Content uploaded by Jialiang Peng
Author content
All content in this area was uploaded by Jialiang Peng on Feb 17, 2021
Content may be subject to copyright.
A novel Binarization Scheme for Real-valued Biometric Feature
Jialiang Peng†‡ and Bian Yang†∗
†Department of Information Security and Communication Technology,
Norwegian University of Science and Technology, 2815, Gjovik, Norway
‡Information and Network Center, Heilongjiang University, 150080, Harbin, China
∗Corresponding author: bian.yang@ntnu.no
Abstract—Biometric binarization is the feature-type transfor-
mation that converts a specific feature representation into
a binary representation. It is a fundamental issue to trans-
form the real-valued feature vectors to the binary vectors
in biometric template protection schemes. The transformed
binary vectors should be high for both discriminability and
privacy protection when they are employed as the input data
for biometric cryptosystems. In this paper, we propose a
novel binarization scheme based on random projection and
random Support Vector Machine (SVM) to further enhance the
security and privacy of biometric binary vectors. The proposed
scheme can generate a binary vector of any given length as
an ideal input for biometric cryptosystems. In addition, the
proposed scheme is independent of the biometric feature data
distribution. Several comparative experiments are conducted
on multiple biometric databases to show the feasibility and
efficiency of the proposed scheme.
1. Introduction
With the increasing demand for personal authentication
in current society, biometric recognition technologies have
been widely applied in the public life. Due to the permanent
correlation between individuals and their biometric traits, the
secure use of biometric data is fundamental in biometric
systems. The leakage of biometric template information
constitutes a serious threatening to personal privacy [1],
because human biometrics is irrevocable and irreplaceable
if it is compromised. Therefore, biometric template protec-
tion (BTP) is essential for both the integrity of biometric
systems and the privacy of users [2]. There are various
research efforts made on BTP in the last decade [3]. These
BTP applications such as biometric cryptosystems [4] can
perform biometric recognition algorithms while maintaining
the privacy of biometric data. In fact, the binary feature
representation is always required to integrate cryptographic
technology with biometrics. Besides being applied in BTP
approaches, the binary feature representation has other two
advantages in practice: the compact storage of biometric
template and the rapid speed of biometric matching [5].
The biometric cryptosystem (BC) is also referred as
a secure sketch with the auxiliary data (AD). With the
help of AD, a secure sketch is derived by error correction
coding techniques. In BC, the secure sketch, e.g., Fuzzy
Commitment Scheme (FCS) [4], is obtained by binding
the binary template with an Error Correction Code (ECC)
generated from a cryptographic key. A cryptographic hash
of the key is stored as pseudonymous identifier (PI) into
the cryptosystem database instead of the original template.
When the query biometric sample sufficiently closes to the
reference template, the same key can be extracted or released
from the reference template with the help of ECC. It is
computationally hard for an adversary to reconstruct the
template from the secure sketch without the genuine user’s
biometric data. As a well-known biometric cryptosystem,
FCS has been applied to secure biometric feature templates
represented in the form of binary vectors [4]. Despite that
some extracted feature representation of iris and palm print
can be in binary representation, other typical biometric
traits (such as face, fingerprint, vein, voice) are always in
real-valued feature representation [6]. Moreover, the real-
valued biometric feature of a user cannot be used directly
as an input to FCS. Thus, a promising solution is to apply
a feature-type transformation approach to obtain an exact
binary input for FCS. In this paper, we work on the bio-
metric binarization scheme that transforms feature vectors
from real-value to binary-value representation, because the
real-valued representation is still the most popular in the
current biometric feature extraction techniques [6]. This
implies that a binarization scheme is needed to effectively
transform real-valued vectors into binary vectors for BTP
approaches without degrading the biometric recognition per-
formance. Therefore, the objective of biometric binarization
is to preserve the original biometric recognition performance
without compromising the security and privacy of users. In
this paper, we are motivated to generate desirable the binary
vectors by means of a random binarization procedure.
The rest of this paper is organized as follows: Section 2
introduces the related work with respect to biometric bina-
rization methods. Section 3 describes the proposed frame-
works of binarization. Experimental results are presented in
Section 4. Finally, Section 5 concludes this paper.
