Content uploaded by Fathi E. Abd El-Samie
Author content
All content in this area was uploaded by Fathi E. Abd El-Samie on Dec 07, 2018
Content may be subject to copyright.
Double random phase encoding for cancelable
face and iris recognition
RANDA F. S OLIMAN,1GHADA M. ELBANBY,2ABEER D. ALGARNI,3MOHAMED ELSHEIKH,4,5
NAGLAA F. S OLIMAN,3,6 MOHAMED AMIN,1AND FATHI E. ABD EL-SAMIE7,*
1Faculty of Science, Mathematics and Computer Science Department, Menoufia University, Shebin El-Koom, 32511, Egypt
2Faculty of Electronic Engineering, Department of Industrial Electronics and Control Engineering, Menoufia University, Menouf 32952, Egypt
3Faculty of Computer and Information Sciences, Princess Nourah Bint Abdulrahman University, Riyadh, Saudi Arabia
4Electrical and Computer Engineering Department, Queen’s University, Kingston, Ontario, Canada
5Faculty of Engineering, Tanta University, Tanta, Egypt
6Faculty of Engineering, Zagazig University, Zagazig, Egypt
7Faculty of Electronic Engineering, Department of Electronics and Electrical Communications Engineering, Menoufia University,
Menouf 32952, Egypt
*Corresponding author: fathi_sayed@yahoo.com
Received 17 September 2018; revised 8 November 2018; accepted 9 November 2018; posted 9 November 2018 (Doc. ID 334851);
published 10 December 2018
Most modern security systems depend on biometrics. Unfortunately, these systems have suffered from hacking
trials. If the biometric databases have been hacked and stolen, the biometrics saved in these databases will be lost
forever. Thus, there is a desperate need to develop new cancelable biometric systems. The basic concept of can-
celable biometrics is to use another version of the original biometric template created through a one-way trans-
form or an encryption scheme to keep the original biometrics safe and away from utilization in the system. In this
paper, the optical double random phase encoding (DRPE) algorithm is utilized for cancelable face and iris rec-
ognition systems. In the proposed cancelable face recognition scheme, the scale invariant feature transform is used
for feature extraction from the face images. The extracted feature map is encrypted with the DRPE algorithm.
The proposed cancelable iris recognition system depends on the utilization of two iris images for the same person
and features are extracted from both images. The features extracted from one of the iris images are encrypted with
the DRPE algorithm, provided that the second phase mask used in the DRPE is generated from the other iris
image features. This trend guarantees some sort of feature fusion between the two iris images into a single can-
celable iris code and increases user privacy. Simulation results show good performance of the two proposed can-
celable biometric schemes even in the presence of noise, especially with the proposed cancelable face recognition
scheme. © 2018 Optical Society of America
https://doi.org/10.1364/AO.57.010305
1. INTRODUCTION
Biometrics are signals or images presented by persons for the
purpose of discrimination between them. The most common
biometrics are fingerprints, faces, iris, and speech signals. The
basic idea of operation of biometric systems is to collect bio-
metrics from authorized persons, extract discriminating features
from the biometrics as a tool for data reduction, and store these
features in a database. This is known as the training phase. In
the other phase of biometric systems, the testing phase, features
are extracted from the incoming biometrics for new persons
and matched to the features in the database. The testing phase
can be performed with or without classifiers [1].
Generally, face recognition depends on the geometric deter-
mination of eyes, nose, and mouth in the face image [2]. On
the other hand, in iris recognition, concentration is first on the
localization of the iris region in the image. After that, the effects
of eyelids and eyelashes are eliminated, as they are represented
as noise in the images. Some transformation is performed on
the iris images from polar to rectangular coordinates, and dis-
criminating features are extracted and coded from the trans-
formed patterns [3]. We can say that the features extracted
from face images are geometric features, while the features ex-
tracted from iris images are transform-domain features.
All traditional biometric systems depend on the acquisition
of a signal or an image from users or subscribers, feature extrac-
tion from these signals or images, and then feature matching
with a prearranged database [1–3]. The main disadvantage
of this traditional trend is that each person has to provide
his or her original biometrics for possible feature extraction.
These biometrics will not be changed again after the training
Research Article Vol. 57, No. 35 / 10 December 2018 / Applied Optics 10305
1559-128X/18/3510305-12 Journal © 2018 Optical Society of America
phase. The extracted features are saved in the database. This
means that, if the database is stolen, the original biometrics will
be lost forever, and hence the system will no longer be suitable
for utilization. Moreover, the users whose biometrics have been
stolen will not be able to use them again in other systems.
A new trend toward more secure biometrics is to use cancela-
ble biometrics for persons. These biometrics can be changed
easily in emergency cases without the need to change the system
at all [4]. The trend of cancelable biometrics is evolving and
promising toward more secure biometric systems. For geometric
biometrics as in face recognition, it is possible to use some one-
way geometric distortion transforms that can change the indices
in the biometric at hand. These transforms can be designed and
changed easily, while maintaining the one-way nature to avoid
inversion. In the case of biometrics that are dependent on trans-
form-domain features, as in the iris recognition case, it is possible
to use techniques such as random projection and bio-hashing
[4]. Moreover, encryption techniques can be used with all types
of biometrics [5–7].
Some attempts have been presented to generate cancelable
biometrics from fingerprints. Wang and Hu presented a blind
cancelable fingerprint recognition system based on creating
binary strings for pair-minutiae vectors, and the utilization of
their frequency samples for identification [8]. This approach
is suitable for resource-limited applications such as smart cards.
