Fingerprint Liveness Detection Using Convolutional Neural Networks

Abstract

With the growing use of biometric authentication
systems in the recent years, spoof fingerprint detection has become
increasingly important. In this study, we use Convolutional
Neural Networks (CNN) for fingerprint liveness detection. Our
system is evaluated on the datasets used in The Liveness Detection
Competition of years 2009, 2011 and 2013, which comprise
almost 50,000 real and fake fingerprints images. We compare
four different models: two CNNs pre-trained on natural images
and fine-tuned with the fingerprint images, CCN with random
weights, and a classical Local Binary Pattern approach. We show
that pre-trained CNNs can yield state-of-the-art results with
no need for architecture or hyperparameter selection. Dataset
Augmentation is used to increase the classifiers performance, not
only for deep architectures but also for shallow ones. We also
report good accuracy on very small training sets (400 samples)
using these large pre-trained networks. Our best model achieves
an overall rate of 97.1% of correctly classified samples - a relative
improvement of 16% in test error when compared with the best
previously published results. This model won the first prize in the
Fingerprint Liveness Detection Competition (LivDet) 2015 with
an overall accuracy of 95.5% [1].

Full text

Fingerprint Liveness Detection using Convolutional
Neural Networks
Rodrigo Frassetto Nogueira1, Roberto de Alencar Lotufo2, and Rubens Campos Machado3
Abstract—With the growing use of biometric authentication
systems in the recent years, spoof fingerprint detection has be-
come increasingly important. In this study, we use Convolutional
Neural Networks (CNN) for fingerprint liveness detection. Our
system is evaluated on the datasets used in The Liveness Detection
Competition of years 2009, 2011 and 2013, which comprise
almost 50,000 real and fake fingerprints images. We compare
four different models: two CNNs pre-trained on natural images
and fine-tuned with the fingerprint images, CNN with random
weights, and a classical Local Binary Pattern approach. We show
that pre-trained CNNs can yield state-of-the-art results with
no need for architecture or hyperparameter selection. Dataset
Augmentation is used to increase the classifiers performance, not
only for deep architectures but also for shallow ones. We also
report good accuracy on very small training sets (400 samples)
using these large pre-trained networks. Our best model achieves
an overall rate of 97.1% of correctly classified samples - a relative
improvement of 16% in test error when compared with the best
previously published results. This model won the first prize in the
Fingerprint Liveness Detection Competition (LivDet) 2015 with
an overall accuracy of 95.5% [1].
Index Terms—Fingerprint recognition, Machine learning, Su-
pervised learning, Neural networks.
I. INTRODUCTION
The basic aim of biometrics is to automatically discriminate
subjects in a reliable manner for a target application based on
one or more signals derived from physical or behavioral traits,
such as fingerprint, face, iris, voice, palm, or handwritten sig-
nature. Biometric technology presents several advantages over
classical security methods based on either some information
(PIN, Password, etc.) or physical devices (key, card, etc.) [2].
However, providing to the sensor a fake physical biometric can
be an easy way to overtake the systems security. Fingerprints,
in particular, can be easily spoofed from common materials,
such as gelatin, silicone, and wood glue [2]. Therefore, a safe
fingerprint system must correctly distinguish a spoof from
an authentic finger (Figure 1). Different fingerprint liveness
detection algorithms have been proposed [3], [4], [5], and
they can be broadly divided into two approaches: hardware
and software. In the hardware approach, a specific device is
added to the sensor in order to detect particular properties of
a living trait such as blood pressure [6], skin distortion [7], or
odor [8]. In the software approach, which is used in this study,
1Rodrigo Frassetto Nogueira is with School of Engineering, Department of
Computer Science, New York University, USA, 11209
2Roberto de Alencar Lotufo is with the Department of Electrical and
Computer Engineering, University of Campinas, Brazil, 13083-852
3Rubens Campos Machado is with Center for Information Technology
Renato Archer, Brazil, 13069-901
Fig. 1. Typical examples of real and fake fingerprint images that can be
obtained from the LivDet2009 database used in the experiments. Figure
extracted from [9].
fake traits are detected once the sample has been acquired with
a standard sensor.
The features used to distinguish between real and fake
fingers are extracted from the image of the fingerprint. There
are techniques such as those in [2] and [9], in which the

features used in the classifier are based on specific fingerprint
measurements, such as ridge strength, continuity, and clarity.
In contrast, some works use general feature extractors such
as Weber Local Descriptor (WLD) [10], which is a texture
descriptor composed of differential excitation and orientation
components. A new local descriptor that uses local amplitude
contrast (spatial domain) and phase (frequency domain) to
form a bi-dimensional contrast-phase histogram was proposed
in [11]. In [12] two general feature extractors are compared:
Convolutional Neural Networks (CNN) with random (i.e.,
not learned) weights (also explored in [13]), and Local Bi-
nary Patterns (LBP), whose multi-scale variant reported in
[14] achieves good results in fingerprint liveness detection
benchmarks. In contrast to more sophisticated techniques that
use texture descriptors as features vectors, such as Local
Phase Quantization (LPQ) [15], LBP with wavelets [16], and
BSIF [17], their LBP implementation uses the original and
uniform LBP coding schemes. Moreover, a variety of op-
tional preprocessing techniques such as contrast normalization,
frequency filtering, and region of interest (ROI) extraction
were attempted without success. Augmented datasets [18] [19]
are successfully used to increase the classifiers robustness
against small variations by creating additional samples from
image translations and horizontal reflections. In this study
we extend the work presented in [12] by using a similar
model from the well known AlexNet [19], pre-trained on the
ILSVRC-2012 dataset[20], which contains over 1.2 million
images and 1000 classes, and then fine-tuned on fingerprint
images. We show that although the pre-trained model was
designed to detect objects in natural images, fine-tuning it to
the task of fingerprint liveness detection yields better results
than if trained the model using randomly initialized weights.