2. Related works
Many research works has been devoted to biometric
binarization schemes over the past decade. In general, the
2017 IEEE 41st Annual Computer Software and Applications Conference
0730-3157/17 $31.00 © 2017 IEEE
DOI 10.1109/COMPSAC.2017.26
724
existing binarization methods can be mainly classified into
two categories [7]: static binarization and dynamic binariza-
tion.
Static binarization: There are three types of methods
for static binarization, i.e., threshold-based, equal-width,
and equal-probable binarization [8]. The threshold-based
binarization methods [9] [10] map each original feature
component into two intervals labeled by ‘0’ or ‘1’ base
on a given threshold. In practice, however, it is difficult
to select the best threshold to achieve optimal performance
for the threshold-based methods. Equal-width binarization
approach [11] maps each feature element to a fixed length
of bits through creating equal partitions in a feature space.
However, the equal-width technique is very sensitive to the
defined width of interval. With respect to privacy, it may be
leaked the range of original features base on the equal-width
method. Compared with the equal-width and threshold-
based methods, the equal-probable binarization method [12]
can offer maximum randomness for the output binary string
because the output binary values are equal-probable. Never-
theless, the probability density function of original features
needs to be stored in the binarization process as the helper
data that may cause privacy threat to users [13].
Dynamic binarization: Since the discriminative compo-
nents of original feature show the relative importance during
the feature matching, dynamic binarization approaches [14]
[15] [16] have been proposed according to the discriminabil-
ity of feature components. In fact, dynamic binarization can
be also regarded as an optimization technique on original
feature space to minimize the false match rate (FMR) and/or
false nonmatch rate (FNMR). Similarly, the dynamic method
[14] is devoted to optimize minimal area under the FNMR
curve for the hamming distance classifier. Given the prob-
abilities of bit errors, the method in [14] dynamically allo-
cates bits to original feature components under FNMR curve
measurement. The detection rate optimized bit allocation
(DROBA) [15] is another representative dynamic method
based on the probability mass in the intra-class interval.
DROBA can allocate more bits to more discriminative fea-
ture components for a higher recognition performance. The
reliability-dependent bit allocation method (RDBA) [16] is
introduced to ensure the binary output string to be as stable
as possible. RDBA uses the reliability of training samples
as the interval-merging measure based on how stable each
bit of the transformed string is. It is expected to be offered
maximum randomness to the out binary strings when equal-
probable quantization is adopted in DROBA or RDBA.
However, there still exists a risk in dynamic binarization
that the helped data would reveal additional information
regarding the importance of each component in feature space
[11].
Although the methods proposed in the literatures [15]
[16] can improve the recognition accuracy of the trans-
formed binary features, these methods have to exploit the
intra-class user feature data distribution or the user-specific
dependencies [8]. Therefore, there is a potential defect that
the exposure of the probability density or mass on intra-
class feature data could be a clue to the adversary. Unlike
the existing methods, the proposed method aims to obtain
an output binary vector of any given length based on ran-
dom projection and random SVM classifier to improve the
security and privacy of biometric binarization, rather than
depending on the intrinsic feature data distribution.
On the other hand, it is difficult to achieve well non-
linkability in BC due to the special properties that any
linear combination of error correction codewords is also
a codeword [2]. To overcome this deficiency of BC, one
way can be applied by projecting the original feature vector
onto a sequence of random subspace before generating
binary input data for a BC. Random projection (RP) is a
promising tool to project the original data vector onto a
random subspace using a projection matrix [17]. Meanwhile,
RP can preserve distances well, namely, the original features
intra-class variation is preserved while exactly amplifying
the projected inter-class variations via projection onto uncor-
related random subspaces [18]. Therefore, RP is a desirable
method for generating unlinkable biometric templates, due
to the fact that original feature vectors can be projected to
random vectors with different projection matrices [19].
In general, the contribution in this paper can be described
as follows:
•The proposed binarization scheme can preserve the
discrimination of original feature components as
much as possible.
•Random projection in the proposed binarization
scheme offers a promising privacy protection for
user biometrics.
•A random SVM classifier is applied to obtain the
binary vectors and improve the randomness of the
binary vectors for biometric cryptosystems.
