 746–761,2015.
[3] J. Komulainen, A. Anjos, A. Hadid, S. Marcel, and M.
Pietikäinen,“Complementary countermeasures for
detecting scenic face spoofing attacks,” in IAPR
International Conference on Biometrics, 2013.
[4] J. Komulainen, A. Hadid, and M. Pietikäinen, “Context
based face antispoofing.”in Proc. International
Conference on Biometrics: Theory, Applications and
Systems (BTAS 2013), 2013.
[5] J. Li, Y. Wang, T. Tan, and A. K. Jain, “Live face detection
based on the analysis of fourier spectra,” in Biometric
Technology for Human Identification, 2004, pp. 296–
303.
[6] S. Tirunagari, N. Poh, D. Windridge, A. Iorliam, N. Suki,
and A. T. S.Ho, “Detection of face spoofing using visual
dynamics,” IEEE Transactions on InformationForensics
and Security, vol. 10, no. 4, pp.M. Young, The Technical
Writer’s Handbook, Mill Valley, CA: University Science,
1989.
[7] S. Bharadwaj, T. I. Dhamecha, M. Vatsa, and S. Richa,
“Computationally efficient face spoofing detection with
motion magnification,” in Proceedings of IEEE
Conference on Computer Vision and Pattern
Recognition, Workshop on Biometrics, 2013.
[8] J. Komulainen, A. Hadid, and M. Pietikäinen, “Face
spoofing detection from single images usingtextureand
local shape analysis”, Biometrics, IET, vol. 1, no. 1, pp. 3
10,March 2012.

IRJET- Face Spoof Detection using Machine Learning with Colour Features

More Related Content

What's hot (20)

Similar to IRJET- Face Spoof Detection using Machine Learning with Colour Features (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Face Spoof Detection using Machine Learning with Colour Features