Modeling Fingerprint Presentation Attack Detection Through Transient Liveness Factor-A Person Specific Approach

This section sums up related work in two parts, close-set and open-set approach to spoof detection. The purpose here is to acknowledge past practices. Other than this we wish to understand chronology in action, Intentions of past work and the opportunity for future.

2.1 FPAD in close-set scenario, a conservative approach

Handling fingerprint presentation attack in close-set scenario identifies the following problem areas [5] in the current spoof detection paradigm:

(1) Non-uniform evaluation scheme for performance measures of FPAD-system.

(2) FPAD-system works in specific condition & depends on external factor.

A. Sensor variation.

B. Spoof-Material variation.

C. Limited-Universe of Dataset

(3) Limited scope and constraint mitigation approach.

We can see the FPAD solution improves every year in comparison of previous year benchmark, which further motivates developers to propose an improved solution in subsequent years. Many researchers are working in this broad area, several good publications coming every year. These trends improve confidence and reflect researcher interest in this problem area. But generalization capability and reproducible results were missing in the past causes interoperability issues.

A large-scale empirical study in the work of Ref. [13, 14] highlights interoperability issues. We often use biometric systems in supervisory, defense services and financial/civilian operations. There is a variety of sensors and matching algorithms available from various vendors. This makes up an aggressive market for these products, which is suitable for the users but shows the relevance of interoperability. Interoperability is the capacity of a biometric system to deal with changes introduced in the biometric sample because of various capture mechanisms. Using various biometric devices may increase failure percentages. It shows the extensive-scale experimental investigation of the condition of interoperability between fingerprint sensors and determines the performance effect when interoperability countermeasure is lacking.

The research community still cannot define a more general FPAD system. There are lots of many generalization attributes that hampers FPAD implementation. We may use one broad generalization; PAD applies across boundaries of Geographical/Political/ Individual beliefs. On the same line, the PAD system expects to work on different hardware conditions too. Most of the techniques having anti-spoofing capability do not perform well with generalization and robustness across geographical boundaries. So, Anjos et al. [15] propound the need for an open-source platform for testing and reporting FPAD. As per them, it is difficult to reproduce results too. Many researchers often overlook this requirement, or rather address it in an optimistic and in biased manner. Natural complications of R&D ‘research and development’ as discussed in the study [16] a probable reason behind such a gap in breakthrough solutions. Other essential reasons are, biometric data is private and protected data in the public domain, so hard to get sample; software frameworks difficult to install and maintain; an extensive set of intricate steps in the test protocol difficult to handle.

In a nutshell, underlying constraint includes research complexity, lack of the universe of data for testing, reproducible results and generalization requirement. If we survey the above discussion in deep, the researcher is working on ‘How to Do’ without addressing ‘What to Do’. We can ask a research question ‘Is it possible to model liveness’, if yes, what to do next. Shift our focus, it is better to identify presentation attack using liveness factors instead of spoof characterization. To generalize the PAD model, learning spoof samples may not be a good idea. Before we propose our approach, how to deal with fingerprint attack detection in its most innate ‘Well formed, organized, natural’ way, it is important to look FAPD approaches in open-set environment. Next subsection summarizes efforts of researchers and issues left by them while handling FPAD in a more liberal way.

2.2 FPAD in open-set scenario, a liberal approach

The need for generalized liveness detection model arises because of very few independent results that are available and out of which largely self-declared [17] in close-set FPAD system. Hence, it is difficult to deploy large scale and wide area based biometric authentication systems integrated with liveness detector. So, it is now time to implement the FPAD algorithm over and above testbed system that works in specific conditions. A FAPD system is a must and integral part within an authentication system. It is generalizing enough as a system and reproduces result too. Thus, we may not argue over which spoof pattern classification technique is best, we focus on a wider context of spoof detection that are universal.

Classification of spoof sample as a live sample, a false positive (FP) response is due to lack of a wider context which is always missing in close-set scenario. These are the recognition problem where full knowledge of spoof and live samples are limited (~2000-4000 sample), at the time of training. If we test this model with a known sample using various cross-validation approaches it give the best results as an average classification error ‘ACE’ of (~3) [18]. On a complementary course to generalize a PAD operation, in an open-set implementation, there is requirements for every type of representative dataset, which may not be a great idea. On contrary, the earlier methods proposed under close-set approach succumbed on a limited dataset, weak to detect unknown sample. In recent past, the world is now opening up, biometrics are being used in many organizations and many applications. Which is why many efforts are being made to create an open set solution. This section summaries them and conclude its limitation.

