Content uploaded by Dario Maio
Author content
All content in this area was uploaded by Dario Maio on Dec 13, 2012
Content may be subject to copyright.
360 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 1, NO. 3, SEPTEMBER 2006
Fake Finger Detection by Skin Distortion Analysis
Athos Antonelli, Raffaele Cappelli, Dario Maio, and Davide Maltoni, Member, IEEE
Abstract—Attacking fingerprint-based biometric systems by
presenting fake fingers at the sensor could be a serious threat for
unattended applications. This work introduces a new approach
for discriminating fake fingers from real ones, based on the anal-
ysis of skin distortion. The user is required to move the finger
while pressing it against the scanner surface, thus deliberately
exaggerating the skin distortion. Novel techniques for extracting,
encoding and comparing skin distortion information are formally
defined and systematically evaluated over a test set of real and
fake fingers. The proposed approach is privacy friendly and does
not require additional expensive hardware besides a fingerprint
scanner capable of capturing and delivering frames at proper rate.
The experimental results indicate the new approach to be a very
promising technique for making fingerprint recognition systems
more robust against fake-finger-based spoofing attempts.
Index Terms—Biometric systems, fake fingers, security, skin
distortion, skin elasticity.
I. INTRODUCTION
B
IOMETRIC systems offer great benefits with respect to
other authentication techniques: in particular, they are
often more user friendly and can guarantee the physical pres-
ence of the user. Thanks to their good performance and to the
growing market of low-cost acquisition devices, fingerprint-
based identification/verification systems are becoming very
popular and are being deployed in a wide range of applications:
from PC logon to electronic commerce, from ATMs to phys-
ical access control [18]. On the other hand, it is important to
understand that, as any other authentication technique, finger-
print recognition is not totally spoof-proof. The main potential
threats for fingerprint-based systems are [28], [29]
• attacking the communication channels, including replay at-
tacks on the channel between the sensor and the rest of the
system;
• attacking specific software modules (e.g., replacing the
feature extractor or the matcher with a Trojan horse);
• attacking the database of enrolled templates;
• presenting fake fingers to the sensor.
Recently, the feasibility of the last type of attack has been
reported by some researchers [19], [25]: they showed that
it is actually possible to spoof some fingerprint recognition
systems with well-made fake fingertips (Fig. 1), created with
the collaboration of the fingerprint owner or from a latent
Manuscript received January 18, 2006; revised May 3, 2006. This work was
supported by the European Commission (BioSec—FP6 IST-2002-001766). The
associate editor coordinating the review of this manuscript and approving it for
publication was Dr. Anil Jain.
A. Antonelli is with the Biometrika s.r.l., Forlì 47100, Italy (e-mail: an-
tonelli@biometrika.it).
R. Cappelli, D. Maio, and D. Maltoni are with the DEIS—Università di
Bologna, Cesena (FO) 47023, Italy (e-mail: cappelli@csr.unibo.it; maio@
csr.unibo.it; maltoni@csr.unibo.it).
Digital Object Identifier 10.1109/TIFS.2006.879289
fingerprint; in the latter case, the procedure is more difficult
but still possible.
A deep study on the feasibility of spoofing some commercial
fingerprint scanners was performed by the authors within the
BioSec project [1], [5], [2]. From the critical review of the
related bibliography (as described in Section II) and from the
24-months experience we accumulated by making hundreds of
fake fingers with different materials and procedures and using
them to spoof existing fingerprint scanners (of different types:
optical, capacitive, thermals, RF-based, etc.), we may draw
some conclusions.
• Forging a fake finger is not as easy as some authors claim,
even when the person whose finger has to be cloned is
cooperative; it is necessary to find the right materials to
mould the cast, learn the right process and handle with care
the artificial finger.
• Creating a fake finger from a latent fingerprint is sig-
nificantly more difficult, requiring a skill comparable to
that of a forensic expert equipped with the appropriate
instrumentation.
• To the best of our knowledge and from the experience
gained testing recent scanners provided with fake detec-
tion mechanisms, nowadays, in spite of the claims of some
fingerprint scanner producers, no commercial fingerprint
scanner (among those we tested) seems to be resistant to
well-made fake fingerprints.
• The lack of satisfactory solutions to reject fake fingers
shows that there are a lot of challenges in fake detection;
more research and investments on fingerprint fake detec-
tion methods are needed.
This work introduces a novel method for discriminating fake
fingers from real ones, based on the analysis of a peculiar char-
acteristic of the human skin: its elasticity. When a real finger
moves on a scanner surface, it produces a significant amount of
distortion, which can be observed to be quite different from that
produced by fake fingers. Usually fake fingers are more rigid
than skin and the distortion is definitely lower; even if highly
elastic materials are used, it seems very difficult to precisely
emulate the specific way a real finger is distorted, because the
behavior is related to the way the external skin is anchored to
the underlying derma and influenced by the position and shape
of the finger bone.
The analysis of skin distortion requires in input a sequence
of frames instead of a single static image. To this purpose,
the fingerprint scanner must be able to deliver a set of frames
(Fig. 2) to the processing unit at a high speed (at least 20 frames
per second). In our study, we used the prototype of a fingerprint
scanner that the company Biometrika developed within the
BioSec project [5] (Fig. 3).
A database of video sequences has been collected, acquiring
images both from real and fake fingers. Systematic experiments
1556-6013/$20.00 © 2006 IEEE
ANTONELLI et al.: FAKE FINGER DETECTION BY SKIN DISTORTION ANALYSIS 361
Fig. 1. Fake fingertips created with different materials. From left to right: gelatin, silicone, and latex.
Fig. 2. Set of frames acquired while a finger was rotating over the surface of a fingerprint scanner.
Fig. 3. Specific version of the scanner Fx3000 (by Biometrika) that allows to acquire and transfer frames to the host at 20 fps.
have been performed to understand how much the proposed
method is capable to discriminate real from fake fingers; the re-
sults achieved are very promising.
The rest of this work is organized as follows. Section II sum-
marizes the state-of-the art in this field, Section III describes
the proposed approach, Section IV reports the experimentation
carried out to validate the new technique, and finally Section V
draws some conclusions.