Furthermore, we train our system using a larger pre-trained
model [21], VGG, the second place in the ILSVRC-2014 [20],
to increase the accuracy of the classifier by another 2% in
absolute values.
Thus, the contributions of this study are three-fold:
•Deep networks designed and trained for the task of object
recognition can be used to achieve state-of-the-art accu-
racy in fingerprint liveness detection. No specific hand-
engineered technique for the task of fingerprint liveness
detection was used. Thus, we provide another success
case of transfer learning for deep learning techniques.
•Pre-trained Deep networks require less labeled data to
achieve good accuracy in a new task.
•Dataset augmentation helps to increase accuracy not only
for deep architectures but also for shallow techniques
such as LBP.
II. METHODOLOGY
Transfer Learning is a research problem in machine learning
that focuses on storing knowledge gained while solving one
problem and applying it to a different but related problem.
In this study, we showed that it is possible to achieve state-
of-the-art fingerprint liveness detection by using models that
were originally designed and trained to detect objects in
Fig. 2. Some images from the ImageNET dataset used to pre-train the
networks. Despite their difference to the fingerprint images, pre-training with
natural images do help in the task of fingerprint liveness detection.
Fig. 3. Illustration of the models used in this study. The boxes in red are the
only layers that are different from the original VGG-19 and Alexnet models.
natural images (such as animals, car, people). The same idea
is explored in [22], for which the authors achieved state of
the art performance in CIFAR-10, Flicker Style Wikipaintings
benchmarks using a pre-trained convolutional network. One
important difference from their experiments to ours is that all
the datasets they used contain similar images to the ImageNET
dataset (Figure 2), such as objects and scenes. In our study,
fingerprint images were used, which differ significantly from
those of other domains.
A. Models
Table I describes the models in this study. All of them use
dataset augmentation. Additionally, we show the architecture
of the models in Figure 3. For CNN-VGG and CNN-Alexnet,
the architecture is the same as described in [20] and [19],
respectively, except that we replaced the last 1000-unit softmax
layer by a 2-unit softmax layer (shown in red in the figure),
so the network can output the 2 classes (if the image is real
or fake) instead of the original 1000 classes that the networks
were designed for. For the CNN-Random the architecture is
different for each dataset and it was chosen via an extensive
grid-search as described in [12].
B. Convolutional Networks
Convolutional Networks [23] have demonstrated state-of-
the-art performance in a variety of image recognition bench-
marks, such as MNIST [24], CIFAR-10 [24], CIFAR-100 [24],
SVHN [24], and ImageNet [25]. A classical convolutional

Model Name Pipeline Description
CNN-VGG 16 Convolutional
Layers + 3 Fully
Connected Layers
Pre-trained model from [20] and finetuned using liveness detection datasets.
CNN-Alexnet 8 Convolutional
Layers + 3 Fully
Connected Layers
Pre-trained model from [18] and finetuned using liveness detection datasets.
CNN-Random CNN-Random +
PCA + SVM
Features are extracted using Convolutional Networks. The feature vector is reduced using PCA
and then fed into a SVM classifier using (Gaussian) RBF kernel.
LBP LBP + PCA + SVM Features are extracted using LBP. The feature vector is reduced using PCA and then fed into
a SVM classifier with (Gaussian) RBF kernel.
TABLE I
SUM MARY O F THE M OD ELS U SE D IN TH IS S TUDY.
network is composed of alternating layers of convolution and
local pooling (i.e., subsampling). The aim of a convolutional
layer is to extract patterns found within local regions of the
inputted images that are common throughout the dataset by
convolving a template over the inputted image pixels and
outputting this as a feature map c, for each filter in the layer.
A non-linear function f(c)is then applied element-wise to
each feature map c:a=f(c). A range of functions can be
used for f(c), with max(0; c)a common choice. The resulting
activations f(c) are then passed to the pooling layer. This
aggregates the information within a set of small local regions,
R, producing a pooled feature map s(normally of smaller size)
as the output. Denoting the aggregation function as pool(), for
each feature map c we have: sj=pool(f(ci))∀i∈Rj, where
Rjis the pooling region jin feature map cand iis the index
of each element within it. Among the various types of pooling,
max-pooling is commonly used, which selects the maximum
value of the region Rj.
The motivation behind pooling is that the activations in the
pooled map sare less sensitive to the precise locations of
structures within the image than the original feature map c.
In a multi-layer model, the convolutional layers, which take
the pooled maps as input, can thus extract features that are
increasingly invariant to local transformations of the input
image [26] [27]. This is important for classification tasks, since
these transformations obfuscate the object identity. Achieving
invariance to changes in position or lighting conditions, ro-
bustness to clutter, and compactness of representation, are all
common goals of pooling.
Figure 4 illustrates the feed-forward pass of a single layer
convolutional network. The input sample is convoluted with
three random filters of size 5x5 (enlarged to make visualization
easier), generating 3 convoluted images, which are then subject
to non-linear function max(x, 0), followed by a max-pooling
operation, and subsampled by a factor of 2.
In this study we compared three different models of convo-
lutional networks.