3. The Proposed binarization scheme
The overview of the proposed binarization scheme is
shown in Fig. 1. The real-valued feature vector of a user’s
biometric sample are first extracted via the feature extrac-
tion step. Then, a random projection approach is applied
to project the extracted real-valued feature vector with a
random matrix related to the user itself. Meanwhile, a SVM
classifier trained in advance by both a random input real-
valued vector and a random output label vector is introduced
to generate binary vectors. This trained SVM classifier can
output a different binary string depending on the input real-
valued feature vector. The user-specific random matrix and
the trained SVM classifier are stored as the helper data
in the database (see Fig. 1). Finally, the output binary
string can be input to a BC, i.e., FC, to generate the
secure biometric template. Thus, the combination of random
projection and SVM classification in the proposed scheme
aims to generate a binary vector with the given length for
different BC applications. Note that we have not made any
assumptions on the specific biometric trait being used in
the proposed scheme. It could be any biometric trait as
long as the extracted feature vectors are in the real-valued
representation that can be matched in Euclidean distance.
725
ŝŽŵĞƚƌŝĐ
ĐůŝĞŶƚ
&ĞĂƚƵƌĞ
ĞdžƚƌĂĐƚŝŽŶ
ŝŽŵĞƚƌŝĐ
ĐƌLJƉƚŽƐLJƐƚĞŵ
;&ƵnjnjLJĐŽŵŵŝƚŵĞŶƚͿ
ŝŶĂƌŝnjĂƚŝŽŶ
ĂƚĂďĂƐĞ
ŝŶĂƌLJǀĞĐƚŽƌ
ZĞĂůͲǀĂůƵĞĚ
ĨĞĂƚƵƌĞǀĞĐƚŽƌ
ZĂŶĚŽŵ
ƉƌŽũĞĐƚŝŽŶ
ZĂŶĚŽŵ^sD
ĐůĂƐƐŝĨŝĞƌ
,ĞůƉĞƌĚĂƚĂ
ŝŽŵĞƚƌŝĐ
ĚĂƚĂ
Figure 1. The framework of the proposed biometric binarization scheme.
Compared with the existing methods, the main advantages of
the proposed binarization scheme are not only independent
of the distribution estimation of biometric feature data, but
also do not rely on the selection of reliable components from
biometric original feature vectors. The binarization parts in
Fig. 1 will be discussed in the following subsections.
3.1. Random projection
Random projection (RP) provides good distinguishabil-
ity by projecting a biometric feature vector onto a set of
orthonormal random vectors. For a n-dimensional biometric
feature vector F=(f1,f
2,··· ,f
n)T∈Rn,fi∈R(i=
1,2,··· ,n), an Gram-schmidt orthonormal random matrix
R=(r1,r
2,··· ,r
m)∈Rn×m,ri∈Rn(i=1,2,··· ,m)
isan-dimensional random vector and each element rij (i=
1,2,··· ,m, j =1,2,··· ,n)in Rfollowing an indepen-
dent and identically distributed (i.i.d) Gaussian distribution
rij ∼N(0,1). The random projection result is a m-
dimensional real-valued vector, shown as Eq. (1).
V=RTF(1)
Theoretical analysis of the random projection technology
has been presented in [17]. The RP theory well addresses
the distance-preserving problem in real-valued domain us-
ing the Johnson-Lindenstrauss Lemma (JL Lemma) [20].
According to JL Lemma, the pairwise distance between
the two projected vectors of the same user who holds the
same projection matrix Ris preserved approximately in the
random subspace.
The mathematical analysis of RP-based methods for
security and privacy-preserving biometrics has been intro-
duced in [21]. The similarity and privacy-preserving proper-
ties of RP are applied to enhance the security of the gener-
ated biometric templates. Therefore, the template’s security
can be guaranteed by keeping secret the projection matrix,
thus it is desirable for generating unlinkable templates with
different user-specific projection matrices. Note that the
introduction of RP in this paper can not only guarantee the
security of the original feature vectors but also generate the
projected vectors with the given dimensionality of the binary
feature vectors for biometric cryptosystems.
