Content uploaded by Hanaa Mohsin Ahmed Salman
Author content
All content in this area was uploaded by Hanaa Mohsin Ahmed Salman on Oct 24, 2017
Content may be subject to copyright.
Al-Mansour Journal/ Issue( 25 ) 2016 رﻮﺼﻨﻤﻟا ﺔﻠﺠﻣ/ دﺪﻌﻟا )25 (
- 1 -
A Proposed Algorithms to Design Support Multimodal
biometric System
Hanaa M. Ahmed*,Ph.D (Asst. Prof.) Bushra J. Abdulkareem**
Abstract
This paper is an attempt to address the biometric security issue and
improve the system accuracy through introducing a design for multimodal
biometric verification system using multiple traits (Iris, fingerprint) and
adding another phase called liveness detection to the phases of
multimodal system the purpose of this phase is to protect the multimodal
biometric systems against spoofing attacks. The system is tested in two
levels, unimodal level and multimodal level (fusion level). in unimodal
level two tests have been performed, one for iris verification phase
performed on two types of database MMU DB (Multi Media University
database) for 180 samples and CASIA DB (Chinese Academy of
Sciences database) for 90 samples. and gave accuracy (99.44%) with
FAR (False Acceptance Rate) of (0.0277) and FRR (False Reject Rate)
(0.0055) for MMU DB, and accuracy (97.77%) with FAR of (0.0333) and
FRR (0.0222) for CASIA DB, and other for fingerprint verification phase
performed on database collected from two types of database for 60
samples and gives accuracy of 95% with FAR of 0.1% and FRR of
0.05%.
In multimodal level the system is tested on database composed of 60
samples for iris images and 60 samples for fingerprint images and gives
an overall accuracy of 100% with FRR of 0%, and FAR of 0.0166%.
Keywords: Liveness detection, Multimodal, Iris, Fingerprint, Anti-
spoofing, Verification, Fusion function.
____________________
*University of Technology
**Al-Mansour university College
Hanaa M. Ahmed*,Ph.D (Asst. Prof.) Bushra J. Abdulkareem
- 2 -
1. Introduction
Reliable verification schemes is require in wide variety of applications
such as secure access to buildings, computer systems, laptops, cellular
phones and ATMs, to confirm the identity of an individual requesting their
service. Unimodal biometric systems establish person verification based
on a single biometric trait, its characteristic needs to meet some basic
requirements like [1 and 2]: Universality, Uniqueness, Permanence, and
Collectability. However, there are a number of other issues that should
be considered with any biometric trait that meets previous criteria, these
are, Performance, Acceptability, and Circumvention.
From the practical applications, no biometric characteristic fully
meets these requisites. Besides it suffers from noisy sensor data, poor
quality biometric traits, continuous threats of spoof attacks, and
unacceptable error rates etc., hence, may not always meet crucial
security requirements. Multimodal biometric systems that consolidate
evidence from multiple biometric sources can be used to overcome or
minimized some of these limitations [3 and 4].
Many previous researches addressed the issue of “anti-spoofing” in
unimodal biometric systems, the common method used for this purpose,
is to insert an additional module, called “liveness detector” that is used as
a countermeasure to spoof attacks to assess if an input biometric sample
acquired by some sensors belongs to a “live” person or is a “spoof”
artifact [6].
Anti-spoofing in multimodal biometric systems, is not a clear concept
as in the unimodal case. Multimodal biometric systems have been
commonly believed to be intrinsically more robust to spoof attacks than
unimodal systems. This confidence is based on the intuitive hypothesis
that evading the multimodal biometric system always requires an attacker
to spoof all the involved traits (or at least more than one). Recently this
belief has been questioned and several works provided clear evidence
that they can be evaded by spoofing a single biometric trait, as in [3, 6, 7,
and 8]. These researches showed that number of vulnerability points will
be increased in multimodal biometric system, and can be explored by an
Al-Mansour Journal/ Issue( 25 ) 2016 رﻮﺼﻨﻤﻟا ﺔﻠﺠﻣ/ دﺪﻌﻟا )25 (
- 3 -
intruder. Hence a multimodal system may be easier to spoof than some
of the unimodal systems that compose it. This question is especially
important when multimodal system combining face traits which can be
easily spoofed, and retina veins traits which is very hard (if not
impossible) to spoof. In this case, an impostor that spoofs only the face
trait may have a very high chance of being falsely accepted [3 and 7].
This paper is an attempt to address the biometric security issue and
improve the system accuracy through introducing a design for multimodal
biometric verification system using multiple traits (Iris, fingerprint) and
adding another phase called liveness detection to the phases of
multimodal system to protect the multimodal biometric systems against
spoof attacks.