II. R
ELATED WORKS
Several papers have been recently devoted to this important
topic: from the analysis of potential weaknesses in generic bio-
metric systems [28], [29], [34], to experiments aimed at in-
vestigating how current fingerprint verification systems can be
spoofed [6], [16], [19], [25], [33]; from proposals of possible
solutions [1], [2], [9], [10], [20], [24], to surveys of the current
state-of-the-art [31].
It is worth noting that the idea of spoofing fingerprint recog-
nition systems by using a fake reproduction of the fingertip is
not a novelty. The idea seems to have been described for the
first time by the mystery writer R. A. Freeman in the book
“The Red Thumb Mark” [12], published in 1907. More re-
cently, James Bond in the film Diamonds are Forever (1971)
was able to spoof a fingerprint check with a thin layer of
latex glued on his fingertip [35]. However, only recently some
researchers published the results of experiments aimed at ana-
lyzing such vulnerability.
• In [25], the authors described two methods for creating
fake fingers: duplication with cooperation and without co-
operation; in both these cases, the material used to create
362 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 1, NO. 3, SEPTEMBER 2006
the fakes was silicone; six different commercial fingerprint
scanners were tested and the authors reported to be able to
spoof all of them at the first or second attempt.
• In [19], it was reported that fakes created with gelatin
were more effective, in particular against scanners based
on solid-state sensors [18]; similar to [25], the authors
described cooperative and noncooperative fake-creation
methods; 11 commercial fingerprint scanners were tested,
with a success rate higher than 67% for both the coopera-
tive and the noncooperative scenarios.
• In [6], three commercial fingerprint scanners were tested:
all of them were spoofed by fake fingers made of gelatin,
with a level of ease depending on the scanner and software
characteristics.
• In [16], the studies reported in [25] and [19] were extended
by testing new scanners that included specific fake detec-
tion measures; the authors concluded that such measures
were able to reject fake fingers made of nonconductive ma-
terials (such as silicone), but were not able to detect con-
ductive materials such as gelatin.
The main fake finger detection techniques that have been pro-
posed to date can be roughly classified as explained in the rest
of this section.
•
Analysis of skin details in the acquired images: using very
high resolution sensors (e.g., 1000 dpi) allows to capture
some details that may be useful for fake detection, such as
sweat pores [18] or coarseness of the skin texture [20]. In
fact, it has been experimentally noted that typical materials
used to make fake fingers (e.g., gelatin) usually consist of
large organic molecules that tend to amalgamate, resulting
in a surface coarser than human skin and where small de-
tails such as pores are not present or poorly reproduced.
• Analysis of static properties of the finger: additional hard-
ware is used to capture information such as temperature
[25], impedance or other electric measurements [15], [32],
odor [2], and spectroscopy [21]. In [2], electronic noses are
used with the aim of detecting the odor of those materials
that are typically used to create fake fingers (e.g., silicone
or gelatin); spectroscopy-based techniques expose the skin
to multiple wavelengths of light and analyze the reflected
spectrum: nonhuman tissues show a spectrum usually quite
different from human ones. Other techniques [7] direct light
to the finger from two or more sources and capture finger-
print images with different illuminations: the authors claim
that it is possible to discriminate between real and fake fin-
gers by comparing such differently illuminated images.
• Analysis of dynamic properties of the finger, such as: skin
perspiration [10], [24], pulse oximetry [23], blood pulsa-
tion [17], [23] and skin elasticity [1], [11], [9]. To date, fake
detection by skin-perspiration is probably the technique
most deeply studied in scientific publications: the idea is
to exploit the perspiration of the skin that, starting from
the pores, diffuses in the fingerprint patter following the
ridge lines, making them appear darker over time. In [24],
the perspiration process is detected through a time-series
of images acquired from the scanner over a time window
of a few seconds. Skin elasticity, which produces distortion
in the acquired fingerprint images [18], has been studied in
some previous works, but mainly focusing on the problems
that such distortion causes to fingerprint matching algo-
rithms [3], [22], [27], [30], or trying to find a mathematical
model to explain its behavior [8]. In [11], it was suggested
that the acquisition of a video sequence of fingerprint im-
ages could be used to define a new type of biometric fea-
ture, which combines a physiological trait (fingerprint) to
behavioral traits (e.g., a particular movement of the finger
on the sensor chosen by the user); the authors underlined
that this new biometric feature, among the other advan-
tages, could be harder to be spoofed, but they did not re-
ported any experiment with fake fingers. In [1], we briefly
introduced a fake detection approach based on skin distor-
tion and reported some preliminary results. In this paper,
the whole technique is described and experiments with a
new prototype scanner are reported and discussed.
III. F
AKE
FINGER
DETECTION APPROACH
The user is required to place a finger onto the scanner surface
and to apply some pressure while rotating the finger in either
clockwise or counter-clockwise direction (this particular move-
ment has been chosen after some initial tests, as it seems quite
easy for the user and it produces the right amount of distortion).
A sequence of frames is acquired at a high frame rate during the
movement and analyzed to extract relevant features related to
skin distortion. Although the finger can be rotated at different
speed, we experimentally found that an angular speed of about
15
per second is optimal for measuring the distortion.
Some constraints are enforced to simplify the subsequent pro-
cessing steps; in particular
• any frame such that the amount of rotation with respect
to the previous one (inter-frame rotation) is less than
is discarded (the inter-frame rotation angle is calculated
as described in Section III-B);
is a parameter whose
optimal value has been experimentally determined as 0.25
(see Section IV-B);
• only frames acquired when the rotation of the finger is
less than
are considered: when angle has been
reached, the acquisition halts (the rotation angle of the
finger is calculated as described in Section III-E-1).
is a parameter that was set to 15 in the experimentations
(see Section IV-B); hence, if we assume an angular speed
of about 15
per second, on the average, the user is required
to rotate the finger for about 1 s before the system informs
her or him that the acquisition process is terminated.
Let
be a sequence of images that satisfies
the above constraints: each frame
, , is segmented
by isolating the fingerprint area from the background; then, for
each frame
, the following steps are per-
formed (Fig. 4):
• computation of the optical flow between the current frame
and the next one;
• computation of the distortion map;
• temporal integration of the distortion map;
• computation of the DistortionCode from the integrated dis-
tortion map.
