Iris Liveness Detection Methods in Mobile Applications
Ana F. Sequeira
1,2
, Juliano Murari
3
and Jaime S. Cardoso
1,2
1
INESC TEC (formerly INESC Porto), Porto, Portugal
2
Faculdade de Engenharia, Universidade do Porto, Porto, Portugal
3
Universidade Federal de S. Paulo, S
˜
ao Paulo, Brazil
Keywords:
Biometrics, Iris, Liveness Detection, Fake Database, Handheld Device.
Abstract:
Biometric systems are vulnerable to different kinds of attacks. Particularly, the systems based on iris are vul-
nerable to direct attacks consisting on the presentation of a fake iris to the sensor trying to access the system as
it was from a legitimate user. The analysis of some countermeasures against this type of attacking scheme is
the problem addressed in the present paper. Several state-of-the-art methods were implemented and included
in a feature selection framework so as to determine the best cardinality and the best subset that conducts to the
highest classification rate. Three different classifiers were used: Discriminant analysis, K nearest neighbours
and Support Vector Machines. The implemented methods were tested in existing databases for iris liveness
purposes (Biosec and Clarkson) and in a new fake database which was constructed for evaluation of iris live-
ness detection methods in the mobile scenario. The results suggest that this new database is more challenging
than the others. Therefore, improvements are required in this line of research to achieve good performance in
real world mobile applications.
1 INTRODUCTION
Biometric systems can offer several advantages over
classical security methods as they rather identify an
individual by what he is instead of based on some-
thing he knows or possesses. However, in spite of
its advantages, biometric systems have some draw-
backs, including: i) the lack of secrecy (e.g. every-
body knows our face or could get our fingerprints),
and ii) the fact that a biometric trait cannot be replaced
(no new iris can be generated if an impostor “steals”
it). Furthermore, biometric systems are vulnerable to
external attacks which could decrease their level of
security. Concerning these vulnerabilities we find in
the literature (Galbally et al., 2007) an analysis of the
eight different points of attack on biometric recogni-
tion systems previously identified (Ratha et al., 2001).
These points are illustrated in Fig. 1.
Figure 1: Architecture of an automated biometric verifica-
tion system. Possible attack points are numbered from 1 to
8, from (Galbally et al., 2007).
These attacks are divided into two main groups:
direct and indirect attacks.
Direct Attacks: the first vulnerability point in a
biometric security system is the possibility to gen-
erate synthetic biometric samples (for instance,
speech, fingerprints or face images) in order to
fraudulently access a system. These attacks at
the sensor level are referred to as direct attacks.
It is worth noting that in this type of attacks no
specific knowledge about the system operation is
needed (matching algorithm used, feature extrac-
tion, feature vector format, etc). Furthermore, the
attack is carried out in the analogue domain, out-
side the digital limits of the system, so the digital
protection mechanisms (digital signature, water-
marking, etc.) can not be used.
Indirect Attacks: this group includes all the re-
maining seven points of attack. Attacks 3 and 5
might be carried out using a Trojan Horse that
bypasses the feature extractor, and the matcher
respectively. In attack 6 the system database
is manipulated (a template is changed, added or
deleted) in order to gain access to the applica-
tion. The remaining points of attack (2, 4, 7 and 8)
are thought to exploit possible weak points in the
communication channels of the system, extract-
22
Sequeira A., Murari J. and Cardoso J..
Iris Liveness Detection Methods in Mobile Applications.
DOI: 10.5220/0004691800220033
In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 22-33
ISBN: 978-989-758-009-3
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
ing, adding or changing information from them.
In opposition to direct attacks, in this case the in-
truder needs to have some information about the
inner working of the recognition system and, in
most cases, physical access to some of the ap-
plication components (feature extractor, matcher,
database, etc.) is required.
Among the different existing biometric traits, iris
has been traditionally regarded as one of the most re-
liable and accurate. This fact has led researchers to
pay special attention to its vulnerabilities and in par-
ticular to analyze to what extent their security level
may be compromised by spoofing attacks. These at-
tacks may consist on presenting a synthetically gen-
erated iris to the sensor so that it is recognized as the
legitimate user and access is granted. The most com-
mon and simple approaches are those carried out with
high quality iris printed images (Ruiz-Albacete et al.,
2008). However, other more sophisticated threats
have also been reported in the literature such as the
use of contact lenses (Wei et al., 2008).
The development of iris liveness detection tech-
niques is crucial for the deployment of iris biometric
applications in daily life. The evolution in the use
of mobile devices in our society also raises the urge
for liveness solutions in the mobile biometric field.
To pursue this goal there is also a need for suitable
databases in which new methods can be tested.
In this work we implemented state-of-the-art
methods conceived to deal with spoofing attacks in
iris recognition, in particular, the use of printed im-
ages and contact lenses. The proposed method com-
prises a feature selection method, in order to deter-
mine the best cardinalities and respective subset of
features, with the use of state-of-the-art classifiers.
This framework intended to achieve the best classifi-
cation rates with only the “necessary” number of fea-
tures Two existing databases were tested, one com-
prising samples of printed iris images and another
comprising images of eyes with contact lenses. Tak-
ing in account the results obtained and the character-
istics of the databases available and the new trend of
performing biometric recognition in mobile scenar-
ios, we constructed a new fake iris database. This
database comprises printed copies of the original im-
ages (after being printed, the images were acquired
with the same device and in similar conditions as the
original ones). We found this new database to be more
challenging than the others.