2. Literature Survey
As any of the traditional security systems, identity verification
systems that use biometric, attempts to attack him by opponents, and
who have the ability to compromising data integrity through alteration so
the system becomes inactive. Many researchers and designers of
biometric systems highlighted the lack of security in regards to biometric
systems through the provision of studies and algorithms to solve this
issue. Therefore, the evaluation of these systems is an open question
whether the investigation will lead to a secure biometric systems design
[5]. To build secure biometric systems it is necessary to understand and
evaluate these threats through the development countermeasures,
design of impervious against these attacks. Many researchers who study
the weaknesses in the biometric systems, possible attack methods and
their countermeasure, here is a collection of previous studies related to
the paper theme:
In 2009, R. N. Rodrigues, L. L Ling, and V. Govindaraju [3]
Propose two original multi-modal biometric fusion's methods that
consider the spoofing assumptions and the security of each uni-modal
biometric being merged. One method which is an extension of the
Likelihood Ratio (LLR), and the other method, is using fuzzy logic. The
Hanaa M. Ahmed*,Ph.D (Asst. Prof.) Bushra J. Abdulkareem
- 4 -
two models follow the same basic ideas, but their details and
implementation are different. The work in these two schemes shows that
when using traditional fusion method (i.e. weighted sum or LLR)
attacker's chances of evading a multi-modal system by spoofing only one
of the biometrics can dramatically increase. Experiments showed, first:
the existence of a trade-off between robustness against spoof attacks,
and recognition accuracy, second: the fuzzy fusion scheme had a best
overall performance compared with the probabilistic fusion scheme.
In 2011, Maruf Monwar et al. [4] Rreliable and robust multi-modal
biometric based security system have been proposed, it composed of
(face, ear, and iris) and use soft biometric identifiers (gender, ethnicity
and eye color). A novel fuzzy fusion technology is used to fuse these
biometric traits. This scheme adopts match score, rank information and
soft biometrics information from unimodal biometrics as the input, and
final identification decision via a fuzzy rule as output. Supplement
information about the identity of a subject has been provided by their
research that makes the operation of human recognition more accurate.
The improvement in recognition performance results is duo to the used of
an optimum weighting scheme has been advanced based on the
distinctive abilities of the primary and the soft biometric traits.
Comparison for the experimental results between the fuzzy fusion
technology and other fusion methods has been conducted, which prove
that the proposed method is not only accurate but also faster beyond
existing technologies.
In 2012, P.U.Lahane, and S.R.Ganorkar [9] suggested Multimodal
biometric identification system, composed of (iris and fingerprint). Each
biometric trait processes its information independently. In this system, the
iris is extracted, removes the influence of the eyelids and eyelashes, and
through a series of operations on the eye image provided. Singularity
region is segmented from input fingerprint image by preprocessing
operations performed on it. Then the Region of Interests (ROI) extracted
and used as input for the normalization. Gabor filter used to extract
features from fingerprint and iris. Then fusion is performed in the feature
Al-Mansour Journal/ Issue( 25 ) 2016 رﻮﺼﻨﻤﻟا ﺔﻠﺠﻣ/ دﺪﻌﻟا )25 (
- 5 -
extraction level by combining the biometric features extracted from
fingerprints and irises images. Finally Euclidean Distance is the matching
algorithm used for computing matching score. Experimental threshold
used to decide whether or not the two representations belong to the
same user by comparing with the result of the measurement. This work
produces efficient security system.
In 2012, Akhtar Z, Fumera G, Marcialis GL and Roli F, [11] this
research presented "comparison based robustness against spoofing
attacks between serial fusion of multi-modal systems, and parallel multi-
modal systems, by empirically analyzing between the robustness of serial
fusion of dual modal systems with the corresponding parallel systems,
using of a fingerprint and a face matcher, against several real spoofing
attacks. Results obtained regarding the level of fusion rules, which were
common in the literature, are not robust to spoofing attacks as believed,
since they can be avoidable by spoofing only one biometric trait. Also,
spoofing the extremely accurate biometrics makes more probably to
avoid a multi-modal system. However, they found confirmations that
serial multi-modal systems are more robust than parallel ones versus
spoofing attacks, and can earn a better trade-off between performance,
verification time, user acceptability and robustness.
In 2013, Dapinder and Gaganpreet [10] Stated the fusion methods
by dividing it into the following three categories:(max, or, product,
majority voting, min, sum, and ) belong to first category (fixed rule-based
methods), and (the Bayesian inference, support vector machine,
maximum entropy model, and neural networks ) belong to the second
category (classification based methods), and the third category
(estimation-based methods) which includes (the particle filter fusion
methods, Kalman filter, and extended Kalman filter). These methods
have been primarily used to better estimate the state of a moving object
based on multimodal data. The basic nature of these methods is the
base of this categorization, and it means the classification of the problem
scope, such as, estimation-based methods solved a problem of
Hanaa M. Ahmed*,Ph.D (Asst. Prof.) Bushra J. Abdulkareem
- 6 -
estimating parameters. And classification-based or rule based methods
solved the problem of obtaining a decision based on specific observation.
3. Individual Recognizers
Iris and fingerprint biometrics perform better as compared to other
available traits due to their accuracy, reliability and simplicity. These
properties make iris and fingerprint recognition particularly promising
solution to the society. Below is theoretical part of these two biometric
traits and the method used for fused these traits.
3.1 Iris Recognition
Iris is an important feature of the human body and unique to each
individual and very stable throughout lifetime of a person. Iris biometric
trait offers many advantages over other human biometric features. The
Iris is the only internal human body organ that is visible from the outside
and is well protected from external modifiers. Due to the richness of the
texture details in the iris image the eyes of an individual contain
completely independent iris patterns, and these minute details are
randomly distributed which make the human iris as one of the most
important biometric characteristics [12].
3.2 Fingerprint Recognition
Fingerprints are one of the most widely used biometric modality
which used in courts of law in all over the world. And increase number of
civilian and commercial applications which used fingerprint-based
identification, because of their three properties. First the character of the
pattern on each finger is permanent and unchanged, second the ridge
details are uniqueness ,and the third point is the feature vector can be
easily extracted from fingerprint and stored in a compact fashion, and
suitable for matching [2 and 13].
