See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/286108098

Liveness detection for biometric authentication in mobile applications

Conference Paper · October 2014

DOI: 10.1109/CCST.2014.6986982

CITATIONS READS
24 2,529

3 authors:

Zahid Akhtar Christian Micheloni

State University of New York Polytechnic Institute University of Udine
129 PUBLICATIONS 2,493 CITATIONS 232 PUBLICATIONS 4,878 CITATIONS

SEE PROFILE SEE PROFILE

G.L. Foresti
University of Udine
328 PUBLICATIONS 6,572 CITATIONS

SEE PROFILE

All content following this page was uploaded by Zahid Akhtar on 17 November 2017.

The user has requested enhancement of the downloaded file.

 Liveness Detection for Biometric Authentication in
Mobile Applications

Zahid Akhtar, Christian Michelon and Gian Luca Foresti

University of Udine, Italy
Email: {zahid.akhtar, christian.micheloni, gianluca.foresti}@uniud.it

Abstract—The vulnerability of biometric authentication sys-

tems to spoofing attacks is now a widely accepted fact. A spoofing
attack occurs when an impostor attempts to masquerade as
genuine user by falsifying biometric data and thereby gaining
illegitimate access. Several liveness detection methods have been
proposed, which consist in determining whether there is a live
person in front of the biometric sensor or an artificial replica.
But, the problem is still unresolved owing to high level difficulty
in determining efficient features with low computational cost to
detect the spoofing attacks. In addition, existing methods are not
particularly targeted for liveness detection in mobile biometric
applications, thus mainly inapplicable for portable devices. Hence,
we present a multi-biometric approach, that can detect face, Fig. 1. Example of images captured from real face, iris and fingerprint (upper
iris and fingerprint spoofing attacks in mobile applications, row) and from spoofing attacks (lower row): (From left to right) face photo
by employing a novel real-time feature description based on attack, iris printed contact lens attack, iris photo attack, fingerprint latex attack.
order permutations, named Locally Uniform Comparison Image
Descriptor (LUCID). LUCID is computable in linear time with to masquerade as genuine user by replicating the biometrics
respect to number of pixels and does not require floating point
of genuine user and thereby gaining illegitimate access and
computation, beside the fact that typical mobile devices perform
poorly for floating point applications. Our approach is therefore advantages [1] [2]. For instance, German Hackers group Chaos
exclusively simple, fast and efficient, making it thus highly suit- Computer Club has exhibited in 2013 that fingerprint scanner
able for mobile devices. Moreover, contrary to existing schemes, in Apple’s iPhone 5s can be fooled by an artificial fingerprint,
our method utilize the same lone image descriptor technique while in Black Hat 2009 conference a research team from
effectively for three biometric traits, i.e. face, iris and fingerprint, University of Hanoi (Vietnam) has shown that how easily they
liveness detection. Additionally, our method uses only one image can spoof facial images of the legitimate user and bypass
for liveness detection, which can also be used for recognition. Lenovo, Asus and Toshiba laptops’ Face Recognition and thus
Experiments on publicly available face, iris and fingerprint data gaining access to laptops. Fig. 1 depicts an example of images
sets with real spoofing attacks show promising results. captured via biometric sensors from ‘live’ and ‘spoofed’ face,
Keywords—Biometrics, Spoofing Attacks, Liveness Detection, iris and fingerprint of genuine user. It can be seen that the
Mobile Devices, Face Recognition, Fingerprint Recognition, Iris spoofing attack images are very similar to the real ones.
Recognition.
Spoofing attacks have a great practical relevance because
they do not require advanced technical skills and hence poten-
I. I NTRODUCTION
tial number of attackers is very large. The effectiveness of these
These days mobile devices with built-in sensors like cam- attacks relies on the fact that biometric sensors are unable to
era have become ubiquitous, which are indeed widely used detect whether the submitted biometric trait is ‘fake/spoofed’
worldwide not only for basic communications but also as a or ‘live’. Moreover, it is well-known fact that attempts to
tool to access and deal with personal affairs and data with the fool biometric systems goes on increasing with the growth
help of biometrics. Biometrics is used as a natural alternative in use of biometric systems. Therefore, it is an urgent need
to the traditional means, such as a password, of access control to address biometric spoofing attacks on mobile devices to
by using biological or behavioral characteristics of the user. enhance security and robustness of the systems, and hence
For instance, iPhone 5s, Motorola Atrix and Fujitsu F505i bringing biometric applications into practical use.