At the beginning of the sequence, the finger is assumed re-
laxed (i.e., nondistorted), without any superficial tension; this is
ANTONELLI et al.: FAKE FINGER DETECTION BY SKIN DISTORTION ANALYSIS 363
Fig. 4. Main steps of the feature extraction approach: a sequence of acquired fingerprint images is processed to obtain a sequence of DistortionCodes.
reasonable since when the finger approaches the sensor platen
there is no skin distortion.
The isolation of the fingerprint area from the background is
performed by computing the gradient of the image block-wise:
let
be a generic pixel in the image and a
square block of frame
centered in : each whose gra-
dient module exceeds a given threshold is associated to the fore-
ground [18] (Fig. 5). Only foreground blocks are considered in
the rest of the algorithm.
A. Computation of the Optical Flow
Block-wise correlation is computed to detect the new position
of each block in frame . For each block ,
the vector
denotes the estimated
Fig. 5. Fingerprint image before and after the segmentation from the back-
ground.
movement of from frame to frame . In the fol-
lowing, for simplifying the notation,
will be indicated as
364 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 1, NO. 3, SEPTEMBER 2006
Fig. 6. From left to right: two consecutive images, their difference (reported to graphically highlight the movement) and the corresponding optical flow.
Fig. 7. Optical flow before (on the left) and after (on the right) the regularization process.
. A graphical representation of the movement vectors (see
Fig. 6), is also known in the literature as optical flow [4].
This method is in theory only translation-invariant but, since
the images are taken at a fast frame rate, for small blocks it is pos-
sible to assume a certain rotation- and deformation-invariance.
The block size (in pixels) is a parameter that should be
adjusted according to the sensor area and resolution. If the
blocks are too small, they do not contain enough information
to univocally identify their positions in the subsequent frame.
Otherwise, if they are too large, two problems may arise: the
algorithm would become computationally expensive and the
distortion could make the matching unfeasible. To increase
the accuracy of the optical flow, the blocks can be also partially
overlapped: in this case the distance among the centers of two
consecutive blocks is smaller than the block size.
In order to filter out outliers produced by noise, by false cor-
relation matches, or by other anomalies, the optical flow is then
regularized as follows.
1) Each
such that
is discarded . This step allows to remove
outliers, under the assumption that the movement of each
block cannot deviate too much from the largest movement
of the blocks of the previous frame;
is a parameter that
should correspond to the maximum expected acceleration
between two consecutive frames.
2) For each
, the value is calculated as the weighted
average of the 3
3 neighbours of , using a 3 3
Gaussian mask; elements discarded by the previous step
are not included in the average: if no valid elements are
present,
is marked as “invalid”.
3) Each
such that is discarded. This
step allows to remove elements that are not consistent with
their neighbours;
is a parameter that controls the strength
of this procedure.
4)
are recalculated as in step 2, but considering only the
elements retained at step 3.
Fig. 7 shows the optical flow before (
vectors) and after
(
vectors) the steps described above.
B. Computation of the Distortion Map
The center of rotation
is estimated as the
weighted average of the positions
of all the foreground blocks
such that the corresponding movement vector is
valid
is valid (1)
where
is the average of the elements in set .
The inter-frame rotation angle
(around the center ) and
the translation vector
are then computed in the
least square sense, starting from all of the average movement
vectors
(2)
ANTONELLI et al.: FAKE FINGER DETECTION BY SKIN DISTORTION ANALYSIS 365
Fig. 8. Graphical representation of a distortion map (A) and of the corre-
sponding integrated distortion map (B). Blocks with a lighter gray color denote
higher distortion values.
using the Gauss–Newton approach to numerically solve the
problem [13].
If the finger were moving solidly, then each movement vector
would be coherent with
and . Even if the movement is not
solid,
and still encode the dominant movement and, for
each block
, the distortion can be computed as the inco-
herence of each average movement vector
with respect to
and . In particular, if a movement vector were computed
according to a solid movement, then its value would be
(3)
and, therefore, the distortion can be defined as the residual
if is valid
otherwise.
(4)
A distortion map is defined as a block-wise image whose
blocks encode the distortion values
[Fig. 8(a)].
C. Temporal Integration of the Distortion Map
The computation of the distortion map, made on just two con-
secutive frames, is affected by three problems.
• The movement vectors are discrete (because of the discrete
nature of the images) and, in case of small movement, the
loss of accuracy might be significant.
• Errors in seeking the new position of blocks could lead to
a wrong distortion estimation.
• The measured distortion is proportional to the amount of
movement between two frames (and, therefore, depends
on the finger speed), without considering previous ten-
sion accumulated/released. This makes it difficult to com-
pare a distortion map against the distortion map of another
sequence.
An effective solution to the above problems is to perform a
temporal-integration of the distortion map, resulting in an in-
tegrated distortion map [Fig. 8(b)]. The temporal integration is
simply obtained by block-wise summing the current distortion
map to the distortion map “accumulated” in the previous frames.
Each integrated distortion element is defined as shown in (5) at
the bottom of the page with
.
The rationale behind the definition is that if the norm of the
average movement vector
is smaller than the norm of the
estimated solid movement
, then the block is moving slower
than expected and this means it is accumulating tension (i.e., dis-
tortion). Otherwise, if the norm of
is larger than the norm
of
, the block is moving faster than expected, thus it is slip-
ping on the sensor surface releasing the tension accumulated.
The integrated distortion map solves most of the previously
listed problems: 1) discretization and local estimation errors are
no longer serious problems because the integration tends to pro-
duce smoothed values; 2) for a given movement trajectory, the
integrated distortion map is quite invariant with respect to the
finger speed. Fig. 9 shows the integrated distortion maps com-
puted for a given image sequence acquired by rotating a real
finger.
D. Distortioncode
Comparing two sequences of integrated distortion maps, both
acquired under the same movement trajectory, is the basis of
this fake finger detection approach. On the other hand, directly
comparing two sequences of integrated distortion maps would
be computationally very demanding and it would be quite dif-
ficult to deal with the unavoidable local changes between the
sequences.