Al-Mansour Journal/ Issue( 25 ) 2016 رﻮﺼﻨﻤﻟا ﺔﻠﺠﻣ/ دﺪﻌﻟا )25 (
- 7 -
3.3 Fusion
Fusion is the procedure which performs integrates information from
multiple biometric traits to consolidate the effectiveness of the biometric
system and making it difficult for an intruder to spoof multiple biometric
traits simultaneously. In this proposed system the fusion is performed in
the matching score level, after the extraction of the features of each
model in the multi-modal system, matching the stored data, and test data
for each biometric feature using the same matching algorithm. Scores
generated from matching module from each biometric trait moved to
score fusion rule at matching score level using weighted sum of fusion
technique. In Fusion: The two resulted scores Na and Nb are fused
linearly using weighting sum rule as [14]:
MS=αX Na+ βXNb, … (1)
Where (αX) and (βX) are two weight values that can be determine by
the training data which considered the degrees of accuracy for each
biometric trait contributed to construct the system.
3.4 Performance Measures used in Biometric System
The fundamental parameters used to measure the performance of
biometric verification systems are explained below:
1. False Acceptance Rate (FAR):
It measures the likelihood of confusing two identities or it is the ratio
of acceptance intruder falsely. Obviously, this measure very affected
by the desirable security degree and the system goodness.FAR can
be defined as [15]:
…(2)
2. False Reject Rate (FRR)
It measures the probability that enrolled person is identified wrongly;
in other word it is the rate of rejecting real user. FRR can be defined
as:
Hanaa M. Ahmed*,Ph.D (Asst. Prof.) Bushra J. Abdulkareem
- 8 -
…(3)
4. The Proposed System
This proposed system has been constructed by using multimodal
schema represented by iris and fingerprint models. The steps of the
proposed system are shown in figure (1):
Algorithm (1): multimodal system
Input:
1- Two samples of eye images
2- Two samples of fingerprint images
Output : Final decision, if person has been genuine or imposter or it is fake
try
Begin
Step 1: Read the two samples of eye image
Step 2: Execute dynamic liveness detection module on the two samples of
eye image to cheek liveness. If the result from this module true
then goes to (step3) otherwise make the decision (the request is
fake) and go to end.
Step 3: one of the two samples of eye image input to the static liveness
detection module. If the result from this module is true go to
(step 4) otherwise make the decision (the request is fake) and
go to end.
Step 4: read two samples of fingerprint image
Step 5: Cheek dynamic liveness for the two finger image if result from this
module is true go to step 6 otherwise make the decision (the
request is fake) and go to end.
Step 6: one of the two samples of fingerprint image input to the static
liveness detection module. If the result from this module is true
go to (step 7) otherwise make the decision (the request is fake)
and go to end.
Step 7: perform iris verification module in this module feature vector is
extracted and comparison is performed between the feature
vector submitted by the person and the one stored in its database
called (template) to produce iris matching score.
Step 8: Fingerprint verification module are executed by extract feature
vector and compute fingerprint matching score
Al-Mansour Journal/ Issue( 25 ) 2016 رﻮﺼﻨﻤﻟا ﺔﻠﺠﻣ/ دﺪﻌﻟا )25 (
- 9 -
Step 9: Fusion for the matching scores result from (step 7) and (step 8) are
performed using weighting sum rule
Step 10: make the final decision according the result from (step 9) if the
result is equal or greater than (85) then the person is genuine else
the person is imposter.
End Algorithm
Figure (1) The algorithm of the proposed system
The algorithm start with very important phase called liveness
detection in both (static sub model and dynamic sub model) for each one
of those two models, which added to the system to protect it from spoof
attack. And then verification phase to produce matching scores from two
biometric traits iris and fingerprint. And the last phase in proposed
system which represent by fusion phase where a person is declared as
genuine or an imposter as following: Figure (2) illustrates the architecture
of the proposed system.
4.1 Liveness Detection Phase
In iris liveness detection module two modules used to detect
liveness. The variation in pupil size caused by acquired eye images for
the same person in different lightness which is restricted in the range (5-
15%) is exploited to detect the liveness in input eye image by dynamic
iris liveness detection module. In static iris liveness detection module the
property of focus degree of the acquired eye image by compute the
sharpening of eye image using high pass filter used to detect liveness in
input eye image. The threshold detecting by training data in this algorithm
to make liveness decision is (34) (if the mean of gradients of eye image
less than 34 then the input image sample is fake else it is live). In
fingerprint model also two module used to detect liveness, Dynamic
module by compute difference in standard deviation between two input
fingerprint image acquired in period of time (3-5 second), and static
module by using number of First order statistical features, and properties
extracted from analyze input fingerprint image. The output from liveness
detection phase decides whether the acquired image come from real or
fake biometric trait. If the decision that the acquired image come from
fake biometric trait, from any one of two models (iris or fingerprint), the
Hanaa M. Ahmed*,Ph.D (Asst. Prof.) Bushra J. Abdulkareem
- 10 -
processing operation will be stopped, and the system points that this
request to enter to the system is spoof attack. Else if the decision that the
acquired images come from real biometric traits, the algorithm will be
continue and moves to verification phase.