This paper is organized as follows. In section 2
the concept of liveness detection in an iris recogni-
tion system is presented. In section 3, we explain
the algorithms implemented. In section 4 is presented
the database constructed with fake printed images for
testing liveness detection methods in iris recognition.
In section 5, the dataset of images is presented in 5.1,
the methodology used is presented in 5.2 and the re-
sults and their discussion are presented in 5.3. Finally,
in section 6 we draw some conclusions and sketch
some ideas for future works.
2 IRIS LIVENESS DETECTION
The problem of liveness detection of a biometric trait
can be seen as a two class classification problem
where an input trait sample has to be assigned to one
of two classes: real or fake. The key point of the pro-
cess is to find a set of discriminant features which per-
mits to build an appropriate classifier which gives the
probability of the sample vitality given the extracted
set of features (Galbally et al., 2012b).
Biometric recognition systems are vulnerable to
be spoofed by fake copies (Daugman, 2004), for in-
stance, fake finger tips made of commonly available
materials such as clay and gelatine. Iris is no excep-
tion. There are potential threats for iris-based sys-
tems, the main are (He et al., 2009):
Eye image: Screen image, Photograph, Paper
print, Video signal.
Artificial eye: Glass/plastic etc.
Natural eye (user): Forced use.
Capture/replay attacks: Eye image, IrisCode tem-
plate.
Natural eye (impostor): Eye removed from body,
Printed contact lens.
The feasibility of some attacks have been reported
by some researchers (Daugman, 1998; Daugman,
2004; Lee et al., 2005) who showed that it is actually
possible to spoof some iris recognition systems with
printed iris and well-made colour iris lens. Therefore,
it is important to detect the fake iris as much as possi-
ble (He et al., 2009).
Several liveness detection methods have been pre-
sented through the past recent years. In fact, anti-
spoofing techniques were presented that use physio-
logical properties to distinguish between real and fake
biometric traits. This is done in order to improve the
robustness of the system against direct attacks and to
increase the security level offered to the final user. Iris
liveness detection approaches can broadly be divided
into: i)software-based techniques, in which the fake
irises are detected once the sample has been acquired
with a standard sensor (i.e., features used to distin-
guish between real and fake eyes are extracted from
the iris image, and not from the eye itself), and ii)
IrisLivenessDetectionMethodsinMobileApplications
23
hardware-based techniques, in which some specific
device is added to the sensor in order to detect partic-
ular properties of a living iris such as the eye hippus
(which is the permanent oscillation that the eye pupil
presents even under uniform lighting conditions) or
the pupil response to a sudden lighting event (e.g.,
switching on a diode) (Galbally et al., 2012b). Ac-
cording to this author, even though hardware-based
approaches usually present a higher detection rate, the
software-based techniques have the advantage of be-
ing less expensive (as no extra device in needed), and
less intrusive for the user (very important character-
istic for a practical liveness detection solution). In
general, a combination of both type of anti-spoofing
schemes would be the most desirable approach to in-
crease the security level of biometric systems. (Gal-
bally et al., 2012b)
In this work we focus on software based tech-
niques since these are more easily and affordable ap-
plicable in real-world applications.
In the literature we found that the methods of live-
ness detection may be classified into four categories
based on the physical features of biometric and live-
ness data and the timing of measurement (Une and
Tamura, 2006). In this framework, the biometric data
are used in the iris recognition and the liveness data
are used in the liveness detection. We can itemize the
four categories:
Perfect matching model: Both biometric and live-
ness data are simultaneously obtained from the
same physical feature.
Simultaneous measuring model: Biometric and
liveness data are simultaneously obtained from
different physical features.
Same biometric measuring model: Biometric and
liveness data are obtained from the same physical
feature with different timings.
Independent measuring model: Biometric and
liveness data are obtained from different features
with different timings.
The ideal configuration of liveness detection for bio-
metrics recognition is represented by the perfect
matching model with the highest ability to distinguish
between live and fake irises (Kanematsu et al., 2007).
The potential of quality assessment to identify real
and fake iris samples acquired from a high quality
printed image has previously been explored as a way
to detect spoofing attacks (Galbally et al., 2012b).
Some quality based features have been used individu-
ally for liveness detection in traits such as iris (Kane-
matsu et al., 2007; Wei et al., 2008) or face (Li et al.,
2004). A strategy based on the combination of sev-
eral quality related features has also been used for
spoofing detection in fingerprint based recognition
systems (Galbally et al., 2012a) as well as in iris live-
ness detection (Galbally et al., 2012b). In this latter
work, a set of quality measures are used as iris live-
ness detection features to aid the classification of fake
or real iris images included in a framework of feature
selection. We find in literature that works concerning
the quality of iris images are often the starting point to
iris liveness detection techniques. One example is the
assessment of the iris image quality based on mea-
sures like occlusion, contrast, focus and angular de-
formation (Abhyankar and Schuckers, 2009), other is
the use of texture analysis of the iris (He et al., 2007),
among others like, for example, the analysis of fre-
quency distribution rates of some specific regions of
iris (Ma et al., 2003).
