Presentation Attack Detection using Laplacian ... - IEEE Xplore

2 downloads 0 Views 2MB Size Report
{kiran.raja; raghavendra.ramachandra; christoph.busch} @hig.no. Norwegian Biometrics Laboratory, Gjøvik University College, 2802 Gjøvik, Norway. Abstract.
Presentation Attack Detection using Laplacian Decomposed Frequency Response for Visible Spectrum and Near-Infra-Red Iris Systems Kiran B. Raja, R. Raghavendra, Christoph Busch {kiran.raja; raghavendra.ramachandra; christoph.busch} @hig.no Norwegian Biometrics Laboratory, Gjøvik University College, 2802 Gjøvik, Norway

Abstract Biometrics systems are being challenged at the sensor level using artefact presentation such as printed artefacts or electronic screen attacks. In this work, we propose a novel technique to detect the artefact iris images by decomposing the images into Laplacian pyramids of various scales and obtain frequency responses in different orientations. The obtained features are classified using a support vector machine with a polynomial kernel. Further, we extend the same technique with majority voting rule to provide the decision on artefact detection for video based iris recognition in the visible spectrum. The proposed technique is evaluated on the newly created visible spectrum iris video database and also Near-Infra-Red (NIR) images. The newly constructed visible spectrum iris video database is specifically tailored to study the vulnerability of presentation attacks on visible spectrum iris recognition using videos on a smartphone. The newly constructed database is referred as ’Presentation Attack Video Iris Database’ (PAVID) and consists of 152 unique iris patterns obtained from two different smartphone - iPhone 5S and Nokia Lumia 1020. The proposed technique has provided an Attack Classificiation Error Rate (ACER) of 0.64% on PAVID database and 1.37% on LiveDet iris dataset validating the robustness and applicability of the proposed presentation attack detection (PAD) algorithm in real life scenarios.

1. Introduction Biometric systems have continued to gain popularity and wider acceptance in many secure authentication scenarios. Among the wide range of biometric modalities, iris recognition in Near-Infra-Red (NIR) domain has proven to perform with high accuracy of verification [4]. Thorough experiments on iris recognition [4], have influenced large-scale biometric deployment such as UIDAI to employ iris as one of the main biometric characteristics [8]. Nonetheless, iris based systems are prone to presentation attacks or spoof at-

tacks at the sensor level. A presentation attacks, which is also widely called as spoof attack, is the process of attacking the biometric capture device by presenting the biometric artefacts to gain access in a secure authentication scenario. Many studies such as [5, 9, 7, 16, 10, 24, 3], have investigated presentation attacks on NIR based iris recognition systems specifically for printed artefact presentation at the sensor level. More recently, iris recognition has been explored in the visible spectrum using mobile devices [6, 15, 14, 13]. Recent works [21, 19, 12, 20, 18] have studied the vulnerability of such visible spectrum iris recognition systems. MobILive 2014 (Mobile Iris Liveness Detection Competition) [21] was specifically designed to address presentation attacks using printed images. One alternative to avoid a presentation attack with still images is to incorporate video based iris recognition at the sensor level. In this work, we propose such a system that employs video based iris recognition in the visible spectrum on smartphones. Further, such a system can be easily attacked by any impostor, if the enrolment video is acquired and replayed using any electronic device with a sufficiently high-quality screen. As there is no existing work on the analysis of vulnerability for video based iris recognition system in the visible spectrum using smartphones, we carry out this work. To the best of our knowledge, this is the first work carrying out video based iris recognition on smartphones and studying presentation attacks of such systems. In the context of the current work, the significant contributions can be outlined as: 1. This work explores video based presentation attack detection for iris recognition on smartphones in the visible spectrum. 2. This work has resulted in the largest video iris database acquired using two different smartphones in the visible spectrum. The database consists of 152 unique iris patterns obtained from 76 subjects. 3. This work explores the vulnerability of video based iris systems to presentation attack, especially regarding the video replay attack using a high-quality display tablet.