4.2 Verification Phase
This phase composed of two modules for produce matching scores
executed in parallel one for iris verification model and other for fingerprint
verification model. Below is the explanation for these two models:-
4.2.1 Iris Verification Model
The iris verification module work as following:
The eye image of the user who claims his identity is input to the iris
verification module and pass through sequence of steps started with iris
segmentation which performed by, pupil localization which done through
eight point ending with compute the radius Rp and center coordinates
(Cpx,Cpy) of the pupil to detect the inner boundary. And then iris
localization to detect the outer boundary which performed by produce the
gradient image using canny edge detector, and then using circular
summation by exploit pupil's radius and center coordinates of the pupil.
Summing the intensities over all circles, pass over all possible radii
starting from pupil's radius +15 to the pupil's radius + 50 and center
coordinates of the pupil . The circle with highest summation corresponds
to the outer boundary. After this the interest region will be detected by
selecting the part of iris region to the left and right and to the bottom of
the pupil; selection the region in this way is to avoid noise caused by
eyelashes and eyelids, then made all gray level values in pupil region
equal to zero to isolate it from selected region. The feature vector will be
extracted from this region using series of second order statistical feature
computed from GLCM for this region. In the matching, these feature set
compared with the enrolled respective feature vector stored in the
database (template) of the claimed identity using (percentage of the
matching) algorithm to produce the iris matching score.
Al-Mansour Journal/ Issue( 25 ) 2016 رﻮﺼﻨﻤﻟا ﺔﻠﺠﻣ/ دﺪﻌﻟا )25 (
- 11 -
Hanaa M. Ahmed*,Ph.D (Asst. Prof.) Bushra J. Abdulkareem
- 12 -
4.2.2 Fingerprint Verification Model
In the other side of the proposed system, the fingerprint image for the
same user input to the fingerprint verification module which pass through
preprocessing operation include: first remove noise by adaptive filter,
second segment the image, third binarization the segmented image, last
thinning this binary image. Feature vector will be extracted from thinning
image using central moment technique. The same operation performed
in the iris matching will be performed in the fingerprint matching, to
produce fingerprint matching score result.
4.3 Fusion Phase
The matching scores produced from phase two for two biometric
model (iris and fingerprint) have been input to the phase three of the
system to be linearly fused by using weighting sum rule illustrated in
equation (1).
Where α, and β are weight values that determined by the training
data which consider the degree of accuracy for each biometric traits
contributed to construct the system. T is threshold previously defined.
Hence if the Fused Score (FS) ≥T, then user is accepted as Genuine
(G), otherwise it is rejected as an Impostor (I).
5. Experiment Result
To show the benefit of designing multimodal biometric system, the
experiment results will be presented in two levels. These levels are
liveness detection and verification.
5.1 Liveness Detection Results
The first level of experimental results is liveness detection. In this
level two field of experiment have been illustrated, these are iris liveness
detection and fingerprint detection:
5.1.1 Iris liveness detection module:
Database that is set up to test the robustness of the proposed
system through iris liveness detection process consists of 15 original
(MMU database) [16] folders each folder contain two eye image samples
represent live tries, and 15 (MMU database) folders of eye images
printed using scanner devise and recapture using specific camera and
Al-Mansour Journal/ Issue( 25 ) 2016 رﻮﺼﻨﻤﻟا ﺔﻠﺠﻣ/ دﺪﻌﻟا )25 (
- 13 -
resaved in computer to represent 15 attempted spoof attack against the
system. Each folder contained two samples of fake eye image. Table (1)
show experiment results for dynamic iris liveness module and Table (2)
show the results of static iris liveness module:
Table (1): The experiment results for dynamic iris liveness module
Results by Appling on original eye images
Results by Appling on recaptured eye images
Person
No.
No. of pixel in
pupil for two
samples with
different
eliminations
Percentage
Difference in
pupil size
decision
No of pixel in pupil
for two samples with
different eliminations
Percentage
Difference in
pupil size
decision
1
291210
325125
11.005
Live
694620
695895
0.183
Fake
2
482460
540090
11.272
=
682380
691050
1.613
Fake
3
457980
525300
13.693
=
1418310
1419330
0.072
Fake
4
330735
372300
11.824
=
961605
974355
1.317
=
5
699210
759900
8.319
=
512295
772395
40.492
=
6
410550
474300
14.409
=
930495
759900
20.184
=
7
408000
431460
5.589
=
1266075
1327530
4.739
=
8
337875
368220
8.595
=
542640
548760
1.121
=
9
248625
288150
11.116
=
482715
685695
34.745
=
10
514845
566355
9.528
=
675750
907800
29.308
=
11
585735
627555
6.894
=
1564425
1195695
26.718
=
12
469455
536010
13.239
=
828750
562785
38.226
=
13
439110
469710
6.734
=
1036065
1171980
12.311
faulty
live
14
336345
336345
5.529
=
818805
842265
2.825
Fake
15
719865
762450
5.746
=
1002915
860880
15.241
Fake
Hanaa M. Ahmed*,Ph.D (Asst. Prof.) Bushra J. Abdulkareem
- 14 -
Table (2) the experiment results for static iris liveness module
Results by Appling static module on
original eye images
Results by Appling static module on recaptured eye
images
Person
No
Mean of gradient for
input original MMU
eye sample images
decision
Mean of gradient for input
recaptured MMU eye sample
images(spoof image)
decision
1
45.867
38.969
Live
=
28.808
32.486
Fake
=
2
36..203
58.622
Live
=
29.355
39.577
Fake
Faulty Live
3
42.561
37.927
Live
=
34.55
30.977
Faulty Live
Fake
4
35.306
35.998
Live
29.382
28.72
Fake
=
5
55.638
49.592
Live
=
32.273
32.884
=
=
6
36.935
41.593
Live
=
26.155
28.721
=
=
7
36.489
36.454
Live
27.464
29.281
=
=
8
35.789
45.489
Live
26.947
32.051
=
=
9
35.49
35.186
Live
=
26.22
25.554
=
=
10
35.511
34.5
Live
=
27.708
23.303
=
=
11
35.122
35.052
Live
=
25.392
28.361
=
=
12
39.363
37.276
Live
=
28.55
28.323
=
=
13
36.702
38.173
Live
=
23.235
28.416
=
=
14
34.363
35.389
Live
=
23.112
27.843
=
=
15
35.55
33.55
Live
=
26.295
24.792
=
=
Al-Mansour Journal/ Issue( 25 ) 2016 رﻮﺼﻨﻤﻟا ﺔﻠﺠﻣ/ دﺪﻌﻟا )25 (
- 15 -
5.1.2 Fingerprint liveness detection results:
Database that is set up to test the robustness of the proposed
system through fingerprint liveness detection process consists of 12
folders of original fingerprint database Each folder contain two fingerprint
images sample represent live tries, 12 folders generated from original
database used in verification phase to represent fake tries. Table (3)
show experiment results for dynamic fingerprint liveness module and
table (4) show the results of static fingerprint liveness module.