It was evaluated over FVC2002 DB1, DB2, and DB3 databases
with good results. Wang et al. presented an approach for can-
celable fingerprints based on the segmentation of the fingerprint
pattern into local zones and the utilization of a Fourier-like trans-
form to change the indices of all minutiae inside the same local
zone [9]. This approach does not require registration of the fin-
gerprints. It succeeded in recognition on four different standard
databases with equal error rates (EERs) in the lost token scenario
ranging from 0.19% to 9%. Wang et al. also presented a can-
celable fingerprint biometric system based on the Hadamard
transform implemented on the binary representation of the
fingerprint minutiae [10]. The Hadamard transform is applied
on the Fourier transform of the binary data series representing
the biometric to yield complex vectors while preserving the dis-
tance between the vectors estimated prior to the application of
the transform. This approach achieved EERs ranging from 1%
to 5% on different standard databases. Some attempts have also
been presented for fingerprint and finger-vein cancelable multi-
biometric systems [11].
For the generation of cancelable iris templates, different
algorithms have been presented in the literature due to the im-
portance and popularity of iris recognition systems [12–22].
A sector random projection (SRP) approach has been presented
in the literature [4,23]. This approach begins with the segmen-
tation of the iris image for iris localization. After that, sectori-
zation is performed on the iris region into four sectors.
A rectangular-to-polar transformation is performed on the sec-
tors, and Gabor features are extracted separately from all sec-
tors. The objective of this process is to minimize the effect of
eyelids and eyelashes on the feature extraction process. The ex-
tracted features from each sector are then subjected to random
projection, which generates the cancelable iris template after a
concatenation process. From this trend, we can deduce that the
cancelable templates are generated after the feature extraction
process. Lai et al. introduced a cancelable iris recognition ap-
proach that is based on hashing [24]. This approach depends
also on Hadamard product and a modulo threshold function. It
guarantees a high degree of accuracy in addition to the security
achieved by the cancelable approach itself. Umer et al. pre-
sented a bio-hashing approach with a spatial pyramidal feature
extraction process for cancelable iris recognition [25]. This ap-
proach depends on a user-specific token and an independent
token. It achieved good results on standard databases.
Cancelable biometric concepts have been adopted for face
recognition by applying geometric deformations on the face
images prior to feature extraction through one-to-many non-
linear functions [4]. Moreover, face encryption through convo-
lution with random kernels has been adopted for generating
cancelable face templates. This approach is called the minimum
average correlation energy filter approach. It adopts the corre-
lation coefficient as a tool for verification without automatic
classifiers [4,26]. Bio-hashing has also been considered in can-
celable face recognition. It depends on the generation of self-
customized tokens for users [4].
This paper introduces two cancelable face and iris recognition
schemes. The first cancelable face recognition scheme is based on
the generation of a secure feature matrix using scale invariant
feature transform (SIFT) followed by double random phase en-
coding (DRPE). This scheme consists of feature extraction, en-
coding, and classification stages. For the feature extraction stage,
SIFT descriptors are estimated. Then, the encryption process of
the feature matrix is performed with the DRPE algorithm, where
two encoding keys are inserted. The final stage is based on es-
timating the correlation coefficient between the encrypted fea-
ture matrix of the original face templates stored in a database
and the encrypted feature matrix of the user face template.
Based on a threshold value, the matcher can decide and classify
the incoming encrypted feature matrix as accepted or rejected.
The quantitative evaluations of the proposed cancelable face rec-
ognition scheme have been performed using EER based on false
acceptance rate and false rejected rate (FRR), area under receiver
operating characteristic (ROC) curve (AROC), and decidability.
A cancelable iris recognition scheme based on DRPE in the
fractional Fourier transform (FrFT) domain is also presented.
In this scheme, we go through a coarse-to-fine stage for iris
localization, a normalization stage, and a feature extraction
stage with a LogGabor filter. The LogGabor filter matrix is
treated with the DRPE. The DRPE encryption is performed
on the feature matrix as in the cancelable face recognition
scheme. Two encryption keys (RPM1and RPM2) are used
to enhance security. We have freedom in selecting the first
phase mask (RPM1). The other phase mask (RPM2) is taken
from the other iris image. After the DRPE stage, we work on
the modulus of the output; hence, we have real values. These
real values are interpreted as binary iris codes, which allows
working with the Hamming distance (HD) as a metric.
The rest of this paper is organized as follows. In Section 2,
the DRPE algorithm is introduced. In Section 3, the proposed
cancelable face recognition scheme based on SIFT algorithm
and DRPE algorithm is presented. Section 4is devoted for the
proposed cancelable iris recognition scheme. Section 5presents
the simulation results of both schemes. Section 6gives the con-
cluding remarks of the paper.
10306 Vol. 57, No. 35 / 10 December 2018 / Applied Optics Research Article
2. DOUBLE RANDOM PHASE ENCODING
Double random phase encoding is one of the most popular op-
tical encryption schemes. Its popularity comes from the ability
to implement it either with an optical setup using lenses or with
software. Figure 1illustrates the optical setup required for the
DRPE algorithm. The system in Fig. 1is known as the 4f
system for optical encryption. It is extended for four focal
lengths, and it requires only two lenses and two random phase
masks. In the traditional DRPE encryption, an input image is
set at the input plane, which is one focal length away from the
input lens, and the Fourier transform is obtained at one focal
length away from the lens on the other side [27]. A random
phase mask is applied at the input plane, and another random
phase mask is applied at the Fourier plane to increase the degree
of security. An inverse Fourier transform is applied in the sec-
ond 2fsection of the setup through the second lens. Thus, it is
guaranteed here that the obtained encrypted image is in the
spatial domain. The DRPE has a large degree of resistance
to attacks [28,29].
It is possible to represent the whole encryption process with
the DRPE algorithm through the following formula [27,30,31]:
ψx,yFT−1fFTffx,yφnx,yg ×φmu,vg,(1)
where FT refers to the Fourier transform, ψx,yrefers to the
encrypted image in the spatial domain, φnx,yrefers to the first
spatial-domain random phase mask (RPM1), and φmu,vrefers
to the second frequency-domain random phase mask (RPM2).