The way forward seems to be the development of
techniques for iris liveness detection that work well
independently of the particular characteristics of the
databases available nowadays. It is required to de-
velop and improve methods as well as to construct
new databases in less constrained conditions.
3 IMPLEMENTED METHODS
Some of the measures are obtained from the entire
eye image but others are extracted only from the iris
region, therefore a segmentation step is required. We
choose to make the segmentation process manually, in
order to ensure reasonable accuracy. The manual seg-
mentation is done by marking three differents points
in the image. The first point is the eye centre, i.e., we
consider a single centre for both pupil and iris. The
second point is marked in the pupil border and the
third in the iris border. With these points is possible
to determine the iris and pupil radius and then approx-
imate the contours as two concentric circles. With
the manual segmentation’s information we are able to
map the regions of interest which will be eventually
used by the liveness detection algorithms.
3.1 Algorithm 1 - High Frequency
Power
The High Frequency Power algorithm, which pro-
vides feature 1, works on the whole image and mea-
sures the energy concentration in the high frequency
components of the spectrum using a high pass con-
volution kernel of 8x8. The application of this con-
volution is a good Fourier Transform approximation
and works as high frequency spectral analysis, which
can be considered an estimator of focus (Daugman,
2002). The focus of a real iris, as it is a 3D volume,
VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications
24
is different from a fake iris focus, which has a 2D sur-
face. For more details on the method see (Galbally
et al., 2012b).
3.2 Algorithm 2 - Local Contrast
The Local Contrast algorithm, which provides fea-
ture 2, is based on bounding box that involves the
iris and the pupil. The bounding box is divided in
blocks of P × P and for each block it is applied the
Fast Fourier Transform (FFT) algorithm to extract the
medium power frequencies, which better represents
the contrast. The final value is given by the num-
ber of blocks with medium values (between 20 and
60) divided by the total number of blocks. This al-
gorithm was inspired in an occlusion estimation tech-
nique (Abhyankar and Schuckers, 2009) and it was
adapted for contrast estimation for iris liveness detec-
tion in (Galbally et al., 2012b) where more details can
be found.
3.3 Algorithm 3 - Global Contrast
The Global Contrast algorithm, which provides fea-
ture 3, explores the fact that parts extremely bright or
dark of the image are not useful and can be consid-
ered as noise. Thus, pixels near medium value (128
in 8-bit image) are considered of best contrast (Ab-
hyankar and Schuckers, 2009). In order to quantify
the contrast, the original pixels values are normalized
between 0 and 25 (Figure 2). Original pixels near
medium value will get higher values in the normalized
scale, as well as very low and very high values (< 10
and > 245) are normalized to 0. This measure was
presented in (Abhyankar and Schuckers, 2009) and
it was adapted for global contrast estimation for iris
liveness detection in (Galbally et al., 2012b) where
more details can be found.
Figure 2: Normalization function of the algorithm 3.
3.4 Algorithm 4 - Frequency
Distribution Rates
The Frequency Distribution Rates algorithm consists
in different mathematical combinations of three dif-
ferent parameters which consider respectively the
power of the low (F
1
), medium (F
2
), and high (F
3
)
frequencies (computed according to the 2D Fourier
Spectrum) from two iris subregions in the horizontal
direction. This subregions are illustrated in Figure 3.
Each subregion is subdivided in three circular concen-
tric region, which determine the three different fre-
quencies, i.e, for the first subregion, F
1
1
refers to the
central circle, F
1
2
refers to the middle circular ring and
F
1
3
refers to the outer circular ring, as depicted in Fig-
ure 4. The final F
1
is given by the average between the
two regions: F
1
=
F
1
1
+F
2
1
2
. The same is done to F
2
and
F
3
. More details on the method can be found in (Ma
et al., 2003; Galbally et al., 2012b).
Figure 3: Example of the subregions used in the algorithm
4 (Galbally et al., 2012b).
Figure 4: One of the regions of interest subdivided to cal-
culate the frequencies (Galbally et al., 2012b).
With the three final frequencies we extract seven
different combinations, represented in Table 1 (Ma
et al., 2003; Galbally et al., 2012b).
3.5 Algorithm 5 - Statistical Texture
Analysis
The Statistical Texture Analysis algorithm was devel-
oped as a contact lens countermeasure. The outer por-
tion of the colour contact lens (corresponding to re-
IrisLivenessDetectionMethodsinMobileApplications
25
Table 1: Extracted measures from the final frequencies.
Features no. Combination
4 F
1
+ F
2
+ F
3
5 F
2
/(F
1
+ F
3
)
6 F
3
7 F
2
8 F
1
9 (F
1
+ F
2
)/F
3
10 (F
1
F
2
)/F
3
gions closer to outer circle) provides the most useful
texture information for fake iris detection since this
section of the fake iris is insensitive to the pupil di-
lation (He et al., 2007). The region of interest is the
lower part of the iris in order to minimize the occlu-
sion by the eyelashes and eyelids, which in general
occurs in the upper iris portion. In order to achieve
invariance to translation and scale, the region of inter-
est is further normalized to a rectangular block of a
fixed size W × H (Figure 5).
(a) Original (b) Normalized
Figure 5: Region of interest used in the algoritm 5 (He et al.,
2007).