The first one, CNN-Random, uses only random filter
weights draw from a Gaussian distribution. Although the
filter weights can be learned, filters with random weights can
perform well and they have the advantage that they do not need
to be learned [28] [29] [30] . The architecture of the model is
the same as that used in [12]. It uses a convolutional network
with random weights as the feature extractor, the dimensions
are further reduced using PCA and a SVM classifier with RBF
kernels used as the classifier. An extensive search for hyper-
parameter fine-tune was performed automatically on more than
2000 combinations of hyper-parameters, listed in table II. The
best hyper-parameters were chosen per sensor and per dataset
(ex. Biometrika 2009, Bimetrika 2011, etc) through a 5x2 cross
validation method [31] which used the training dataset of each
sensor in each LivDet dataset (2009, 2011, 2013).
Pipeline
Element
Hyper-parameter Range
CNN-Random # Layers 1, 2, 3, 4, 5
CNN-Random # Filters (in each layer) 32, 64, ..., 2048
CNN-Random Filter Size Convolution 5x5, 7x7, ..., 15x15
CNN-Random Filter Size Pooling 3x3, 5x5, 7x7, 9x9
CNN-Random Stride (reduction fac-
tor)
2, 3, ..., 7
LBP Coding Standard or Uniform
LBP # Images Divisions 1x1 (no division), 3x3,
5x5, 7x7
PCA # Components 30, 100, 300, 500, 800,
1000, 1300
SVM Regularization Parame-
ter C
0.1, 1, ..., 105
SVM Kernel coefficient γ10−7,10−6, ..., 10−1
TABLE II
RAN GE OF HY PE R-PAR AME TER S SE ARC HE D FOR T HE CNN-RANDOM
AND LBP PIPELINES.
The second model, CNN-Alexnet, is very similar to AlexNet
[19], pre-trained on the ILSVRC-2012 dataset. This model
won both classification and localization tasks in the ILSVRC-
2012 competition. Their trained model has been used to
improve accuracy in a variety of other benchmarks such as
CIFAR-10, CIFAR-100. The pre-trained network provides a
good starting point for learning the network weights for other
tasks, such as fingerprint liveness detection.
The third model, CNN-VGG, is very similar to the one used
in [21], a 19 layer CNN which achieved the second place in
the detection task of the ImageNet 2014 challenge.
For CNN-ALEXNET and CNN-VGG models, the last
1000-unit soft-max layer (originally designed to predict 1000
classes) was replaced by a 2-unit softmax layer, which assigns
a score for true and fake classes. The pre-trained model was
further trained with the fingerprint datasets.
The algorithm used to train CNN-Alexnet and CNN-VGG
is the Stochastic Gradient Descent (SGD) with a minibatch of
size 5, using momentum [32] [33] 0.9 and a fixed learning

Fig. 4. Illustration of a sequence of operations performed by a single layer
convolutional network in a sample image.
rate of 10-6.
C. Local Binary Patterns
Local Binary Patterns (LBP) are a local texture descriptor
that have performed well in various computer vision applica-
tions, including texture classification and segmentation, image
retrieval, surface inspection, and face detection [34]. It is a
widely used method for fingerprint liveness detection [14] and
it is used in this work as a baseline method.
In its original version, the LBP operator assigns a label to
every pixel of an image by thresholding each of the 8 neigh-
bors of the 3x3-neighborhood with the center pixel value and
considering the result as a unique 8-bit code representing the
256 possible neighborhood combinations. As the comparison
with the neighborhood is performed with the central pixel, the
LBP is an illumination invariant descriptor. The operator can
be extended to use neighborhoods of different sizes [35].
Another extension to the original operator is the definition
of so-called uniform patterns, which can be used to reduce the
length of the feature vector and implement a simple rotation-
invariant descriptor [35]. An LBP is called uniform if the
binary pattern contains at most two bitwise transitions from 0
to 1 or vice versa when the bit pattern is considered circular.
The number of different labels of LBP reduces from 256 to
just 10 in the uniform pattern.
The normalized histogram of the LBPs (with 256 and 10
bins for non-uniform and uniform operators, respectively)
is used as a feature vector. The assumption underlying the
computation of a histogram is that the distribution of patterns
matters, but the exact spatial location does not. Thus, the
advantage of extracting the histogram is the spatial invariance
property. To investigate if location matters to our problem,
we also implemented the method presented in [36], for face
recognition, where the LBP filtered images are equally divided
in rectangles and their histograms are concatenated to form a
final feature vector.
In this study, the histogram of the LBP image was further
reduced using PCA, and a SVM with RBF kernel is used as the
classifier. Similarly to the CNN-Random models, the hyper-
parameters, such as the number of PCA components and SVM
regularization parameter, where found using an extensive brute
force search on more than 2000 combinations, listed in table
II.
D. Increasing the Classifiers Generalization through Dataset
Augmentation
Dataset Augmentation is a technique that involves artifi-
cially creating slightly modified samples from the original
ones. By using them during training, it is expected that the
classifier will become more robust against small variations
that may be present in the data, forcing it to learn larger (and
possibly more important) structures. It has been successfully
used in computer vision benchmarks such as in [19], [37],
and [38]. It is particularly suitable to out-of-core algorithms
(algorithms that do not need all the data to be loaded in
memory during training) such as CNNs trained with Stochastic
Gradient Descent. Our dataset augmentation implementation is
similar to the one presented in [19]: from each image of the
dataset five smaller images with 80% of each dimension of the
original images are extracted: four patches from each corner
and one at the center. For each patch, horizontal reflections
are created. As a result, we obtain a dataset that is 10 times
larger than the original one: 5 times are due to translations
and 2 times are due to reflections. At test time, the classifier
makes a prediction by averaging the individual predictions on
the ten patches.
