ON THE ROBUSTNESS OF FINGERPRINT LIVENESS DETECTION
ALGORITHMS AGAINST NEW MATERIALS USED FOR
SPOOFING
Emanuela Marasco and Carlo Sansone
Dipartimento di Informatica e Sistemistica, University of Napoli Federico II, via Claudio 21 80125, Napoli, Italy
Keywords:
Biometrics, Liveness detection.
Abstract:
Fingerprint biometric systems may be deceived by attacks at sensor level that use fake fingers. A secure
fingerprint scanner is required to possess the ability to determine if the image comes from a living individual
or not. Recently, several liveness detection approaches have been proposed to address this problem. At present
performances of the existing software-based solutions have been assessed with different sensors and using
small data sets. Moreover, it is assumed that fake fingerprints are produced by adopting the same materials
used for training the system. This paper looks at the cases where the test spoof finger is made by employing a
material new for the fingerprint sensor. We propose an experimental comparison among the current fingerprint
liveness detection approaches accomplished by adopting materials for training different than those used for
testing. Experiments have been performed by using standard databases taken from the LivDet09 Competition.
1 INTRODUCTION
Fingerprint biometric recognition systems are consid-
ered the most efficient and widely adopted technique
for secure authentication. Recent research has shown
that it is not difficult to deceive a fingerprint recog-
nition system using fake fingertips (Maltoni et al.,
2003). Common spoofing techniques realize fake fin-
gers employing inexpensive materials, e.g., gelatin,
clay, play-doh and silicon. Then, it is required that a
fingerprint sensor has the ability to determine if an in-
put biometric signal comes from a live person or not
(Schuckers, 2002). Recently, in order to address this
problem, several approaches for liveness detection
have been proposed. The state-of-the-art concern-
ing these solutions consists of two main categories:
software-based and hardware-based approaches. The
methods belonging to the first category make use
of additional hardware to measure temperature and
heartbeat characteristics resulting in an expensive im-
plementation. The software-based approaches exploit
intrinsic vitality properties that are extracted directly
from the fingerprint images acquired by the sensor
resulting in a cheaper device. At present, the main
software-based vitality detection approaches analyze
the skin perspiration through the pores, the elastic
properties of the skin and the morphology of the fin-
gerprint (Coli et al., 2007) (Schuckers et al., 2006).
The main limitation of most of those methods re-
sides in the device-dependence, in fact the resolution
of the fingerprint image varies across different tech-
nologies, and subsequently, the extracted characteris-
tics of vitality are not universal enough (Coli et al.,
2008). Moreover, in the previous works the systems
are tested under the assumption that fake fingerprints
are realized by employing one of the materials used
for training. This assumption makes optimistic the
results showed by the authors. It is worth noting
that this aspect is a challenging problem in fingerprint
liveness detection, since nowadays materials used for
fraudulent spoof attacks are going to become very so-
phisticated.
So, in order to assess the influence of this as-
pect on the performance of liveness detection algo-
rithms, in this paper we analyze the cases where spoof
fingers are realized with materials that are new for
the fingerprint scanner. We propose an experimen-
tal comparison among the current fingerprint liveness
detection approaches accomplished by adopting ma-
terials for training different than those adopting for
testing. Experiments were carried out by using stan-
dard databases taken from Liveness Detection Com-
petition 2009 (LivDet09). The paper is organized as
follows. Section 2 presents the main steps of each
methods we have considered for our study. Section 3
describes the datasets we employed to carry out our
553
Marasco E. and Sansone C..
ON THE ROBUSTNESS OF FINGERPRINT LIVENESS DETECTION ALGORITHMS AGAINST NEW MATERIALS USED FOR SPOOFING.
DOI: 10.5220/0003270505530558
In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (MPBS-2011), pages 553-558
ISBN: 978-989-8425-35-5
Copyright
c
2011 SCITEPRESS (Science and Technology Publications, Lda.)
experiments. Section 4 reports results about the ro-
bustness of the approaches to new materials used for
spoofing. Section 5 draws our conclusions.