Reject / Alarm Attack Presentat ion Spoof Detect ion

Normal Presentat ion

Feat ure Ext ract ion

Image Capt ure

Iris Feat ures Localized Eye Region

Enrollment Database

Comparat or

Aut hent icated/ Failed

Periocular Feat ures

Figure 1: The proposed scheme for PAD is illustrated in the figure. The block enclosed in the dotted blue line indicates a conventional verification scheme. In the due course of analysing the vulnerability of the iris based systems in the visible spectrum, we have devised a novel Presentation Attack Detection (PAD, a.k.a, spoof attack detection) algorithm. The proposed algorithm is based on exploring the Laplacian pyramid decomposition at various scales for each frame in the video followed by localising the Short Term Fourier Transform (STFT) frequency response in different orientations. Further, we employ a Support Vector Machine (SVM) classifier to distinguish the normal presentation versus attack presentation. The classification output is considered for various frames and the final decision for the entire video is made using majority voting of classification scores obtained on individual frames. As the proposed algorithm is based on frequency localization using Laplacian Pyramids at different scales, we evaluate the robustness of the proposed algorithm to detect the presentation attack in NIR domain also by employing wellknown Warsaw iris dataset from LiveDet 2013 competition [24, 3]. The proposed algorithm has indicated its agnostic nature that is independent of the visible spectrum and NIR domain. In the rest of the paper, Section 2 presents the proposed scheme for presentation attack detection and also discusses

Image/ Video

Laplacian Pyramid Decomposit ion

STFT Response M aps

the proposed technique for PAD. Section 3 presents the details of the database constructed in this work. Section 4 describes the experimental protocols associated with the newly constructed database and results obtained using the proposed method. Finally, Section 5 gives the summary and conclusion of the current work.

2. Proposed Scheme for Presentation Attack Detection Conventional authentication systems in biometrics employ the framework that is indicated by the dotted blue line in Figure 1. In many of the existing systems, once the image is acquired, the eye region is localized, and the features are extracted from iris and/or periocular region. These features are compared against the biometric reference in the database. Based on the comparison score, a subject is either authenticated or rejected by the biometric system. With the advancement in artefact generation techniques, such systems are highly prone to presentation attack as they do not gauge the liveness of the subject that is presenting the biometric characteristic. Thus, in this work, we propose to incorporate the spoof detection or presentation attack detection module to check, if the subject is presenting in person

Hist ogram Feat ure Vect or

SVM Classifier

M ajorit y Vot ing

Figure 2: Proposed algorithm for presentation attack detection

Attack Presentat ion Normal Presentat ion

rather than an impostor using the artefact samples. The presentation attack detection module takes the captured video of predefined frames and checks if the presentation is normal (i.e., not spoof). If the PAD module (Spoof Detection module in Figure 1) detects the presented subject as impostor using the artefact, then the PAD module rejects the capture attempt and triggers the alarm. If the presentation is classified as normal presentation, we obtain the periocular features to perform the verification. Although, iris features can be used in this context, we limit our work to periocular based verification such that the algorithm can be generalized to a larger population irrespective of the color of the iris and the spectrum of the sensor.

2.1. Proposed Method for PAD The proposed algorithm for the presentation attack detection is depicted in the Figure 2. Given the image/video of the subject under verification scenario, we first decompose each image/frames into Laplacian pyramids of multiple scales. Each of the resulting images at the specific scale is used to obtain a STFT response at four different orientations. The response corresponding to 4 different orientation is encoded as a single response image, and the features are obtained by taking the histogram as described in this section. The features are used to classify the presentation category as normal or attack presentation using a SVM classifier. The motivation and intuition behind the proposed technique are thoroughly discussed in this section. 2.1.1