To simplify this task, a feature vector (called DistortionCode
for the analogy with the FingerCode introduced in [14]) is ex-
tracted from each integrated distortion map:
circular annuli
of increasing radius (
, , where is the radius of
the smaller annulus) are centered in
and superimposed to the
map. For each annulus
, a feature is computed as the av-
erage of the integrated distortion elements of the blocks falling
inside it (Fig. 10)
belongs to annulus (6)
The number of annuli
and the radius are parameters
that must be chosen to optimally cover a typical fingerprint, ac-
cording to the sensor area and resolution.
A DistortionCode
is obtained from each frame ,
if is valid and
if is not valid
if
(5)
366 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 1, NO. 3, SEPTEMBER 2006
Fig. 9. Sequence of integrated distortion maps.
Fig. 10. Integrated distortion map with the annuli superimposed. Note that
background blocks are discarded at the beginning of the process (see Section III)
and therefore they are not taken into account in (6).
The DistortionCodes are invariant to rotation since distortion
values are averaged over the circular annuli; they are also in-
variant with respect to the position of the fingerprint in the image
(translation), since they are centered in
; in any case, it should
be noted that translation accuracy is not critical because of the
integrated and global nature of the features adopted.
A DistortionCode sequence
is then defined by normalizing
the distortion codes
where
(7)
The obtained DistortionCode sequence (Fig. 11) characterizes
the distortion of a particular finger under a specific movement.
Further sequences from the same finger do not necessarily lead to
the same DistortionCode sequence: the overall length might be
different, because the user could produce the same trajectory (or
a similar trajectory) faster or slower. While a minor rotation ac-
cumulates less tension, during a major rotation the finger could
slip and the tension be released in the middle of the sequence.
Therefore, a straightforward comparison of DistortionCode se-
quences is not feasible and an alignment technique like those
introduced in Sections III-E1 and III-E2 is necessary.
E. Distortion Match Function
In order to discriminate a real finger from a fake one, the
DistortionCode sequence acquired at verification/identifica-
tion time (current sequence) is compared with a reference
sequence obtained from a real finger. The reference sequence
may be a sequence acquired from the finger of the same user
during an “enrolment” session (similarly to what happens in
biometric recognition), or a predefined “ideal” sequence to be
adopted for all users (in this case the fake-detection system
does not require an enrolment stage, see Section IV-C). Let
be the reference sequence and
the current sequence; a distor-
tion match function DMF
compares the reference and
the current sequence and returns a score in the range
,
indicating how much the current sequence is similar to the
reference sequence (1 means maximum similarity).
A Distortion Match Function must define how to do the
following.
Step 1) Calculate the similarity between two Distortion-
Codes.
Step 2) Align the elements by establishing a correspondence
between the DistortionCodes of the two sequences
and .
Step 3) Measure the similarity between the two aligned
sequences.
As to Step 1), a simple Euclidean distance between two Dis-
tortionCodes has been adopted, since it is a good metric and also
very efficient to be computed, having the vectors a very small di-
mensionality. As to Step 2), two different approaches have been
experimented.
• Aligning the sequences according to the accumulated inter-
frame rotation (Section III-E1).
• Aligning the sequences using dynamic time warping
(DTW) [26] (Section III-E2).
In both cases, the result of Step 2) is a new DistortionCode
sequence
, obtained from
during the alignment process with ; has the same car-
dinality of
and the final similarity can be simply computed
(Step 3) as the average Euclidean distance of corresponding Dis-
tortionCodes in
and .
ANTONELLI et al.: FAKE FINGER DETECTION BY SKIN DISTORTION ANALYSIS 367
Fig. 11. Sequence of DistortionCodes calculated on the integrated distortion maps in Fig. 9.
Fig. 12. Example of DTW alignment. On the left, the mapping function , which maps each DistortionCode in the current sequence to a DistortionCode
in the reference sequence. On the right, a graphical representation of the same mapping: note that the same DistorsionCode in the reference sequence c
an be
associated twice or more times (or not associated at all), not only to deal with different lengths but, more in general, to
find the optimal alignment.
1) Aligning the Sequences According to the Accumu-
lated Inter-Frame Rotation: Any DistortionCode
can
be associated to the rotation angle
obtained by accumu-
lating the inter-frame rotation angles
(see Section III-B):
. This approach determines optimal pairing
between the DistortionCodes in
and according to rota-
tion angles
; interpolation is used to deal with discretization
effects. The new sequence
is obtained by calculating, for
each pair
in the current sequence, a new distor-
tion code
from the two consecutive DistortionCodes in
the reference sequence
and (where
) as follows:
(8)
Equation (8) simply estimates
as the linear interpolation
of the distortion codes corresponding to the two closest rotation
angles.
2) Aligning the Sequences Using Dynamic Time Warping:
The main limitation of the previous alignment approach is that
the distortion is not only related to the amount of rotation, but
also to the pressure applied while rotating the finger, and more
in general to the movement performed, hence aligning only on
the basis of the rotation angle may be not always a good choice.
An alternative approach to align the two DistortionCode se-
quences is based on DTW [26]. Using DTW with constrained
endpoints, slope three and the Euclidean distance as a cost func-
tion, each DistortionCode
in is associated to a Distor-
tionCode
(see Fig. 12). This allows to warp
the time dimension of the reference sequence
to obtain the
new sequence
.
The DTW algorithm aligns the two sequences according to
the less expensive path. If the two sequences are similar, the
resulting path will have a total cost low and will be quite close
to the diagonal path (Fig. 12).
3) Computation of the Final Score: Once the new sequence
has been obtained (using one
of the approaches described above), the final score can be com-
puted as follows:
(9)
The normalization coefficient
ensures that the
score is always in the range
. In fact, for any Distortion-
Code sequence
and for any of its ele-
ments
, it is easy to prove that
(10)
and
(11)
Constraint (10) follows directly from the definition of Distor-
tionCode sequence (7), constraint (11) from the definitions of
integrated distortion map (5) and DistortionCode (6).
It is worth noting that the transformations performed to obtain
the new sequence
do not violate the two constraints in both
the proposed approaches, since:
• in the first one,
, thus (10)
is guaranteed by the triangular inequality and (11) by the
definition of
;
• in the DTW approach,
, thus (10) and
(11) are trivially verified.