III. EXP ER IM EN TS
A. Datasets
The datasets provided by the Liveness Detection Competi-
tion (LivDet) in the years of 2009 [39], 2011 [40], and 2013
[41] are used in this study.
LivDet 2009 comprises almost 18,000 images of real
and fake fingerprints acquired from three different sensors
(Biometrika FX2000, Crossmatch Verifier 300 LC, and Identix
DFR 2100). Fake fingerprints were obtained from three differ-
ent materials: Gelatin, Play Doh, and Silicone. Approximately
one third of the images of the dataset are used for training and
the remaining for testing.
LivDet 2011 comprises 16,000 images acquired from four
different sensors (Biometrika FX2000, Digital 4000B, Italdata
ET10, and Sagem MSO300), each having 2000 images of fake
and real fingerprints. Half of the dataset is used for training
and the other half for testing. Fake fingerprints were obtained
from four different materials: Gelatin, Wood Glue, Eco Flex,
and Silgum.
LivDet 2013 comprises 16,000 images acquired from four
different sensors (Biometrika FX2000, Crossmatch L SCAN
GUARDIAN, Italdata ET10, and Swipe), each having approx-
imately 2,000 images of fake and real fingerprints. Half of
the dataset is used for training and the other half for testing.
Fake fingerprints were obtained from five different materials:
Gelatin, Latex, Eco Flex, Wood Glue, and Modasil.

In all datasets, the real/fake fingerprint ratio is 1/1 and they
are equally distributed between training and testing sets. The
sizes of the images vary from sensor to sensor, ranging from
240x320 to 700x800 pixels, but they were all resized according
to the input size of the pre-trained models, which is 224x224
for the CNN-Alexnet model and 227x227 pixels for the CNN-
VGG model.
B. Performance Metrics
The classification results were evaluated by the Average
Classification Error (ACE), which is the standard metric for
evaluation in LivDet competitions. It is defined as
ACE =S F P R +S F NR
2(1)
where SFPR (Spoof False Positive Rate) is the percentage of
misclassified live fingerprints and SFNR (Spoof False Negative
Rate) is the percentage of misclassified fake fingerprints.
C. Implementation Details
CNN-VGG and CNN-Random were trained using the Caffe
package [42], which provides very fast CPU and GPU im-
plementations and a user-friendly interface in Python. For
the CNN-Random and LBP models, we wrote an improved
cross-validation/grid-search algorithm for choosing the best
combination of hyper-parameters, in which each element of
the pipeline is computed only when its training data is changed
(the term element refers to operations such as preprocessing,
feature extraction, dimensionality reduction or classification).
This modification speeded-up the validation phase by approx-
imately 10 times, although the gain can greatly vary as it
depends on the element types and number of hyper-parameters
chosen. An important aspect of this work is that the algorithms
were run on cloud service computers, where the user can
rent virtual computers and pay only for the hours that the
machines are running. To train the algorithms, we used the
GPU instances that allowed us to run dataset augmented
experiments in a few hours; using traditional CPUs the training
would take weeks.
IV. RES ULT S
The average error for each testing dataset is shown on
Table III. Along with the models used in this study, we also
show the error rate of the state-of-the-art method for each
dataset, of which most of them were found in the compilation
made by [43].
Particularly interesting results are for the Crossmatch 2013
dataset. As commented by [43], most techniques have prob-
lems in this dataset. For example, the LBP presents error rates
close to zero at validation time and around 50% at test time.
It can be noticed from LivDet 2013 competition results that
this dataset is particularly difficult to generalize, since nine of
the eleven participants presented error rates greater than 45%.
Contrary to these results, CNN models perform very well in
this dataset, with error rates between 3.2%-4.7%.
It is important to highlight that CNN-Random did require an
exhaustive hyper-parameter finetune (number of layers, filter
Dataset State-of-
the-Art
CNN-
VGG
CNN-
Alexnet
CNN-
Random
LBP
Crossmatch 2013 7.9 [13] 3.4 4.7 3.2 49.4
Swipe 2013 2.8 [43] 3.7 4.3 7.6 3.3
Italdata 2013 0.8 [41] 0.4 0.5 2.4 2.3
Biometrika 2013 1.1 [17] 1.8 1.9 0.8 1.7
Italdata 2011 11.2 [43] 8.0 9.1 9.2 12.3
Biometrika 2011 4.9 [11] 5.2 5.6 8.2 8.8
Digital 2011 2.0 [43] 3.2 4.6 3.6 4.1
Sagem 2011 3.2 [11] 1.7 3.1 4.6 7.5
Biometrika 2009 1.0 [11] 4.1 5.6 9.2 10.4
Crossmatch 2009 3.3 [43] 0.6 1.1 1.7 3.6
Identix 2009 0.5 [43] 0.2 0.4 0.8 2.6
Average 3.5 2.9 3.7 4.7 9.6
TABLE III
AVERA GE CLASSIFICATION ERRO R ON TES TI NG DATASE TS.
size, number of filters, etc.) in order to get a model with
good accuracy. On the other hand, the architectures of CNN-
Alexnet and CNN-VGG, which were already carefully selected
for the ImageNet object detection task, are general enough
to be reused for the fingerprint liveness detection task and
yield excellent accuracy. Another interesting aspect is that the
CNN-VGG performed better than the CNN-Alexnet in both
object detection from ILSVRC-2012 and fingerprint liveness
detection tasks. This suggests that further improvements in
models for object recognition can be applied to increase
accuracy in fingerprint liveness detection.