Both random phase masks are 2D matrices of the same size as
fx,yhaving values uniformly distributed between 0 and 2π.
Randomness of the generated phase masks is required to perform
efficient encryption of the images. Different possible random dis-
tributions such as uniform and Gaussian distributions can be
used in the phase masks.
With the development of Fourier analysis in two dimen-
sions, concepts such as FrFT have come into application in
the field of image encryption [32,33]. The FrFT is just a ro-
tation of the time/frequency plane of the 2D FT with certain
fractional orders. It is represented as follows [32]:
Fα,βu,vX
X−1
x0X
Y−1
y0
fx,yRα,βx,y;u,v,(2)
where Xand Yare the dimensions of the image.
The inverse FrFT is represented as follows [32]:
fx,yX
X−1
x0X
Y−1
y0
Fα,βu,vR−α,−βx,y;u,v,(3)
where Rα,βx,y,u,vis the basis function of the FrFT, and α
and βare the fractional orders in both dimensions.
The FrFT can take place of the FT in the DRPE encoding
for image encryption [34,35,36,37]. The new contribution in
this paper is to make use of the DRPE in the encryption of
feature matrices extracted from biometric images in order to
generate cancelable templates that can be used to secure bio-
metric systems. Two types of biometrics are considered, i.e.,
face and iris biometrics, and two versions of DRPE are also
considered based on FT and FrFT.
3. PROPOSED CANCELABLE FACE
RECOGNITION SCHEME
The proposed cancelable face recognition scheme begins
by extracting features from the face images using the SIFT.
This transform has been used in the literature in traditional
face recognition schemes [38–40]. The SIFT algorithm de-
pends on scale-space analysis for keypoint detection and de-
scription. It achieves high robustness to scaling, rotation,
and illumination variations, which usually occur during the face
imaging process.
This descriptor consists of multiple steps: constructing a
scale space; performing a Laplacian of Gaussian approximation;
finding keypoints; eliminating bad keypoints; determination of
the orientation of selected keypoints; and, finally, generation of
a feature vector of each dominant keypoint [41–43].
In the initial step to begin the scale-space processing, blurred
versions of the face images are generated from the original im-
ages in different octaves. The blurred faces are obtained by con-
volution with a Gaussian kernel of scale σas follows [41–43]:
Lx,y,σGx,y,σIx,y,(4)
where Lx,y,σis the blurred face at scale σ,Gx,y,σis the
Gaussian kernel of scale σ, and Ix,yis the original face image.
The Gaussian kernel is defined as follows:
Gx,y,σ 1
2πσ2e−x2y2∕2σ2:(5)
The next step is the detection of keypoints, in which two con-
secutive images in the same octave are subtracted from each
other. This process is repeated for all octaves as an approxima-
tion of the scale-invariant Laplacian of Gaussian. Figure 2
illustrates a fast and efficient implementation of the Laplacian
of Gaussian, which is known as the difference of Gaussian
(DoG) and an example of application on a face image.
After calculating the DoG, maxima and minima keypoints
are determined by comparing the neighboring pixels in the cur-
rent scale, scale above, and scale below. After this step, some
keypoints are rejected because they have low contrast or are
considered as edges. Then, the orientations of the accepted key-
points are estimated. The keypoint orientation depends on gra-
dient magnitude and direction estimated as
mx,y
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Lx1,y−Lx−1,y2Lx,y1−Lx,y−12,
q
(6)
Fig. 1. Proposed setup of the DRPE encryption.
Research Article Vol. 57, No. 35 / 10 December 2018 / Applied Optics 10307
θx,ytan−1Lx,y1−Lx,y−1
Lx1,y−Lx−1,y,(7)
where Lx,yis the gradient at x,y.
To generate the feature vector representing the keypoint,
128 distinctive values are generated from a 16 ×16 window
of gradients around the keypoint. This window is divided into
4×4windows to generate a histogram of eight bins for each
window with each bin representing 45°, as shown in Fig. 3. The
most dominant gradient orientation of a keypoint is deter-
mined by another angle histogram, in which the 360° of ori-
entations are divided into 36 bins. The main orientation of a
keypoint is determined in the 80% bandwidth around the
angle histogram peak.
In the proposed cancelable face recognition scheme, key-
points are extracted from faces; hence, feature vectors are esti-
mated for these keypoints. These feature vectors include the bin
values of the angle histograms of gradients in the neighbor-
hoods of the keypoints. The feature vectors are concatenated
to each other to form a feature matrix, which is encrypted with
the DRPE algorithm to generate a random kernel. The original
face is encrypted by convolution with the random kernel gen-
erated from the same face to increase the degree of security and
privacy of the cancelable face generation process. The ability to
change the generated face codes depends on the ability to
change the random phase masks used in the DRPE algorithm.
Figure 4illustrates the steps of the proposed cancelable face
recognition scheme.
4. PROPOSED CANCELABLE IRIS
RECOGNITION SCHEME
Similar to the proposed cancelable face recognition scheme, the
proposed cancelable iris recognition scheme illustrated in Fig. 5
begins with the feature extraction step. To increase privacy of
the users, the scheme performs some sort of merging between
the two iris images of the same person in the extraction of the
cancelable iris codes. First, for the generation of iris codes, there
is a need to go through some steps, including iris localization
and iris normalization in a coarse-to-fine algorithm. We have
adopted the coarse-to-fine algorithm in [44–46] for extracting
the pupil and iris circles in the segmentation step.
In the coarse stage, prior knowledge about the image acquis-
itionsystemisassumed.Thisstagedependsonthresholdingof
the gray-scale image to separate the dark regions. This approach is
not sensitive to the threshold value used due to the morphological
processing made before and after the thresholding process. The
histogram analysis of CASIA-IrisV3-Interval database images has
shown that an adaptive three-level thresholding system is efficient.