come with touch sensor which reads user’s fingerprint and
Liveness detection [3][4][5] has been suggested as a typ-
automatically unlocks the phone. Also, Android KitKat mobile
ical countermeasure against spoofing attacks, which aims at
OS and Lenovo, Asus and Toshiba laptops come with embeded
detecting whether the submitted biometric trait is a live or an
biometric systems that authenticate users by their faces. A
artificial one, by observing physiological signs of life such as
recent study estimates that by 2018 almost 3.4 billion users
eye blinking, precipitation, etc. Liveness detection is performed
will have biometric features on their mobile devices.
by either software module-based on signal processing or
While many efforts are being focused on the performance hardware module-embedded into input device itself. Software-
improvement of biometrics on mobile platforms, the issue based solutions are the most interesting and challenging ones,
related to their vulnerabilities to spoofing attacks are generally since they do not employ any additional and possibly invasive
overlooked. A spoofing attack occurs when an impostor tries measurements such as blood pressure etc. Therefore, this work
is also focused on the software-based systems. For instance, Pan et al. [9] exploited the fact that human eye-
blink occur once every 2-4 seconds and proposed a technique
Several countermeasures have been proposed in the litera-
for photo-spoofing using an undirected conditional random
ture so far, but none of them have shown to reach a very low
field framework to model eye-blinking. Kollreider et al. [10]
error rates. Also, existing methods are mainly trait dependent,
developed a livness detection approach based on a short
such that feature descriptors proposed for face spoofing may
sequence of images using a binary detector, which captures and
not work effectively if used for iris or fingerprint spoofing and
tracts the subtle movements of different selected facial parts
vice versa. Likewise, most of the methods are not applicable
using a simplified optical flow analysis followed by an heuristic
for real-time or mobile applications, since they generally use
classifier. The same authors presented in [11] a method to
complex features and/or have high computational cost. To the
combine scores from different experts systems which concur-
best of our knowledge, no liveness detection work (except
rently observe the 3-D face motion approach introduced in the
[5]) has been particular targeted for mobile applications. Thus,
former work as liveness attributes like eye-blinks or mouth
there is a need and room for method proposals to make
movements. In a similar fashion, Bao et al [12] also used
biometric applications on mobile devices secure.
optical-flow to estimate motion for detecting attacks produced
Based on the above motivations, in this paper we propose with planar media such as prints or screens. In another work,
a novel software-based liveness detection scheme built on Pan et al. [13] augmented several countermeasures to include
image descriptor classification algorithm, the LUCID (Locally a scene context matching that helps preventing video-spoofing
Uniform Comparison Image Descriptor) [6], which can be used in stationary face-recognition systems.
in multiple biometric systems on mobile devices. LUCID is
Texture analysis liveness detection methods examine the
simple, fast and a novel approach to real-time image feature
skin properties, such as skin texture and skin reflectance, under
description based on linear time permutation obtained by the
the assumption that surface properties of real faces and prints,
ordering of RGB values of the image patches. LUCID does not
e.g. pigments, are different. Examples of detectable texture
require floating point computation, which is desirable condition
patterns due to artifacts are printing failures or blurring. Li
for mobile processors. LUCID is insensitive to monotonic
et al. [14] described a method for print-attack face spoofing
photometric transformations and noise, thus we believe it could
by exploiting differences in the 2-D Fourier spectra of live
be useful in determining the differences between a live and a
and spoof images. The method only works well for down-
fake trait due to the information loss during the spoofing attack
sampled photos of the attacked identity, but likely fails for
production.
higher-quality samples. Tan et al. [15] considered Lambertian
The proposed countermeasure analyzes local features of reflectance model with difference-of-Gaussians (DoG) to de-
face, fingerprint and iris images using LUCID and encodes rive differences of motion deformation patterns between 2-D
local patterns into enhanced feature vector. The results are face photos presented during spoofing attacks and 3-D live
then fed to Support Vector Machine (SVM) classifier that faces. Kose et al. [16] exploited Retinex reflectance models