Laplacian Pyramids

Laplacian pyramid decomposition of the image was initially developed with the idea of encoding the image using local operators at many scales with identical basis functions [1]. The significance of the Laplacian pyramid decomposition comes from the fact that the elements of an image are localized both in space and frequency domain. Further, the Laplacian pyramid can be used effectively to represent images as a series of band-pass filtered images that are sampled successively at sparser representations [1]. Although, the frequency content of the image is well localized using Laplacian Pyramids, the orientation information of each frequency content is not obtained. 2.1.2

Proposed Algorithm for PAD

In the case of any natural image or video frames, there exists a substantial amount of edge information which contributes to the frequency information of that image or frame [22]. However, when the same images are printed using low/highresolution printers or when the same images are displayed on an electronic screen, the images present frequency information which is different from the original frequency distribution of the image. This additional frequency information

is inherently present in the artefacts generated by printing the live samples or replaying the live samples in the context of biometric samples. Intuitively, localizing this frequency makes the separation of normal presentation versus attack presentation. In order to localize this frequency, we employ Laplacian pyramids at 5 different scales with binomial filter kernel of size 9 1 . Laplacian pyramid decomposition separates the lower and higher frequency in a well-defined components. We have employed a scale of n = 5 in this work. The difference of the obtained low pass and high pass filter at each scale is used to localize the frequency information further by analyzing STFT response corresponding to four different orientations φ = {0◦ , 45◦ , 90◦ , 135◦ }. If an image at a particular scale s of the Laplacian pyramid is represented by Is , we obtain the STFT response of the image. The STFT of the image at scale s, which is represented by Fs is the image resulting to response of frequency components in four different orientations such that φ = {0◦ , 45◦ , 90◦ , 135◦ }. The filter response obtained from each orientation are separated for real and complex values subsequently. Each of the responses denoted by b is finally encoded to form the final response map as given by F Rs where i corresponds to different orientation angles given by φ = {0◦ , 45◦ , 90◦ , 135◦ }. F Rs = Re(

4 X

(bi )∗(2(i−1) ))+Im(

4 X

(bi )∗(2(i−1) )) (1)

i=1

i=1

The feature vector F V , of the image at a particular scale s is formed by obtaining the histogram of the response map at scale F Rs . 255 X F Vs = {F Rs }i (2) i=0

The final feature vector for the frame or image is formed by concatenating the feature vectors of images from scale 1 to n and orientation φ = {0◦ , 45◦ , 90◦ , 135◦ }. The final feature vector F Vf can be represented as : F Vf = {F Vs=1,φ=0◦ , F Vs=1,φ=45◦ , F Vs=1,φ=90◦ , F Vs=1,φ=135◦ , ..., F Vs=n,φ=0◦ , F Vs=n,φ=45◦ , F Vs=n,φ=90◦ , F Vs=n,φ=135◦ }

(3)

The final feature vector given by Equation 3 is used to represent the image for classification purposes. Figure 3 presents the Laplacian pyramid decomposition at five different scales and its corresponding STFT response maps. Figure 3(a) presents a multi-scale Laplacian pyramid images 1 Various kernels with different size were evaluated and the size of the kernel was fixed based on the empirical trials conducted on the development database. We have also explored the Steerable Pyramid and Gaussian pyramid decomposition which resulted in similar but lower accuracy as compared to Laplacian pyramid decomposition

(a) Decomposed Laplacian pyramids for live frame at different scales ( Scale 1 to 5 from left to right)

(b) STFT response maps for live frame at different scales corresponding to (a)

(c) Decomposed Laplacian pyramids for replay frame corresponding to live frame shown in (a) at different scales ( Scale 1 to 5 from left to right)

(d) STFT response maps for spoof frame at different scales corresponding to (c) Figure 3: Decomposition of images into Laplacian pyramids of scale 5 and corresponding STFT response maps in each scale. [*Note: Images from all scales are resized to uniform size for the purpose of illustration only]. for a sample frame from live video and Figure 3(b) presents the corresponding STFT response map of live frame at different scales. Figure 3(c) and Figure 3(d) illustrate the Laplacian pyramid and STFT response map of replay attack video frame. It can be observed from the figure that subtle changes in the frequency information along various orientations of the frame of replay attack video can be highly enhanced at different scales in STFT response maps.