2 A TAXONOMY OF THE
EXISTING FINGERPRINT
LIVENESS DETECTION
APPROACHES
Liveness methods belong to two main categories. The
first one exploits characteristics as the temperature of
the finger, the electrical conductivity of the skin and
the pulse oximetry. They can be detected by using
additional hardware in conjunction with the biomet-
ric sensor. This makes costly the device. The sec-
ond category performs an extra software process of
the biometric sample in order to detect the vitality
information directly from the fingerprint images. In
this paper, we focus on this second category of ap-
proaches, known as software-based (Schuckers et al.,
2006). The existing software-based solutions may in-
clude dynamic or static methods.
2.1 Dynamic Approaches
Dynamic features derive from the analysis of multiple
frames of the same fingers. A typical dynamic prop-
erty of a live finger is the perspiration phenomenon
that starts from the pores and evolves in time across
the ridges. This distinctive spatial moisture pattern
can be detected by observing multiple fingerprint im-
ages acquired in two appropriate different times. An
interesting method based on perspiration changes in
live fingers was presented by Abhyankar and Schuck-
ers in (Abhyankar and Schuckers, 2009). In this
method, the changing perspiration pattern is isolated
through a wavelet analysis of the entire fingerprint im-
age. For an image processing algorithm, to quantify
the sweating pattern is challenging. Since this pattern
is a physiological phenomenon, it is variable across
subjects. Further, it presents a certain sensitivity to
the environment, the pressure of the finger, the time
interval and the initial moisture content of the skin
(Derakhshani et al., 2003). Its effectiveness requires
an efficient extraction of the evolving pattern from im-
ages.
2.2 Static Approaches
Static features can be extracted from a single finger-
print impression or as difference between different
impressions. Generally, static measurements may be
altered by factors as the pressure of the finger on the
scanner surface.
According to the taxonomy proposed in (Coli
et al., 2008), features extracted by different impres-
sions can be skin deformation-based or morphology-
based, while features extracted by a single impres-
sion can be perspiration-based or morphology-based.
Elastic deformations due to the contact, the pressure
and the rotation of the fingertip on the plane surface
of the sensor, are more evident in fake fingerprints
made using artificial materials than in live finger-
prints. Deformation-based methods detect liveness
by comparing these distortions through static features
(Chen et al., 2005). The elastic behavior of live and
fake fingers has been analyzed by extracting a specific
set of minutiae points. The second type of static fea-
tures using multiple impressions relies on a morpho-
logic investigation which exploits the thickness of the
ridges that is modified after producing the fingerprint
replica.
Methods which exploit intrinsic properties of a
single impression study the skin perspiration phe-
nomenon. The vitality indication can be found by
using Wavelet Transform and Fast Fourier Transform
(Coli et al., 2007). Wavelet analysis is able to cap-
ture the non-regular shape typical of the ridges in an
image acquired from a live finger. Images taken from
artificial fingers show a more regular shape. Fourier
Transform is employed to study the regular periodic-
ity of pores on the ridges in live fingerprints. Such
a regularity is not present in signals corresponding to
spoof fingerprints.
Liveness detection methods which search for mor-
phological characteristics of fingerprint images, are
significantly efficient when based on the surface
coarseness.
3 EXISTING METHODS
EXPLOITED FOR OUR STUDY
In this section, we describe the methods employed
for our comparative analysis. Firstly, we describe the
three static morphology-based methods which exploit
a single fingerprint image for vitality information ex-
traction.
Moon et al. (Moon et al., 2005) proposed a
method based on analyzing the surface coarseness in
high resolution (1000dpi)fingertipimages. It has been
observed that the surface of a fake finger is much
coarser than that one of the human skin. The coarse-
ness feature is measured by computing the standard
deviation of the residual noise of the fingerprint im-
age. This algorithm is fast and convenient but it
BIOSIGNALS 2011 - International Conference on Bio-inspired Systems and Signal Processing
554
works well only in presence of an high resolution
sensor (1000dpi, while the common commercial sen-
sors present a resolution of about 500dpi) (Coli et al.,
2007).
An interesting texture-based approach using a sin-
gle fingerprint image was proposed by (Nikam and
Agarwal, 2009). They analyzed liveness of a fin-
gerprint image by using the gray level associated to
the fingerprint pixels. The gray level distribution in
a fingerprint image changes when the physical struc-
ture changes. Then, real and fake fingerprint images
are expected to present different textural properties.