In case of capable intruders, a new fake sample presented to the sensor is always there. The work by argued that one cannot set our aim to model the fake samples in totality. With two reasons, first-fake sample are available in limited numbers to get them train and we do not know it beforehand [19]. So, they proposed a semi-supervised approach for live samples. The work by Kho et al. [20] propounded the identical scheme handle this obstacle, they come up with an existing spoof detector that should learn in steps to sustain this problem. A strong observation by Evans et al. [21] emphasized that the research community had taken one step forward to develop generalized countermeasure approach, but two-step backwards in terms of data homogeneity and reproducible research. So, the need for a de facto standard on datasets, metrics and protocols for spoof detection is a must and still missing. A recent survey by Galbally et al. [22] has also addressed a generalized approach to counter liveness detection. They enlisted with an example of how to design, develop and validate FPAD approaches based on benchmarks and what should be a systematic and replicable protocol.

There is an increasing number of investigations that apply the Convolutional Neural Network (CNN) in recent past [8], [23-26]. CNN based approach has high categorization performance for detecting fake fingerprints with best results in known material as average classification error of 0.25 [9]. But, for unknown-material and cross-sensor data-set CNN approach fallen on the classification error. Apart from the above performance variance, CNN has following accompanying issues in past for fake fingerprint detection approach. CNN model learns parameter from images of all kinds through a pre-trained network. Thus, parameters tuned here cannot be considered as optimized parameters for fake fingerprint detection. They give most of the concentration on increasing classification performance rather than looking at the size of data-set while training the model. Many investigations [27, 28] have shown that improved performance is attained when a deeper and wider network is used. But, when examining the expandability FPAD technique, the model must incorporate only narrow parameters so it can run in a slow machine. Further, many pre-processing steps such as cropping, segmentation, resize of the input image before applying it to CNN, thus an FPAD through a CNN cannot be proposed in generalization and with isolation.

One of the earliest citations of learning liveness is available in Refs. [29-32] which argues “Behavioral biometrics for active authentication program” that can learn and adapt over a period. Behavioral biometrics attributed for liveness aspect. In a work by [10] showed the feasibility of an optimized Android app for FPAD that can run on the smartphone which can predict PA ‘presentation attack’ in less than 300 milliseconds. A patented and commercial solution by Systems [33] performs user authentication with PAD as a service. As claimed, it has security, easy-to-install, easy-to-use, add biometrics anywhere like feature. Even though private firms are using it in a close organization, so they bound the user to use it. But an expansion of the solution in the public domain has less acceptance. So, it cannot be accepted as it is. It will create opposition as similar to ‘Unique Identification Authority of India’ AADHAR project [34] row in India, even though the private sector is using the above solution under their service condition. This public acceptance is still missing in biometric solution because of the lack of liveness solutions universality. Another solution by Aware et al. [35] is a biometric solution, available as cloud-based software-as-a-service (SaaS). They design it to deliver a useful functionality and powerful biometric matching performance, without requiring integration and configuration. Work by Talreja et al. [36] showed a cloud-based model for platform-as-a-service that enables the recognition. Above three examples are enough to motivate us to architect a liveness detection model as service.

The solution as a service has several advantages, it is easy to update algorithm, we can mature varying algorithm(s) of liveness detection in incremental conditions, thus the developer gets motivated for continuous innovation. We may link developer’s motivation to acceptance of an algorithm as a social contribution or some rewards. This architecture has shown its feasibility as a case study based on recognition problem, so applicable for liveness services too. Interoperability issue that deals with generalization capability and reproducible results are missing in the conventional approach.

The proposed approach is being tested on data of two persons from LivDet, crossmatch 2015. We have created the dataset based on the subject ID, the dataset contains 30 live images of both persons. It also contains 74 spoof samples of first person named ‘A’ and 64 spoof samples of second person named ‘B’ for testing. The composition of the dataset is: Person A (Live: 30, BodyDouble: 15, EcoFlex: 15, Gelatin: 15, OO:14, PlayDoh:15) and Person B (Live: 30, Body Double: 12, Ecoflex: 13, Gelatin: 13, OO: 12, PlayDoh: 14). Now from these 30 images we have created 1000 images out of which 800 is for training and we use rest 200 for testing of the approach. We now test our trained model for both the persons as discussed in the previous section using test data. The test data for Person A use 200 Live images and rest 74 spoof samples. We got the following results for our Person-A.

The confusion matrix see Table 2 for our proposed model for Person A shows following outcome. Out of 274 sample, in 200 live images, we predicted 123 as live and rest 77as spoof. In 74 spoof images, we predicted all as spoof, this an essential requirement for secure access control in the critical system. Thus, the main motive of our proposed approach, use of only the live images of a person we can create an authentication system which rejects all type of spoof samples. The Table 3 explains more about the different performance scores. We are getting F1-score of 0.7616 which is decent score meaning that our model is not just making random guesses. The average precision score is also around 90% which is good. Our model lacks on the overall accuracy as we can see from the confusion matrix, that 77 out of 200 live samples classed as spoof. But it completes the primary motive of detecting spoof samples as we are getting a false positive rate of zero, no spoof is ever called live. Next, we discuss our approach in a much more elaborate manner. The results of different classifier are shown in next section, they are playing an important role in the overall calculation of our results. We are using three different classifiers for the learning purposes.