Table 1: Configurations of the SVM employed in this work

Name Package svm type kernel degree

Parameter LIBSVM [2] nu-svc Polynomial Kernel 3

2.2. Classification of features In order to classify the features of normal presentation versus the features of the attack presentation (aka, spoof attack), we employ a Support Vector Machine (SVM) with a polynomial kernel [2]. The detailed configurations of the SVM classifier employed in this work is presented in Table 1.

complete details of databases are provided in this section.

3. Database

As this work explores video based iris recognition in the visible spectrum on smartphones, we construct a new iris video database which is specifically tailored to study the presentation attack and hence the newly constructed database is termed as ’Presentation Attack Video Iris Database’ (PAVID).

This section provides a brief description of the databases employed in this work namely: PAVID - our newly constructed database and LiveDet Iris 2013 [24] database to evaluate the robustness of the proposed PAD method. The

3.1. Presentation Attack Iris Video Database in the visible spectrum using smartphones

3.1.1

PAVID - Real Iris Video Database

As there exists no standard large-scale iris video database acquired using smartphones for research in biometrics, we construct a new iris video database in visible spectrum using two smartphones in this work. The significance of the newly constructed database lies in three folds: (1) To provide the research community with the video iris database collected using smartphones in visible spectrum, (2) To provide a database with relatively higher number of subjects, (3) To provide a presentation attack database to assess the robustness of the proposed PAD technique. To fulfill three main Table 2: Distribution of real iris video database in current work

Number of subjects Unique eye instances Enrolment video Probe Video

Smartphone Nokia Lumia 1020 iPhone 5S 76 76 152 152 152 152 152 152

goals discussed above, we have created the PAVID database by capturing 152 unique eye instances from 76 subjects using two new smartphones - Nokia Lumia 1020 and iPhone 5S. The total distribution of iris videos is presented in the Table 2. Each unique eye instance is captured in two different sessions that correspond to enrolment and verification. In each of the sessions, a video of duration 1 − 3 seconds is acquired for each subject. The captured eye videos are processed such that at least 25 frames are obtained between the blinks. In the second stage of pre-processing, we register each frame in a video using mutual information of frames to mitigate the effect of involuntary head motion. Finally, the registered video frames are used to localize the eye region using the Haar cascade based eye detector [23]. This is essential to remove the background captured due to the larger field of view of the camera. The located boundary of eye region from the first frame is propagated across other frames in the video to localize the eye region. Finally, these videos for each subject are used as the enrolment iris videos. In a similar manner, the iris video is obtained for a verification session. 3.1.2

more, as the baseline system in this work employs video for enrolment, it becomes important to analyze video based attacks. Although, many of the earlier works have successfully explored presentation attack using the print attacks [7, 3, 24] in NIR domain, this work addresses the presentation attack in the visible spectrum iris recognition using video-based recognition. To simulate a realistic attacks in a verification scenario, we create an artefact database under the assumption that the video from the enrolment database is available to the impostor. In this scenario, the impostor can use the enrolment video to generate the printed iris artefacts or replay the iris video. Going by such a scenario, we create the artefact database by replaying the iris video on the high-quality display enabled iPad and presenting it to smartphones (i.e., the biometric sensor in our case). The replay attack database consists of 4 different attack subsets such that the enrollment videos obtained from iPhone is replayed to iPhone and also to Nokia. Similarly, the enrolment videos obtained from Nokia are replayed to Nokia and iPhone as well. Under each replay attack subset, a total of 152 iris videos are present, which make a total of 608 artefact iris videos in total in the PAVID database. Table 3 provides an overview of the different subsets in the PAVID database. Table 3: PAVID artefact database composition Source obtained from iPhone Nokia

3.1.3

Sensor Attacked

Number of videos

iPhone

152

Nokia

152

iPhone

152

Nokia

152

Availability of PAVID Real and Artefact Database

In the light of no existing video iris database captured in visible spectrum using smartphones available for research, we intend to make the PAVID database (Real and Artefact database) available for non-profitable scientific and academic research. The details of the database can be obtained www.nislab.no/biometrics_lab/pavid_db.

