A presentation attack consists in submitting to the fingerprint capture device an artificial replica of the finger of the targeted client. If the sensor is not equipped with an appropriate algorithm aimed to detect the fingerprint spoof, the system processes the obtained image as a one belonging to a real fingerprint. In order to face this problem, several presentation attacks detection (PAD) algorithms have been proposed so far. Current methods heavily rely on features extracted from a large data set of fake and real fingerprint images, and an appropriate classifier trained with such data to distinguish between live (real) and fake (spoof) fingerprint images. Building such data set requires a significant effort for fabricating samples of fake fingerprints, with the most effective materials used to circumvent the sensor. Interesting and promising results have been obtained, but they also suggest that the PAD is tailored on the particular sensor. Small and significant differences also occur when a novel version of the same sensor is released, and this may affect the PAD. Therefore, making a PAD interoperable is among the main current issues when considering fingerprints as the first level of protection and security of logical or physical resources. This paper is a first attempt to assess at which extent the sensor interoperability can be an issue for fingerprint PADs and to eventually propose a solution to this limitation. In particular, textural features will be under focus and a feature space transformation method based on the least square is proposed.
Fingerprint capture devices suffer from the presentation attacks problem [
Fingerprint verification systems, without an appropriate detector of such spoofs, may process
the fingerprint image as belonging to an authentic one. If the image quality and the replica are
good enough, very similar to the original fingeprint, they can be circumvented [
This problem has been firstly pointed out in [
Although many promising results have been obtained, especially by features extracted with
textural algorithms, there are significant issues still open. Among others, it is normal in
fingerprint verification systems to substitute the capture device with a newest and better one,
1See http://livdet.diee.unica.it for further information on the competition and data sets.
as a sort of upgrade. Unfortunately, this is not possible without updating templates as well2.
In other words, there is not interoperability among fingerprint sensors, due to the different
characteristic of image captured and, in some cases, on pre-processing steps operated before
making available the image to the verification step [
In this paper, we answer the question: “is the interoperability also a problem for fingerprint presentation attacks detectors?”. By an extensive experimental analysis on the data sets presented at LivDet 2011, we show that the problem exists (Section 2). On the other hand, we also show that the problem can be mitigated by an appropriate domain transformation algorithm, thus allowing to preserve the performance on the system during upgrade and downgrade (Sections 3-4). Conclusions are drawn in Section 5. 2
Each fingerprint scanner model is usually different from the others in both hardware and
software. The main hardware difference is the sensor type: optical, solid state and ultrasound [
Software based differences are at API level, there can be many pre-processing steps in order
to highlights ridges and valleys. Every step changes the dynamic of gray levels. The fingerprint
matcher (comparator) based on the minutiae position, does not depend on these operations, as
opposed to the performance of liveness detector that is based on a high frequency analysis. As
a matter of fact the most important algorithms in FPAD like LBP [
In Fig. 1 we are able to appreciate the differences of geometric distortion with two optical 2The template is a model stored in the data base and related to the registered subject, that is, authorized to pass the fingerprint verification test 0.01 fingerprint capture devices, namely, Biometrika FX2000 and Italdata, adopted in the second edition of the International Competition on Figerprint Liveness Detection. The Biometrika capture device is more rounded than the Italdata one (both images belong to same fingerprint of the same person). The background is quite different, in the Italdata is almost white whereas in the Biometrika some shadows appear in the corners.
Another peculiarity is the gray level histogram. The example on the left side of Fig. 2 shows the image gray scale average histogram of Biometrika for the two subsets released for LivDet2011 and named training and test sets. In order to remove background effect, each fingerprint image is cropped, finding the center, in a ROI of 200x200 pixel. The trend of the four plots is quite similar. The main difference is in the highest gray levels between live and fake images. Moreover, on the right side of Fig. 2, differences among the four fingerprint scanners by plotting the average histogram of gray levels for all images (Biometrika, Digital Persona, Sagem, Italdata) are shown. As a consequence of those differences among sensors, we may expect that the feature vectors, calculated using for example textural algorithms, are different as well. The textural algorithms work very well in FPAD, but they can be used independently from the scanner. In other words, it may be expected no interoperability among scanners, because textural algorithms capture both small details of the fingerprint which are useful for our target and scanner-specific information.
In the following experiment we demonstrate this assertion using the four LivDet 2011 data sets [19], divided into a subset of 2000 live and 2000 fake samples.