After the normalization, the GLCM (Gray Level
Co-occurence Matrix), one of the most proeminent
approaches used to extract textural features (Haral-
ick et al., 1973), is calculated. Four measures are
extracted: the mean (µ) and standard deviation (σ),
direct from the normalized region of interest, and the
contrast (con) and the energy (e) from the GLCM ma-
trix. These measures will provide features 11 to 14
and its values are given, respectively, by the equations
below:
µ =
1
W H
H
i=1
W
j=1
I(i, j) (1)
σ =
v
u
u
t
1
W H
H
i=1
W
j=1
(I(i, j)µ)
2
(2)
con =
N
i=1
N
j=1
(i j)
2
P(i, j) (3)
e =
N
i=1
N
j=1
P(i, j)
2
(4)
Where I denotes the normalized iris image, W
is the width of the normalized iris image, H is the
height of the normalized iris image. P is the co-
occurrence matrix and N denotes the dimension of the
co-occurrence matrix. For more details on the method
see (He et al., 2007).
3.6 Feature Selection
The algorithms implemented originated 14 different
features. Due to this dimensionality it is possible that
the best classification results are not obtained using all
the features, but a subset of them. It is convenient to
search for the optimum number and set of features.
To exhaustively test all possibilities is not feasible.
Therefore we use the “Sequential Forward Floating
Selection” (SFFS) (Pudil et al., 1994) to perform fea-
ture selection. The SFFS is basically a combination
of search methods such as “Plus-l-Minus-r” (Stearns,
1976) and Sequential Forward Search (SFS) (Whit-
ney, 1971). The appearance of “floating” comes from
the fact that the values l and r are not fixed, i.e., they
can “float”. Another aspect is the dominant direction
of search, including (forward) or excluding (back-
ward) characteristics (Pudil et al., 1994). We use
the Mahalanobis distance as criterion function. The
SFFS has shown to be competitive when compared to
other selection techniques (Jain and Zongker, 1997).
3.7 Classification
The classification results were obtained using three
classification methods: Discriminant Analysis (DA),
k-Nearest Neighbour (kNN) and Support Vector Ma-
chine (SVM).
4 MobBIOfake: IRIS IMAGES
CAPTURED WITH A
HANDHELD DEVICE
The MobBIOfake database was constructed upon the
MobBIO Multimodal Database (Blind Ref, 2013).
The MobBIO Multimodal Database comprises the
biometric data from 105 volunteers. Each individ-
ual provided samples of face, iris and voice. The
equipment used for the samples acquisition was an
Asus Transformer Pad TF 300T, with Android ver-
sion 4.1.1. The device has two cameras, one frontal
and one back camera. The camera used was the back
camera, version TF300T-000128, with 8 MP of reso-
lution and autofocus.
The iris images were captured in two different
lighting conditions, in a room with both natural and
VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications
26
artificial sources of light, with variable eye orienta-
tions and occlusion levels, so as to comprise a larger
variability of unconstrained scenarios. Each volun-
teer contributed with 16 images (8 of each eye) with a
300 × 200 resolution. Some examples of iris images
are depicted in Figure 6.
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 6: Iris images from MobBIO database illustrating
different kinds of noise: a) Heavily occluded; b) Heavily
pigmented; c) Glasses reflection; d) Glasses occlusion; e)
Off-angle; f) Partial eye; g) Reflection occlusion and h)
Normal.
MobBIOfake
The MobBIOfake is composed by a subset of 800
iris images from MobBIO and its corresponding fake
copies, in a total of 1600 iris images. The fake sam-
ples were obtained from printed images of the original
ones captured with the same handheld device and in
similar conditions. From the original dataset of im-
ages
The aim of constructing such a database is, on one
hand, to fulfil the necessity of databases and, on the
other hand to broad the acquisition conditions of the
images. The number and variety of databases for iris
liveness detection is somewhat limited so the fact that
these images were captured with a portable device and
are RGB images come as a novelty and makes it possi-
ble to evaluate liveness methods in this new upcoming
scenario.
The construction of the MobBIOfake upon the
MobBIO iris images subset comprised several steps.
The images of each volunteer were joined in a single
image, as shown in Figure 7.
A preprocessing (contrast enhancement) was ap-
plied using GIMP software (GIMP, 2008) to the im-
age. This enhancement is believed to improve the
quality of the fake sample (Ruiz-Albacete et al.,
2008). After this, the images were printed in a profes-
sional printer using high quality photographic paper.
At this point we were able to capture the images. Each
individual image (a image of one single eye) was ac-
quired using the same portable device and in similar
lighting conditions as the original ones were captured,
as illustrated in Figure 8.
Figure 7: MobBIOfake construction: joint images of one
volunteer.
Figure 8: MobBIOfake construction: fake samples acquisi-
tion.
Finally, the individual eye images were cropped
and resized to fix dimensions. An example of a real
image and its copy is depicted in Figure 9.
(a) Real image (b) Fake image
Figure 9: Corresponding real and fake images of MobBIO.
5 EXPERIMENTAL SETUP
5.1 Datasets
The implemented methods were tested in three
datasets. One was a database constructed within our
work, the MobBIOfake, comprised of 800 iris images
IrisLivenessDetectionMethodsinMobileApplications
27
and its corresponding fake copies, captured with the
same portable device and in similar conditions as the
original ones. The description of the construction of
this dataset was detailed in section 4.