PAVID - Artefact Iris Video Database

As the current work intends to study the vulnerability of iris recognition in the visible spectrum using the videos of iris, in this work we create the database for presentation attack. One can think of image based attacks (either electronic or printed attacks) and video based attacks to surpass the biometric system. As most of the works have concentrated on the print attacks [5, 9, 7, 16, 10, 24, 3], we intend to explore the vulnerability of sensor to video replay attacks. Further-

3.2. LiveDet Iris 2013 The Warsaw database [3] has been well explored for the iris liveness detection (LivDet 2013) competition due to the challenging nature of the artefact images. The Warsaw subset of the LiveDet 2013 database consists of 852 images obtained from 284 distinct eye instances, and 815 biometric images of paper printouts for 276 distinct eye instances. The database is divided into training and testing subsets in

40

40

Baseline Performance Baseline Performance under attacks

20

False Reject Rate (in %)

False Reject Rate (in %)

20

Baseline Peformance Baseline Performance under attacks

10

5

2

10

5

2

1

1

0.5

0.5

0.2

0.2 0.1

0.1 1

2

5

10

20

40

False Acceptance Rate (in %)

0.1

0.2

0.5

1

2

5

10

20

40

False Acceptance Rate (in %)

(a) DET Curve indicating the system performance for iPhone 5S (b) DET Curve indicating the system performance for Nokia 1020

Figure 4: Baseline scores and performance of the systems employing iPhone 5S and Nokia Lumia 1020 as capture sensors which the training subset consists of high-resolution printouts only. However, the testing subset consists of images obtained using both low and high-resolution printouts to test the robustness of algorithms under varying resolutions of printout [3]. The detailed protocol for evaluating the proposed method is provided the Section 4.

4. Experiments and Results The detailed description of the experiments carried out in the current work is presented in this section. We report the performance of the presentation attack detection (PAD) algorithm in accordance to ISO/IEC CD 30107-3 [11]. The metrics used in this work are: (1) Attack Presentation Classification Error Rate (APCER), which is defined as a proportion of attack presentation incorrectly classified as normal (or real) presentation (2) Normal Presentation Classification Error Rate (NPCER) which is defined as proportion of normal presentation incorrectly classified as attack presentation [11]. Finally, the overall performance of the PAD algorithm is presented in terms of Average Classification Error Rate (ACER) such that: AP CER + N P CER (4) 2 The lower values of ACER indicate the superior performance of a PAD technique. The ACER for various attacks are compared against the standard state-of-the-art techniques which are based on Image Quality Metrics with SVM classifier (IQM-SVM) [7], Local Binary Pattern with SVM classifier (LBP-SVM) [17] and more recently proposed Binarized Statistical Image Features with SVM (BSIF-SVM) [19]. ACER =

4.1. Vulnerability analysis on PAVID database Table 6: Division of our database for experiments using iris videos acquired from each smartphone Smartphone Nokia Lumia 1020 iPhone 5S Real Iris Videos Development 20 20 Training 50 50 Testing 82 82 Artefact Iris Videos for each attack Development 40 (20 x 2) 40 (20 x 2) Training 40 (20 x 2) 40 (20 x 2) Testing 224 (112 x 2) 224 (112 x 2)

