Presentation Attack Detection Algorithms for Finger Vein Biometrics: A Comprehensive Study

215 11th International Conference on Signal-Image Technology & Internet-Based Systems Presentation Attack Detection Algorithms for Finger Vein Biometrics: A Comprehensive Study R. Raghavendra Christoph Busch Norwegian Biometrics Laboratory Gjøvik University College Gjøvik, Norway Email: raghavendra.ramachandra, christoph.busch @ hig.no Abstract Finger vein biometric systems have shown a vulnerability to presentation attacks (or direct attack or spoof attack). The presentation attacks have raised crucial security issues for finger vein biometric systems mainly due to the widespread application of these biometric sensors in the financial sector. In this paper, we present a novel algorithm to detected the presentation attack on a finger vein recognition system. The proposed method relies on the texture analysis using steerable pyramids, and the final decision is carried out by classifying the textures features using a Support Vector Machines (SVM). Furthermore, we have also introduced a new kind of attack by using a smartphone display for the first time. Extensive experiments are carried out on a large scale database of 3 unique finger veins collected from 1 subjects. We introduce a new finger vein artefact database collected using three different kinds of artefacts namely: (1) ink jet print (2) laser jet print and (3) smartphone display. An extensive comparison of the proposed scheme with ten different Presentation Attack Detection (PAD) methods demonstrated the efficacy of the proposed scheme. I. INTRODUCTION Biometric technology has emerged as an alternative security access control method when compared with a traditional access control mechanism using a password, pin code or smart card. Biometrics is widely used to identify or recognise the person of interest based on their physical and behaviour characteristics. Since biometric systems can verify the person based on what he/she has the user of the biometric system is exempted from remembering the complex passwords or the access cards for various security application. Similar to any other security systems, biometric systems are also vulnerable to various attacks that can be broadly classified into two types [3]: (1) Direct attack: That involves presenting a biometric artefact directly to the sensor. (2) Indirect attack: That involves attacking the biometric system components through a Trojan horse or another kind of malware and virus. Direct attacks on biometric systems are of higher interest, as the attacker needs no specific knowledge of the biometric system under attack. which is relevant for the assessment of the attack potential for instance in a security evaluation. Among the various biometrics modalities, the finger vein modality has rapidly gained popularity because of its high accuracy, reliability, user convince, permanence and difficulty to spoof the sensor. Moreover unlike for fingerprints data subjects can not leave latent analog Figure 1: Example of finger vein artefact captured using smartphone (a) Real image from a normal presentation (b) Artefact image from an attack presentation samples unintentionally. For this reason, the finger vein biometrics are widely used in the financial sector especially in banking applications in many countries [9]. As the popularity of the finger vein modality increases, the vulnerability of the finger vein biometric systems also increases [6]. Even though the finger vein patterns are not visible to the naked eyes and the risk of latent prints does not exist, it is possible to attack directly the finger vein biometric system in case the finger samples stored in such systems are stolen or compromised. Biometric researchers have successfully demonstrated presentation (or spoofing) attacks on the finger vein recognition sensor using a printed image[12]. In [12], a finger vein artefact database comprised of 1 unique finger vein images from 5 subjects is collected by printing a real finger vein images using the LaserJet printer. The vulnerability analysis presented in [12] indicate a Spoof False Acceptance Rate (SFAR) of 86%. These results strongly indicate the vulnerability of certain finger vein biometric systems for such a direct attack. This motivates researchers to introduce the Presentation Attack Detection (PAD) or spoof detection as a countermeasure technique that can mitigate the attacks and thereby improve the reliability and trustworthiness of the finger vein biometric system. There exist various Presentation Attack Detection (PAD) algorithms for finger vein biometrics. In [5], Frequency and Time-Frequency analysis using Fourier and Wavelet Transform to detect the artefact (or fake/spoof) finger vein images were proposed. The artefact finger veins are captured using HP LaserJet printer on three different papers namely normal A4 paper, matte paper and OHP film. Experiments are presented to a small database of 7 subjects that indicates that the use of frequency information can detect the presence of finger vein artefacts. Considering the importance of the finger vein spoofing, the 1st competition was organised in conjunction with 978-1-4673-9721-6/15 $31. 215 IEEE DOI 119/SITIS.215.74 628 628