The other two databases, described below, are the
Biosec database (Fierrez et al., 2007), composed by
real iris images and the corresponding fake printed
images; and the Clarkson database (S. Schuckers and
Yambay, 2013) comprising real iris images and fake
ones obtained by the use of contact lenses.
Biosec
The Biosec database was created at the Polytech-
nic University of Madrid (UPM) and the Universitat
Polit
`
ecnica de Catalunya (UPC). The images were ac-
quired in an office room with a large table for hard-
ware of biometric recognition system and two chairs,
one for the donor and one for the supervisor of the ac-
quisition process. Environmental conditions such as
lighting and noise, were not controlled to simulate a
real situation (Fierrez et al., 2007). To construct the
false database original images are pre-processed and
printed on paper using a commercial printer. Then the
printed images were presented to the iris sensor, ob-
taining the fake copy. This study considered different
combinations of pre-processing, printing equipment
and paper type (Ruiz-Albacete et al., 2008). There-
fore, the database used consists of real and fake iris
images and follows the same structure as the original
database. Biosec dataset comprises a total of 1600
images: 800 real images and its corresponding 800
fake samples. All images are in greyscale and its di-
mensions are 640 × 480 (Galbally et al., 2012b). The
two eyes of the same individual are considered as dif-
ferent users. The acquisition of both real and fake
samples were made using the sensor LG IrisAccess
EOU3000 (Ruiz-Albacete et al., 2008).
Clarkson
The subset of Clarkson database that we used was
made available under request and contains 270 real
iris images and 400 fake iris images. The fake sam-
ples are images of eyes with contact lenses compris-
ing 14 models of contact lenses. There are two differ-
ent lighting conditions in the database, which was ac-
quired by video (capturing 100 frames and with vari-
ation of focus). The Clarkson database was made
available to participants of the LivDet-2013 compe-
tition (S. Schuckers and Yambay, 2013) .
5.2 Methodology
The proposed method is depicted in Figure 10.
Figure 10: Steps of the proposed method.
The first step of the method is the segmentation, in
this case it was made manually so as to purge our re-
sults from the errors associated with automatic meth-
ods of iris segmentation. Although it has to be noted
that in a real world application this step needs to be
necessarily automatized.
The second step was the feature extraction. This
comprises the application of the methods described in
subsections 3.1, 3.2, 3.3, 3.4 and 3.5.
Next step was the feature selection. This com-
prises the application of the method Sequential For-
ward Floating Search, described in subsection 3, be-
fore applying the classifiers to evaluate the proposed
methods. We ran the SFFS to obtain the best subset
for each cardinality from = 2 to = 12 features.
This range is defined by the selection method when
considering a set amount of 14 features.
The last step was the classification. We used
three state-of-the-art classifiers: Discriminant Anal-
ysis (DA), k-Nearest Neighbours (kNN) and Sup-
port Vector Machines (SVM). For each cardinality,
(= 2,...,12), the results of classification were ob-
tained calculating the average of the results of 50 runs
for classification of the images based on the corre-
sponding best features. The results were obtained
by randomly dividing, in each run, the 1600 samples
in two sets: 1000 samples for training and 600 for
testing. The parameter k in kNN was optimized using
cross-validation, and tested in the interval [1,20] by
steps of 1. For the SVM, we used a polynomial ker-
nel and also used cross-validation for optimization of
the parameters. It was performing a ”grid-search” on
the parameters of the models. Exponentially growing
sequences of C were tested: C = 2
N
with N varying
between 1 and 15. For the polynomial degree, d,
values tested were: d = 1,2,3,4,5.
For the evaluation of the accuracy of the features
extracted in discriminating between fake and real im-
ages, we used the Equal Error Rate (EER). The EER
is obtained when the false acceptance rate (FAR)
VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications
28
and the False Rejection Rate (FRR) are equal. For
the classification results we use the missclassification
rate averaged over the 50 runs.
5.3 Experimental Results and
Discussion
In this section we present the results obtained by the
proposed method for iris liveness detection. The al-
gorithms applied return a set of 14 different features.
The first step was to analyse individually each of
the 14 different features, for each image dataset. By
the analysis of the histogram obtained for fake and
real images we can from that moment have a hint
about which features will be good discriminative be-
tween fake and real images. For each histogram, the
threshold obtained considering equal error rate (EER)
allow us to determine the minimum error associated
with that feature, for the considered dataset. In Ta-
ble 2 are shown the minimum error values associated
with each feature for each database.
Table 2: Minimum error associated with each feature for
each database.
Associated Error (%)
Feature no. Biosec MobBIOfake Clarkson
1 (alg1) 31.3 31.8 35.4
2 (alg2) 17.2 31.2 26.3
3 (alg3) 21.1 21.9 26.7
4 (alg4) 15.1 40.8 31.5
5 (alg4) 43.6 27.9 36.0
6 (alg4) 14.4 42.9 31.5
7 (alg4) 15.6 33.0 31.3
8 (alg4) 15.2 30.8 31.9
9 (alg4) 39.9 26.4 36.3
10 (alg4) 15.8 29.3 31.8
11 (alg5) 22.9 27.2 32.2
12 (alg5) 17.2 35.7 37.8
13 (alg5) 13.2 35.7 39.1
14 (alg5) 25.8 29.3 37.9
It is clear from the Table 2 that each dataset has a
variable behaviour concerning the features obtained.