Even before measuring the performance of a PAD algorithm, it is important to determine the baseline performance with normal iris videos and the baseline performance of a system, when iris videos are used for attacking the sensor of the system. Such an analysis is based on a trivial idea of measuring the genuine and impostor scores under normal presentation. Furthermore, it comprises of measuring the genuine and impostor scores when the artefacts are used to attack the sensor. The genuine score is obtained by comparing the reference frame from enrolment video with all the frames from probe video. These scores are used to obtain the Detection Error Trade-off (DET) curves of the system. To simplify the number of comparisons, we consider one frame from iris video of enrolment as the reference image and 25 frames from probe video as the probe samples. A similar approach is used to compute the genuine and impostor scores when the replay attack video is used. Figure

Table 4: Classification error rates obtained using various schemes for PAVID databse Reference Video iPhone Nokia

Presentation Attack Video

IQM-SVM [7]

LBP -SVM [17]

BSIF-SVM [19]

Proposed Method

NPCER

APCER

ACER

NPCER

APCER

ACER

NPCER

APCER

ACER

NPCER

APCER

ACER

iPhone

57.31

11.6

34.45

4.87

0.89

2.88

6.09

9.82

7.955

1.21

1.78

1.49

Nokia

76.92

10.71

43.81

3.84

3.54

3.69

2.56

8.92

5.74

1.28

0

0.64

iPhone

76.92

4.5

40.71

3.84

4.51

4.175

2.56

10.81

6.68

1.28

4.46

2.87

Nokia

57.31

3.57

30.44

4.87

2.67

3.77

6.09

0.89

3.49

1.21

2.68

1.95

4(a) and Figure 4(b) present the DET curves for systems using iPhone 5S and Nokia Lumia 1020 as the capture sensor. A closer look at Figure 4(a) and (b) indicates the vulnerability of the proposed baseline system for the attacks without any countermeasures or PAD module. The DET curves for attacks are very close to the baseline system pushing the system to risk of being compromised by an attack. Thus, we evaluate the proposed method to overcome such a weak performance against presentation attacks in the subsequent sections.

4.2. Experiments on PAVID database In order to effectively evaluate the proposed algorithm for presentation attack detection on our newly constructed database, we have divided the whole database of 152 unique eye patterns (i.e. instances) obtained using a particular smartphone into three sets: Training set, Development set and Testing set. The training set comprises of 50 unique eye patterns that were used only for training the SVM classifier. The development dataset comprises of 20 unique eye patterns that are used to tune any parameters associated with the presentation attack detection algorithms. The development set was used to determine the filter kernel for the Laplacian pyramid, the size of the window and scales for the pyramid. The testing dataset comprises of 82 unique eye patterns that are solely used to evaluate the presentation attack detection algorithm proposed in this work. The detailed division is provided in Table 6. Additionally, we analyze the vulnerability of the video based visible iris system that involves dividing the development and testing dataset into two independent groups that correspond to reference and probe set. From the set of reference videos, we consider features obtained from one frame as the reference and features ob-

tained from all the frames of probe video as the probe sample to obtain the baseline performance of video based visible iris recognition algorithm.

4.2.1

Results on PAVID database

The Table 4 presents results obtained on the PAVID database using various state-of-the-art methods such as (IQM-SVM) [7], LBP-SVM [17] and BSIF-SVM [19]. It can be observed in the Table 4 that the proposed method has emerged as the best technique for PAD out of all the techniques available in state-of-the-art schemes. The best ACER is obtained consistently across all different attacks from the proposed technique. The best ACER of 1.49% is obtained when the system employing iPhone as the primary sensor is attacked using enrolment videos captured using iPhone. Similarly, an ACER of 0.64% is obtained when enrolment video captured using Nokia phone is used to attack the iris recognition system employing iPhone as capture sensor. The obtained results support the applicability of the proposed approach for detecting presentation attacks in real life verification scenarios when adapted to video based iris recognition systems.

4.3. Experiments on LiveDet Iris 2013 To measure the robustness of the proposed algorithm to detect presentation attack detection in NIR domain, we employ the Warsaw dataset that is captured with a NIR imaging device [3]. To report the ACER obtained from the proposed approach and compare them against already published results, we follow the protocol suggested for this database [24].