Table (3) The experiment results for dynamic fingerprint liveness module
Results by Appling on live fingerprint samples
Results by Appling on fake fingerprint
samples
Person
No
Standard deviation
for two input
samples computed
based on GLCM
normalized matrix
Percentage
Difference in
two standard
decision
Standard
deviation for
two input
samples
Percentage
Difference in
two standard
decisio
n
1
0.788
0.725
8.333
Live
0.763
0.788
3.209
Fake
2
0.732
0.78
6.418
=
0.817
0.786
3.884
Fake
3
0.525
0.411
24.319
=
0.525
0.513
2.394
Fake
4
0.855
0.668
24.584
=
0.655
0.668
2.02
=
5
0.706
0.757
6.997
=
0.758
0.799
5.447
faulty
live
6
0.701
0.802
13.556
=
0.58
0.58
0
Fake
7
0.743
0.851
13.521
=
1.232
1.243
0.838
=
8
0.829
0.769
7.405
=
0.829
0.809
2.412
=
9
1.306
1.501
13.919
=
1.477
1.501
1.668
=
10
1.006
1.495
39.101
=
1.006
1.007
0.105
=
11
1.418
1.278
10.4
=
1.418
1.456
2.647
=
12
1.869
1.681
10.653
=
1.936
1.871
3.399
=
Hanaa M. Ahmed*,Ph.D (Asst. Prof.) Bushra J. Abdulkareem
- 16 -
Table (4) the experiment results for static fingerprint liveness module
No.
Original/captured
Rao1
Rao2
mean
Energy
variance
Kurtosis
1.
Original
14.694245
0.0004947
2.005844
5469.826
21870917.77
0.002067
Captured
0
0
1.4218
2455.869
2274110.788
0.00322
2.
Original
6.31323490
0.1907496
1.49317
5100.496
16832942.407
0.00291
Captured
0
0
1.4218
1921.318
2054730.515
0.00354
3.
Original
26.155145
7.4634910
1.840148
4112.509
19312355.021
0.00224
Captured
0
0
1.4218
2377.84
1514958.28
0.00338
4.
Original
2.9797069
0.00030003
1.678756
2754.41
15374683.366
0.002785
Captured
0
0
1.4218
1767.457
1748983.244
0.0038460
5.
Original
0.478353
0.563804
1.421875
30618.22
13840099.733
0.003754
Captured
0
0
1.4218
2107.362
2377095.068
.0045012
6.
Original
5.6075581
.20576589
1.421875
24902.71
16718770.69
0.003090
Captured
0
0
1.421875
1.421875
3002221.079
0.003486
7.
Original
1.4583712
6.1065050
4.6875
1065839.9
69272205.80
0.000881
Captured
0
0
1.421875
4784.820
2516529.38
0.003273
8.
Original
0.526416
3.656012
4.6875
885777.3
63127761.85
0.000964
Captured
0
0
1.421875
3045.649
3140497.75
0.004039
9.
Original
0.607280
0.359060
1.421875
20955.06
12589930.35
0.003897
Captured
2.14633E-05
0
1.421875
1628.664
2819894.415
0.004536
10.
Original
0.341605
0.412820
1.421875
30254.21
12531497.963
0.003869
Captured
0
0
1.421875
2154.626
2221829.152
0.004986
11.
Original
0.011108
0.003
0.3973388
104.1590
1013907.64
0.013428
Captured
0.0053482
0.0010496
0.39733
97.0417
1048356.372
0.015515
Al-Mansour Journal/ Issue( 25 ) 2016 رﻮﺼﻨﻤﻟا ﺔﻠﺠﻣ/ دﺪﻌﻟا )25 (
- 17 -
5.2 Verification Results
The second level of experimental result is verification. In this level
three folds of experiment have been illustrated. These are, iris verification
model, fingerprint verification module, and finally fusion module.
5.2.1 Iris verification Experimental Results
Two types of iris database used to training and testing the proposed
system, these are: MMU Iris database and CASIA-IrisV1database [17].