Simply observing the minimum errors (emphasized in
the table) we may conclude that the “best” feature for
one database is not necessarily the best for any of the
others.
To enlighten a bit more how the discriminative
power of each feature was analysed we show in Fig-
ure 11 the best and worse feature for each database.
This histograms illustrate clearly the efficiency of
each feature in discriminating real images from fake
images. For some features the lines for fake and real
images are well separated while for others this lines
are too much coincident compromising the separabil-
ity between the two classes.
The next step was to perform feature selection as
to avoid possible redundancies in the set of features.
Reducing the number of features to the strictly nec-
essary will improve the computational efficiency of
the method. In Table 3 are shown the best subset of
features for each cardinality, from 2 to 12, for each
database.
Again we observe the diversity of the results ob-
tained for each database. Another relevant aspect is
the combinations of features, in some cases we ob-
serve that features that individually do not have a
good performance when combined provide the best
subsets. This fact reinforces the pertinence of using a
method for feature selection.
Finally, Tables 4, 5 and 6 show the classification
results for each cardinality, for each database, ob-
tained using the best subset determined by the feature
selection (averaged over the 50 runs).
Table 4: Classification results for Biosec (classification er-
rors in %).
DA kNN SVM
µ σ µ σ µ σ
2 10.24 0.99 10.34 1.23 10.16 1.17
3 4.36 0.92 4.43 0.83 4.68 0.67
4 0.52 0.24 0.76 0.27 0.77 0.34
5 1.14 0.48 0.89 0.34 0.78 0.29
6 0.85 0.38 0.56 0.27 0.54 0.26
7 0.92 0.32 0.40 0.20 0.57 0.31
8 0.93 0.30 0.47 0.25 0.56 0.33
9 1.80 0.47 1.11 0.34 0.87 0.28
10 1.28 0.39 0.46 0.28 0.73 0.26
11 1.30 0.28 0.40 0.24 0.52 0.33
12 1.68 0.59 0.37 0.24 0.50 0.26
Table 5: Classification results for MobBIOfake (classifica-
tion errors in %).
DA kNN SVM
µ σ µ σ µ σ
2 18.03 1.31 16.52 1.47 17.29 1.13
3 29.29 1.54 17.34 1.25 20.69 1.83
4 17.50 1.42 12.62 1.18 14.36 0.94
5 18.03 1.30 13.00 1.26 14.18 1.20
6 18.29 1.44 12.82 1.02 14.33 1.15
7 18.88 1.35 13.27 1.35 14.55 1.27
8 18.31 1.11 13.52 1.21 13.74 1.28
9 18.34 1.28 14.44 1.25 14.11 2.22
10 17.58 1.38 13.92 1.19 13.39 1.01
11 17.15 1.35 14.51 1.44 12.53 1.39
12 17.25 1.12 14.45 1.21 12.50 1.21
IrisLivenessDetectionMethodsinMobileApplications
29
(a) Biosec - Feature 13 (best result). (b) Biosec - Feature 5 (worse result).
(c) MobBIOfake - Feature 3 (best result). (d) MobBIOfake - Feature 6 (worse result).
(e) Clarkson - Feature 2 (best result). (f) Clarkson - Feature 13 (worse result).
Figure 11: Histograms for the best result/smallest minimum error (left) and worse result/biggest minimum error (right) for
each database.
The overall best results were obtained for Biosec
database and the worst overall results were obtained
for MobBIOfake. This is not a surprising result
since we expected this latter to be a more challeng-
ing database due to its characteristics. It was noto-
rious from the study of features individually that this
database presented the worse results. We interpret this
fact as a sign that new databases were needed for the
research of liveness in new scenarios.
Comparing the classifiers, we conclude that DA
led to worse results. This fact is also not surprising
since this classifier may be considered simpler than
VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications
30
Table 3: Best subset of features for each cardinality, for each database.
Subset of features
Biosec MobBIOfake Clarkson
2 [1 6] [3 10] [3 14]
3 [1 2 11] [5 8 10] [9 11 14]
4 [1 2 6 11] [3 5 8 10] [3 9 11 14]
5 [1 2 6 11 13] [3 4 7 8 10] [2 3 9 11 14]
6 [1 2 6 11 12 13] [3 4 7 8 9 10] [1 2 3 9 11 14]
7 [1 2 5 6 11 12 13] [3 4 5 7 8 9 10] [1 2 3 9 11 12 14]
8 [1 2 5 6 7 11 12 13] [3 4 5 7 8 9 10 13] [1 2 3 5 9 11 12 14]
9 [1 2 5 6 7 9 10 11 13] [3 4 5 7 8 9 10 12 13] [1 2 3 5 9 11 12 13 14]
10 [1 2 5 6 7 9 10 11 12 13] [3 4 5 7 8 9 10 11 12 13] [1 2 3 4 5 6 9 11 12 14]
11 [1 2 3 5 6 7 9 10 11 12 13] [2 3 4 5 7 8 9 10 11 12 13] [1 2 3 4 5 6 9 11 12 13 14]
12 [1 2 3 5 6 7 9 10 11 12 13 14] [1 2 3 4 5 7 8 9 10 11 12 13] [1 2 3 4 5 6 7 9 11 12 13 14]
Table 6: Classification results for Clarkson (classification
errors in %).