Table 5: Performance of proposed method on LiveDet Iris 2013 Warsaw Dataset Proposed techniques by team ATVS Federico Porto BSIF-SVM [19] Proposed Method

APCER (FerrFake) (%) 7.60 0.6 11.95 2.14 1.95

NPCER (FerrLive) (%) 25.25 21.15 5.25 0.4 0.80

ACER (Average Error) (%) 16.42 10.87 8.60 1.27 1.37

4.3.1

Results on LiveDet Iris 2013

The terminology reported earlier is rather different than the standard harmonized biometric vocabulary in accordance to ISO/IEC cD 30107-3 [11]. Thus, in this work we report the results in terms of NPCER, APCER and ACER and present the corresponding terms, which were described in the original paper, in brackets [24]. Table 5 presents the performance of the proposed PAD algorithm in comparison with four different existing schemes of which three were evaluated during LiveDet 2013 competition and one recently published results [24]. As indicated in the Table 5, the proposed PAD scheme shows the performance with an ACER of 1.37% which is better than most of state-of-the-art schemes. However, the performance of the proposed method is by 0.1% higher than the best available result based on BSIF-SVM [19]. This trade-off of performance can be attributed to the classifier level decision fusion employed in [19].The obtained performance of the proposed PAD algorithm validates the robust performance for detecting attacks in NIR iris systems as well.

5. Conclusions Presentation attacks pose high level of threats to existing biometric systems. Innovative ways of presentation attacks have emerged with the advancement of technology. Although iris systems are well known for robust verification they provide in biometric systems, they are still prone to attacks at various levels. In this work, we have explored presentation attacks on video based iris recognition system in visible spectrum using smartphones. In the course of this work, we have constructed a large scale video iris database consisting of 152 unique patterns acquired using 2 different smartphones - iPhone 5S and Nokia Lumia 1020. Furthermore, to analyse the vulnerability of such a video based system for iris recognition, we create video replay artefacts and to overcome such challenges, we propose a novel technique for detecting video replay attacks using features obtained from frequency response maps of images obtained through Laplacian pyramid decomposition and a SVM classifier along with majority voting. The proposed method has provided the best ACER of 0.64% for video replay attacks on iris recognition systems employing iPhone as a capture sensor. The proposed technique has also performed well on the NIR based artefact detection proving the applicability of proposed PAD algorithm in real life biometric systems.

Acknowledgments The authors wish to express thanks to Morpho (Safran Group) for supporting this work, and in particular to Morpho Research & Technology team for the fruitful technical and scientific exchanges related to this particular work.