The training of iris verification algorithm consists of three experiments as
follow:
The first experiment conducted to test of the proposed iris verification
algorithm by applied on 180 eye image of MMU database for 30 persons
for left and right eyes. Three samples for left eye and three sample for
right eye for each person. The second experiment testing of the
proposed iris verification algorithm phase by applied on 90 image of
CASIA-IrisV1 database for 30 persons. Three samples of eye image for
each person. In the third experiment the eye images database collected
for testing the proposed multimodal verification algorithm as completed
system. Started from liveness detection phase and ending with fusion
phase consists of 15 folders of (MMU database) for 15 person each
folder contained four samples for left eye image. Table (5) clarify the
results of testing operation for verification phase of iris model and the
accuracy computed and The FAR and FRR.
Hanaa M. Ahmed*,Ph.D (Asst. Prof.) Bushra J. Abdulkareem
- 18 -
Table (5) experiment results of the iris verification phase
Eye image
Database
No of people
No.of
samples
No. of
samples
successfully
verification
No. of
samples
faulty
accepted
No. of
samples
faulty
rejected
Tame of
verification
Time for
matching
FRR %
FAR%
Average
accuracy
MMU DB
30
180
179
5
1
20
second
0.8
second
0.0055
0.0277
99.44%
MMU DB
15
60
59
2
1
20
second
0.8
second
0.0166
0.0333
98.33%
CASIA
(Version
1.0)
30
90
88
3
2
24
second
0.8
second
0,0222
0.0333
97.77%
5.2.2 Fingerprint Verification Experimental Results
The database collected for training and testing fingerprint verification
model is consists of 15 folders for 15 persons each folder contain four
fingerprint image samples for same person. The database taken by the
internet from the database of University of Bologna and all of them are
gray scale of a tiff image file format. The results of testing operation for
verification phase of fingerprint model are showed in table (6) which
illustrates the accuracy and FAR and FRR.
Al-Mansour Journal/ Issue( 25 ) 2016 رﻮﺼﻨﻤﻟا ﺔﻠﺠﻣ/ دﺪﻌﻟا )25 (
- 19 -
Table (6) experiment results of the fingerprint verification phase
fingerprint
image Database
No.of people
No.of samples
No.of samples
successfully
verification
No.of samples
faulty accepted
No of samples
faulty rejected
Tame of
verification in
seconds
Time for
matching in
seconds
FRR
%
FAR
%
Average
Accuracy
University of
Bologna
15
60
57
6
3
24
0.8
0.05
0.1
95
6.Fusion Experimental Results
In this fold of experiment result the matching scores of the iris and
fingerprint are combined and total accuracy is computed, as shown in
Table (7). The combination of two databases, Iris database which
composed of 15 folders as maintained in iris verification phase, and
fingerprint database which is the same database used in fingerprint
verification phase used for whole system which ending by fusion phase.
From the accuracy results showed previously in iris verification phase
and fingerprint verification phase, the accuracy of iris verification is
highest than the accuracy of fingerprint verification, on this basis, the
values of ( = 0.7 : which represent the weight given to iris matching
score) and ( = 0.3 :which represent the weight given to fingerprint
matching score) and perform fusion operation by apply fusion equation.
Table (7) experiment results of (fusion phase)
Data set used
No of people
No of
matching
scores
No of fusion
operations
No of fusion
successfully
verification
Tame of
verification
For complete
system in
seconds
Time for
fusion
operation
in seconds
Average
accuracy
MMU iris
database
and
University
of Bologna
for
fingerprint
15
60 for iris
60 for
fingerprint
60
60
45.5
0.5
100%
Hanaa M. Ahmed*,Ph.D (Asst. Prof.) Bushra J. Abdulkareem
- 20 -
Table (8) shows the accuracy obtained of unimodal and combined
system. The whole performance of the system has increased in
multimodal production accuracy of 100% with FAR of 0% and FRR of 0%
consecutive receivers.
Table (8) experiment results iris and fingerprint (fusion phase)
Biometric
verification
module
Database
used
No of
person
No of
verification
operation
No of
successfully
verification
operation
Time for
verification
operation in
seconds
Average
accuracy
Iris
verification
MMU
15
60
59
20
98.33%
Fingerprint
verification
University
of Bologna
15
60
57
25
95%
Multimodal
iris and
fingerprint
verification
system
MMU and
University
of Bologna
15
60
60
45.5
100%
The accuracy result of proposed system is compared with the
accuracy results of other existing methods. This comparison are showed
in table (9)
Table (9) experiment results of the proposed system and other existing
system accuracies.
Author
Biometric system
Database
Algorithm
FAR
FRR
accura
cy
Hunny, Ajita,
Phalguni
Multimodal (iris,
fingerprint)
database collected
by the authors for
200 samples
Haar +
Minutiae
1.58
6.43
96.04%
P.U.Lahane,
Prof.
S.R.Ganorkar
Multimodal (Iris +
Fingerprint)
ten users with five
iris image & five
Fingerprint image
of each person
Gabor filter
0.3
0.5
99.5
Proposed
algorithm
Multimodal (Iris +
Fingerprint)
fifteen users with
four iris image &
four Fingerprint
image of each
person
Second order
statistical
feature and
central
moment
features
0.0166
0
100%
Al-Mansour Journal/ Issue( 25 ) 2016 رﻮﺼﻨﻤﻟا ﺔﻠﺠﻣ/ دﺪﻌﻟا )25 (
- 21 -
7. Conclusions
By extensive and hard work in the design of the proposed system
which composed of multimodal biometric system to improve the security
and accuracy of these types of system through adding another module
(liveness detection) to it, number of conclusions have been reached:-
1. The dynamic method which used to detect liveness is more effective,
more accuracy and success than static method. Static method need
to detect fixed threshold, extracted from original properties of the
biometric trait to accept image sample as real sample. Choosing this
threshold be influenced by the variance of these original properties
from person to another, such as degree of sharpening of eye image
and the first order statistical features which extracted from fingerprint
image. And the way used to collect the original image database and
manufacture spoof database which is, cause ratio of error in liveness
decision.