In this system, we select the threshold as follows [44–46]:
Fig. 2. Difference of Gaussian (DoG). (a) Implementation. (b) Face
image results.
Fig. 3. Keypoint description.
SIFT
Transform
Input face
RPM
1
FrFT
RPM
2
IFrFT
Cancelable
face image
*
Fig. 4. Generation of cancelable face images.
10308 Vol. 57, No. 35 / 10 December 2018 / Applied Optics Research Article
Threshold 8
<
:
115,P255
i150 hi>0.75XY
50,P100
i0hi>0.3XY
85,Otherwise
,(8)
where hirepresents the histogram value at the pixel level i,andX
and Yare the iris image dimensions.
The objective of the adaptive thresholding is to accommo-
date for variable illumination conditions. Hence, the proposed
scheme will be suitable for bright-, dark-, and medium-inten-
sity conditions. In the coarse stage and prior to this threshold-
ing process, gray-scale closing with a disk structuring element is
performed to reduce the effect of eyelashes on the iris localiza-
tion process. The thresholding strategy leads to the determina-
tion of the pupil region and some residuals of the eyelashes. To
reduce the effect of specular reflections resulting from the near-
infrared illuminators, holes in the pupil region are filled, and
hence the whole connected area is marked as the pupil region.
In the fine iris localization stage, the Daugman integro-
differential operator in Eq. (9)[44–46] is used twice for the
determination of both the pupil boundary circle and the iris
boundary circle. Both circles are not concentric. This operator
is accurate, because it searches over the image domain for the
global maximum [44–46]. However, this operator suffers from
the computational cost of searching in a 3D parameter space for
the centers and radii of pupil and iris circles, as shown in Eq. (9):
max
r,xo,yo
Gr d
dr Ir,xo,yo
Ix,y
2πrds
,(9)
where Ix,y,r,G,andsrepresent the eye image, radius of
search for the iris region, Gaussian smoothing kernel, and con-
tour of the circle, respectively.
In order to simplify this 3D search problem, the image is
resized to one quarter of its size. Moreover, the radius search
range can be reduced based on prior knowledge about the iris
acquisition system. The initial center is estimated from the
coarse stage in a neighborhood of 10 ×10 pixels only around
the center. This range is sufficient to accurately localize the iris
and pupil centers. In the iris boundary search in the proposed
scheme, the search is restricted to two sectors not the whole
360°. Both sectors extend from −30° to 30°, one on the left
and the other on the right to reduce the search cost and the
noise created by eyelids and eyelashes.
Figure 6shows an example of the iris localization process. The
input eye image is a 256 ×256 gray-scale image. The coarse
localization stage starts with the gray-scale closing, as shown
in Fig. 6(b), followed by the thresholding step to obtain the ini-
tial searching center, as shown in Fig. 6(c). The result of the fine
localization stage in Fig. 6(d) shows the segmented iris. Due to
the differences in the distance between the subject and the cam-
era, the segmented iris size might vary, and here comes the nor-
malization step. The normalization step utilizes Daugman’s
rubber sheet model [44]. This model transforms the iris pixels
from the Cartesian coordinates to a dimensionless polar coordi-
nate system. Therefore, the normalized output image in Fig. 6(e)
has a fixed size independent of the pupil dilation and iris size
variations. To reduce the effects of the eyelashes and eyelids, only
Fig. 5. Generation of cancelable iris codes.
Fig. 6. Iris localization steps. (a) Input eye image (256 ×256). (b) Eye image after gray-scale closing. (c) Thresholded image. (d) Segmented iris.
(e) Normalized iris (32 ×256). (f) Upper half of the normalized iris (16 ×256).
Research Article Vol. 57, No. 35 / 10 December 2018 / Applied Optics 10309
the upper half of the normalized image, shown in Fig. 6(f ),is
passed to the feature extraction stage.
Features are extracted from the upper half of the normalized
iris with the LogGabor filter. A convolution process is per-
formed between the unwrapped iris and the 1D LogGabor filter
[45–47]. The LogGabor filter is complex in nature and gives
complex outputs. In traditional iris recognition systems, the
phase of the complex output is binarized for obtaining the iris
codes. In the proposed cancelable iris recognition scheme, the
phase of the LogGabor filter output constitutes the feature ma-
trix to be used as input to the modified DRPE stage, which is
based on the FrFT, as shown in Fig. 5.
To enhance the privacy of the proposed cancelable iris rec-
ognition scheme, a mixing process is performed between fea-
tures extracted from the right iris and those extracted from the
left iris. We have two possibilities for this mixing process
through RPM1or through RPM2. It is required in a random
phase mask to span a 2πdynamic range. The extracted Gabor
feature values from the left iris give a large dynamic range
through the variations of their values to cover the whole 2π
range. The distribution of these feature values may not be
uniform. An issue that needs to be considered here is that
all cancelable biometric recognition systems need to generate
noninvertible outputs, which means that we do not need to
go through a decryption process. Thus, the constraints that
may be imposed on the two random phase masks RPM1
and RPM2in the traditional DRPE algorithm are not necessary
in our case. This means that the perfect uniformity of the sec-
ond random phase mask is not a must.
The main objective is only how to generate distorted or de-
formed versions of the iris features that represent the subject
without the ability to revert the process. The variations of
the iris Gabor features extracted from the left iris were found
to be sufficient to mask the original feature pattern of the right
iris. Another issue that deserves consideration here is that the
ability to cancel the iris pattern is preserved through the inde-
pendence of the first random phase mask of the left iris feature
pattern. The last step in the proposed cancelable iris recogni-
tion scheme is to transform the encrypted mixed features into
binary sequences in order to measure the distance between
these sequences with a numerical metric such as the HD.