Let us refer for example to two data sets coming from different sensors as A and B. We
further subdivide these sets into train set T rainA and test set T estA such that A = T rainA ∪
T estA and T rainB and T estB such that B = T rainB ∪ T estB. Let us then define the set
trainAB = T rainA ∪ T rainB and the set as testAB = testA ∪ testB. In this experiment,
we train a classifier to discriminate among the two scanners images. In order to estimate the
capability to discriminate the device, the training process is repeated with an increasing number
of samples. The used classifier is a linear SVM [
The used features are LBP [
Fig. 3 shows the mean accuracy rate for all the possible scanner pairs by increasing the number of training patterns. As expected, the performance grows up with the number of pattern, but even with a few samples the accuracy is more than 85%. Therefore, just a few patterns 100 Accuracy Dataset LivDet2011 algorithm LBP
Italdata−Biometrika Italdata−Digital Italdata−Sagem Biometrika−Digital Biometrika−Sagem
Digital−Sagem 85 101 number of samples per dataset 103 102 cy 95 ca r cu A n a eM90 85
Italdata−Biometrika Italdata−Digital Italdata−Sagem Biometrika−Digital Biometrika−Sagem
Digital−Sagem 101 number of samples per dataset 103 102 cy 95 ca r cu A n a eM90 85 Accuracy Dataset LivDet2011 algorithm bsif win=5x5 12bit 100 Italdata−Biometrika
Italdata−Digital Italdata−Sagem Biometrika−Digital Biometrika−Sagem
Digital−Sagem 101 number of samples per dataset 103 102 are sufficient to identify the capture device. This shows the existence of the interoperability problem.
The second evidence is reported in Tables 1,2,3 where the liveness detection accuracy and cross-accuracy among fingerprint capture devices is reported. For example, the row 3 and the column 2 reports the liveness detection accuracy on features extracted from images coming from the Biometrika sensor submitted to the classifier trained on features extracted from images coming to the Italdata sensor. These values indicate that the features space of live and fake images is differently distributed for all scanners.
Based on these evidences, we can conclude that there is not interoperability among fingerprint capture devices even when liveness of fingerprint traits must be assessed. In other words, the liveness detector should be tailored on the specific sensor, by using an appropriate set of live and fake fingerprint images.
test 91.95% 70.40% 50.00% 53.05%
test 50.00% 86.15% 49.80% 50.00%
In the previous sections we have shown, from a high level viewpoint, that it is not possible to substitute a sensor with a novel and better one without updating the whole liveness detection algorithm. In other words, the interoperability problem, already pointed out for fingerprint matching, also exists when upgrading a fingerprint liveness detection system. According to our results, this mainly affects the feature space distribution of live and fake images, especially if we adopt textural features, thus making impossible to keep the same classifier trained on features extracted from the “old” sensor. The practical consequence is that, if the new sensor must be used, a novel data set of live and fake fingerprints must be captured and a novel classifier must be trained. This can be quite expensive and difficult to obtain in a short time. Therefore, we propose here a method that can be adopted to mitigate the problem. “Mitigating” the problem means to allow users to adopt the new sensor whilst keeping the old classifier, by paying a drop of performance less significant that the one shown in the previous Section, until a novel data set is fully available.
To this aim, we used the least square method [17] that is adopted to minimize the error in the ill posed problem A × x = b. In this problem A is an n × m matrix, x and b are two m and n elements vectors. Since we want to transform one n × m matrix into another, our problem can be divided in m steps:
A × xi = bi
i = 1, 2, ..., m
In each of these step we calculate the column vector xi that, from the matrix A, generate the i − th column vector of B.
Basically we select the features extracted from the datasets corresponding to two different sensors and calculate the matrix X that allows, with a simple product, to transform one feature set into the other by minimizing the squared error. In other words, we adapt the domain of the new sensor to that of the old one. We use an approximation algorithm instead of an interpolation algorithm because we want to avoid the overfitting of the new space on the old one.
Given the two matrices trainA and trainB, according to the terms given in the previous Section, we calculate the matrix X such that:
trainB = trainA × X
With this method we can turn every feature space A in any other feature space B by paying attention to the fact that we want to maintain the distinction between lives and fakes, that is, we want the live features of A moving on the live features of B and the fake features of A moving on the fake features of B. Thus, if each row of trainA and trainB contains a feature (1) (2) vector, we have to be sure that for every row to each live (or fake) in trainA correspond a live (or fake) in trainB. Clearly the two matrices must have the same number of rows (number of feature vectors) and columns (feature vector elements).
Let us suppose we have a large number of collected images with sensor B with which we
trained a classifier and to have a new sensor A with which we want to build another dataset. If
the number of images acquired with the new sensor is limited, can we exploit the other classifier
moving the feature space of A in that of B? Can we just use live fingerprints or is it required
to also collect fake samples? How many feature vectors do we need to get satisfying results?