3.2. Random SVM classifier
Support Vector Machine (SVM) is a machine learning
technique widely used to obtain a supervised learning model
for the classification problem in one of two classes. SVM
is applied to classify two class problems by determining
the optimal decision hyperplane. Through applying the k-
ernel trick, the SVM classifier implicitly maps data into
a high-dimensional space and finds an optimal separating
hyperplane that maximizes the margin between two classes
of data in the kernel induced feature space. After solving
the SVM problem in its dual form, the support vectors are
identified and SVM separates the training data vector in
a high-dimensional space. Some of these support vectors
belong to the positive class while others belong to the
negative class. Using these support vectors, the decision
function or the classification result for any testing vector
xcan be expressed as
y(x)=sign #SV
k=1
akykK(xk,x)+z(2)
where xk∈RNdenotes a support vector and yk∈{−1,1},
#SV stands for the number of support vectors, and z∈Ris
the bias that can be calculated from KKT-complementarity
conditions with respect to the optimization problem. Here,
the function Kin Eq. (2) is known as kernel function. To
facilitate the proposed scheme, we reformulate Eq. (2) into
the following expression:
y(x)=⎧
⎪
⎪
⎨
⎪
⎪
⎩
1,if
#SV
k=1
akykK(xk,x)+z>0
0,if
#SV
k=1
akykK(xk,x)+z≤0
(3)
A general view of the proposed scheme using SVM
is shown in Fig. 2. The main role of the SVM classifier
in this work is to map a RP projected vector to a bina-
ry output vector. In the first stage, a random real-valued
vector X=(x1,x
2,...,x
m)T,xi∈Rand a random
class label vector Y=(y1,y
2,...,y
m)T,yi∈{−1,1}
(i=1,2,··· ,m)are organized as the training data for
obtaining the SVM classifier S. The training set of pairs
is defined as {xi,y
i}(i=1,2,··· ,m), where xiis an
element of the random vector Xfollowing an i.i.d Gaus-
sian distribution N(0,1) and yiis a random class label
following a uniform distribution. Moreover, xiis labeled
corresponding to the random label yifor training SVM
classifier S(see Fig. 2). Obviously, the classification process
depends strictly on the SVM classifier S, and it means
that a binary vector corresponding to a real-valued feature
vector is unknown until the SVM classifier Sis defined.
Therefore, the training vectors Xand Ycan be discarded to
preserve the security and user privacy after training the SVM
classifier Sin practical implementations. It is important to
notice that the SVM classifier Sis trained by the random
feature vectors Xand Ythat are independent of the genuine
biometric feature vectors. In the second stage, an original
feature vector F=(f1,f
2,...,f
n)T,fi∈R(1 ≤i≤n)
is extracted from a biometric sample, and then the projected
feature vector Vis obtained by V=RTFwhere Ris a
random projection matrix. Finally, the classifier Sis used
to output the binary vector Baccording to the input vector
726
^sD
ůĂƐƐŝĞƌ
ŝŽŵĞƚƌŝĐ
ĨĞĂƚƵƌĞǀĞĐƚŽƌ
ZĂŶĚŽŵ
ƉƌŽũĞĐƚŝŽŶ
ŵĂƚƌŝdž
&ĞĂƚƵƌĞ
džƚƌĂĐƚŝŽŶ
ŝŽŵĞƚƌŝĐĂƚĂ
WƌŽũĞĐƚĞĚ
ĨĞĂƚƵƌĞ
sĞĐƚŽƌ
ŝŶĂƌLJ
sĞĐƚŽƌ
/ŶƉƵƚ KƵƚƉƵƚ
7
Q
)II I !
QP
5
u
7
95)
6
%
^sD
ůĂƐƐŝĞƌ
6
ZĂŶĚŽŵƌĞĂůͲ
ǀĂůƵĞĚǀĞĐƚŽƌ
ZĂŶĚŽŵďŝŶĂƌLJ
ĐůĂƐƐůĂďĞůǀĞĐƚŽƌ
7
P
;
[[ [ !
7
P
<\\ \ !
dƌĂŝŶŝŶŐ^sDĐůĂƐƐŝĞƌ
6
ŝŽŵĞƚƌŝĐĨĞĂƚƵƌĞďŝŶĂƌŝnjĂƚŝŽŶƵƐŝŶŐ^sD
ķ
ĸ
ķ&ŝƌƐƚƐƚĂŐĞƚƌĂŝŶŝŶŐĂ^sDĐůĂƐƐŝĞƌ^
ďĂƐĞĚŽŶďŽƚŚƌĂŶĚŽŵŝŶƉƵƚǀĞĐƚŽƌ yĂŶĚ
ƌĂŶĚŽŵďŝŶĂƌLJĐůĂƐƐůĂďĞůǀĞĐƚŽƌz
ĸ^ĞĐŽŶĚƐƚĂŐĞƚŚĞǁŽƌŬĨůŽǁŽĨďŝŽŵĞƚƌŝĐ
ĨĞĂƚƵƌĞďŝŶĂƌŝnjĂƚŝŽŶƵƐŝŶŐ^sDĐůĂƐƐŝĞƌ^
9
'
Figure 2. The general block diagram of biometric binarization using SVM.