The HD is the sum of non-equivalent bits (exclusive-OR)
between binary stored and query templates. For the two binary
templates A,B, each having a length of Nbits, the HD will be
estimated as
HD 1
NX
N
j1
Aj⊕Bj,(10)
where Ajand Bjare jth the bits of the query and stored tem-
plates, respectively, and Nis the total number of bits in the
template. To accommodate for the rotation effect on iris im-
ages, circular shifts are tested on the normalized patterns to be
matched, and the minimum HD is recorded.
5. SIMULATION RESULTS
Performance assessment of a cancelable biometric recognition
system depends on the evaluation of metric parameters to
estimate the distance between the codes generated for new sub-
jects and those stored in the database. Because the codes gener-
ated in the proposed cancelable face recognition system are in the
form of real-valued images, the correlation coefficient between
the codes of new subjects and those stored in the database
are adopted for performance evaluation. On the other hand,
the generated iris codes for the proposed cancelable iris recogni-
tion scheme are of binary nature; hence, the HD is selected as the
evaluation metric. Figure 7illustrates the matching strategy in
both proposed cancelable face and iris recognition schemes.
The genuine and imposter distributions are estimated for
both proposed cancelable biometric schemes based on the se-
lected evaluation metric. The performance efficiency of the
proposed cancelable biometric schemes is evaluated through
calculating the false positive rate (FPR), the false negative rate
(FNR), and the equal error rate (EER). The FPR is the prob-
ability that an authorized attempt is incorrectly identified as an
unauthorized one (incorrect reject). The FNR is the probability
that an unauthorized attempt is incorrectly identified as an au-
thorized one (incorrect accept). The EER is estimated at the
intersection of the genuine and imposter distributions. At
the intersection point between these distributions, the incorrect
reject and incorrect accept errors are equal. The lower the EER
value, the more efficient the security of the system. The ROC
curve is a parametric relation between the true positive rate
(TPR) (T) and the FPRTwith Tas a varying discrimination
threshold parameter. The ROC curve will be adopted in this
paper for performance evaluation of biometric systems. The
AROC will be used as an indicator for the system efficiency.
A high AROC value means better performance.
A. Cancelable Face Recognition Results
In order to test and evaluate the performance of the proposed
cancelable face recognition scheme, we have carried out the
proposed algorithm on images taken from the ORL database
of faces, which was built between 1992 and 1994 at the labo-
ratories of Cambridge University [48]. This database comprises
10 different images for 40 distinctive subjects taken at different
times, illumination conditions, and facial expressions.
For the validation of the proposed cancelable face recogni-
tion scheme, we have worked on 20 images from the ORL data-
base that were selected randomly. The cancelable templates are
extracted from these images with two different scenarios for
comparison: the full proposed cancelable face recognition
Fig. 7. Matching strategy.
10310 Vol. 57, No. 35 / 10 December 2018 / Applied Optics Research Article
scenario illustrated in Fig. 4with a rotation angle 0 for the
FrFT, and another scenario implementing bio-convolving with
a random convolution kernel with finite extent, which resem-
bles a lowpass kernel that blurs the images, as shown in Fig. 8.
The histograms of the images shown in Fig. 8are given in Fig. 9
revealing unnoticeable changes in histograms of the blurred
versions of the faces, which allows recognition of biometric im-
ages based on histograms. On the other hand, the cancelable
face templates generated with the proposed scheme are shown
in Fig. 10. It is clear that the proposed scheme succeeds to hide
the details of the images. Moreover, the histograms given in
Fig. 11 reveal closeness to uniformity over a certain bandwidth,
which is a desired characteristic for a high degree of security.
For numerical evaluation of the bio-convolving and the pro-
posed cancelable schemes, the genuine and imposter distribu-
tions for both schemes have been estimated, as shown in
Fig. 12, and the ROC curves have also been estimated, as shown
in Fig. 13.Table1gives a comparison between both schemes
from a numerical perspective considering the AROC, EER,
and decidability metrics. It is clear that the proposed cancelable
face recognition scheme outperforms the bio-convolving scheme.
Thedecidabilityisametricfortheabilitytodistinguishbe-
tween genuine and imposter distributions. It is defined as follows:
´
djμi−μgj
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
σ2
iσ2
g∕2
q,(11)
where μiand μgare the means, and σ2
iand σ2
gare the variances of
the impostor and genuine distributions, respectively. Larger values
of decidability reflect better performance of the biometric system.
It is not possible to leave this discussion without consider-
ation of the sensitivity of both cancelable face recognition
schemes to the presence of noise. Tables 2and 3give the
numerical values extracted from the sensitivity-to-noise study.
It is clear that the proposed cancelable face recognition scheme
has low sensitivity to noise as the SIFT itself eliminates the
effect of noise during the feature extraction process.
B. Iris Recognition Results
For performance evaluation of the proposed cancelable iris rec-
ognition scheme, images from the CASIA-IrisV3-Interval data-
base (CASIA-V3) [49] have been considered (see Table 4). This
database has been used because the pupil region in the iris
Fig. 8. Samples of bio-convolution results. (a) Original faces. (b) Random convolution kernels with finite extent (stretched only for view).
(b) Output blurred faces.
Fig. 9. Histograms of (a) original faces and (b) bio-convolution results.
Research Article Vol. 57, No. 35 / 10 December 2018 / Applied Optics 10311
images is unmasked in contrary to the uniform pupil circles in
CASIA-V1. This in fact is a real challenge of specular reflections
and nonuniform intensity. For the analysis purpose, for each
subject, the first five iris images were acquired from the left
eye, while the remaining five images were acquired from the
right eye. For each subject, five pairs of iris images are formed
for genuine comparisons, and a pair of iris images is used for
imposter comparisons.
Fig. 10. Samples of cancelable face templates generated with the proposed scheme. (a) Original faces. (b) Convolution kernels generated with
SIFT. (c) Output encrypted face templates.