determines whether input biometric trait is coming from live to differentiate spoofed and live faces. In [1][17][18], authors
person or not. Experiments on publicly available data sets developed micro-texture analysis based methods to detect
containing several real and fake faces, irises and fingerprints printed photo-attacks. One limitation of presented methods
show promising results. is the requirement of reasonably sharp input image. Other
countermeasures against face spoofing attacks include multi-
The paper is organized as follows. Section II summarizes spectral imaging, which analyze the reflectance of object
works on face, fingerprint and iris spoofing and anti-spoofing. surfaces and thus discriminate live faces from fake ones [19].
The proposed approach is described in Section III. The data
sets, experimental protocol and results are reported in Section
IV. Preliminary conclusions are then drawn in Section V. B. Iris Spoofing
It is generally conceded that iris recognition is the most
II. L ITERATURE S URVEY accurate. However, iris recognition systems may be deceived
Biometric systems are known to perform weak against by (i) an iris photo of a genuine user; (ii) an iris video of a
spoofing attacks for a long time [2][7][8]. In this short survey, genuine user; (iii) a printed contact lens of a genuine user;
we present overview of face, iris and fingerprint spoofing with (iv) an artificial eye made of glass or plastic of a genuine
their liveness detection methods. user; (v) a reverse-engineered iris image from the template of
a genuine user; (vii) a natural eye removed from the body of
a genuine user. Iris liveness detection schemes can be broadly
A. Face Spoofing
categorized in four categories: frequency spectrum analysis,
Despite of great advancement in face recognition systems, reflectance analysis, dynamics analysis and texture analysis.
face spoofing still poses a serious threat. They may be spoofed
Frequency spectrum analysis based liveness detection tech-
by (i) a photo of a genuine user; (ii) a video of a genuine user;
niques use the frequency spectrum information assuming the
(iii) a 3D face model of a genuine user. Typical countermea-
existence of artefacts in the images of spoofing attacks.
sure, i.e., face liveness detection, approaches can be coarsely
Daugman et al. [7] and Ma et al. [20] suggested detecting
classified in two categories based on clues used for spoofing
the printed iris via spectrographic analysis based on FFT
attack detection: motion analysis and texture analysis.
(Fast Fourier Transform). However, these methods have some
Motion analysis methods broadly try to detect spontaneous serious drawbacks, originating from Shanon’s theory. First,
movement clues generated when two dimensional counterfeits they fail if the resolution of the printing device being used
are presented to the system, such as photographs or videos. for counterfeit fabrication is more than twice the resolution of
the biometric image acquisition camera. Second, if the input based approach for detecting fingerprint liveness by analyzing
counterfeit iris image is defocused and blurred purposely, the perspiration pattern from two fingerprints captured at zero and
spoofed iris may be accepted as live one. He et al. [21] five seconds. In [30] authors stated that perspiration based
designed a method to analyze statistical properties of 2-D approaches are time-consuming, since the user is required to
Fourier spectra together with iris image quality assessment present his finger twice thus unusable in real-time applications.
Reflectance analysis based approaches involves illuminat- Skin deformation-based methods exploits flexibility prop-
ing the eye with multiple different wavelengths of light and erties of skin (skin elasticity). Zhang et al. [31] developed
comparing the relative response in the sclera and iris regions. a novel method to capture finger distortion using Thin-plate
Lee et al. [22] proposed to detect live and fake irises based on Spline model at different rotated angles. Jia et al. [32] used
reflectance attributes using the theoretical reflectance model. two features (correlation coefficient and standard deviation)
for liveness detection based on skin elasticity. The main
Dynamics analysis liveness methods acquire several images
drawbacks of deformation-based schemes are that users need
while manipulating the illumination level to check for a change
to have special training and the scanners should be capable of
in pupil dilation. Pacut et al. [23] introduced a composite
delivering frames at a proper rate.
iris liveness detection method which analyzes dynamic (i.e.
behavioral) eye features such as pupil dynamics along with Image quality-based approaches use image quality proper-
image frequency spectrum and controlled light reflection from ties or texture. Nikam et al. [30] utilized ridgelet transform
the cornea. Galbally et al. [24] adopted motion features with ensemble classifier to capture texture information from