V. We further describe the whole process of the proposed
binarization scheme in Algorithm 1.
Algorithm 1 The proposed binarization scheme based on
SVM.
Input:
The biometric data D, the random vectors Xand Y.
Output:
The binary vector B.
1: Client collects biometric data Dfrom the user’s sample;
2: Feature vector Fis computed from Dvia feature ex-
traction;
3: Generate the random projection matrix Rn×mand com-
pute the projected vector V=RTF;
4: Train the SVM classifier Swith Xand Y;
5: Output the binary vector Busing SVM classier Swith
the vector V.
4. Experimental results
4.1. Experiment Setup
To evaluate effectiveness of the proposed scheme, we
conduct the experiments on several biometric datasets, in-
cluding face image dataset [22], fingerprint image dataset
[23], fingervein image dataset [24], and finger-knuckle-print
image dataset [25]. For the face image dataset, it is the
total Faces94 database [22], containing 3060 images from
153 subjects and 20 samples per-subject. A commercial-off-
the-shelf (COTS) software [26] is applied to extract face
feature as a 256 dimensional real-valued vector. For the
fingerprint image dataset, it consists of 67 subjects selected
with high-quality images from FVC2002 Db1A database
[23], and six selected fingerprint impressions per-subject.
In the experiments, a bank of Gabor filters-based algorithm
(FingerCode) [27] is utilized to extract fingerprint feature
as a 128-dimensional real-valued vector. For the fingervein
image dataset, it stems from Hong Kong Polytechnic U-
niversity Finger Image Database [24]. There are 100 sub-
jects selected with high-quality fingervein images, and six
images per-subject are employed in the experiments. The
fingervein features are extracted by Gabor wavelet and local
binary pattern (GLBP) [28]. All the GLBP feature vectors
conducted in the experiments are 128-dimensional real-
valued vectors. For the finger-knuckle-print image dataset,
it consists of 100 subjects selected with high-quality images
from PolyU Finger-Knuckle-Print database [25], and six
knuckles images per-subject. The finger-knuckle-print (FKP)
features are extracted by the log-Gabor phase congruency
(PC) model [29]. All the finger-knuckle-print feature vectors
conducted in the experiments are also 128-dimensional real-
valued vectors. To lower the impact of different biometric
feature data distribution, all the extracted feature vectors
from the above four biometric modal traits are normalized
into the same data distribution (zero mean and unit vari-
ance).
To simulate a real application, there is no overlap be-
tween the training and the testing subjects for the per-
formance evaluation in our experiments. The evaluation is
performed in such a manner that each single image is used
as a template once and the rest of the images as the probe
set. In the experiments, the verification decision is made
upon Hamming distance (HD, the number of bit differences)
between the query binary vector and the reference template
vector. The evaluation protocols of false match rate (FMR),
false nonmatch rate (FNMR), and Receiver Operator Char-
acteristic (ROC) curve are used in this paper. The Equal
Error Rate (EER), which is the point where FMR is equal
to FNMR, is used to evaluate the verification performance in
our experiments. The lower value of EER shows the better
verification performance. To facilitate the performance eval-
uation, EER obtained by the original feature presentation is
called as the Baseline-1 performance in the experiments.
Admittedly, the binary vector can also be obtained by the
original feature vector and a random SVM classifier without
applying the RP procedure. It means that a normalized
original feature vector can be directly input to a random
SVM classifier for obtaining a binary vector. The random
SVM classifier should certainly be trained by the random
vectors X and Y that have the same dimensionality as the
original feature vector. We call the verification performance
of the above binarization without the RP procedure as the
Baseline-2 performance in the experiments. Note that both
the RP matrix Rand the SVM classifier Sare different
for each subject during the experimental evaluation. To
minimize the effect of randomness, all the experiments are
performed five times and the results are averaged to reduce
the statistical fluctuation.