In fact, due to the presence of sweat pores and the
perspiration phenomenon, authentic fingerprints ex-
hibit non-uniformityof gray levels along ridges, while
due to the characteristics of artificial material surface,
such as gelatin or silicon, spooffingers show high uni-
formity of gray levels along ridges.
In (Abhyankar and Schuckers, 2006), Abhyankar
and Schuckers proposed an approach based on multi-
resolution texture analysis and the inter-ridge fre-
quency analysis of fingerprint images. They used
different texture features to quantify how the gray
level distribution in a fingerprint image changes when
the physical structure changes. First order statistics
model the gray level distribution of the single pix-
els by using histograms, while second order statis-
tics refer to the joint gray level function between pair
of pixels. Two secondary features were used, Clus-
ter Shade and Cluster Prominence, based on the co-
occurrence matrix. All these features have been com-
bined with features derived from fingerprint local-
ridge frequency analysis.
Secondly, we describe a static method which com-
bines characteristics describing the morphologyof the
fingerprint and characteristics describing the perspira-
tion phenomenon (Marasco and Sansone, 2010). The
approach relies on static features derived from the vi-
sual texture of the fingerprint image. In particular,
first order statistics and residual noise standard de-
viation are exploited as morphology-based features,
while ratios between gray level values and individual
pore spacing are are exploited as perspiration-based
features.
The standard deviation of the residual noise mea-
sures the coarseness of the fingerprint image. Ma-
terials used to make fake fingers such as silicon or
gelatin consist of organic molecules which tend to ag-
glomerate, thus the surface of a fake finger is gen-
erally coarser than a live one (Moon et al., 2005).
The residual noise indicates the difference between
the original and the de-noised image, in which the
noise components are due to the coarseness of the
fake finger surface (Abhyankar and Schuckers, 2006).
In fact, according to the approach proposed by Moon
et al.(Moon et al., 2005), the surface coarseness has
been treated as a kind of gaussian white noise added
to the image.
First order statistics measure the likelihood of ob-
serving a gray value at a randomly-chosen location
in the image. The gray level associated to each pixel
is exploited to determine a vitality degree of the fin-
gerprint image. They can be computed from the his-
togram of pixel intensities in the image. The goal is
to quantify the variations of the gray level distribution
when the physical structure changes. The distinction
between a fake and a live finger is based on the differ-
ence of these statistics.
Individual pore spacing characteristics are ex-
tracted after analyzing the occurrence of pores that
causes a gray value variability in the fingerprint im-
age. In (Marasco and Sansone, 2010), according to
the algorithm proposed in (Derakhshani et al., 2003),
the 2-dimensional fingerprint image was mapped to
1-dimensional signal which represents the gray-level
values along the ridges. The gray-level variations in
the signal correspond to variations in moisture due
to the pores and the presence of perspiration. By
transforming the signal in the Fourier domain lets to
measure this static variability in gray-level along the
ridges. In particular, the focus is on frequencies corre-
sponding to the spacial frequencies of the pores. The
FFT was computed and the total energy associated to
the spacial frequency of the pores was obtained as
static feature. The coefficients of interest are from
11 to 33, since these values correspond to the spacial
frequencies (0.4 - 1.2 mm) of pores.
Intensity-based features are based on the assump-
tion that, the spoof and cadaver fingerprints images
are distributed in the dark (<150), among the 256 dif-
ferent possible intensities (Tan and Schuckers, 2005).
They have computed two particular features: i) gray
level 1 ratio, corresponding to the ratio between the
number of pixels having a gray level belonging to the
range (150, 253) and the number of pixels having a
gray level belonging to the range (1, 149); ii) gray
level 2 ratio, corresponding to the ratio between the
number of pixels having a gray level belonging to the
range (246, 256) and the number of pixels having a
gray level belonging to the range (1, 245). More-
over, they have analyzed the uniformity of gray lev-
els along ridge lines and the contrast between valleys
and ridges. Real fingerprints exhibit non-uniformity
of gray levels and high ridge/valley contrast values.