(spontaneous dynamic features) with image quality properties fingerprint image. Galbally et al. [33] formulated an scheme
caused by either motion of the iris or of the sensor. based on ten different quality measures such as: ridge strength,
ridge continuity and ridge clarity. While, Ghiani et al. [34] ex-
Texture analysis mechanisms essentially analyze and clas-
ploited local phase quantization features for liveness detection.
sify the image texture features to determine the iris spoofing
attacks. He et al. [25] presented use of gray level co-occurrence Pore detection-based techniques detect pores as a sign of
matrix (GLCM) with statistical texture analysis for spoof fingerprint vitality. Manivanan et al. [35] evolved a method
contact lens detection. Tan et al. [26] employed multi-scale using two filtering techniques: highpass and correlation; former
local binary pattern texture features with Adaboost to learn one to extract active sweat pore and latter to locate pores’
effective spoof detection. However, this approach requires position. Espinoza et al. [36] proved that pores can be used as
samples of the contact lens patterns to be detected as part of a liveness sign and suggested to use pore quantity differences
its training data. Wei et al. [27] devised a scheme where Iris- between query image (real or fake) and a reference image.
Textons (i.e. texture representation) are used to detect fake iris.
Lee et al. [28] proposed a new method for fake iris detection Combined mechanisms generally integrate the above men-
via calculating the theoretical positions and distances between tioned schemes. Tan et al. [37] combined ridge signal algo-
the Purkinje images (using collimated IR-LED) based on the rithm and valley noise analysis to detect perspiration patterns
human eye model. This method fails for contact lenses and along ridges in live fingers and noise patterns along valleys in
requires additional hardware, a series of iris images and needs spoofs. Jai et al. [38] applied skin elasticity and perspiration
user’s full cooperation. A comparative analysis of state-of-the- pattern via five features from a sequence of fingerprint images.
art iris liveness detection in [23] shows that none of the method The above outlined survey states that majority of face, iris
achieves acceptable error rate and computational cost. Thus, it and fingerprint spoofing detections are either very complex
is imperative to develop new iris liveness detection methods. (and thus not very practical for mobile applications requiring
fast processing) or using non-conventional imaging methods
C. Fingerprint Spoofing (e.g. multi spectral imaging) and devices (e.g. thermal cam-
Fingerprint spoofing is quite an old practice [2]. A finger- eras). Besides, existing methods are mainly trait dependent.
print recognition system can be fooled by (i) a 2-D (flat) fake Namely, image descriptors proposed for iris spoof detection
fingerprint of a genuine user; (ii) a 3-D fake fingerprint of may not effectively work if employed for face or fingerprint
a genuine user. Fake fingerprint can be fabricated either by liveness detection and vice versa. Hence, we propose in this
‘consensual/cooperative/direct casts’ or ‘non-consensual/non- paper a single approach for face, iris and fingerprint liveness
cooperative/indirect casts’ method using easily available mate- detection, which is very simple, computationally fast and uses
rials like latex etc. In consensual method, the fake fingerprints conventional images plus requires no user-cooperation.
are created directly from real fingers with person’s consent,
while in non-consensual method fake fingerprints are fabri- III. T HE PROPOSED METHOD
cated from latent finger-marks on daily use product or sensors;
hence, the cooperation of the user is not required. Software- Though, face, iris and fingerprint images captured from
based fingerprint liveness detection can be grouped in 5 groups: spoof attacks may visually look very similar to the images
perspiration-based, skin deformation-based, image quality- captured from live person (see Fig. 1), a close look reveals
based, pore detection-based, and combined approaches. that spoof attack images contain some particular artefacts.
Thus, motivated by image quality assessment and charac-
Perspiration-based schemes analyze the perspiration pat- terization of artefacts, we propose a novel software-based
tern present in the fingerprint. Parthasaradhi et al. [8] used multi-biometric method, which can be used in mobile and
static patterns and dynamic changes in the moisture struc- real-time applications. In particular, we derive a face, iris
ture of skin around sweat pores caused by perspiration for and fingerprint representation (or a feature space) that is
liveness detection. Abhyankar et al. [29] framed a wavelet capable of capturing distinctive characteristics of real and
 [~, desc] = sort (p(:));