Figure 2: Block diagram of the proposed scheme the International Conference of Biometrics (ICB) 215 [11]. This competition uses the publicly available finger vein artefact database collected from 5 subjects [12]. Three different algorithms were introduced namely[11]: (1) Binarized Statistical Image Features (BSIF) (2) Rieze Transform (3) Local Binary Pattern (LBP), Local Phase Quantitation (LPQ). All these three methods have used the Support Vector Machine (SVM) as a classifier to categorize the presented finger vein object as an artefact (or spoof) or live. Thus, it is noted from the existing schemes that, finger vein PAD is limited only to detect the finger vein artefact generated using a laser printer. In this work, we propose a novel scheme for finger vein PAD using a steerable pyramid as a feature extraction scheme and subsequently a linear Support Vector Machine (SVM) as a classifier. The steerable pyramid uses a bank of steerable pyramids at each level of the pyramid rather than a single Laplacian or Gaussian. As a consequence, the use of steerable pyramids will capture more efficient texture information that can be further processed using Support Vector Machine (SVM) to make the final decision on whether the presented finger vein is real or spoof. We also present a new attack on the finger vein sensor by presenting finger vein images to the sensor using a smartphone (see Figure 1). The following are the main contributions of this paper: (1) Novel scheme for finger vein presentation attack detection using steerable pyramids (2) New attack on finger vein sensor by presenting a finger vein sample using a smartphone (3) New finger vein artefact database with 3 unique finger vein identities generated using three different kinds of artefact generation using (a) laser print (b) Ink jet print (c) smartphone presentation (4) Extensive experiments to evaluate the vulnerability of finger vein sensor to three different kinds of artefacts (5) Comprehensive study on the new finger vein database by comparing the proposed method with ten different presentation attack detection algorithms that are widely used in the biometric community. The rest of the paper is structured as follows: Section II presents the proposed scheme for a reliable finger vein biometric system, Section III presents the finger vein artefact data collection, Section IV presents the experimental results and discussion. Finally, Section V draws the conclusion of this work. II. PROPOSED SCHEME Figure 2 shows the block diagram of the proposed PAD scheme for the finger vein recognition system. The proposed scheme can be structured in two functional subsystems namely: (1) PAD Module (2) Finger vein recognition module. Given the finger vein Image F I,we first perform the Region Of Interest (ROI) extraction. Since our in-house sensor has a dedicated space to place the finger, the ROI extraction is carried out by setting a pre-assigned rectangular area to cover the interesting region of the finger. The obtained ROI finger vein image Fr I is further re-sized to have a dimension of 1 3 pixels before passing it to the PAD module. A. PAD Module This section describes the proposed presentation attack detection module introduced in this work. The proposed PAD module has two main functional blocks namely: (1) Feature extraction using steerable pyramids and (2) Classification using linear Support Vector Machine (SVM). In this work, we explore the applicability of the steerable pyramid features to detect the finger vein artefact. The steerable pyramids are a set of oriented filters in which a filter of arbitrary orientation is synthesised as a linear combination of a set basis function [1]. Thus, the use of steerable pyramids allows one to calculate the filter response at different orientations. Furthermore, the steerable pyramids exhibit the property of both rotation and translation invariance. Our motivation in using steerable pyramids for finger vein artefact detection (or spoof detection) is motivated by the fact that they have demonstrated prominently in extracting the texture features from the image. Thus, the texture features extracted using steerable filter are useful in identifying the presence of dot patterns in the case of laser print artefact or in identifying the glare patterns in case of inkjet print artefact or in identifying the additional frequency components in case of smartphone display attack. Given the ROI of the finger vein image Fr I the corresponding steerable pyramid decomposition is defined as: SF mn (x, y) = x F r (x, y)d mn (x x1,y y1) (1) y 629629