DA kNN SVM
µ σ µ σ µ σ
2 29.25 2.48 18.86 2.38 21.63 2.65
3 23.38 2.25 18.38 2.06 16.29 2.55
4 20.15 2.61 10.64 1.85 9.20 2.16
5 17.53 2.47 7.82 1.90 7.03 1.62
6 15.74 2.08 8.89 1.71 7.45 2.10
7 14.36 2.01 8.32 2.16 6.77 1.41
8 14.55 1.88 9.50 1.63 7.57 1.75
9 12.99 2.49 8.88 1.74 6.92 2.22
10 11.03 1.92 7.86 1.65 5.89 1.86
11 11.02 2.11 7.17 1.32 5.74 1.56
12 14.33 3.17 7.51 1.82 5.69 1.65
the others. The kNN achieved the overall best results.
Now, analysing each database per se, we observe
for the Biosec database that the best average classi-
fication rate was obtained with kNN. In terms of the
cardinality of features, we note that the best average
result, 0.37%, obtained with a subset of 12 features, is
followed closely by the value 0.4% with only a cardi-
nality of 7. And this again encourages the use of fea-
ture selection since the computational time and com-
plexity is lowered if we lower the number of features.
Concerning the MobBIOfake, undoubtedly the
classification errors obtained are higher than the other
databases, what is not unexpected as we already re-
ferred. The best average results were obtained with
the SVM classifier, 12.50% , but corresponding to a
high cardinality, 12. Not very far form this value we
find a subset with much lower cardinality, 4, for the
kNN, with an average error of 12.62%.
Analysing the Clarkson results, we note that the
combination of features improved considerably the
results when compared with the performance of the
features individually. The best average result was ob-
tained with SVM, 5.69%, this value is not as good
as the Biosec best result but is better than the Mob-
BIOfake one, but unfortunately it reffers to a subset
of high cardinality, 12. However, we may find a 3
rd
-
best value with a cardinality of 7.
6 CONCLUSIONS AND FUTURE
WORK
The actuality of the iris liveness detection topic is un-
questionable. As the field of application of iris recog-
nition broads, to embrace the demands of a society
highly dependant on mobile and portable devices, the
necessity of improving the security urges. To achieve
new methods it is also necessary to explore new sce-
narios of image acquisition and this leads to the neces-
sity of adequate, freely, public available databases.
In this work, we constructed a new database for
iris liveness detection purposes with images acquired
in unconstrained conditions and with a handheld de-
vice. This database was tested for state-of-the-art
methods and the results were compared with the re-
sults obtained for two existing and tested databases.
The MobBIOfake database proved itself to be more
challenging and our results may not be considered sat-
isfactory but lead to a new more challenging scenario.
Published works present methods tested with ex-
isting databases which achieve excellent results, (0%
error classification rate). However, we note that some
of these methods are closely connected with the par-
ticular database characteristics. The results with this
database did not achieve that excellent accuracy, but
we consider this justifiable by the fact that we avoided
the use of methods strongly dependent on the images
used, such as ratios of iris and pupil radius or areas,
among others.
For future work, we foresee the necessity of im-
proving the existing methods and develop new ones
IrisLivenessDetectionMethodsinMobileApplications
31
more suitable to the new imaging scenarios. Another
aspect to invest is the segmentation step which prefer-
ably should be automatic, however, the iris segmenta-
tion problem constitutes by itself a whole new set of
challenges.
We participated in an iris liveness competition, the
“LivDet Competition 2013” (Clarkson University and
of Technology, 2013a), held as part of the IEEE BTAS
2013
1
. We applied this methodology combined with
an automatic segmentation method (Monteiro et al.,
2013; Monteiro et al., 2014) and achieved the first
place
2
.
ACKNOWLEDGEMENTS
The first author would like to thank the Fundac¸
˜
ao
para a Ci
ˆ
encia e Tecnologia (FCT) - Portugal for the
financial support for the PhD grant with reference
SFRH/BD/74263/2010. The second author would
like to thank the National Council for Scientific and
Technological Development (CNPq) - Brazil.
REFERENCES
Abhyankar, A. and Schuckers, S. (2009). Iris quality assess-
ment and bi-orthogonal wavelet based encoding for
recognition. Pattern Recognition, 42(9):1878 – 1894.
Blind Ref, B. R. (2013). Reference removed for blind re-
view.
Clarkson University, N. D. U. and of Technol-
ogy, W. U. (2013a). Liveness Detection-
iris competition 2013. IEEE BTAS 2013.
http://people.clarkson.edu/projects/biosal/iris/.
Clarkson University, N. D. U. and of Technol-
ogy, W. U. (2013b). Liveness Detection-
iris competition 2013. IEEE BTAS 2013.
http://people.clarkson.edu/projects/biosal/iris/results.php.
Daugman, J. (1998). Recognizing people by their iris
patterns. Information Security Technical Report,
3(1):33–39.