......
......

Real Face /
Iris / Fingerprint
R G B
SVM

......
......
Classifier

Fake Face /
Iris / Fingerprint

......

......
Pre-processed Concatenated
User Sensors Input image image using LUCID features Classification Decision
features
blur kernel
(a) (b)

Fig. 2. (a) Top: LUCID feature construction method in 1 line of Matlab. Note: ∼ is used to ignore first return value of sort; and second value is the order
permutation. Bottom: An illustration of an image patch split into its RGB color channels, vectorized and then sorted; inducing a permutation on the indices.
(b) The proposed approach. The captured image of the respective biometric trait is preprocessed with averaging blur kernel and the resulting image is divided
into blocks according to LUCID patch size. Features are computed from each patch and are concatenated to devise a single feature vector. Finally, linear SVM
classifier is used to determine whether the input image belongs to a live user or not.

fake face, iris and fingerprint images. Our method learns Face Spoofing. We used five publicly available datasets.
the fine differences between images of real and fake faces,
Print Attack [18]. The data set is composed of 200 video
irises and fingerprints via Locally Uniform Comparison Image
clips of printed-photo attack attempts to 50 users, under
Descriptor (LUCID) [6]. LUCID is a novel approach to feature
different lighting conditions, and of 200 real-access attempts
description based on order permutations, that is computable in
from the same users.
linear time with respect to the number of pixels and does not
require floating point computation, beside the fact that typical Replay Attack [18]. It is consist of 1300 video clips of
mobile devices perform poorly for floating point applications. photo and video attack attempts to 50 clients, under different
Moreover, LUCID is surprisingly simple and efficient for lighting conditions. As we need to operate on images, we
feature construction, which implicitly encapsulate all possible extracted the ‘live’ and ‘spoofed’ face images from the corre-
intensity comparisons in a local area of an image. sponding videos of Print Attack and Replay Attack databases.
In particular, for each client, we extracted 20 ‘live’ face images
Let p be n × n image patch with c color channels. We can
and 20 ‘spoofed’ face images from each video clip.
compute LUCID descriptor for the patch in one line of Matlab
as shown in Fig. 2(a). Here desc is the order permutation NUAA Photograph Imposter Database [15]. The data set is
representations for p. LUCID has three parameters, blur kernel made of real-accesses and photo-attacks of 15 users utilizing
width, image patch size, and the option to use color or both photo-quality and laser-quality prints. On average, for
grayscale images. Before LUCID construction an averaging each user 340 ‘live’ and 500 ‘spoofed’ face images are
blur is applied to the entire image to remove noise that may available. The ‘live’ and ‘spoofed’ face images were collected
perturb the order permutation. Let m = cn2 , then native into three sessions, with a time interval of about 2 weeks
implementation makes use of a stable comparison-free linear between them, under different changing environmental and
time sort and thus takes O(m) time and space. Descriptor illumination conditions. The live subjects were asked to look
construction is depicted in Fig. 2(a). like a static by minimizing the movements and eye-blinking,
while attacks were with varying motions.
Our method adopts the LUCID descriptor to compute
feature vector of input image. The vectors in the feature space Personal Photo Attack. We collected ‘live’ face images of
are then fed to linear Support Vector Machine (SVM) classifier 40 clients into two sessions under different facial expressions.
which determines whether feature description corresponds to We then created the spoofed face images using the ‘photo
a live person or not. The proposed method for face, iris and attack’ method described in [39]. It consists in displaying a
fingerprint liveness detection is depicted in Fig. 2(b). photo of the targeted client on a laptop screen, which is then
put in front of the camera. We used personal photos of live
face data set’s clients taken from the Web, for instance from
IV. E XPERIMENTS
social networks. On average, we collected 5 photos per client.
In this section we provide an experimental evaluation of
Yale Recaptured Database [39]. This data set is composed
the proposed liveness detection method for face, iris and
of 640 real faces (taken from 10 subjects with 64 different
fingerprint biometrics.
illumination conditions) and 1920 LCD spoofs by displaying
the images from the Yale Face Database B on three LCD
A. Data sets monitors and captured using a DELL Inspiron 1545 notebook.
Since no spoofing attack datasets are publicly available that Iris Spoofing. ATVS-FIr DB [24]. This is also a publicly
are collected using mobile device, we utilized existing datasets. available data set comprising 50 subjects × 2 eyes × 4 images
 TABLE I. P ERFORMANCE (H ALF T OTAL E RROR R ATE (HTER)-%) OF
× 2 sessions = 800 fake iris images and its corresponding THE PROPOSED METHOD USING LUCID DESCRIPTOR ALONG WITH
original samples. In the experiments both eyes of same subject EXISTING IRIS [24], FACE [16] AND FINGERPRINT [34] METHODS IN
are considered as different users (i.e., 50 × 2 = 100 users). DISCRIMINATING LIVE FACE , IRIS AND FINGERPRINT IMAGES FROM FAKE
ONES . T HE AVERAGE TIME REQUIRED FOR FEATURE EXTRACTION IS
Notre Dame [40]. The data set contains 2800 live (without REPORTED IN THE LAST ROW.
contact lenses) and 1400 spoofed (with cosmetic contact System Proposed method Existing methods
lenses) iris images acquired using an LG 4000. Iris (ATVS) 1.03±0.34 4.66±1.15
Iris (Notre Dame) 0.07±0.67 1.64±0.34
Fingerprint Spoofing. ATVS-FFp DB [33]. This database, Face (Print Attack) 2.88±0.88 4.54±1.35
which comprises the index and middle fingers of both hands Face (NUAA) 1.54±0.16 0.54±0.10
of 17 users (17 × 4 = 68 different fingers), is available Face (Personal) 0.01±0.03 4.76±1.51
Face (Yale Recaptured) 1.90±0.20 0.80±0.11
for public as well. For each real finger, two fake imitations Face (Replay Attack) 5.46±0.55 7.30±3.61
were generated using silicon with two procedures (i.e., with Fingerprint (ATVS) 7.17±1.97 14.22±4.10
and without the user’s cooperation). Four samples of each Fingerprint (LivDet09-Crossmatch) 7.94±0.61 9.01±0.47
Fingerprint (LivDet09-Biometrika) 0.14±0.06 0.30±0.14
fingerprint (fake and real) were captured in one acquisition Fingerprint (LivDet09-Identix) 0.49±0.16 7.16±0.52
session with 3 sensors. Thus, the database comprises 68 fingers Fingerprint (LivDet11-Sagem) 8.54±0.75 9.71±0.71
× 4 samples × 3 sensors = 816 real image samples and as Fingerprint (LivDet13-Swipe) 12.24±0.86 16.17±0.77
many fake images for each procedure. Time (s) 0.0025 0.0814