Fig. 11. Histograms of (a) original faces and (b) cancelable face templates generated with the proposed scheme.
Fig. 12. Genuine and impostor distributions for the (a) bio-convolving and the (b) proposed cancelable face recognition scheme.
10312 Vol. 57, No. 35 / 10 December 2018 / Applied Optics Research Article
For our experiments, different snapshots of the right and left
eye images of all the 249 subjects in the dataset have been con-
sidered. For the analysis purpose, for each subject, the first five
iris images were acquired from the left eye, while the remaining
five images were acquired from the right eye. In intra-class
comparisons, each template is matched against the templates
generated from other samples of the same class leading to
600 genuine comparisons. In inter-class comparisons, each tem-
plate is matched against other templates obtained from different
samples of different classes yielding 600 impostor comparisons,
totally. Two scenarios have been considered in this study: the
traditional unprotected iris recognition scenario, and the pro-
posed cancelable iris recognition scenario with rotation angle
10° for the FrFT. The HD defined in Eq. (10) is the adopted
distance metric between any two compared binary iris codes.
The genuine and imposter distributions for both scenarios are
given in Fig. 14.Figure15 shows the ROC curves of both sce-
narios. From the results in both figures, it is clear that the pro-
posed cancelable iris recognition scheme succeeds in achieving
high performance with a larger degree of security.
For better analysis of the proposed cancelable iris recogni-
tion scheme, the TPR is estimated as (1-FRR). The sensitivity
is equal to the TPR. In addition, the specifity is estimated as the
true negative rate (TNR). For better understanding of the sys-
tem evaluation metrics, the FRR measures the probability of
falsely rejecting an iris as an imposter (intra-class) iris pattern,
and the FPR measures the probability of falsely accepting an
imposter iris pattern as a genuine (inter-class) iris pattern.
Negative and positive predictive values (NPV and PPV) have
been utilized to estimate the matching performance using
Eqs. (12) and (13):
PPV Number of true positives
Number of true positives Number of false positives ,
(12)
NPV Number of true negatives
Number of true negatives Number of false negatives:
(13)
With these metrics, we obtained 0.83% EER value for the un-
protected iris codes and 0.63% EER value for the proposed
cancelable iris recognition scheme. In addition, decidability
Fig. 13. Genuine and impostor ROC curves for the (a) bio-convolving and the (b) proposed cancelable face recognition schemes.
Table 1. Evaluation Metrics for Both Cancelable Face
Recognition Schemes
Evaluation
Metric
Bio-convolving
Scheme
Proposed Cancelable Face
Recognition Scheme
AROC 0.88 0.993
EER 0.005 0.0017
Decidability 4.2 26.5
Table 2. Evaluation Metrics for the Bio-convolving Face
Recognition Scheme in the Presence of Noise
Noise Variance EER AROC
0.01 0.005302 0.875
0.02 0.00511 0.877
0.03 0.005 0.88
0.04 0.00511 0.88
0.05 0.005 0.88
Table 3. Evaluation Metrics for the Proposed Cancelable
Face Recognition Scheme in the Presence of Noise
Noise Variance EER AROC
0.01 0.00166 0.993
0.02 0.0019 0.994
0.03 0.0013 0.995
0.04 0.0008 0.995
0.05 0.0008 0.9995
Table 4. CASIA-IrisV3-Interval Properties [48]
# Subjects # Classes
Total Number
of Tested Images
Image
Resolution
249 395 1817 320 ×280
Research Article Vol. 57, No. 35 / 10 December 2018 / Applied Optics 10313
values for both schemes are in favor of the proposed cancelable
scheme.
Table 5gives a tabulation of all results obtained with the pro-
posed cancelable iris recognition scheme in the absence and pres-
ence of noise with different noise levels. It is clear from these
results that the proposed scheme can tolerate moderate noise
levels. In addition, Fig. 16 shows the ROC curves in the presence
of noise at different levels. The results in Table 5and Fig. 16 show
that the noise effect at moderate noise levels is acceptable. Finally,
a comparison between the proposed cancelable iris recognition
scheme and some state-of-the-art schemes is given in Table 6.
These results reveal the superiority of the proposed scheme.
Fig. 14. Hamming distance for genuine and impostor distributions for the (a) unprotected and the (b) proposed cancelable iris recognition
schemes.
Fig. 15. ROC curves for the (a) unprotected and the (b) proposed cancelable iris recognition schemes.
Table 5. Performance Metrics of the Iris Recognition Systems
Performance Metric
Unprotected
IrisCodes
Protected
IrisCodes
Protected IrisCodes
[Noise var 0.01]
Protected IrisCodes
[Noise var 0.02]
Protected IrisCodes
[Noise var 0.03]
Sensitivity 99.5% 100% 97% 96% 74.17%
Specificity 99% 99.5% 92% 91% 58.33%
AROC 0.99971 0.99984 0.96999 0.96427 0.70261
NPV 99.50% 100% 96.84% 95.79% 69.31%
PPV 99.01% 99.50% 92.38% 91.43% 64.03%
EER 0.0083 0.0063 0.0650 0.0746 0.3558
Accuracy 99.25% 99.75% 94.5% 93.5% 66.25%
Decidability 4.31 5.18 3.02 2.88 0.8650
10314 Vol. 57, No. 35 / 10 December 2018 / Applied Optics Research Article
6. CONCLUSION
This paper presents two efficient implementations for the
DRPE algorithm in cancelable face and iris recognition
schemes. The common thread between the two proposed
schemes is that they adopt the same concept of feature matrix
encryption with the DRPE algorithm. The encrypted feature
matrix is used in the proposed cancelable face recognition
scheme in a bio-convolving strategy to guarantee the privacy
of the users. On the other hand, the left and right iris images
are used in an interactive manner in the proposed cancelable
iris recognition scheme to enhance the privacy of users.