368 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 1, NO. 3, SEPTEMBER 2006
IV. EXPERIMENTATION
A. Measuring Fake Detection Errors
A fingerprint scanner that embeds a fake-finger detection
mechanism has to decide, for each transaction, if the current
sample comes from a real finger or from a fake one. This deci-
sion will be unavoidably affected by errors, that should be as
low as possible: in particular, a scanner could reject real fingers
and/or accept fake fingers, independently of the user’s identity.
In the rest of this section, we assume a system operating in
verification mode. Let
be the proportion of fake-finger
transactions where the system incorrectly considered the input
to come from a real finger. Let
be the proportion of
real-finger transactions where the system incorrectly considered
the input to come from a fake sample.
and must
not be confused with identity verification errors typical of any
biometric system; in the following, to avoid confusion we will
denote with
and the identity verification error
rates. Under the simplifying hypothesis of no correlation be-
tween the two classes of errors (fake detection errors and iden-
tity verification errors), and assuming the identify verification
performance is not significantly decreased by the fake-detection
mechanism, the overall FRR error can be estimated as
•
(for an au-
thorized user trying to be authenticated normally using the
real enrolled finger).
Depending on the hypotheses (Real or Fake finger, Enrolled
or Nonenrolled fingerprint) under which the transaction is per-
formed, the overall FAR error can be estimated as
•
-
(for an
attacker trying to be authenticated using a real finger, dif-
ferent from the enrolled one);
•
-
(for an attacker
trying to be authenticated using a fake reproduction of a
finger which is not the enrolled one);
•
-
(for an attacker trying to
be authenticated using a fake reproduction of the enrolled
finger), where
, since, even if a fake
fingerprint is created by using professional equipments, its
quality is usually lower than the real finger it is designed
to imitate and therefore the chance that the identity ver-
ification algorithm does not match it with the user’s real
template is higher.
Actually, the two classes of errors (fake detection errors and
identify verification errors) could be correlated in some cases:
for instance, a low-quality finger may determine both a high
(due, for example, to the difficulty of calculating the
correct optical flow) and a high
(due to the few number
of minutiae that can be reliably found in its fingerprint images).
It should be also considered that the adoption of a fake-detection
approach may affect the performance of the identity verifica-
tion system. For instance, due to the need of measuring specific
features for fake-detection, it could be more difficult to acquire
good quality images, thus increasing
(e.g., in the case of
fingerprint distortion, due to the need of producing distorted im-
ages it could be more difficult to acquire good quality images,
at the beginning of the image sequence, which are not affected
by distortion). Anyway, studying such correlation is beyond the
scope of this work, and will be better investigated in the future.
The experiments carried out in this study consider only fake-
detection errors (
and ), to avoid reporting per-
formance indicators depending on the identity verification ac-
curacy of a specific biometric algorithm. There is obviously a
strict trade-off between
and : both are functions
of a fake-detection threshold
. depends also on how
much skilled the attacker is, which technologies the attacker is
able to implement, how much time and money (s)he can invest,
etc. In the experimentation performed in this work we assumed
that
• the attackers were experts of the application domain and
skilled in manufacturing fake fingers (the fake fingers man-
ufactured in our tests were made by people with 24-month
experience);
• attacks were carried out using some known methods (e.g.,
fake fingers made of silicone, gelatin and other materials
commercially available);
• the attackers were aware of the particular fake-detection
technique adopted and did their best to defeat it (in our
tests fake fingers were created trying to emulate as much
as possible the human skin deformation);
• attacks had to be performed in a short time and without live
feedback from the device.
In Sections IV-B and C,
and errors measured
in the experimentation are reported, together with the
(the value such that ).
B. Database
In order to evaluate the proposed approach, a database of
image sequences was collected using a prototype fingerprint
scanner by Biometrika. No public available benchmark database
could be used, due to the specific requirements of the fake de-
tection algorithm (each sample must consist of a sequence of
images acquired by a scanner while the user is rotating her/his
finger and producing distortion, and samples from both real and
fake fingers acquired by the same device must be available). The
database was collected at the Biometric System Laboratory of
the University of Bologna acquiring, from each of 45 volun-
teers, two fingers (thumb and forefinger of the right hand); 10
image sequences were recorded for each finger. 40 fake fingers
were manufactered (10 made of RTV silicone, 10 of gelatin, 10
of latex, and 10 of wood glue). Instead of making whole 3D
fake fingers, we manufactured just thin layers reproducing the
fingertips (see Fig. 1 for some sample pictures): this allowed
to better imitate genuine finger movements when trying to at-
tack the system. For each fake finger, 10 image sequences were
recorded. The prototype scanner produces 400
560 fingerprint
images at 569 DPI and captures images at 20 fps. In Fig. 13 and
in Fig. 14 some sample fingerprint images are shown.
The volunteers received a brief training before the first acqui-
sition. Sequences having a total finger rotation angle less than
15
were discarded, and the user was asked to repeat the acqui-
sition; no other quality check was adopted during the collection
of the data (for instance ensuring that a minimum amount of dis-
tortion was produced in the sequence).
ANTONELLI et al.: FAKE FINGER DETECTION BY SKIN DISTORTION ANALYSIS 369
Fig. 13. Sample images from the database of sequences: a real finger and four fake fingers (the first image of each sequence is shown).
Fig. 14. Some images from two sequences in the database: a real finger (top row) and a fake finger (bottom row).
TABLE I
P
ARAMETER VALUES USED IN THE
EXPERIMENTATION
The acquisition of the image sequences from the fake fingers
was performed by experts, trying to emulate as much as pos-
sible the deformation of the skin in real fingers and choosing
the optimal conditions for each material; for instance, image se-
quences from fakes made of gelatin were acquired a hour after
their creation, when their elasticity is similar to the human skin,
and not later, when they become rigid and easier to be discrim-
inated from real fingers.
The parameters of the approach (see Table I) were adjusted
on a totally disjoint dataset that was collected using a different
acquisition sensor (see [1]). The only different parameter is the
block size, which here was set to 16
16 pixels to increase the
processing speed.
C. Results
As introduced in Section III-E, the fake detection approach
here proposed may be used in two different modalities:
• per-user reference sequence: for each user, during an en-
rolment stage, a sequence of frames is acquired from
the selected finger, the corresponding DistortionCode se-
quence
is calculated and stored as the reference se-
quence for that user (similar to what happens with the
fingerprint template to be used in a biometric recognition
process);
• predefined reference sequence: a single reference sequence
is adopted for all of the users and no enrolment stage is
required for the fake detection system.