LivDet09 [34]. It contains three subsets named Biometrika, than existing schemes on all data sets (except Yale and
CrossMatch and Identix. Each subset has 142 subjects with NUAA). Not only is it possible to detect face, iris and finger-
20 live and 20 spoofed fingerprint impressions. The gummy print spoofing attack from single image but also, contrary to the
fingers were generated using three different materials: silicone, state-of-the-art, presented individual image feature descriptor,
gelatine and playdoh under consensual procedure. i.e. LUCID, seems to deliver encouraging liveness detection
LivDet11-Sagem and LivDet13-Swipe [34]. Each data set performance for all three modalities, namely face, iris and
contains 2000 live and 2000 fake fingerprint samples. fingerprint. Likewise, the average time required for feature
extraction in proposed technique is much lesser than in existing
B. Experimental protocol methods.

For each above mentioned data set, we randomly selected We can notice that on average the LUCID descriptor
40% of the users (real and fake) as training set, whereas the performed best for iris liveness detection across the whole
remaining 60% users (real and fake) as testing set. We run the range of data sets. For instance, the HTER is 0.07% on Notre
above procedure five times. Reported results are average values Dame database, though the fake samples taken into account are
over the 5 runs with standard deviations. We applied LUCID- high quality spoofed contact lenses. While the method in [24],
24-RGB for face, LUCID-24-Gray for iris and fingerprint, which combines two focus (IQF15 and IQF16), two occlusion
which respectively are LUCID on image patches that are (IQF3 and IQF19) and one pupil dilation (IQF22) features,
24×24 in RGB color, 24×24 in grayscale. Before LUCID achieved 1.64% HTER. The HTERs of face spoof detection on
construction a 5×5 averaging blur is applied to the entire NUAA data set by proposed method and reflectance analysis
image. A liveness detection method is subject to two types based approach in [16] are 1.54% and 0.54%, respectively.
of errors, either the real access is rejected (false rejection) or In real-world, due to illumination a significant difference
a spoofing attack is accepted (false acceptance). We evaluated between real and spoofed face (iris) is that the spoofs can pos-
performance of liveness detection systems as in [18] using Half sibly consist specular reflections. Additionally, spoof artefacts
Total Error Rate (HTER), that combines the False Rejection may be clearly found locally, for instance on uniform surface
Rate (FRR) and the False Acceptance Rate (FAR) as: like cheek. LUCID therefore performs better for face and iris in
F AR(τ, D) + F RR(τ, D) comparison with fake fingerprint, since the order permutation
HT ER(τ, D) = [%] (1) is invariant to monotonic intensity transformations.
2
The HTERs on fingerprint data sets (except LivDet09-
where D denotes the used dataset. We have chosen thresh- Biometrika and Identix) are considerable, since they are more
old τ on the equal error rate (EER) at training set and reported challenging than face and iris databases. Namely, the error
HTER using the test set data. The experiments were conducted rate may be occurring because of less quality difference in
on Asus K52F laptop with 2.13 GHz Intel dual-core CPU with fake and real fingerprint images. LUCID and local phase
3 GB of RAM using un-optimized Matlab code. quantization [34] features thus may contain redundant or noisy
information, which contributes to the overlap between fake
C. Experimental results and live classes and hence their bad performance. But, it is
noteworthy that contrary to existing studies (which use only
In Table I, we report the performance attained by the
cooperative method), we have also used data set (i.e., ATVS)
proposed liveness detection system along with the existing
which is composed of fake fingerprints fabricated by both
methods corresponding for face, iris and fingerprint biometrics.
cooperative and non-cooperative method, since in a realistic
It can be discerned from Table I that the proposed method
scenario we may not know with what type of fake fingerprint
presents a very high potential as a simple, fast and novel
the attacker is trying to fool the system.
method to detect spoofing attacks, which reaches a great