The higher performance of our CNN-VGG solution was
confirmed as this model won the first place in the Finger-
print Liveness Detection 2015 Competition (LivDet) 2015 [1],
with an overall accuracy of 95.51%, while the second place
achieved an overall accuracy of 93.23%.
A. Effect of dataset augmentation
Table IV compares the effect of dataset augmentation in
our proposed models. Despite its longer training and running
times, the technique helps to improve accuracy: the error was
reduced by a factor of 2 in some cases. More importantly,
the technique is not only effective on deep architectures, as
commonly known, but also in shallow architectures, such as
LBP.
Model No Augmentation With Augmentation
CNN-VGG 4.2 2.9
CNN-Alexnet 5.0 3.7
CNN-Random 9.4 4.7
LBP 21.2 9.6
TABLE IV
AUG MEN TATION V S NOAUGMENTATION: AVE RAG E ERRO R ON A LL
DATASE TS.
B. Cross-dataset Evaluation
We would like to verify how a classifier would perform
when unseen samples acquired from spoofy materials and
individuals during training are presented at test time. Addi-
tionally, we want to test the hypothesis that the images share
common characteristics for distinguishing fake fingerprints
from real ones, that is, the important features for classification

Train Dataset Test Dataset CNN-VGG CNN-Alexnet CNN-Random LBP
Biometrika 2011 Biometrika 2013 15.5 15.9 20.4 16.5
Biometrika 2013 Biometrika 2011 46.8 47.0 48.0 47.9
Italdata 2011 Italdata 2013 14.6 15.8 21.0 10.6
Italdata 2013 Italdata 2011 46.0 49.1 46.8 50.0
Biometrika 2011 Italdata 2011 37.2 39.8 49.2 47.1
Italdata 2011 Biometrika 2011 31.0 33.9 46.5 49.4
Biometrika 2013 Italdata 2013 8.8 9.5 47.9 43.7
Italdata 2013 Biometrika 2013 2.3 3.9 48.9 48.4
TABLE V
AVERA GE CLASSIFICATION ERRO R ON CRO SS-D ATAS ET E XPE RIM EN TS.
Dataset Materials - Train Materials - Test CNN-VGG CNN-Alexnet CNN-Random LBP
Biometrika 2011 EcoFlex, Gelatine, Latex Silgum, Wood Glue 10.1 12.2 13.5 17.7
Biometrika 2013 Modalsil, Wood Glue EcoFlex, Gelatine, Latex 4.9 5.8 10.0 8.5
Italdata 2011 EcoFlex, Gelatine, Latex Silgum, Wood Glue, Other 22.1 25.8 26.0 30.9
Italdata 2013 Modalsil, Wood Glue EcoFlex, Gelatine, Latex 6.3 8.0 10.8 10.7
TABLE VI
AVERA GE CLASSIFICATION ERRO R ON CRO SS- MATERIAL EXPERIMENTS.
are independent from the acquisition device. For that, Cross-
dataset experiments were performed, which involve training
a classifier using one dataset and testing on another. For
instance, a cross-dataset experiment would involve training a
classifier using Biometrika-2011 dataset and testing it using
Italdata-2013. In summary, these experiments should reflect
how well the classifier is able to learn relevant characteristics
that distinguish real from fake fingerprints when samples ac-
quired from different environments and sensors are presented.
We chose to use only Biometrika and Italdata sensors from
datasets of years of 2011 and 2013 of the LivDet competition,
since executing all possible dataset combinations would be
almost impractical to run under the current computer archi-
tecture. All the models evaluated use dataset augmentation.
Table V shows the testing error. CNN-Alexnet and CNN-
VGG clearly outperform CNN-Random and LBP in most
cases. However, the testing error is still high (>20%) in 4
out of 8 of the experiments, indicating that the models fail to
generalize when the type of sensor used for testing is different
from the one used in training. Similarly, [14] reported that
their multi-resolution LBP technique had poor results in cross-
device experiments, with errors of around 40-50%.
C. Cross-Material Evaluation
Additionally to the influence of training and testing with
different sensors (section IV-B), we investigated the perfor-
mance of the classifiers when they are tested with spoofing
materials never seen during training. The results are shown
in Table VI. The error rates are lower than Cross-dataset
experiments, which suggests that most of the generalization
error can be attributed to different sensors and not to different
materials.
D. Training All Datasets at Once
In this experiment we report the error rates when training
and testing a single classifier using all datasets (2009, 2011,
2013), except for Swipe-2013 whose images are very different
from the rest. The testing error rates, shown in Table VII, are
compared with the results obtained when individual classifiers
are trained per dataset, which are reported in Table III. The
results show that training a single classifier with all datasets
yields comparable error rates when individual classifiers are
trained per dataset, which suggests that the effort to design
and deploy a liveness detection system can be considerably
reduced if all datasets are trained together, as the hyper-
parameter fine tuning needs to be performed for only one
model.
Model One Classifier
trained with
All Training
Datasets
One Classifier
per Dataset
CNN-VGG 3.4 2.9
CNN-Alexnet 4.1 3.7
CNN-Random 6.0 4.7
LBP 10.0 9.6
TABLE VII
AVERA GE CLASSIFICATION ERRO R WHE N A SI NGL E CL ASS IFI ER IS
TR AIN ED U SIN G ALL D ATASET S VS ON E CL ASS IFI ER PE R DATASE T.