2. in matching module of this proposed system there is a specific
percentage rate (0.1) for similarity between each element in test
feature vector with corresponding element in template feature vector
allowed to it to enter into the comparison process to produce the final
matching score , this to avoid FAR in this system.
3. In biometric verification system, multimodal is better, and more
accuracy than single model, because the biometric traits for these
two model (iris and fingerprint) are fused using specific fusion
function, through merge matching score coming from fingerprint
based on invariants central moment feature and iris matching score
based on radius of pupil and second texture feature from the
normalized GLCM, and consider the degree of accuracy of these two
models.
Hanaa M. Ahmed*,Ph.D (Asst. Prof.) Bushra J. Abdulkareem
- 22 -
References
[1] Mohamed Soltane , Mimen Bakhti, " Multi-Modal Biometric Authentications:
Concept Issues and Applications Strategies" , International Journal of
Advanced Science and Technology, Vol. 48, November, 2012.
[2] Ebtesam Najim AL_Shimmary, "Fingerprint Image Enhancement and
Recognition Algorithms", Ph.D thesis, University of Technology, May, 2007.
[3]. R. N. Rodrigues, L. L Ling, and V. Govindaraju," Robustness of multimodal
biometric fusion methods against spoof attacks", Journal of Visual
Languages and Computing, Vol. (20), Issue (3), June, 2009.
[4]. Md. Maruf Monwar, Marina Gavrilova, " A Novel Fuzzy Multimodal
Information Fusion Technology for Human Biometric Traits
Identification",IEEE, International Conference on biometric compendium ,
20 Aug. 2011
[5]. Marta GomezBarrero, Javier Galbally, Julian Fierrez , Javier OrtegaGarcia,
"Progress in Pattern Recognition, Image Analysis, Computer Vision, and
Applications", Springer Berlin Heidelberg, November, 2013.
[6]. Giorgio Fumera, Gian Luca Marcialis, Battista Biggio, Fabio Roli and
Stephanie Caswell Schuckers, "Handbook of Biometric Anti-spoofing" ,
Springer London, July 2014.
[7]. P. A. Johnson, B. Tan, S. Schuckers, "Multimodal Fusion Vulnerability to
Non-Zero Effort (Spoof) Imposters, IEEE International workshop on
information, December, 2010.
[8]. Ricardo N. Rodrigues, Niranjan Kamat and Venu Govindaraju, "Evaluation
of Biometric Spoofing in a Multimodal System", IEEE International
Conference on biometric compendium, Sept. 2010
[9]. P.U.Lahane, Prof. S.R.Ganorkar, “Fusion of Iris & Fingerprint Biometric for
Security Purpose", International Journal of Scientific & Engineering
Research Volume 3, Issue 8, August-2012.
[10]. Marfella L, Marasco E, Sansone C, “Liveness-Based Fusion pproaches in
Multibiometrics", IEEE International workshop, Sept. 2012.
[11]. Akhtar Z, Fumera G, Marcialis GL, Roli F, "Evaluation of serial and parallel
multibiometric systems under spoofing attacks" In: Proc. IEEE Fifth
International Conference on biometric compendium. Washington DC, USA,
Sept. 2012.
[12]. Babasaheb G. Patil, Shaila Subbaraman, “SVD-EBP Algorithm for Iris
Patten Recognition”, International Journal of Advanced Computer Science
and Applications, Vol. 2, No. 12, 2011.
Al-Mansour Journal/ Issue( 25 ) 2016 رﻮﺼﻨﻤﻟا ﺔﻠﺠﻣ/ دﺪﻌﻟا )25 (
- 23 -
[13]. M.Mani Roja, Sudhir Sawarkar, Ph.D, " Fingerprint Verification System – A
Fusion Approach", published in International Journal of Computer
Application (IJCA), 2011.
[14]. Battista Biggio, Zahid Akhtar, Giorgio Fumera, Gian Luca Marcialis, and
Fabio Roli, "Security Evaluation of Biometric Authentication Systems Under
Real Spoofing Attacks ", IEEE, Biometrics compendium, Volume:1 , Issue:
1, march, 2012.
[15]. Rajesh M. Bodade, Sanjay N. Talbar, " Iris Analysis for Biometric
Recognition Systems", Springer, 2014.
[16]. MMU1: http://pesona.mmu.edu.my/~ccteo/.