classification accuracy for different biometric traits (“multi- To sum up, the preliminary results provide some evidence
biometric”). Moreover, the proposed method performed better that the feature descriptor used is exceptionally simple, fast
and effective, which makes it highly suitable for real-time [11] K. Kollreider, H. Fronthaler, J. Bigun: Verifying liveness by multiple
or mobile devices. As the method does not deploy any trait- experts in face biometrics. In IEEE CVPR Workshops, pp. 1-6, 2008.
specific property (e.g. minutiae points, face detection or iris [12] W. Bao et al.: A liveness detection method for face recognition based
position), thus the computational load is minimized. Moreover, on optical-flow field. In IEEE IASP, pp. 233-236, 2009.
contrary to the existing approaches, we propose to employ only [13] G. Pan et al.: Monocular camera-based face liveness detection by
combining eye-blink and scene context. J. of Telecom. Sys., 2009.
one image descriptor to gain notable accuracy for three bio-
[14] J. Li, Y. Wang, T. Tan, and A. K. Jain: Live face detection based on the
metric traits, viz face, iris and fingerprint, liveness detection. analysis of fourier spectra. In Bio. Tec. for Hu. Id., pp. 296-303, 2004.
Additionally, our method uses only one image for liveness
[15] X. Tan, Y. Li, J. Liu, and L. Jiang: Face liveness detection from a single
detection, which can also be used for biometric recognition. image with sparse low rank bilinear discriminative model. In European
Conf. on Computer Vision (ECCV), vol. 6316, pp. 504-517, 2010.
V. C ONCLUSION [16] N. Kose and J. L. Dugelay: Reflectance analysis based countermeasure
technique to detect face mask attacks. In DSP, pp. 1-6, 2013.
We proposed a novel software-based liveness detection [17] J. Bai, T. Ng, X. Gao, and Y. Shi: Is physics-based liveness detection
method that can be used in multiple biometric systems. In truly possible with a single image?. In ISCAS, pp. 3425-3428, 2010.
particular, we presented an approach for face, iris and fin- [18] I. Chingovska, A. Anjos, S. Marcel: On the effectiveness of local binary
gerprint spoofing attack detection in mobile applications, by patterns in face anti-spoofing. In Biometrics Special Interest Group
employing a novel real-time feature description based on order (BIOSIG), pp. 1-7, 2012.
permutations, named Locally Uniform Comparison Image De- [19] Z. Zhang, D. Yi, Z. Lei, and S. Z. Li: Face liveness detection by learning
multispectral reflectance distributions. In FG, pp. 436-441, 2011.
scriptor (LUCID). The feature descriptor is exclusively simple,
[20] L. Ma, T. Tan, Y. Wang, D. Zhang: Personal identification based on iris
fast and effective, which makes it highly suitable for real- texture analysis. IEEE TPAMI, vol. 25, no. 12, pp. 1519-1533, 2003.
time or mobile applications. Since, LUCID is computable in
[21] X. He, Y. Lu, P. Shi: A fake iris detection method based on FFT and
linear time with respect to number of pixels and does not quality assessment. In Chinese Conf. on Patt. Reco., pp. 316-319, 2008.
require floating point computation, beside the fact that typical [22] S. Lee et al.: A Study on Fake Iris Detection based on the Reflectance
mobile devices perform poorly for floating point applications. of the Iris to the Sclera for Iris Recognition. In ITC-CSCC, 2005.
Moreover, unlike state-of-the-art, our method applies same [23] A. Pacut, A. Czajka: Aliveness Detection for Iris Biometrics. In IEEE
lone image descriptor to deliver encouraging liveness detection International Carnahan Conf. on Security Technology, pp. 122-129, 2006.
performances effectively for three modalities, namely face, iris [24] J. Galbally et al.: Iris liveness detection based on quality related
and fingerprint. Experimental results on publicly available data features. In Int’l. Conference on Biometrics (ICB), pp. 271-276, 2012.
sets with real spoofing attacks showed promising results. [25] X. He, S. An, P. Shi: Statistical texture analysis based approach for fake
iris detection using support vector machine. In ICB, pp. 540-546, 2007.
For image based biometric authentication, we intuitively [26] Z. He, Z. Sun, T. Tan, and Z. Wei: Efficient Iris Spoof Detection via
believe that the inherent characteristics of spoofing attack im- Boosted Local Binary Patterns. In Advances in Biometrics, 2009.