4.2. Performance evaluation
A random SVM classifier is used for binary classification
in this work. After the random projection, the projected
feature vector is binarized by a random SVM classifier. For
any one of input projected feature vectors, the random SVM
classifier outputs a binary vector. Obviously, the dimension-
ality of an output binary vector is equal to that of an input
projected feature vector. Hence, mis also referred to the
727
í í í
)05
í)105
%DVHOLQHí
%DVHOLQHí
)DFHí
)DFHí
)DFHí
)DFHí
)DFHí
)DFHí
)DFHí
(a)
í í í í í
)05
í)105
%DVHOLQHí
%DVHOLQHí
)LQJHUYHLQí
)LQJHUYHLQí
)LQJHUYHLQí
)LQJHUYHLQí
)LQJHUYHLQí
)LQJHUYHLQí
)LQJHUYHLQí
(b)
í í í í
)05
í)105
%DVHOLQHí
%DVHOLQHí
)LQJHUSULQWí
)LQJHUSULQWí
)LQJHUSULQWí
)LQJHUSULQWí
)LQJHUSULQWí
)LQJHUSULQWí
)LQJHUSULQWí
(c)
í í í í í
)05
í)105
%DVHOLQHí
%DVHOLQHí
).3í
).3í
).3í
).3í
).3í
).3í
).3í
(d)
Figure 3. ROC curves of the binary vectors generated by the proposed
scheme for: (a) Face (b) Fingervein (c) Fingerprint (d) Finger-knuckle-
print (the expression α−βin the legend denotes the biometric trait αand
the corresponding binarization dimensionality β).
dimensionality of the binary feature vector in the subsequent
parts of the performance evaluation
As to the implementation of SVM, the kernel function
of SVM should be defined firstly. Although the multiple
kernel functions have been put forward for SVM classifiers,
there has not been a theoretical instruction but choose the
best kernel function by trial due to the various feature data
for classification. We find that the Multilayer Perceptron
kernel function K(xi,x
j)=tanh
γxT
ixj+ris in favor
of the proposed SVM-based scheme, where the default
scales γ=1and r=−1are set by the SVM training
function named svmtrain in Matlab. In addition, the bias
zof Eq. 3 in the kernel function mainly determines the
verification performance of the binary vectors generated
by a SVM classifier. When a SVM kernel function with
|z|≤0.5is generated by iterative random trials, it can obtain
the optimal verification performance for the binary vectors.
This is because that each element of all original feature
vectors is normalized to a zero mean and unit variance data
distribution prior to being input to a SVM classifier. The
smaller |z|is randomly generated by Xand Y, the more
robust binarization can be obtained by a SVM classifier.
We compare the verification performances when the
different dimensionality mis applied in the experiments.
For clear comparison, Table. 1 shows that the obtained
EER with the different dimensionality mby the proposed
scheme. Fig. 3 also shows the ROC curves of the binary
feature vectors with the different dimensionalities. As can
be seen from both Table. 1 and Fig.3, the EER becomes
smaller and the consequent performance is improved in
general with the increasing dimensionality of the binary
vector. It is evident that the proposed scheme constantly
outperforms both Baseline-1 and Baseline-2 methods when
the dimensionality of the binary feature vector is not less
1RUPDOL]HG+DPPLQJ'LVWDQFH
3HUFHQW
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
(a)
1RUPDOL]HG+DPPLQJ'LVWDQFH
3HUFHQW
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
(b)
1RUPDOL]HG+DPPLQJ'LVWDQFH
3HUFHQW
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
(c)
1RUPDOL]HG+DPPLQJ'LVWDQFH
3HUFHQW
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
,QWUDíXVHUP
,QWHUíXVHUP
(d)
Figure 4. The intra-user/inter-user normalized HD distribution for: (a) Face
(b) Fingervein (c) Fingerprint (d) Finger-knuckle-print.
than that of the original feature vector in Fig. 3. It also
implies that the large mensures a clear separation between
the intra-user and inter-user HD distributions. As illustrated
in Fig. 4, the normalized HD distributions between the
intra-user and inter-user tend to separate better as the di-
mensionality mincreases. It can be seen that the intra-
class and inter-class distributions can be better separated
as mincreases, and hence, the lower EER and the better
the verification performance are obtained. The all inter-user
normalized HD distributions are shifted to the right and
centered at 0.5, indicating a high level of randomization
in these distribution, as shown in Fig. 4. On the other
hand, the intra-user distribution is shrunken when mis
large. This is accomplished through the random projection
effect of preserving intra-user variations while amplifying
inter-user variations via mapping onto uncorrelated random
binary sequences with the help of SVM classification. The
clear separation indicates that the proposed scheme results
in dramatically reduced error rates when mis not less than
the dimensionality of the original feature vector. It means
that the proposed scheme not only enables the binarization
of the biometric real-valued feature vector, but also achieves
higher verification performance.