In order to answer these questions we performed a number of experiments by analyzing the
LivDet 2011 datasets [19]. Features were extracted with three different textural algorithms:
LBP[
In our first experiment we used all possible pairs of LivDet2011 datasets. For each couple of datasets A and B, given the trainA and trainB matrices, we calculate the transformation matrix X (see Eq. 2) that allows us to move a vector from the A feature space to the B one. With the X and the testA matrices we are easily able to compute a pseudoT estB matrix moving the testA feature vectors in the B feature space: pseudoT estB = testA × X (3)
In Table 4 we present the obtained results in terms of accuracy. The values in the diagonal, since A and B represent the same dataset, are obtained with the classical procedure and without any transformation: we just trained a classifier with trainB and we used it to classify the feature vectors in testB. Conversely, for each off-diagonal value, we still trained a classifier with trainB (rows), but then we calculated the matrix X by Eq. 2 and the pseudoT estB by Eq. 3. Finally, we classified the feature vectors in pseudoT estB (columns).
pseudo-test 88.85% 86.50% 86.55% 83.90%
pseudo-test 80.10% 81.35% 79.65% 77.15%
pseudo-test 84.15% 85.60% 89.40% 89.35%
pseudo-test 86.25% 88.80% 85.07% 91.65%
The presented results clearly show that the accuracies obtained by transforming testA in a pseudoT estB are comparable (but almost always lower) with those obtained by simply training the original dataset trainB. In other words, the effect of the passage from a sensor to another one is mitigated or reduced. 3.2
The results in Section 3.1 are obtained using both lives and fakes feature vectors to calculate the transformation matrix. Since the acquisition of new fakes is not as easy and fast as collecting live samples, which is not a trivial problem as well, we repeated the experiments using the live samples to calculate the transformation matrix.
Unfortunately, as shown in Table 5 the results are much worse, but this is also explainable since we have willingly reduced the information available for the transformation, by avoiding the use of fake samples. Whilst this is an indirect evidence that textural algorithms are able to extract live and fake fingerprint characteristics, this also show that it is not possible to avoid this information when designing the domain adaptation function. As a matter of fact, the majority of fake fingerprints were classified as live ones in our experiments. In the previous subsections the transformation matrices have been calculated using all the feature vectors (lives and fakes in Section 3.1 and just lives in Section 3.2) extracted from the images in the train parts of the LivDet 2011 datasets [19] (approximately 2000, 1000 lives and 1000 fakes). In order to simulate a limited number of acquisition, we replicate the experiments using just a randomly selected subset of those feature vectors (with the same number of lives and fakes in both subset). Given the two matrices trainA and trainB we extract the two subset subSetA and subSetB and calculate the new matrix X such that: subSetA = subSetB × X (4)
In the experiments presented in Tables 6 - 9 the transformation matrices were calculated with subsets containing the 20%, 40%, 60% and 80% of the feature vectors of the original train datasets. train seems to be sufficient to obtain satisfying results. This means that it is possible to keep the old classifier without an appreciable loss of performance and using at the same time the new sensor, until a good number of live and fake samples has been collected. 4
In this paper, we analyzed the level of interoperability between sensors in the field of fingerprint presentation attacks (liveness) detection. We showed that the problem exists and that it is difficult to overcome since the sensor characteristics heavily affect the image characteristics and the related feature space when using textural algorithms. These features contain so much “sensor-specific” information that the related images can be localized in different regions of the features space itself, almost without overlapping.
Starting from these observations we proposed a method to strongly reduce these sensorspecific features distributions. This method is based on the least square algorithm, pros and cons were also discussed in the paper. Although the pointed out limitations, it allowed achieving a considerable level of interoperability among fingerprint capture devices.
In the future we will focus our efforts on the improvement of the performances but also on the reduction of the number of feature vectors required to calculate the transformation matrix. In particular, calculating the domain transformation equation without the need to use fake samples would be of great help to design an effective algorithm for capture devices interoperability in fingerprint liveness detection.
Biometrika pseudo-test 80.64% 80.83% 82.00% 80.34%
pseudo-test 72.67% 75.38% 75.22% 77.11% [14] Shankar Bhausaheb Nikam and Suneeta Agarwal. Local binary pattern and wavelet-based spoof fingerprint detection. International Journal of Biometrics, 1(2):141–159, 2008. [15] Arun Ross and Anil Jain. Biometric Sensor Interoperability: A Case Study in Fingerprints, pages 134–145. Springer Berlin Heidelberg, Berlin, Heidelberg, 2004. [16] S. Schuckers. Spoofing and anti-spoofing measures. Inf. Sec. Techn. Report, 7(4):56–62, 2002. [17] L. Ridgway Scott. Numerical Analysis. Princeton University Press, Princeton, NJ, USA, 2011. [18] Craig I. Watson, Michael D. Garris, Elham Tabassi, Charles L. Wilson, R. Michael Mccabe, Stanley
Janet, and Kenneth Ko. User’s guide to nist biometric image software (nbis). [19] D. Yambay, L. Ghiani, P. Denti, G. L. Marcialis, F. Roli, and S. Schuckers. Livdet 2011 - fingerprint liveness detection competition 2011. In 2012 5th IAPR International Conference on Biometrics (ICB), pages 208–215, March 2012.