ages captured by biometric sensors may be analogues, there is [27] Z. Wei., X. Qiu, Z. Sun, T. Tan: Counterfeit iris detection based on
thus a need for a generic image feature based liveness detection texture analysis. In Int’l Conf. on Patt. Recog. (ICPR), pp. 1-4, 2008.
which can detect all traits’ spoofing attacks, no matter on [28] E. C. Lee, K. R. Park, J. Kim: Fake iris detection by using purkinje
which biometric trait images they are trained. For instance, a image. In Int’l Conference on Biometrics (ICB), pp. 397-403, 2005.
countermeasure formulated for face spoofing should also work [29] A. Abhyankar, S. Schuckers: Integrating a wavelet based perspiration
effectively for iris or fingerprint spoofing and vice versa. To liveness check with fingerprint recognition. Pattern Recognition, vol. 42,
no. 5, pp. 452-464, 2008.
this aim, presented work in this paper is an introductory step.
[30] S. Nikam, S.Agarwal: Ridgelet-based fake fingerprint detection. Neu-
rocomputing, vol. 72, pp.2491-2506, 2008.
R EFERENCES [31] Y. Zhang, X. Tian, X. Chen, X. Yang, P. Shi: Fake Finger Detection
Based on Thin-Plate Spline Distortion Model. Advances in Biometrics
[1] J. Komulainen, A. Hadid, M. Pietikäinen: Face spoofing detection using
Lecture Notes in Computer Science, vol. 4642, pp. 742-749, 2007.
dynamic texture. In ACCV Work. on Comput. Vision, pp. 146-157, 2013.
[32] J. Jia et al., D. Chen: A New Approach to Fake Finger Detection Based
[2] Z. Akhtar: Security of Multimodal Biometric Systems against Spoof
on Skin Elasticity Analysis. Advances in Biometrics, pp. 309-318, 2007.
Attacks. PhD thesis, DIEE, University of Cagliari, Italy, 2012.
[33] J. Galbally, F. Alonso-Fernandez, J. Fierrez, J. Ortega Garcia: A high
[3] R. Chen, X. Lin, T. Ding: Liveness Detection for Iris Recognition Using
performance fingerprint Iiveness detection method based on quality
Multispectral Images. PRL, vol. 33, no. 1, pp. 1513-1519, 2012.
related features. Future Gen. Comp. Systems, vol. 28, pp. 311-321, 2012.
[4] A. Rattani and A. Ross: Minimizing the Impact of Spoof Fabrication
[34] L. Ghiani, G. Marcialis, F. Roli: Fingerprint liveness detection by local
Material on Fingerprint Liveness Detector. In Proc. IEEE Int’l Conf. on
phase quantization. In Int’l Conf. Pattern Recognition, pp. 537-540, 2012.
Image Processing (ICIP), 2014.
[35] N. Manivanan, S. Memon, W. Balachandran: Automatic detection of
[5] Z. Akhtar, C. Micheloni, C. Piciarelli and G. L. Foresti: MoBio LivDet:
active sweat pores of fingerprint using highpass and correlation filtering.
Mobile Biometric Liveness Detection. IEEE Int’l Conf. on Advanced
Electronic Letters, vol. 46, vo. 18, pp. 1268-1269, 2010.
Video and Signal-Based Surveillance (AVSS), 2014.
[36] M. Espinoza, C. Champod: Using the Number of Pores on Fingerprint
[6] A. Ziegler, E. Christiansen, D. Kriegman, and S. Belongie: Locally
Images to Detect Spoofing Attacks. In IC on Hand Bio., pp. 1-5, 2011.
Uniform Comparison Image Descriptor. In NIPS, pp. 1-9, 2012.
[37] B. Tan, S. Schuckers: Spoofing Protection for Fingerprint Scanner by
[7] J. Daugman: Demodulation by Complex-Valued Wavelets for Stochastic
Fusing Ridge Signal and Valley Noise. Pattern Recognition, 2010.
Pattern Recognition. Int’l J. of W. Multireso. and Info. Proc., 2003.
[38] J. Jia, L. Cai: Fake Finger Detection Based on Time-Series Fingerprint
[8] S. Parthasaradhi, R. Derakhshani, L. Hornak, S. Schuckers: Time-Series
Image Analysis. Advanced Intelligent Computing Theories and App.
Detection of Perspiration as a Liveness Test in Fingerprint Devices. IEEE
With Aspects of Theo. and Methodo., vol. 4681, pp. 1140-1150, 2007.
Trans. on System, MAN, and Cybernetics-Part C, vol. 35, no. 3, 2005.
[39] B. Peixoto, C. Michelassi and A. Rocha: Face liveness detection under
[9] G. Pan, L. Sun, Z. Wu, S. Lao: Eyeblink-based anti-spoofing in face
bad illumination conditions. IEEE Int. Conf. on Image Processing, 2011.
recognition from a generic webcamera. In ICCV, pp. 1-8, 2007.
[40] J. S. et al.: Variation in accuracy of textured contact lens detection based
[10] K. Kollreider, H. Fronthaler, J. Bigun: Non-intrusive liveness detection
on sensor and lens pattern. In IEEE BTAS, 2013.
by face images. Im. and V. Computing, vol. 27, no. 3, pp. 233-244, 2009.

View publication stats