Both of these operating modalities were experimented by
using the same test set described in the previous section.
In the per-user reference sequence modality, the following
transactions were performed on the test set:
• 4050 genuine attempts (each sequence was matched
against the remaining sequences of the same finger, ex-
cluding the symmetric matches to avoid correlation, thus
performing 45 attempts for each of the 90 real fingers);
• 36 000 impostor attempts (each of the 400 fake sequences
was matched against the first sequence of each real finger).
370 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 1, NO. 3, SEPTEMBER 2006
Fig. 15. Integrated distortion maps from the predefined reference sequence used in the experimentation; it is worth noting that the shape of the deformed region
is almost elliptical and distortion is mainly confined to an elliptical annulus around the center of rotation, as discussed in [8].
TABLE II
P
ERFORMANCE OF THE TWO ALIGNMENT APPROACHES
IN THE
TWO DIFFERENT MODALITIES
Note that, since only fake-detection performance was eval-
uated (not combined with identity verification) and con-
sidering that the proposed approach is based on the elastic
properties of real/fake fingers and not on the ridge-line pat-
tern, it is not necessary that a fake finger corresponding to
the real finger is used in the impostor attempts: any fake
finger can be matched against any real finger without sig-
nificantly affecting the results.
In the predefined reference sequence modality, a sequence ac-
quired from a well-trained user (not included in the test data-
base) was selected as the predefined sequence (Fig. 15) and the
following transactions were performed on the test set:
• 900 genuine attempts (each sequence was matched against
the reference sequence, thus performing 10 attempts for
each of the 90 real fingers);
• 400 impostor attempts (each of the 400 fake sequences was
matched against the reference sequence).
Table II reports the
obtained for the two alignment
approaches (Sections III-E1 and III-E2) in the two modalities,
respectively; Fig. 16 compares the ROC graphs.
An error analysis was performed by visually inspecting the
100 real finger sequences that obtained the lowest scores in
the predefined reference sequence modality with the DTW
alignment. In Table III, each sequence is labeled according to
the most evident error cause: 70% of the errors were due to an
incorrect movement (e.g., moving the finger in a nonuniform
way, translating instead of rotating,…) or a too fast movement.
Table IV analyzes the distribution of false rejection errors
among the different users; since 10 sequences were acquired
from two fingers of each user, the maximum number of errors
for each user is 20. It is worth noting that all of the users were
able to provide good sequences with both the fingers (only one
user had more than 10 errors among the 100 examined: 8 with
the first finger and 4 with the second).
On a Pentium IV PC at 3.2 GHz, the feature extraction takes
about 100 ms for each frame: the most demanding step (80% of
the feature extraction time) is the correlation, whose complexity,
in the worst case, is proportional to the square of the number of
foreground pixels in the image. However, thanks to an MMX
optimization of the correlation routine, an efficient implemen-
tation has been achieved. The matching step proved to be very
efficient: the average time is less than 1 ms for both the align-
ment approaches. The average transaction time is about two sec-
onds, including acquisition, feature extraction, and matching.
V. C
ONCLUSIONS AND
FUTURE WORK
Attacks to fingerprint-based biometric systems using fake re-
productions of the finger may be a serious threat, in particular
for nonsupervised access control applications and remote au-
thentication applications.
ANTONELLI et al.: FAKE FINGER DETECTION BY SKIN DISTORTION ANALYSIS 371
Fig. 16. ROC graphs of the two alignment approaches: per-user reference sequence modality (on the left) and prede
fined reference sequence modality (on the
right).
TABLE III
E
RROR ANALYSIS OF THE
100 REAL-FINGER SEQUENCES
THAT OBTAINED THE LOWEST
SCORE IN THE
PREDEFINED
REFERENCE SEQUENCE
MODALITY (DTW ALIGNMENT)
TABLE IV
D
ISTRIBUTION OF THE
FALSE REJECTION ERRORS
AMONG THE DIFFERENT
USERS
(THE SAME SEQUENCES
ANALYZED IN TABLE
III ARE CONSIDERED)
This work introduced a fake finger detection approach based
on skin elasticity: novel techniques for extracting skin distortion
information, for encoding it as DistortionCodes, and for nor-
malizing and comparing DistortionCode sequences have been
formally defined and experimentally evaluated over a test set of
real and fake fingers. Two different operating modalities have
been proposed: the former (per-user reference sequence) where
the user is required to perform an “enrollment” before using
the system, the latter (predefined reference sequence) where no
enrollment is required for the fake-detection (obviously enroll-
ment is still necessary for fingerprint recognition).
Contrary to what one may expect, the performance of the pre-
defined modality was better than that of the per-user modality.
The analysis of the main error causes for both the modalities
suggested that this behavior could probably be ascribed to these
facts.
• The reference DistortionCode sequence (which all the cur-
rent sequences were compared to) was obtained from a
well-trained user with a uniform and smooth movement,
resulting in a sequence that was able to correctly represent
most of real finger distortions and was very difficult to em-
ulate using fake fingers.
• During the database collection, the volunteers received
only a quick training and no specific quality control mea-
sure was enforced (except the minimum amount of finger
rotation, see Section IV-B). For this reason, a good portion
of the users did not produce enough distortion and their
corresponding DistortionCode sequences, when used as
the reference sequence in the per-user modality, were not
enough dissimilar from the fake finger sequences.
It is also worth noting that, in the per-user modality, the inter-
frame rotation angle alignment approach achieved better results
than the DTW-based one. This may be explained by considering
that, if on the one hand DTW is more flexible in adapting to a
given reference sequence (potentially decreasing
), on
the other, if no minimum quality is enforced for the reference
sequences, the greater flexibility is likely to affect
more
than
.
We may conclude that the predefined modality, besides being
simpler to be deployed in a final system, achieves better results
with nonhabituated users; on the other hand, the performance
with the per-user modality may be increased if users are well-
trained and habituated.