where, D mn denotes the directional bandpass filters at stage m =, 1, 2,..., S 1, and orientation n =, 1, 2,...,K 1. In this work, we consider only the residual highfrequency band from the steerable pyramid for the features selection process. Figure 3 shows the visual illustration of the residual high-frequency features obtained using steerable pyramids on both real and different kinds of artefact finger vein images. We then employ the linear SVM classifier to determine whether the captured sample belongs to the real (i.e. normal presentation) or artefact (i.e. attack presentation) category. The SVM classifier is first trained using a set of positive (real finger vein) and negative (finger vein artefact) samples according to the protocol described in the Section IV-A. B. Finger vein verification The Finger vein verification system (or baseline system) employed in this work is based on the Maximum Curvature Points (MCP) [4] as a feature extraction method and correlation as a comparator. This choice is made by considering the high performance and few computation characteristics exhibited by the MCP method [7]. III. FINGER VEIN ARTEFACT DATABASE COLLECTION In this section, we discuss the newly collected finger vein artefact database using our in-house finger vein sensor [8] [1]. Our database is comprised of 1 subjects for which we capture the finger vein from four different fingers namely right index, right middle, left index and left middle. However, some of the subjects were not able to provide all four fingers due to various reasons. Thus, the collected database has 3 unique finger vein instances that correspond to 1 subjects. The collected data is dived into two groups namely: (1) Real finger vein capture and (2) Artefact finger vein capture. The real finger vein capture attempts are considered to be normal presentations and were carried out in two sessions to obtain two images respectively. In the case of the artefact finger vein capture attempts, which are considered as attack presentations, we have used the real finger vein captures in the second session to generate various artefacts. The finger vein artefacts are generated using three different kinds of techniques namely: (1) Artefact 1: This artefact is generated by printing the real finger vein image using an ink jet (HP Photo smart 552) printer on highquality glossary paper. (2) Artefact 2: This artefact is generated by printing the real finger vein images on normal quality paper using a laserjet printer. (3) Artefact 3: This artefact is generated by presenting the real finger vein sample using the smartphone. For this purpose the real finger vein image is loaded to the Samsung S5 smartphone and then displayed to the sensor. This attack is introduced on the finger vein modality. However, we have carried out a series of pre-processing operations on the real finger vein images before obtaining the artefacts. Thus, we first perform the pre-processing by extracting the Region of Interest (ROI) and then rescaling the ROI finger vein to have a dimension of 1 3 pixels. We then carry out the contrast enhancement using histogram equalisation to improve the contrast of the real image before obtaining the corresponding artefact. Figure 3 shows the examples of the real and corresponding artefact images. IV. EXPERIMENTAL RESULTS This section describes the experimental protocols and the experimental results obtained by comparing the proposed scheme with ten different state-of-the-art finger vein presentation attack detection algorithms. A. Performance evaluation protocol For the experiments, each finger is considered as a unique instance and thus in total we have observed 3 unique finger vein instances. We then divided the whole database of 3 unique instances into two independent subsets namely: training and testing set. The training set is comprised of the first 5 unique instances and the testing set is comprised of the remaining 25 instances. The training set is used for multiple purposes in this work that includes (1) To set the threshold value based on the Equal Error Rate (EER%) for vulnerability analysis (2) used as the training set to evaluate the state-of-the-art finger vein PAD algorithms. The testing set is solely used to evaluate the performance of the proposed as well as state-of-the-art finger vein PAD schemes. B. Results and discussion We first present the vulnerability study on our in-house finger vein sensor to three