E. Pre-training Effect
In this experiment the effect of using pre-trained networks is
investigated. Table VIII compares the accuracy for the CNN-
VGG and CNN-Alexnet models trained using only fingerprint
images and when they are first pre-trained with the ImageNet
dataset and then finetuned with fingerprint images. It can be
seem that pre-training is necessary for those large networks
as training them using only the fingerprint images results in
overfitting.
Model Training on LivDet
datasets Only
Training on ImageNet
then LivDet datasets
CNN-VGG 49.4 (0.0) 2.9 (1.5)
CNN-Alexnet 48.1 (0.0) 3.7 (1.2)
TABLE VIII
AVERA GE CLASSIFICATION ERRO R FOR T ES TIN G DATASE T COM PARI NG
TH E EFFIC ACY O F PRE -TR AIN ED M ODE LS W ITH T HE O NES S OL ELY
TRAINED ON THE LIV DET DATASE TS . THE TRAINING ERROR IS SHOWED
IN PARENTHESIS

Fig. 5. Number of training samples vs Avg. Classification Error on all testing
datasets.
F. Number of Training Samples vs Error
Deep learning techniques require large number of labeled
training data in order to achieve a good performance when the
models are initialized with random weights, since there are a
lot of parameters that must be learned, thus requiring many
samples. However, when the weights were already learned
from another task, the number of required samples can be
surprisingly low in order to achieve good accuracy.
Figure 5 shows the number of training samples versus the
average classification error in the test set for all datasets. Using
only 400 training samples, CNN-VGG has almost the same
performance as LBP using all the 18,800 training images. This
suggests that less samples are needed when pre-trained models
are used.
G. Processing Times
In real applications, a good fingerprint liveness detection
system must be able to classify images in a short amount of
time. Table IX shows the average testing/classification times
for a single image (no augmentation) on a single core machine
(1.8 GHz, 64-bit, with 4GB memory).
We also show the times for training all datasets together.
The pre-trained CNN models (CNN-Alexnet and CNN-VGG)
take around 5-40 hours to converge using a Nvidia GTX Titan
GPU. The CNN-Random and LBP models take around 5-10
hours to converge on a 32-Cores machine (the larger portion
of these times are required for dimensionality reduction using
PCA).
Technique Training all Datasets Testing per
Image (1-core
CPU)
CNN-VGG 20-40 hours (GPU) 650ms
CNN-Alexnet 5-10 hours (GPU) 230ms
CNN-Random 5-10 hours (32-core CPU) 110ms
LBP 5-10 hours (32-core CPU) 50ms
TABLE IX
AVERA GE TRAINING AND TES TIN G TIMES.
V. CONCLUSIONS
Convolutional Neural Networks were used to detect false vs
real fingerprints. Pre-trained CNNs can yield state-of-the-art
results on benchmark datasets without requiring architecture
or hyperparameter selection. We also showed that these models
have good accuracy on very small training sets (˜
400 samples).
Additionally, no task-specific hand-engineered technique was
used as in classical computer vision approaches.
Despite the differences between images acquired from
different sensors, we show that training a single classifier
using all datasets helps to improve accuracy and robustness.
This suggests that the effort required to design a liveness
detection system (such as hyper-parameters fine tuning) can
be significantly reduced if different datasets (and acquiring
devices) are combined during the training of a single clas-
sifier. Additionally, the pre-trained networks showed stronger
generalization capabilities in cross-dataset experiments than
CNN with random weights and the classic LBP pipeline.
Dataset augmentation plays an important role in increasing
accuracy and it is also simple to implement. We suggest
that the method should always be considered for the training
and prediction phases if time is not a major concern. Given
the promising results provided by the technique, more types
of image transformations should be included, such as color
manipulation and multiple scales described in [44] and [45].
ACKNOWLEDGMENT
We would like to thank Nvidia for the donation of the GPUs
used in this study. Roberto A Lotufo thanks Conselho Nacional
de Desenvolvimento Cientfico e Tecnolgico (311228/2014-3)
for sponsorship.
REFERENCES
[1] V. Mura, L. Ghiani, G. L. Marcialis, F. Roli, D. A. Yambay, and
S. A. Schuckers, “Livdet 2015 fingerprint liveness detection competition
2015,” in Biometrics Theory, Applications and Systems (BTAS), 2015
IEEE 7th International Conference on. IEEE, month 2015, pp. 1–6.
[2] J. Galbally, F. Alonso-Fernandez, J. Fierrez, and J. Ortega-Garcia, “A
high performance fingerprint liveness detection method based on quality
related features,” Future Generation Computer Systems, vol. 28, no. 1,
pp. 311–321, 2012.
[3] Y. Chen, A. Jain, and S. Dass, “Fingerprint deformation for spoof
detection,” in Biometric Symposium, 2005, p. 21.
[4] B. Tan and S. Schuckers, “Comparison of ridge-and intensity-based
perspiration liveness detection methods in fingerprint scanners,” in
Defense and Security Symposium, vol. 6202. International Society for
Optics and Photonics, 2006, pp. 62 020A–62 020A.
[5] P. Coli, G. L. Marcialis, and F. Roli, “Fingerprint silicon replicas: static
and dynamic features for vitality detection using an optical capture
device,” International Journal of Image and Graphics, vol. 8, no. 04,
pp. 495–512, 2008.
[6] P. D. Lapsley, J. A. Lee, D. F. Pare Jr, and N. Hoffman, “Anti-fraud
biometric scanner that accurately detects blood flow,” Apr. 7 1998, uS
Patent 5,737,439.