References [1] P. J. Burt and E. H. Adelson. The laplacian pyramid as a compact image code. Communications, IEEE Transactions on, 31(4):532–540, 1983. [2] C.-C. Chang and C.-J. Lin. LIBSVM: A library for support vector machines. ACM Transactions on Intelligent Systems and Technology, 2:27:1– 27:27, 2011. Software available at http://www.csie.ntu.edu.tw/ ˜cjlin/libsvm. [3] A. Czajka. Database of iris printouts and its application: Development of liveness detection method for iris recognition. In Methods and Models in Automation and Robotics (MMAR), 2013 18th International Conference on, pages 28– 33, Aug 2013. [4] J. Daugman. How iris recognition works. IEEE Transactions on Circuits and Systems for Video Technology, 14(1):21–30, 2004. [5] J. Daugman. Iris recognition and anti-spoofing countermeasures. In 7-th International Biometrics conference, 2004. [6] M. De Marsico, C. Galdi, M. Nappi, and D. Riccio. Firme: face and iris recognition for mobile engagement. Image and Vision Computing, 32(12):1161– 1172, 2014. [7] J. Galbally, S. Marcel, and J. Fierrez. Image quality assessment for fake biometric detection: Application to iris, fingerprint, and face recognition. Image Processing, IEEE Transactions on, 23(2):710–724, 2014. [8] Goverment of India. Aadhaar-a unique identification number. Research Cell: An International Journal of Engineering Science, 4(2):169–176, 2011. [9] X. He, Y. Lu, and P. Shi. A fake iris detection method based on fft and quality assessment. In Pattern Recognition, 2008. CCPR’08. Chinese Conference on, pages 1–4. IEEE, 2008. [10] K. Hughes and K. W. Bowyer. Detection of contact-lens-based iris biometric spoofs using stereo imaging. In System Sciences (HICSS), 2013 46th Hawaii International Conference on, pages 1763–1772. IEEE, 2013. [11] ISO/IEC TC JTC1 SC37 Biometrics. ISO/IEC CD 30107-3. Information Technology – Biometric Performance Testing and Reporting – Part 3: Testing and reporting. International Organization for Standardization and International Electrotechnical Committee, 2015. [12] Kiran B. Raja., R. Raghavendra, and C. Busch. Video presentation attack detection in visible spectrum iris recognition using magnified phase information. Information Forensics and Security, IEEE Transactions on, PP(99):1–1, 2015. [13] Kiran B. Raja, R. Raghavendra, M. Stokkenes, and C. Busch. Smartphone authentication system using periocular biometrics. In 2014 International Conference on Biometrics Special Interest Group, pages 27–38. IEEE, 2014. [14] Kiran B. Raja, R. Raghavendra, M. Stokkenes, and C. Busch. Multi-modal authentication system for smartphones using face, iris and periocular. In IEEE International Conf. Biometrics (ICB), Phuket, Thailand, 2015. [15] Kiran B. Raja, R. Raghavendra, V. K. Vemuri, and C. Busch. Smartphone based visible iris recognition using deep sparse filtering. Pattern Recognition Letters, 57(0):33 – 42, 2015. [16] S. J. Lee, K. R. Park, and J. Kim. Robust fake iris detection based on variation of the reflectance ratio between the iris and the sclera. In Biometric Consortium Conference, 2006 Biometrics Symposium: Special Session on Research at the, pages 1–6. IEEE, 2006. [17] J. Maatta, A. Hadid, and M. Pietikainen. Face spoofing detection from single images using micro-texture analysis. In Biometrics (IJCB), 2011 international joint conference on, pages 1–7. IEEE, 2011. [18] R. Raghavendra and C. Busch. Presentation attack detection algorithm for face and iris biometrics. In Signal Processing Conference (EUSIPCO), 2014 Proceedings of the 22nd European, pages 1387–1391. IEEE, 2014. [19] R. Raghavendra and C. Busch. Robust scheme for iris presentation attack detection using multiscale binarized statistical image features. Information Forensics and Security, IEEE Transactions on, 10(4):703–715, April 2015. [20] R. Raghavendra, Kiran B. Raja, and C. Busch. Ensemble of statistically independent filters for robust contact lens detection in iris images. In Proceedings of the 2014 Indian Conference on Computer Vision Graphics and Image Processing, page 24. ACM, 2014. [21] A. F. Sequeira, H. P. Oliveira, J. C. Monteiro, J. P. Monteiro, and J. S. Cardoso. Mobilive 2014-mobile iris liveness detection competition. In Biometrics (IJCB), 2014 IEEE International Joint Conference on, pages 1–6. IEEE, 2014. [22] E. P. Simoncelli and B. A. Olshausen. Natural image statistics and neural representation. Annual review of neuroscience, 24(1):1193–1216, 2001. [23] P. Viola and M. Jones. Robust real-time face detection. International Journal of Computer Vision, 57:137–154, 2004. [24] D. Yambay, J. S. Doyle, K. W. Bowyer, A. Czajka, and S. Schuckers. Livdetiris 2013-iris liveness detection competition 2013. In Biometrics (IJCB), 2014 IEEE International Joint Conference on, pages 1–8. IEEE, 2014.