We believe the experimental results are very promising; in
fact, although the system did make errors (the best
achieved was 11.24%), we must underline that what we mea-
sured in our experimentation was not the robustness with
respect to zero-effort attempts, but the robustness to attacks
carried out by experts that were aware of the specific fake-de-
tection technology and did their best to emulate the human skin
deformation. The same experts achieved a very high success
372 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 1, NO. 3, SEPTEMBER 2006
rate (comparable to those reported in [19]) in spoofing all the
commercial devices they tested [5], including devices with
specific fake-finger countermeasures. However, it should be
pointed out that the proposed system, as any other fake-detec-
tion mechanisms, trades usability for security: for some large
scale low-security applications it may not be worth adopting
a fake-finger detection system, while, for other high-security
applications, the fake-detection operating threshold may be
adjusted to meet the given constraints.
Although in the experiments performed we were not able to
find a way to make the proposed system ineffective, as for any
other similar system it cannot be totally excluded that someone
might find a combination of techniques and materials that sig-
nificantly decrease its efficacy. Combining this fake-detection
system with other methods based on uncorrelated features (e.g.,
impedance, odor [2]) could make the resulting system even more
reliable.
The proposed fake detection approach is not privacy invasive,
since it does not collect any information (such as, for instance,
blood pulsation or blood pressure) that may reveal medical dis-
eases; it has the further advantage of not requiring expensive ad-
ditional hardware, provided that the fingerprint scanner is able
to acquire images at a proper frame rate.
Future work will be mainly dedicated to
• implementation and evaluation of alternative alignment
techniques for the DistortionCode sequences;
• experimentation on a larger user population;
• implementation of quality control measures for the enroll-
ment stage in the per-user modality;
• better understanding the relation between fake detection
errors and identity verification errors.
While this paper is being written, a usability study is being
conducted by Prof. Bente’s team at the University of Cologne
where the Biometrika fingerprint scanner equipped with our
fake detection approach is being experimented outside of lab-
oratory environments. The feedback from that experimentation
will help to improve the approach here introduced, thank to
the complementary information that a user-centered perspective
may provide.
A
CKNOWLEDGMENT
The authors would like to thank G. Alboni from Biometrika
(Italy) and J.-F. Mainguet from Atmel (France) for their fruitful
cooperation on the fake-finger detection topic within the scope
of the BioSec project.
R
EFERENCES
[1] A. Antonelli, R. Cappelli, D. Maio, and D. Maltoni, “A new approach
to fake finger detection based on skin distortion,” in Proc. Int. Conf.
Biometric Authentication, Hong Kong, China, Jan. 2006.
[2] D. Baldisserra, A. Franco, D. Maio, and D. Maltoni, “Fake fingerprint
detection by odor analysis,” in Proc. Int. Conf. Biometric Authentica-
tion, Hong Kong, China, Jan. 2006.
[3] A. M. Bazen and S. Gerez, “Fingerprint matching by thin-plate spline
modeling of elastic deformations,” Pattern Recognit., vol. 36, no. 8, pp.
1859–1867, Aug. 2003.
[4] S. S. Beauchemin and J. L. Barron, “The computation of optical flow,”
ACM Comput. Surv., vol. 27, no. 3, pp. 433–467, 1995.
[5] BioSec European Research Project—FP6 IST-2002-001766 [Online].
Available: http://www.biosec.org.
[6] J. Blommé, “Evaluation of Biometric Security Systems Against Artifi-
cial Fingers,” M.S. thesis, Linköping Univ., Linköping, Sweden, 2003.
[7] K. Brownlee, “Method and Apparatus for Distinguishing a Human
Finger From a Reproduction of a Fingerprint,” U.S. Patent 6 292 576,
2001.
[8] R. Cappelli, D. Maio, and D. Maltoni, “Modelling plastic distortion in
fingerprint images,” in Proc. 2nd Int. Conf. Advances in Pattern Recog-
nition (ICAPR2001), Rio de Janeiro, Brazil, Mar. 2001, pp. 369–376.
[9] Y. Chen and A. Jain, “Fingerprint deformation for spoof detection,” in
Proc. Biometrics Symp., Crystal City, VA, Sep. 19–21, 2005.
[10] R. Derakhshani, S. A. C. Schuckers, L. A. Hornak, and L. O. Gorman,
“Determination of vitality from a non-invasive biomedical measure-
ment for use in fingerprint scanners,” Pattern Recognit., vol. 36, pp.
383–396, 2003.
[11] C. Dorai, N. K. Ratha, and R. M. Bolle, “Dynamic behavior analysis
in compressed fingerprint videos,” IEEE Trans. Circuits Syst. Video
Technol., vol. 14, no. 1, pp. 58–73, Jan. 2004.
[12] R. A. Freeman, The Red Thumb Mark. London, U.K.: Collingwood,
1907.
[13] M. T. Heath, Scientific Computing: An Introductory Survey, 2nd ed.
New York: McGraw-Hill, 2002.
[14] A. K. Jain, S. Prabhakar, and L. Hong, “A multichannel approach to
fingerprint classification,” IEEE Trans. Pattern Anal. Machine Intell.,
vol. 21, no. 4, pp. 348–359, Apr. 1999.
[15] P. Kallo, I. Kiss, A. Podmaniczky, and J. Talosi, “Detector for Recog-
nizing the Living Character of a Finger in a Fingerprint Recognizing
Apparatus,” U.S. Patent 6 175 641, Jan. 16, 2001.
[16] H. Kang, B. Lee, H. Kim, D. Shin, and J. Kim, “A study on performance
evaluation of the liveness detection for various fingerprint sensor mod-
ules,” in Proc. KES, 2003, pp. 1245–1253.
[17] P. D. Lapsley, J. A. Less, D. F. Pare Jr, and N. Hoffman, “Anti-Fraud
Biometric Sensor That Accurately Detects Blood Flow,” U.S. Patent
5 737 439, Apr. 7, 1998.
[18] D. Maltoni, D. Maio, A. K. Jain, and S. Prabhakar, Handbook of Fin-
gerprint Recognition. New York: Springer, 2003.
[19] T. Matsumoto, H. Matsumoto, K. Yamada, and S. Hoshino, “Impact
of artificial “Gummy”fingers on fingerprint systems,” Proc. SPIE, vol.
4677, Jan. 2002.