Cancelability is guaranteed in both schemes through the gen-
eration of at least one random phase mask that is independent
of the users’biometrics. Simulation and comparison results ob-
tained for both cancelable biometric schemes ensure low EER
values, high decidability values, and high AROC values com-
pared with those of other traditional schemes. It is possible to
reach a conclusion that the DRPE is an efficient candidate
when implemented on feature matrices to obtain cancelable
biometric templates. Different variants based on different trans-
forms can be investigated for DRPE-based cancelable biometric
schemes.
REFERENCES
1. A. K. Jain, A. A. Ross, and K. Nandakumar, Introduction to Biometrics
(Springer, 2017).
2. S. Z. Li and A. K. Jain, Handbook of Face Recognition, 2nd ed.
(Springer, 2011).
3. M. J. Burge and K. W. Kevin, Handbook of Iris Recognition (Springer,
2013).
4. V. M. Patel, N. K. Ratha, and R. Chellappa, “Cancelable biometrics: a
review,”IEEE Signal Process. Mag. 32,54–65 (2015).
5. S. V. K. Gaddam and M. Lal, “Efficient cancelable biometric key gen-
eration scheme for cryptography,”Int. J. Netw. Secur. 11,61–65
(2010).
6. B. Choudhury, P. Then, V. Raman, B. Issac, and M. K. Haldar,
“Cancelable iris biometric based on data hiding schemes,”in IEEE
Student Conference on Research and Development, Kuala Lumpur,
Malaysia, 2016, pp. 1–6.
7. H. Kaur and P. Khanna, “Non-invertible biometric encryption to gen-
erate cancelable biometric templates,”in Proceedings of the World
Congress on Engineering and Computer Science, San Francisco,
California, 2017, Vol. I, pp. 1–4.
8. S. Wang and J. Hu, “A blind system identification approach to cancela-
ble fingerprint templates,”Pattern Recogn. 54,14–22 (2016).
9. S. Wang, W. Yang, and J. Hu, “Design of alignment-free cancelable
fingerprint templates with zoned minutia pairs,”Pattern Recogn. 66,
295–301 (2017).
10. S. Wang, G. Deng, and J. Hu, “A partial Hadamard transform
approach to the design of cancelable fingerprint templates containing
binary biometric representations,”Pattern Recogn. 61, 447–458
(2017).
11. W. Yang, S. Wang, J. Hu, G. Zheng, and C. Valli, “A fingerprint and
finger-vein based cancelable multi-biometric system,”Pattern Recogn.
78,242–251 (2018).
12. J. Zuo, N. K. Ratha, and J. H. Connel, “Cancelable iris biometric,”in
19th International Conference on Pattern Recognition, Tampa,
Florida, 2008, pp. 1–4.
13. A. B. J. Teoh, D. C. L. Ngo, and A. Goh, “Biohashing: two factor au-
thentication featuring fingerprint data and tokenised random number,”
Pattern Recogn. 37, 2245–2255 (2004).
14. C. Rathgeb, F. Breitinger, H. Baier, and C. Busch, “Towards Bloom
filter-based indexing of iris biometric data,”in 15th IEEE International
Conference on Biometrics (2015), pp. 422–429.
15. C. Rathgeb, F. Breitinger, C. Busch, and H. Baier, “On the application
of Bloom filters to iris biometrics,”IET Biometrics 3, 207–218 (2014).
16. O. Ouda, N. Tsumura, and T. Nakaguchi, “On the security of bioen-
coding based cancelable biometrics,”IEICE Trans. Inf. Syst. E94-D,
1768–1777 (2011).
Fig. 16. ROC curves for the proposed cancelable iris recognition scheme in the presence of Gaussian noise. (a) Noise variance 0.01. (b) Noise
variance 0.02. (c) Noise variance 0.03.
Table 6. Summarized Results of EER for Algorithms on
CASIA-IrisV3-Interval Database
Scheme Performance (EER %)
Salting [12] 4.41
Remapping and image warping [18] 1.30
Haar wavelet and LogGabor filter [21] 1.48
Bioencoding [16] 5.54
Tokenless scheme [17] 6.27
Biorthogonal triplet half-band
filter bank [22]
1.91
Bloom filter [14] 1.14
Bloom filter [15] 8.98
Neural networks [19] 3.56
Bidirectional memory scheme [20] 2.001
Proposed scheme 0.63
Research Article Vol. 57, No. 35 / 10 December 2018 / Applied Optics 10315
17. O. Ouda, N. Tsumura, and T. Nakaguchi, “A reliable tokenless can-
celable biometrics scheme for protecting iriscodes,”IEICE Trans. Inf.
Syst. E93-D, 1878–1888 (2010).
18. J. Hammerle-Uhl, E. Pschernig, and A. Uhl, “Cancelable iris biomet-
rics using block re-mapping and image warping,”in International
Conference on Information Security, Pisa, Italy, 2009, pp. 135–142.
19. M. Tarek, O. Ouda, and T. Hamza, “Robust cancelable biometrics
scheme based on neural networks,”IET Biometrics 5,220–228 (2016).
20. M. Tarek, O. Ouda, and T. Hamza, “Pre-image resistant cancelable
biometrics scheme using bidirectional memory model,”Int. J. Netw.
Secur. 19, 498–506 (2017).
21. A. Kumar and A. Passi, “Comparison and combination of iris matchers for
reliable personal authentication,”Pattern Recogn. 43,1016–1026 (2010).
22. A. D. Rahulkar and R. S. Holambe, “Half-Iris feature extraction and
recognition using a new class of biorthogonal triplet half-band filter
bank and flexible k-out-of-n,”IEEE Trans. Inf. Forensics Secur. 7,
230–240 (2012).