Then, the general variation in gray-level values of in
a spoof fingerprint is less than a live one. To capture
this information the gradient of the gray-level matrix
of the image have been computed, too.
ON THE ROBUSTNESS OF FINGERPRINT LIVENESS DETECTION ALGORITHMS AGAINST NEW MATERIALS
USED FOR SPOOFING
555
Finally, we describe a perspiration-based method
using both static and dynamic features. Tan and
Schuckers (Tan and Schuckers, 2005) have experi-
mented the joint contribution of dynamic and static
features. They studied the perspiration phenomenon
from the intensity distribution perspective, by observ-
ing that live fingers present a distinctive contrast be-
tween white (>250, ASCII gray level range 0:255)
and dark (<20) gray level, while spoof images have
very small contrast difference. The decision rules to
perform liveness classification is generated after con-
sidering static and dynamic features. The static fea-
tures used in this work are based on the following pa-
rameters:
S1 =
sum(151 : 254)
sum(0 : 150)
(1)
and
S2 = sum(151 : 254) (2)
The dynamic features are based on the difference in
the histogram distribution between zero and fifth sec-
ond that is larger in live finger compared to spoof
subjects. In the live fingers, perspiration makes dry
(white) regions between the pores moister (darker) in
time. This approach may present some limitations in
cases of fingers too dry or too moist and other perspi-
ration disorders.
4 EXPERIMENTAL RESULTS
Our experimental phase was carried out by using two
databases composed by live and spoof fingerprint im-
ages. Each one refers to a different sensor (i.e.,
CrossMatch and Identix). They have been taken from
the Liveness Detection Competition 2009 (Marcialis
et al., 2009) and each one of them is composed by two
subsets, one for training and the other one for testing
the algorithm. Details about the data collection are
shown in the tables 1 and 2. In both the cases, the
subjects using for training are different with respect
to those considered for testing. Table 3 reports details
about the sensors used for LivDet 2009 Competition.
Note that the LivDet dataset has also another database
(namely, Biometrika) whose spoof fingerprints have
been produced by employing only one material (Sili-
con). So, for carrying out the current comparison, we
have adopted only CrossMatch and Identix databases
in which spoof fingerprints have been made by em-
ploying three different materials.
The classification performance evaluation was
performed by adopting the same parameters used dur-
ing the LivDet09, defined as follows:
Ferrlive: rate of misclassified live fingerprints.
Table 1: Datasets of images used for training.
Database Sub Live Fake Frames
Identix 35 375 375 0 and 2 sec
CrossMatch 63 500 500 0 and 2 sec
Table 2: Datasets of images used for testing.
Database Sub Live Fake Frames
Identix 125 1125 1125 0 and 2 sec
CrossMatch 191 1500 1500 0 and 2 sec
Table 3: Fingerprint sensors used for LivDet 2009.
Scanners Model No. (dpi) Size
Biometrika FX2000 569 (312x372)
Identix DFR2100 686 (720x720)
CrossMatch Verifier 300 LC 500 (480x640)
Ferrfake: rate of misclassified fake fingerprints.
In particular, performance is measured by using the
value e averaged on the two database CrossMatch and
Identix. The value e is computed as follows:
e =
Ferrlive + Ferr f ake
2
(3)
In our first experiments, each system was trained
by using features extracted from fake samples made
with all the materials available in each database. In
particular, in both Identix and CrossMatch databases,
the materials employed are Gelatin, Silicon and Play-
Doh. Then, we have carried out a further evalua-
tion, in order to study the robustness of the exist-
ing liveness detection approaches with respect to un-
known materials used for producing fake fingers. In
this experiment, each system was trained by using
spoof fingerprints realized with all but one of the
available materials, while the excluded material was
used for testing. Table 4 reports the performance of
the method proposed by Marasco and Sansone. In
presence of high resolution images, taken from the
Identix database, the testing performed using Gelatin
and Silicon, when the training is performed by em-
ploying fake fingers made in Play-Doh, gives rise
to a good spoofing recognition rate. Table 5 shows
that the method proposed by Moon et al. wrongly
classifies the majority of the fake fingerprints taken
from CrossMatch database, while for a higher res-
olution factor, such a method presents a better be-
havior in presence of unknown materials using for
spoofing. Table 6 and 7 show that the variation in
fake materials does not significantly affect the per-
formance of both Nikam-Agarwal and Abhyankar-
Schuckers approaches, when the training set is only
composed by samples made with Gelatin. On the con-
trary, as reported in Table 8, the performance of the
BIOSIGNALS 2011 - International Conference on Bio-inspired Systems and Signal Processing
556
Table 4: Performance of the method proposed by Marasco and Sansone on CrossMatch and Identix databases.