[20] Y. S. Moon, J. S. Chen, K. C. Chan, K. So, and K. C. Woo, “Wavelet
based fingerprint liveness detection,” Electron. Lett., vol. 41, no. 20,
pp. 1112–1113, 2005.
[21] K. Nixon, “Novel spectroscopy-based technology for biometric and
liveness verification,” Proc. SPIE, vol. 5404, 2004.
[22] Proc. Biometric Authentication ECCV Int. Workshop , ser. Lect. Notes
Comput. Sci., D. Maltoni and A. K. Jain, Eds. Prague, Czech Re-
public: Springer, May 15, 2004, vol. 3087.
[23] D. Osten, H. M. Carim, M. R. Arneson, and B. L. Blan, “Biometric,
Personal Authentication System,” U.S. Patent 5 719 950, Feb. 17, 1998.
[24] S. T. V. Parthasaradhi, R. Derakhshani, L. A. Hornak, and S. A. C.
Schuckers, “Time-series detection of perspiration as a liveness test in
fingerprint devices,” IEEE Trans. Syst., Man, Cybern. C, vol. 35, no. 3,
pp. 335–343, Aug. 2005.
[25] T. Putte and J. Keuning, “Biometrical fingerprint recognition: Don’t
get your fingers burned,” in Proc. IFIP TC8/WG8.8, 4th Working Conf.
Smart Card Research and Adv. App., 2000, pp. 289–303.
[26] L. Rabiner and B. H. Juang, Fundamentals of Speech Recognition.
Englewood Cliffs, NJ: Prentice-Hall, 1993.
[27] N. K. Ratha and R. M. Bolle, “Effect of controlled acquisition on fin-
gerprint matching,” in Proc. Int. Conf. Pattern Recognit., 1998, vol. 2,
pp. 1659–1661.
[28] N. K. Ratha, J. H. Connell, and R. M. Bolle, “Enhancing security and
privacy in biometrics-based authentication systems,” IBM Syst. J., vol.
40, no. 3, pp. 614–634, 2001.
[29] ——, “An analysis of minutiae matching strength,” in Proc. 3rd Int.
Conf. Audio- and Video-Based Biometric Person Authentication, 2001,
pp. 223–228.
[30] A. Ross, S. C. Dass, and A. K. Jain, “Fingerprint warping using ridge
curve correspondences,” IEEE Trans. Pattern Anal. Mach. Intell., vol.
28, no. 1, pp. 19–30, Jan. 2006.
[31] S. Schuckers, “Spoofing and anti-spoofing measures,” Inform. Security
Tech. Rep., vol. 7, no. 4, pp. 56–62, 2002.
[32] D. R. Setlak, “Fingerprint Sensor Having Spoof Reduction Features
and Related Methods,” U.S. Patent 5 953 441, 1999.
[33] L. Thalheim and J. Krissler, “Body check: Biometric access protection
devices and their programs put to the test,” C’tMag., Nov. 2002.
ANTONELLI et al.: FAKE FINGER DETECTION BY SKIN DISTORTION ANALYSIS 373
[34] U. Uludag and A. K. Jain, E. J. Delp, III and P. W. Wong, Eds.,
“Attacks
on biometric systems: A case study in fingerprints,” in
Proc. SPIE, Jun.
2004, vol. 5306, Security, Steganography, and Watermarking of Multi-
media Contents VI, pp. 622–633.
[35] Diamonds Are Forever (Film) 1971.
Athos Antonelli received the Laurea degree in
computer science from the University of Bologna,
Bologna, Italy, in 1998.
From 1998 to 2002, he led the research team of
CT Software, Cesena, Italy, for sonar data processing
and underwater image analysis. From 2003 to 2005,
he was a Research Fellow with the Biometric System
Laboratory, University of Bologna, where he studied
new fake detection approaches within the BioSec Eu-
ropean Project. Currently, he is with Biometrika srl,
Forlì, Italy, where he leads innovative algorithms de-
velopment for biometric solutions.
Raffaele Cappelli received the Laurea degree
(Hons.) in computer science from the University of
Bologna, Bologna, Italy, in 1998, and the Ph.D. de-
gree in computer science and electronic engineering
from the University of Bologna in 2002.
Currently, he is an Associate Researcher at the
University of Bologna. He is a member of the Bio-
metric System Laboratory, University of Bologna.
His research interests include pattern recognition,
image retrieval by similarity and biometric systems
(fingerprint classification and recognition, synthetic
fingerprint generation, fingerprint analysis, face recognition, and performance
evaluation methodologies).
Dario Maio is a Full Professor at the University of
Bologna, Bologna, Italy. He is Chair of the Cesena
Campus and Director of the Biometric System
Laboratory, Cesena. He has published many papers
in numerous fields, including distributed computer
systems, computer performance evaluation, data-
base design, information systems, neural networks,
autonomous agents, and biometric systems. He is
author of the books Biometric Systems, Technology,
Design and Performance Evaluation (Springer,
2005) and The Handbook of Fingerprint Recognition
(Springer, 2003) which received the PSP award from the Association of
American Publishers. Before joining the University of Bologna, he received a
fellowship from the Italian National Research Council (C.N.R.) for working
on the air-traffic-control project. He is with DEIS and IEIIT-C.N.R. where he
teaches database and information systems.
Davide Maltoni (M’05) is an Associate Professor
with the Department of Electronics, Informatics,
and Systems, University of Bologna, Bologna,
Italy. He teaches computer architectures and pattern
recognition in the Computer Science Department,
University of Bologna, Cesena. His research in-
terests are in the area of pattern recognition and
computer vision. In particular, he is active in the
field of biometric systems (fingerprint recognition,
face recognition, hand recognition, performance
evaluation of biometric systems). He is co-director of
the Biometric System Laboratory, Cesena, which is internationally known for
its research and publications in the field. He is author of two books Biometric
Systems, Technology, Design and Performance Evaluation (Springer, 2005)
and The Handbook of Fingerprint Recognition (Springer, 2003) which received
the PSP award from the Association of American Publishers.
Dr. Maltoni is an Associate Editor of the IEEE T
RANSACTIONS ON
INFORMATION
FORENSICS AND SECURITY and Pattern Recognition.