5. Conclusion
In this paper, a novel scheme for biometric binarization
is proposed to improve the randomness of the binary vectors
based on random projection and random SVM classification.
The main advantage of our approach is independent of the
intrinsic feature data distribution. Compared to the existing
binarization approaches, the proposed scheme can generate
a binary vector of any given length as an ideal input for
biometric cryptosystems without the discriminative feature
selection. By adopting the appropriate dimensionality of the
728
TABLE 1. THE EER PERFORMANCES OF THE BINARY FEATURE VECTORS WITH THE DIFFERENT DIMENSIONALITY m.
Baseline-1 Baseline-2 m=32 m=64 m= 128 m= 256 m= 512 m= 1024 m= 2048
Face 0.00122 0.01772 0.03119 0.01141 0.00076 0.00016 0 0 0
Fingervein 0.00434 0.18170 0.06179 0.01971 0.00389 0.00290 0.00022 0 0
Fingerprint 0.02179 0.04924 0.02302 0.00710 0.00458 0.00432 0.00348 0.00134 0.00093
FKP 0.01955 0.25422 0.07132 0.03392 0.01607 0.01078 0.00306 0.00297 0.00213
binary vector, the verification performance of the proposed
scheme is superior to the original baseline performance.
Therefore, the proposed scheme achieves better biometric
recognition performance regarding binarization security.
References
[1] A. Jain, K. Nandakumar, and A. Nagar, “Biometric template security,”
EURASIP Journal on Advances in Signal Processing, vol. 2008, no. 1,
p. 579416, 2008.
[2] K. Nandakumar and A. K. Jain, “Biometric template protection:
Bridging the performance gap between theory and practice,” IEEE
Signal Processing Magazine, vol. 32, no. 5, pp. 88–100, 2015.
[3] S. Rane, Y. Wang, S. C. Draper, and P. Ishwar, “Secure biometrics:
Concepts, authentication architectures, and challenges,” IEEE Signal
Processing Magazine, vol. 30, no. 5, pp. 51–64, 2013.
[4] A. Juels and M. Wattenberg, “A fuzzy commitment scheme,” in
Proceedings of the 6th ACM Conference on Computer and Commu-
nications Security, ser. CCS ’99, 1999, pp. 28–36.
[5] C. Rathgeb, F. Breitinger, C. Busch, and H. Baier, “On application
of bloom filters to iris biometrics,” IET Biometrics, vol. 3, no. 4, pp.
207–218, 2014.
[6] S. Z. Li, Encyclopedia of Biometrics. NJ, USA: Springer, 2009.
[7] H. Lee, A. B. J. Teoh, H. G. Jung, and J. Kim, “A secure biometric
discretization scheme for face template protection,” Future Genera-
tion Computer Systems, vol. 28, no. 1, pp. 218 – 231, 2012.
[8] M. H. Lim, A. B. J. Teoh, and K. A. Toh, “An analysis on equal
width quantization and linearly separable subcode encoding-based
discretization and its performance resemblances,” EURASIP Journal
on Advances in Signal Processing, vol. 2011, no. 1, p. 82, 2011.
[9] A. B. J. Teoh, D. C. L. Ngo, and A. Goh, “Personalised cryptographic
key generation based on facehashing,” Comput. Secur., vol. 23, no. 7.
[10] E. Verbitskiy, P. Tuyls, D. Denteneer, and J. P. Linnartz, “Reliable
biometric authentication with privacy protection,” in 24th Benelux
Symp. on Info. Theory, 2003, pp. 125–132.
[11] M. Lim, A. B. J. Teoh, and J. Kim, “Biometric feature-type trans-
formation: Making templates compatible for secret protection,” IEEE
Signal Processing Magazine, vol. 32, no. 5, pp. 77–87, 2015.