CrossMatch Identix
Gelatin Play Doh Silicon Gelatin Play Doh Silicon
Ferrlive 6.5% 5.7% 12.6% 3.8% 19.2% 9.7%
Ferr fake 25.9% 16.7% 10.0% 42.3% 5.5% 30.6%
e 16.2% 11.2% 11.3% 23.05% 12.35% 20.15%
Table 5: Performance of the method proposed by Moon et al. on CrossMatch and Identix databases.
CrossMatch Identix
Gelatin Play Doh Silicon Gelatin Play Doh Silicon
Ferrlive 12.30% 15.00% 35.70% 45.20% 79.60% 40.80%
Ferr fake 63.10% 61.80% 47.30% 31.80% 4.20% 36.80%
e 37.70% 38.40% 41.50% 38.50% 41.90% 38.80%
Table 6: Performance of the method proposed by Nikam and Agarwal on CrossMatch and Identix databases.
CrossMatch Identix
Gelatin Play Doh Silicon Gelatin Play Doh Silicon
Ferrlive 27.20% 43.70% 24.20% 23.50% 29.30% 20.00%
Ferr fake 22.00% 32.90% 31.60% 16.00% 28.80% 31.50%
e 24.60% 38.30% 27.90% 19.75% 29.05% 25.75%
Table 7: Performance of the method proposed by Abhyankar and Schuckers on CrossMatch and Identix databases.
CrossMatch Identix
Gelatin Play Doh Silicon Gelatin Play Doh Silicon
Ferrlive 45.80% 29.80% 58.60% 65.50% 61.60% 37.90%
Ferr fake 12.20% 24.40% 17.00% 2.40% 46.40% 27.70%
e 29.00% 27.10% 37.80% 33.45% 54.00% 32.80%
Table 8: Performance of the method proposed by Tan and Schuckers on CrossMatch and Identix databases.
CrossMatch Identix
Gelatin Play Doh Silicon Gelatin Play Doh Silicon
Ferrlive 38.60% 24.40% 54.80% 64.10% 36.00% 38.80%
Ferr fake 32.20% 39.20% 43.00% 28.70% 42.40% 13.20%
e 35.40% 31.80% 48.90% 46.40% 39.20% 26.00%
Table 9: Performance of the analyzed approaches in terms of the average error e on Identix and CrossMatch databases.
Marasco-Sansone Moon et al. Nikam-Agarwal Abhyankar-Schuckers Tan-Schuckers
Gelatin 19.63% 38.10% 22.18% 31.23% 40.90%
Play-Doh 11.78% 40.15% 33.68% 40.55% 35.50%
Silicon 15.73% 40.15% 26.83% 35.30% 37.45%
Avg 15.71% 39.47% 27.53% 35.79% 37.45%
All materials 12.45% 30.85% 24.53% 39.37% 29.20%
Tan-Schuckers method seems quite dependent on the
material as well as on the considered dataset.
As resumed in Table 9, when the material used to
attack the system is not known during the training,
most of the algorithms decrease in performance. This
confirms our claim that the performance of liveness
detection algorithms reported by the authors typically
represents an overestimate of that obtainable in real
scenarios. Among the considered methods, the one
based on a single feature (Moon et al., 2005) is the
most dependent on the use of unknown materials for
testing. Also the dynamic method proposed in (Tan
ON THE ROBUSTNESS OF FINGERPRINT LIVENESS DETECTION ALGORITHMS AGAINST NEW MATERIALS
USED FOR SPOOFING
557
and Schuckers, 2005) had a significant decrement in
performance when classifying fake fingerprints real-
ized with materials different from those present in the
training set. The other methods are instead more ro-
bust, and the one proposed by Marasco and Sansone,
which is based on a combination of multiple features,
exhibited the best average error e when the material
used for testing is unknown at training time.