Table 2. Overall confusion matrix, Person-A

Table 3. Overall other metrics, Person-A

4.1 One-class SVM

The first model used is a one class SVM classifier. Support Vector Machine (SVM) is a supervised classification algorithm proposed by Laplace and Gauss in the year 1930 based on statistical learning theory. This algorithm creates a line or a hyperplane such that the data is partitioned into various classes based on the line or the hyperplane defined earlier. Here the algorithm tries to create an optimized hyperplane covering the labeled data. In Table 4, we can see that the accuracy of prediction of live has increased compared to combine-model, but the false negatives have also increased. All other metrics according for SVM, see Table 5 finds that the F1-score and the accuracy of this model alone is higher than that of our model, but the precision and false positive rate are less. The primary concern is that the false positive rate is 0.25 meaning that in every 4 spoofs 1 ‘spoof’ will be called as ‘live’ which is a major concern for our authentication system. As we cannot deploy a model which accepts 1 spoof as live of every 4 spoofs. Next, try to understand this notion using a graphical approach, Figure 2. Here a line on the positive side means we label the image as ‘Live’ by our classifier and if it is in the negative side meaning that the image is labelled as ‘Spoof’. It shows all the 274 test images and to the class it labels them by our model. The major takeaways from this graph is major of the spoof samples are quite a good outlier to the model.

Table 4. SVM, confusion matrix, Person-A

Table 5. SVM metrics, Person-A

4.2 Isolation forest

Isolation Forest algorithm is an unsupervised algorithm used for detection of anomalies in our data. This algorithm tries to isolate the outliers existing in our data rather than focusing on giving larger weights or focusing more on the normal points. We have used this algorithm for making our model much more global as isolation forest is a different algorithm from normal outlier detection algorithms. We want our model to be strict toward spoof, the outlier detection by this method saves us from such type of errors. The confusion matrix shown in Table 6 indicates effectiveness of isolation forest for outlier.

The results shown in Table 7 for isolation forest are much better than SVM. Our approach becomes universal and robust with this training. Thus, we get a generalized solution to problems and not just some overlay fit solution. The graphical plot for each sample, Figure 3, shows the class boundary.

Table 6. Isolation forest, confusion matrix, Person-A

Table 7. Isolation forest, Person-A

4.3 Local outlier factor

Local outlier factor used for finding anomalous data points by measuring the local deviation for data point with its neighbors. This algorithm uses the concept of local density for outlier detection in data. Here we are comparing the densities of the k-nearest neighbors, and based on this we find areas with similar density. Data points, which are having a much lesser density than their neighbors, considered outliers to the data. Thus, the algorithm uses distance from neighbors as a major metric for calculation of density. The confusion matrix and other metrics in Tables 8, 9, and Figure 4, a week feature boundary is visible between type-I and type-II error. We can see from the graph distance metrics used for deciding are similar and some are quite a great outlier.

A similar result for person-B is shown in Tables 10, 11, and Figure 5. Out of 264 sample, in 200 live images, we predicted 122 as live and rest 78 as spoof. Out of 64 spoof images, all are predicted as spoof, this fulfills our essential requirement for secure access control to critical system.

Table 8. Local outlier factor, confusion matrix, Person-A

Table 9. Local outlier factor, Person A

Table 10. Confusion matrix, Person-B

Table 11. Other metrics, Person-B

2.png

Figure 2. Class boundary, SVM

3.png

Figure 3. Class boundary, isolation forest

4.png

Figure 4. Class boundary, local outlier factor

5.png

Figure 5. Class boundary with fusion approach using SVM, IF and LOF, Person-B

Table 12. Comparison with state of art

4.4 Critical review with state of art

Consider a (Positive Classes-Live presentation), the poor recall with a reasonable F-measure score is acceptable and not a major challenge in biometric authentication. But, for (Negative Classes-Spoof presentation) the high specificity levels are demanded-a major challenge in high-security establishment. The specificity finds True Negative Rate (TNR) which is important, but more important metrics is False Negative Rate (FNR). Detecting spoof presentation as a live fingerprint is a major challenge. In this proposal, we experimented a scheme to reduce FNR to zero. The proposal works well when the system’s acceptance comes under question mark it affects ‘trust’ on biometric system, like application of financial transaction and entry to restricted area. In our case we able to show (see Table 12) that proposed approach reduces FNR to zero, which better than several close-set and open-set FPAD solution which assumes (0.2%) FNR [10] is acceptable for non-critical (civil) application. Mission critical system related to security of country and privacy preservation requirement cannot assume such a relaxation. is acceptable for non-critical (civil) application.