Daugman, J. (2002). How iris recognition works. In Inter-
national Conference on Image Processing, volume 1,
pages I–33 – I–36.
Daugman, J. (2004). Iris recognition and anti-spoofing
countermeasures. In 7-th International Biometrics
conference.
Fierrez, J., Ortega-Garcia, J., Torre Toledano, D., and
Gonzalez-Rodriguez, J. (2007). Biosec baseline cor-
pus: A multimodal biometric database. Pattern
Recognition, 40(4):1389–1392.
1
http://www.btas2013.org/
2
http://people.clarkson.edu/projects/biosal/iris/results.php
Galbally, J., Alonso-Fernandez, F., Fierrez, J., and Ortega-
Garcia, J. (2012a). A high performance finger-
print liveness detection method based on quality re-
lated features. Future Generation Computer Systems,
28(1):311–321.
Galbally, J., Fierrez, J., and Ortega-Garcia, J. (2007). Vul-
nerabilities in biometric systems: attacks and recent
advances in liveness detection. DATABASE, 1(3):4.
Galbally, J., Ortiz-Lopez, J., Fierrez, J., and Ortega-Garcia,
J. (2012b). Iris liveness detection based on quality
related features. In 5th IAPR International Conference
on Biometrics (ICB), pages 271–276. IEEE.
GIMP, G. (2008). Image manipulation program. User Man-
ual, Edge-Detect Filters, Sobel, The GIMP Documen-
tation Team.
Haralick, R. M., Shanmugam, K., and Dinstein, I. H.
(1973). Textural features for image classification.
Systems, Man and Cybernetics, IEEE Transactions,
(6):610–621.
He, X., An, S., and Shi, P. (2007). Statistical texture
analysis-based approach for fake iris detection using
support vector machines. In Advances in Biometrics,
pages 540–546. Springer.
He, X., Lu, Y., and Shi, P. (2009). A new fake iris detec-
tion method. In Advances in Biometrics, pages 1132–
1139. Springer.
Jain, A. and Zongker, D. (1997). Feature selection: Evalua-
tion, application, and small sample performance. Pat-
tern Analysis and Machine Intelligence, IEEE Trans-
actions, 19(2):153–158.
Kanematsu, M., Takano, H., and Nakamura, K. (2007).
Highly reliable liveness detection method for iris
recognition. In SICE, 2007 Annual Conference, pages
361–364. IEEE.
Lee, E., Park, K., and Kim, J. (2005). Fake iris detection
by using purkinje image. In Advances in Biometrics,
volume 3832 of Lecture Notes in Computer Science,
pages 397–403. Springer Berlin / Heidelberg.
Li, J., Wang, Y., Tan, T., and Jain, A. K. (2004). Live face
detection based on the analysis of fourier spectra. In
Defense and Security, pages 296–303. International
Society for Optics and Photonics.
Ma, L., Tan, T., Wang, Y., and Zhang, D. (2003). Per-
sonal identification based on iris texture analysis. Pat-
tern Analysis and Machine Intelligence, IEEE Trans-
actions, 25(12):1519–1533.
Monteiro, J. C., Oliveira, H. P., Sequeira, A. F., and Car-
doso, J. S. (2013). Robust iris segmentation under un-
constrained settings. In Proceedings of International
Conference on Computer Vision Theory and Applica-
tions (VISAPP), pages 180–190.
Monteiro, J. C., Sequeira, A. F., Oliveira, H. P., and Car-
doso, J. S. (2014). Robust iris localisation in challeng-
ing scenarios. In CCIS Communications in Computer
and Information Science. Springer-Verlag.
Pudil, P., Novovi
ˇ
cov
´
a, J., and Kittler, J. (1994). Floating
search methods in feature selection. Pattern recogni-
tion letters, 15(11):1119–1125.
Ratha, N. K., Connell, J. H., and Bolle, R. M. (2001). An
analysis of minutiae matching strength. In Audio-and
VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications
32
Video-Based Biometric Person Authentication, pages
223–228. Springer.
Ruiz-Albacete, V., Tome-Gonzalez, P., Alonso-Fernandez,
F., Galbally, J., Fierrez, J., and Ortega-Garcia, J.
(2008). Direct attacks using fake images in iris verifi-
cation. In Biometrics and Identity Management, pages
181–190. Springer.
S. Schuckers, K. Bowyer, A. C. and Yambay, D.
(2013). Liviness Detection - Iris Competition 2013.
http://people.clarkson.edu/projects/biosal/iris/.
Stearns, S. D. (1976). On selecting features for pattern clas-
sifiers. In Proceedings of the 3rd International Joint
Conference on Pattern Recognition, pages 71–75.
Une, M. and Tamura, Y. (2006). liveness detection tech-
niques. IPSJ Magazine, 47(6):605–608.
Wei, Z., Qiu, X., Sun, Z., and Tan, T. (2008). Counterfeit
iris detection based on texture analysis. In ICPR 2008.
19th International Conference on Pattern Recogni-
tion., pages 1–4. IEEE.
Whitney, A. W. (1971). A direct method of nonparametric
measurement selection. Computers, IEEE Transac-
tions, 100(9):1100–1103.
IrisLivenessDetectionMethodsinMobileApplications
33