23. J. K. Pillai, V. M. Patel, R. Chellappa, and N. K. Ratha, “Sectored ran-
dom projections for cancelable iris biometrics,”in IEEE International
Conference Acoustics, Speech, Signal Processing, Dallas, Texas,
2010, pp. 1838–1841.
24. Y. Lai, Z. Jin, A. B. J. Teoh, B. Goi, W. Yap, T. Y. Chai, and C.
Rathgeb, “Cancelable iris template generation based on indexing-
first-one hashing,”Pattern Recogn. 64, 105–117 (2017).
25. S. Umer, B. C. Dhara, and B. Chanda, “A novel cancelable iris rec-
ognition system based on feature learning techniques,”Inf. Sci. 406-
407, 102–118 (2017).
26. M. Savvides, B. Kumar, and P. Khosla, “Cancelable biometric filters
for face recognition,”in International Conference of Pattern
Recognition (2004), Vol. 3, pp. 922–925.
27. P. Refregier and B. Javidi, “Optical image encryption based on input plane
and Fourier plane random encoding,”Opt. Lett. 20,767–769 (1995).
28. Y. Frauel, A. Castro, T. J. Naughton, and B. Javidi, “Resistance of the
double random phase encryption against various attacks,”Opt.
Express 15, 10253–10265 (2007).
29. X. Peng, P. Zhang, H. Wei, and B. Yu, “Known-plaintext attack on op-
tical encryption based on double random phase keys,”Opt. Lett. 31,
1044–1046 (2006).
30. A. M. Elshamy, A. N. Z. Rashed, A. A. Mohamed, O. S. Faragallah, Y.
Mu, S. A. Alshebeili, and F. E. Abd El-Samie, “Optical image encryp-
tion based on chaotic Baker map and double random phase encod-
ing,”J. Lightwave Technol. 31, 2533–2539 (2013).
31. A. M. Elshamy, F. E. Abd El-Samie, O. S. Faragallah, E. M. Elshamy,
H. S. El-Sayed, S. F. El-Zoghdy, A. N. Z. Rashed, A. A. Mohamed,
and A. Q. Alhamad, “Optical image cryptosystem using double ran-
dom phase encoding and Arnold’s Cat map,”Opt. Quantum
Electron. 48,1–18 (2016).
32. S. C. Pei and M. H. Yeh, “Two dimensional discrete fractional Fourier
transform,”Signal Process. 67,99–108 (1998).
33. H. M. Ozaktas, “Fractional Fourier domains,”Signal Process. 46, 119–
124 (1995).
34. S. K. Rajputa and N. K. Nishchal, “Optical double image security using
random phase fractional Fourier domain encoding and phase-retrieval
algorithm,”Opt. Commun. 388,38–46 (2017).
35. G. Unnikrishnan, J. Joseph, and K. Singh, “Optical encryption by
double-random phase encoding in the fractional Fourier domain,”
Opt. Lett. 25, 887–889 (2000).
36. J. W. Goodman, Introduction to Fourier Optics, 2nd ed. (McGraw-Hill,
1996).
37. M. Joshi, C. Shakher, and K. Singh, “Color image encryption and de-
cryption for twin images in fractional Fourier domain,”Opt. Commun.
281, 5713–5720 (2008).
38. C. Geng and X. Jiang, “Face recognition using sift features,”in 16th
IEEE International Conference on Image Processing (ICIP), Cairo,
Egypt, 2009, pp. 3313–3316.
39. R. Zhou, J. Wu, Q. He, C. Hu, and Z. Yu, “Approach of human face
recognition based on SIFT feature extraction and 3D rotation model,”
in IEEE International Conference on Information and Automation,
Shenzhen, China, 2011, pp. 476–479.
40. L. Lenc and P. Kral, “Automatic face recognition system based on the
SIFT features,”Comput. Electr. Eng. 46, 256–272 (2015).
41. D. Giveki, M. A. Soltanshahi, and G. A. Montazer, “A new image fea-
ture descriptor for content based image retrieval using scale invariant
feature transform and local derivative pattern,”Optik 131, 242–254
(2017).
42. T. P. Shiji, S. Remya, and V. Thomas, “Computer aided segmentation
of breast ultrasound images using scale invariant feature transform
(SIFT) and bag of features,”Procedia Comput. Sci. 115, 518–525
(2017).
43. R. Rajkumar and K. M. Singh, “Digital image forgery detection using
SIFT feature,”in International Symposium on Advanced Computing
and Communication, Silchar, India, 2015, pp. 1–6.
44. J. Daugman, “How iris recognition works,”IEEE Trans. Circuits Syst.
Video Technol. 14,21–30 (2004).
45. M. Essam, M. Abd Elnaby, M. Fikri, and F. E. Abd El-Samie, “Afast
accurate algorithm for iris localization using a coarse-to-fine approach,”
in Japan-Egypt Conference on Electronics, Communications and
Computers (JEC-ECC), Alexandria, Egypt, 2012, pp. 63–67.
46. M. Essam, M. Fikri, M. A. Elnaby, and F. E. A. El-Samie, “An efficient
iris localization algorithm,”in 29th National Radio Science Conference
(NRSC), Cairo, Egypt, 2012, pp. 285–292.
47. N. F. Soliman, E. Mohamed, F. Magdi, F. E. Abd El-Samie, and M.
AbdElnaby, “Efficient iris localization and recognition,”Optik 140,
469–475 (2017).
48. “ORL database,”https://www.cl.cam.ac.uk/research/dtg/attarchive/
facedatabase.html.
49. “CASIA-IrisV3 database,”http://www.cbsr.ia.ac.cn/english/IrisDatabase
.asp.
10316 Vol. 57, No. 35 / 10 December 2018 / Applied Optics Research Article
A preview of this full-text is provided by Optica Publishing Group.
Content available from Applied Optics
This content is subject to copyright. Terms and conditions apply.