5 CONCLUSIONS
In this paper, we have analyzed the impact of a new
material on the performance of the existing liveness
detection algorithms. This analysis has been per-
formed by considering three different sensors.
Our experiments showed that the performance of
liveness detection approaches in which only few fea-
tures are exploited, significantly decreases in pres-
ence of spoof attacks realized by employing materi-
als different from those using for training the system.
This weakness can be reduced by combining multiple
vitality features. In particular, the more robust ap-
proach was given by the joint usage of morphology-
and perspiration-based features.
ACKNOWLEDGEMENTS
The authors would like to thank Prof. Stephanie
Schuckers of the Clarkson University (USA) for her
useful talks, Prof. Luisa Verdoliva of the University
of Naples (Italy), for supporting the wavelet analy-
sis techniques and Prof. Davide Maltoni of the Uni-
versity of Bologna (Italy), for his valuable comments
about the existing core detection algorithms.
REFERENCES
Abhyankar, A. and Schuckers, S. (2006). Fingerprint live-
ness detection using local ridge frequencies and mul-
tiresolution texture analysis techniques. IEEE Inter-
national Conference on Image Processing, pages 321–
324.
Abhyankar, A. and Schuckers, S. (2009). Integrating a
wavelet based perspiration liveness check with finger-
print recognition. Pattern Recognition, 42:452–464.
Chen, Y., Jain, A., and Dass, S. (2005). Fingerprint defor-
mation for spoof detection. Biometric Symposium.
Coli, P., Marcialis, G., and Roli, F. (2007). Vitality detec-
tion from fingerprint images: a critical survey. Lecture
Notes in Computer Science, 4642:722–731.
Coli, P., Marcialis, G., and Roli, F. (2008). Fingerprint
silicon replicas: static and dynamic features for vi-
tality detection using an optical capture device. In-
ternational Journal of Image and Graphics (IJIG),
8(4):495–512.
Derakhshani, R., Schuckers, S., Hornak, L., and O’Gorman,
L. (2003). Determination of vitality from non-invasive
biomedical measurement for use in fingerprint scan-
ners. Pattern Recognition, 36:383–396.
Maltoni, D., Maio, D., Jain, A., and Prabhakar, S. (2003).
Handbook of Fingerprint Recognition. Springer.
Marasco, E. and Sansone, C. (2010). An anti-spoofing
technique using multiple textural features in finger-
print scanners. IEEE Workshop on Biometric Mea-
surements and Systems for Security and Medical Ap-
plications (BioMs), pages 8–14.
Marcialis, G., Lewicke, A., Tan, B., Coli, P., Grimberg,
D., Congiu, A., Tidu, A., Roli, F., and Schuckers, S.
(2009). First international fingerprint liveness detec-
tion competition - livdet 2009. Lecture Notes in Com-
puter Science, 5716:12–23.
Matsumoto, T., Matsumoto, H., Yamada, K., and Hoshino,
S. (2002). Impact of artificial gummy fingers on fin-
gerprint systems. Optical Security and Counterfait
Deterrence Techniques IV, 4677:275–289.
Moon, Y. S., Chen, J. S., Chan, K. C., So., K., and Woo,
K. S. (2005). Wavelet based fingerprint liveness de-
tection. Electronic Letters, 41(20):1112–1113.
Nikam, S. B. and Agarwal, S. (2009). Curvelet-based fin-
gerprint anti-spoofing. Signal, Image and Video Pro-
cessing, 4(1):75–87.
Schuckers, S. (2002). Spoofing and anti-spoofing measures.
Information Security Technical Report, 7(4):56–62.
Schuckers, S., Derakhshani, R., Parthasaradhi, S., and Hor-
nak, L. (2006). Liveness detection in biometric de-
vices.
Tan, B. and Schuckers, S. (2005). Liveness detection using
an intensity based approach in fingerprint scanner. In
Proceedings of Biometrics Symposium (BSYM).
BIOSIGNALS 2011 - International Conference on Bio-inspired Systems and Signal Processing
558