[12] E. J. C. Kelkboom, G. G. Molina, J. Breebaart, R. N. J. Veldhuis,
T. A. M. Kevenaar, and W. Jonker, “Binary biometrics: An analytic
framework to estimate the performance curves under gaussian as-
sumption,” IEEE Transactions on Systems, Man, and Cybernetics -
Part A: Systems and Humans, vol. 40, no. 3, pp. 555–571, 2010.
[13] J. Galbally, J. Fierrez, J. Ortega-Garcia, C. McCool, and S. Marcel,
“Hill-climbing attack to an eigenface-based face verification system,”
in 2009 First IEEE International Conference on Biometrics, Identity
and Security (BIdS), 2009, pp. 1–6.
[14] C. Chen and R. Veldhuis, “Extracting biometric binary strings with
minimal area under the frr curve for the hamming distance classifier,”
in 2009 17th European Signal Processing Conference, Aug 2009, pp.
50–54.
[15] M. H. Lim, A. B. J. Teoh, and K. A. Toh, “Dynamic detection-
rate-based bit allocation with genuine interval concealment for binary
biometric representation,” IEEE Transactions on Cybernetics, vol. 43,
no. 3, pp. 843–857, 2013.
[16] M.-H. Lim, A. B. J. Teoh, and K.-A. Toh, “An efficient dynamic
reliability-dependent bit allocation for biometric discretization,” Pat-
tern Recognition, vol. 45, no. 5, pp. 1960 – 1971, 2012.
[17] E. Bingham and H. Mannila, “Random projection in dimensionality
reduction: Applications to image and text data,” in Proceedings of
the Seventh ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, ser. KDD ’01. New York, NY, USA:
ACM, 2001, pp. 245–250.
[18] A. T. B. Jin, “Cancellable biometrics and multispace random projec-
tions,” in 2006 Conference on Computer Vision and Pattern Recog-
nition Workshop (CVPRW’06), June 2006, pp. 164–164.
[19] B. Yang, D. Hartung, K. Simoens, and C. Busch, “Dynamic random
projection for biometric template protection,” in 2010 Fourth IEEE
International Conference on Biometrics: Theory, Applications and
Systems (BTAS), Sept 2010, pp. 1–7.
[20] S. Kaski, “Dimensionality reduction by random mapping: fast simi-
larity computation for clustering,” in 1998 IEEE International Joint
Conference on Neural Networks Proceedings. IEEE World Congress
on Computational Intelligence, vol. 1, May 1998, pp. 413–418 vol.1.
[21] Y. Wang and K. N. Plataniotis, “An analysis of random projection
for changeable and privacy-preserving biometric verification,” IEEE
Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics),
vol. 40, no. 5, pp. 1280–1293, 2010.
[22] “Collection of facial images: Faces94,” 2017. [Online]. Available:
http://cswww.essex.ac.uk/mv/allfaces/faces94.html
[23] A. K. Jain and D. Maltoni, Handbook of Fingerprint Recognition.
Secaucus, NJ, USA: Springer-Verlag New York, Inc., 2003.
[24] “The hong kong polytechnic university finger image database
(version 1.0),” 2013. [Online]. Available: http://www4.comp.polyu.
edu.hk/∼csajaykr/fvdatabase.htm
[25] “Polyu finger-knuckle-printdatabase,” 2010. [Online]. Available:
http://www.comp.polyu.edu.hk/∼biometrics/FKP.htm
[26] “Viulib,” 2017. [Online]. Available: http://www.viulib.org/index.html
[27] A. K. Jain, S. Prabhakar, L. Hong, and S. Pankanti, “Filterbank-based
fingerprint matching,” Trans. Img. Proc.
[28] J. PENG, Q. LI, A. A. A. EL-LATIF, N. WANG, and X. NIU, “Finger
vein recognition with gabor wavelets and local binary patterns,”
IEICE TRANS. INF.SYST, vol. E96-D, no. 8, pp. 1886–1889, 2013.
[29] L. Zhang, L. Zhang, D. Zhang, and Z. Guo, “Phase congruency
induced local features for finger-knuckle-print recognition,” Pattern
Recogn., vol. 45, no. 7, pp. 2522–2531, 2012.
729