[7] A. Antonelli, R. Cappelli, D. Maio, and D. Maltoni, “Fake finger de-
tection by skin distortion analysis,” Information Forensics and Security,
IEEE Transactions on, vol. 1, no. 3, pp. 360–373, 2006.
[8] D. Baldisserra, A. Franco, D. Maio, and D. Maltoni, “Fake fingerprint
detection by odor analysis,” in Advances in Biometrics. Springer, Berlin,
Heidelberg, 2005, pp. 265–272.
[9] A. K. Jain, Y. Chen, and M. Demirkus, “Pores and ridges: High-
resolution fingerprint matching using level 3 features,” Pattern Analysis
and Machine Intelligence, IEEE Transactions on, vol. 29, no. 1, pp.
15–27, 2007.

[10] D. Gragnaniello, G. Poggi, C. Sansone, and L. Verdoliva, “Fingerprint
liveness detection based on weber local image descriptor,” in Biometric
Measurements and Systems for Security and Medical Applications
(BIOMS), 2013 IEEE Workshop on. IEEE, 2013, pp. 46–50.
[11] ——, “Local contrast phase descriptor for fingerprint liveness detection,”
Pattern Recognition, vol. 48, no. 4, pp. 1050–1058, 2015.
[12] R. Frassetto Nogueira, R. de Alencar Lotufo, and R. Campos Machado,
“Evaluating software-based fingerprint liveness detection using convolu-
tional networks and local binary patterns,” in Biometric Measurements
and Systems for Security and Medical Applications (BIOMS) Proceed-
ings, 2014 IEEE Workshop on. IEEE, 2014, pp. 22–29.
[13] D. Menotti, G. Chiachia, A. Pinto, W. Robson Schwartz, H. Pedrini,
A. Xavier Falcao, and A. Rocha, “Deep representations for iris, face,
and fingerprint spoofing detection,” Information Forensics and Security,
IEEE Transactions on, vol. 10, no. 4, pp. 864–879, 2015.
[14] X. Jia, X. Yang, K. Cao, Y. Zang, N. Zhang, R. Dai, X. Zhu, and
J. Tian, “Multi-scale local binary pattern with filters for spoof fingerprint
detection,” Information Sciences, vol. 268, pp. 91–102, 2014.
[15] L. Ghiani, G. L. Marcialis, and F. Roli, “Fingerprint liveness detection
by local phase quantization,” in Pattern Recognition (ICPR), 2012 21st
International Conference on. IEEE, 2012, pp. 537–540.
[16] S. B. Nikam and S. Agarwal, “Local binary pattern and wavelet-based
spoof fingerprint detection,” International Journal of Biometrics, vol. 1,
no. 2, pp. 141–159, 2008.
[17] L. Ghiani, A. Hadid, G. L. Marcialis, and F. Roli, “Fingerprint liveness
detection using binarized statistical image features,” in Biometrics:
Theory, Applications and Systems (BTAS), 2013 IEEE Sixth International
Conference on. IEEE, 2013, pp. 1–6.
[18] P. Y. Simard, D. Steinkraus, and J. C. Platt, “Best practices for convo-
lutional neural networks applied to visual document analysis,” in null.
IEEE, 2003, p. 958.
[19] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Advances in neural infor-
mation processing systems, 2012, pp. 1097–1105.
[20] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,
Z. Huang, A. Karpathy, A. Khosla, M. Bernstein et al., “Imagenet large
scale visual recognition challenge,” International Journal of Computer
Vision, pp. 1–42, 2014.
[21] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” CoRR, vol. abs/1409.1556, 2014.
[22] S. Karayev, M. Trentacoste, H. Han, A. Agarwala, T. Darrell, A. Hertz-
mann, and H. Winnemoeller, “Recognizing image style,” arXiv preprint
arXiv:1311.3715, 2013.
[23] Y. LeCun et al., “Generalization and network design strategies,” Con-
nections in Perspective. North-Holland, Amsterdam, pp. 143–55, 1989.
[24] L. Wan, M. Zeiler, S. Zhang, Y. L. Cun, and R. Fergus, “Regularization
of neural networks using dropconnect,” in Proceedings of the 30th
International Conference on Machine Learning (ICML-13), 2013, pp.
1058–1066.
[25] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep
network training by reducing internal covariate shift,” arXiv preprint
arXiv:1502.03167, 2015.
[26] Y.-L. Boureau, J. Ponce, and Y. LeCun, “A theoretical analysis of feature
pooling in visual recognition,” in Proceedings of the 27th International
Conference on Machine Learning (ICML-10), 2010, pp. 111–118.
[27] M. D. Zeiler and R. Fergus, “Stochastic pooling for regularization of
deep convolutional neural networks,” arXiv preprint arXiv:1301.3557,
2013.
[28] Y. LeCun, K. Kavukcuoglu, and C. Farabet, “Convolutional networks
and applications in vision,” in Circuits and Systems (ISCAS), Proceed-
ings of 2010 IEEE International Symposium on. IEEE, 2010, pp. 253–
256.
[29] K. Jarrett, K. Kavukcuoglu, M. Ranzato, and Y. LeCun, “What is the best
multi-stage architecture for object recognition?” in Computer Vision,
2009 IEEE 12th International Conference on. IEEE, 2009, pp. 2146–
2153.
[30] A. Saxe, P. W. Koh, Z. Chen, M. Bhand, B. Suresh, and A. Y. Ng,
“On random weights and unsupervised feature learning,” in Proceedings
of the 28th International Conference on Machine Learning (ICML-11),
2011, pp. 1089–1096.