[17]. CASIA V1 :http://biometrics.idealtest.org/dbDetailForUser.do?id=1
Hanaa M. Ahmed*,Ph.D (Asst. Prof.) Bushra J. Abdulkareem
- 24 -
ﺔﯾﻮﮭﻟا ﻰﻠﻋ فﺮﻌﺘﻠﻟ ﻂﺋﺎﺳﻮﻟا دﺪﻌﺘﻣ ﻦﯿﺼﺣ مﺎﻈﻧ ﻢﯿﻤﺼﺘﻟ تﺎﯿﻣزراﻮﺧ حاﺮﺘﻗا
ﺔﯿﺼﺨﺸﻟا
أ.م.د.ﺪﻤﺣا ﻦﺴﺤﻣ ءﺎﻨھ*ﻢﯾﺮﻜﻟاﺪﺒﻋ رﺎﺒﺟ ىﺮﺸﺑ**
ﺺﻠﺨﺘﺴﻤﻟا
ﺔﺠﻟﺎﻌﻤﻟ ﺔﻟوﺎﺤﻣ ﻲھ ﺚﺤﺒﻟا اﺬھ
) (
ﺔﯾﻮﯿﺤﻟا ﺪﯾﺪﺤﺗ ﺔﻠﺣﺮﻣ ﻰﻋﺪﺗ ىﺮﺧا ﺔﻠﺣﺮﻣ)Liveness Detection (
ﮭﻟا ﻦﻣ ﻖﻘﺤﺘﻟا مﺎﻈﻧﻟا ةدﺪﻌﺘﻤﻟا يﻮﯿﺤﻟا ﺔﯾﻮﻂﺋﺎﺳﻮﻞﯾﺎﺤﺘﻟﺎﺑ قاﺮﺘﺧﻻا تﺎﻤﺠھ ﻦﻣ.
ﻦﯿﯾﻮﺘﺴﻣ ﻲﻓ مﺎﻈﻨﻠﻟ ﻖﻘﺤﺘﻟا ةءﺎﻔﻛ تﺮﺒﺘﺧُأ ، دﺪﻌﺘﻤﻟا ىﻮﺘﺴﻤﻟاو ﻂﺋﺎﺳﻮﻟا يدﺎﺣﻻا ىﻮﺘﺴﻤﻟا
ﻂﺋﺎﺳﻮﻟا . ةﺪﻋﺎﻗ تﺎﻧﺎﯿﺒﻟا ﺪﻋاﻮﻗ ﻦﻣ ﻦﯿﻋﻮﻨﻟو ﻦﯿﻌﻟا ﺔﯿﺣﺰﻗ ماﺪﺨﺘﺳﺎﺑ يدﺎﺣﻻا ىﻮﺘﺴﻤﻟا ﻦﻤﺿ مﺎﻈﻨﻟا ﺮﺒﺘﺧُأ ذإ
ﺋﺎﺳﻮﻟا ﺔﻌﻣﺎﺟ تﺎﻧﺎﯿﺑ ةدﺪﻌﺘﻤﻟا ﻂ)MMU DB (ﺎھدﺪﻋ نﺎﻛو180 ـﻟ ﺔﻨﯿﻋ30 تﺎﻧﺎﯿﺑ ةﺪﻋﺎﻗو ،ًﺎﺼﺨﺷ
ﺔﯿﻨﯿﺼﻟا ﺔﯿﻤﯾدﺎﻛﻻا)CASIA DB ( ﺎھدﺪﻋ نﺎﻛ90 ـﻟ ﺔﻨﯿﻋ30ًﺎﺼﺨﺷ . ﺔﻗد ﻰﻄﻋأو)99.44٪ ( ﻊﻣ
FAR)0.0277 (وFRR)0.0055 (ﻞﻟMMU DB ﺔﻗدو ،)97.77٪ ( ﻊﻣFAR)0.0333 ( و
FRR0.0222) (ﻞﻟCASIA DBرﺎﺒﺘﺧﻻاو ، ﻦﻋ ﺺﺨﺸﻟا ﻦﻣ ﻖﻘﺤﺘﻠﻟ يدﺎﺣﻻا ىﻮﺘﺴﻤﻟا ﻦﻤﺿ ﺮﺧﻵا
ﻦﻣ نﻮﻜﺘﺗ تﺎﻧﺎﯿﺒﻟا ةﺪﻋﺎﻗ ﻰﻠﻋ ﺖﯾﺮﺟأ ﻲﺘﻟا ﺔﻤﺼﺒﻟا ﻖﯾﺮﻃ60 ـﻟ ﺔﻨﯿﻋ15 ﺔﻗد ﺖﺤﻨﻣو ﺎﺼﺨﺷ95 ٪ ﻊﻣ
FAR(0.1%) وFRR(0.05%).
ﻦﻣ نﻮﻜﺘﺗ تﺎﻧﺎﯿﺑ ةﺪﻋﺎﻗ ﻰﻠﻋ ﻂﺋﺎﺳﻮﻟا دﺪﻌﺘﻤﻟا ىﻮﺘﺴﻣ ﻲﻓ مﺎﻈﻨﻟا ﺮﺒﺘﺧا60 ﺔﯿﺣﺰﻗ رﻮﺼﻟ ﺔﻨﯿﻋ
و ﻦﯿﻌﻟا60 ﻰﻟا رﺎﺒﺘﺧﻻا ﺞﺋﺎﺘﻧ ﺐﺴﺣ مﺎﻈﻨﻟا ﺔﻗد ﺖﻠﺻو ﺪﻗو ﻊﺒﺻا ﺔﻤﺼﺑ رﻮﺼﻟ ﺔﻨﯿﻋ100% ﻊﻣFAR
)0.0166 ( وFRR)0%.(
:
جﺎﻣﺪﻧﻻا.
____________________
*ﺔﯿﺟﻮﻟﻮﻨﻜﺘﻟا ﺔﻌﻣﺎﺠﻟا
**ﺔﻌﻣﺎﺠﻟا رﻮﺼﻨﻤﻟا ﺔﯿﻠﻛ