[31] T. G. Dietterich, “Approximate statistical tests for comparing supervised

classification learning algorithms,” Neural computation, vol. 10, no. 7,
pp. 1895–1923, 1998.
[32] B. T. Polyak, “Some methods of speeding up the convergence of iter-
ation methods,” USSR Computational Mathematics and Mathematical
Physics, vol. 4, no. 5, pp. 1–17, 1964.
[33] I. Sutskever, J. Martens, G. Dahl, and G. Hinton, “On the importance
of initialization and momentum in deep learning,” in Proceedings of the
30th international conference on machine learning (ICML-13), 2013,
pp. 1139–1147.
[34] A. Hadid, M. Pietik¨
ainen, and T. Ahonen, “A discriminative feature
space for detecting and recognizing faces,” in Computer Vision and
Pattern Recognition, 2004. CVPR 2004. Proceedings of the 2004 IEEE
Computer Society Conference on, vol. 2. IEEE, 2004, pp. II–797.
[35] T. Ojala, M. Pietik¨
ainen, and T. M¨
aenp¨
a¨
a, “Multiresolution gray-scale
and rotation invariant texture classification with local binary patterns,”
Pattern Analysis and Machine Intelligence, IEEE Transactions on,
vol. 24, no. 7, pp. 971–987, 2002.
[36] T. Ahonen, A. Hadid, and M. Pietik¨
ainen, “Face recognition with local
binary patterns,” in Computer vision-eccv 2004. Springer, Berlin,
Heidelberg, 2004, pp. 469–481.
[37] D. C. Cires¸an, U. Meier, J. Masci, L. M. Gambardella, and J. Schmidhu-
ber, “High-performance neural networks for visual object classification,”
arXiv preprint arXiv:1102.0183, 2011.
[38] D. Ciresan, U. Meier, and J. Schmidhuber, “Multi-column deep neural
networks for image classification,” in Computer Vision and Pattern
Recognition (CVPR), 2012 IEEE Conference on. IEEE, 2012, pp. 3642–
3649.
[39] G. L. Marcialis, A. Lewicke, B. Tan, P. Coli, D. Grimberg, A. Congiu,
A. Tidu, F. Roli, and S. Schuckers, “First international fingerprint
liveness detection competitionlivdet 2009,” in Image Analysis and
Processing–ICIAP 2009. Springer, Berlin, Heidelberg, 2009, pp. 12–23.
[40] D. Yambay, L. Ghiani, P. Denti, G. L. Marcialis, F. Roli, and S. Schuck-
ers, “Livdet 2011fingerprint liveness detection competition 2011,” in
Biometrics (ICB), 2012 5th IAPR International Conference on. IEEE,
2012, pp. 208–215.
[41] L. Ghiani, D. Yambay, V. Mura, S. Tocco, G. L. Marcialis, F. Roli, and
S. Schuckcrs, “Livdet 2013 fingerprint liveness detection competition
2013,” in Biometrics (ICB), 2013 International Conference on. IEEE,
2013, pp. 1–6.
[42] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick,
S. Guadarrama, and T. Darrell, “Caffe: Convolutional architecture for
fast feature embedding,” in Proceedings of the ACM International
Conference on Multimedia. ACM, 2014, pp. 675–678.
[43] D. Gragnaniello, G. Poggi, C. Sansone, and L. Verdoliva, “An investiga-
tion of local descriptors for biometric spoofing detection,” Information
Forensics and Security, IEEE Transactions on, vol. 10, no. 4, pp. 849–
863, 2015.
[44] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
V. Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,”
arXiv preprint arXiv:1409.4842, 2014.
[45] A. G. Howard, “Some improvements on deep convolutional neural
network based image classification,” arXiv preprint arXiv:1312.5402,
2013.

Rodrigo Frassetto Nogueira received his B.S. in
Electrical Engineering from University of Campinas
(Unicamp), Brazil, in 2009 and obtained his Masters
degree in Computer Engineering in 2014 from that
same university. He is a PhD student at the School
of Engineering, New York University, USA, since
2014. His principal research interests are in the areas
of Machine Learning, Computer Vision and Natural
Language Processing.
Roberto A. Lotufo Roberto A. Lotufo received the
B.S. degree from Instituto Tecnologico de Aeronau-
tica, Brazil, in 1978, and the Ph.D. degree from the
University of Bristol, U.K., in 1990, in Electrical
Engineering. He is a full professor at the School
of Electrical and Computer Engineering, University
of Campinas (Unicamp), Brazil, were he has worked
for since 1981. His principal research interests are in
the areas of Image Processing and Analysis, Pattern
Recognition and Machine Learning. Prof. Lotufo has
published over 150 refereed international journal and
full conference papers. He was awarded the Innovation Personality in 2008 and
the Zeferino Vaz Academic Recognition in 2011 from University of Campinas.
Rubens Campos Machado received the BSc degree
in Electonic and Telecommunications Engineering
from the Catholical University of Minas Gerais,
Brazil, in 1978. In 2002 he received the MSc degree
in Electrical Engineering from the University of
Campinas, Brazil. He works for the Renato Archer
Center of Information Technology, CTI, Campinas,
Brazil, since 1983. In CTI, he headed the Realtime
Processing Division from 1984 to 1987, from 1987
to 1996, he was the head of the Process Con-
trol Department and headed the Applied Control
Methodologies Division from 1996 to 2000. During these periods, he develops
advanced automation projects for several brazilian industries. Currently, he
is a Senior Researcher of the CTI Robotics and Computer Vision Division.
His main interest areas are Computer Vision and Machine Learning/Deep
Learning.