different kinds of finger vein print artefacts. The main goal of this vulnerability study is to obtain the Spoof False Acceptance Rate (SFAR%) that indicates the applicability of the artefact samples collected in our work to spoof the sensor. To this extent, we consider the baseline finger vein recognition system that operates in two modes namely: normal mode and attack mode to obtain the comparison scores. In the normal mode, we have used the real finger vein image collected from the first session as the reference and we have used real finger vein image collected in the second session as a probe. For the attack mode, we have used an artefact finger vein image that is generated using the real finger vein image used as the probe in the normal mode. Since we have generated three different artefacts in this work, we use one artefact at a time in the attack mode of the finger vein system.note that, in the attack mode, the reference image corresponds to the real finger vein image captured in session 1 and the probe image corresponds to the artefact image. The only difference in the operation of the attack mode is the use of artefact finger vein as a probe and then to measure the SFAR to measure the vulnerability of the finger vein sensor to the generated artefact. Figure 4 shows the comparison score distribution obtained using the finger vein baseline system when both real and artefact samples were used as the probe. Figure 4 (a) shows the distribution of genuine scores (in green color), zero effort impostor score (in blue color) and artefact (or spoof) scores (in pink color). The black vertical line in 6363

Figure 3: Illustration of residual high frequency features on (a) Real finger vein image (b) Artefact 1 (Ink jet print artefact) (c) Artefact 2 (Laser jet print ) (3) Artefact 3 (smartphone display).4.35.3 Zero Effort Imposter Scores Artefact 1 scores Threshold obtained on Development Database.4.35.3 Zero-Effort Imposter Scores Artefact 2 Scores Threshold obtained on development dataset.4.35.3 Zero-effort Impooster Scores Artefact 3 Scores Threshold obtained on development dataset.25.25.25.2.2.2.2.3.4.5.6.7.8.9.2.3.4.5.6.7.8.9.2.3.4.5.6.7.8.9 Figure 4: Comparison score distributions of finger vein baseline system (a) with Artefact 1 (b) with Artefact 2 (c) with Artefact 3 Figure 4 (a) indicates the threshold value obtained on the training dataset that corresponds to the EER value of the baseline finger vein system operating in the normal mode (i.e. with real presentations). It can be noticed from the Figure 4 (a) that, the comparison scores obtained using artefacts will lie between genuine and zero effort impostor scores. Thus, with this threshold value more than 78% (SFAR) of the artefact 1 samples are accepted as the real presentation. A similar observation is also noted with Figure 4 (b) on artefact 2 that shows an SFAR of 76.4% and from Figure 4 (c) on artefact 3 indicating a SFAR of 1%. Thus, based on the high SFAR obtained from the vulnerability analysis, it is clear that the finger vein systems can be spoofed (or attacked) with the artefacts presented in this study. Table I: Performance results for the proposed PAD algorithm on the Artefact 1 Ensemble of BSIF SVM 13.2 8.4 1.8 GLCM-SVM 5.6 4 22.8 LBP(3X3) u2 SVM 22.4 6 14.2 LPQ SVM 7.6 6 6.8 LBP Variance SVM 21.6 36.4 29 2D FFT SVM 47.2 9.6 28.4 2D Cepstrum SVM 9.6 43.2 26.4 Local Entropy Map SVM 12.4 7.6 1 Quality based Techniques Block-wise Sharpness - SVM 1.8 19.2 15 Block-wise standard deviation - SVM 13.2 9.6 11.4 Proposed Method Steerable Pyramids - SVM 4.4 2.8 3.6 Table I - Table III indicates the quantitative performance of the proposed scheme on three different artefacts. Further, we also present the comprehensive comparison by comparing the proposed method with ten different presentation attack detection algorithms that are widely used in the biometric community. The quantitative performance of the presentation attack detection algorithms are presented according to the metric developed in ISO/IEC 317-3 [2] in terms of: (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. Finally, the performance of the overall PAD algorithm is presented in terms of Average Classification Error (AP CER+NPCER) 2 Rate (ACER) such that, ACER = the lower the values of ACER, the better is the PAD performance. Table II: Performance results for the proposed PAD algorithm on the Artefact 2 Ensemble of BSIF SVM 15.6 11.2 13.4 GLCM-SVM 18.4 36 27.2 LBP(3X3) u2 SVM 1.2 2.8 11 LPQ SVM.4 2.4 1.4 LBP Variance SVM 14 47.2 3.6 2D FFT SVM 14 82 48 2D Cepstrum SVM 3.6 27.6 15.6 Local Entropy Map SVM 6.8 52 29.4 Quality based Techniques Block-wise Sharpness - SVM 3.2 41.2 22.2 Block-wise standard deviation - SVM 4.4 42.8 23.6 Proposed Method Steerable Pyramids - SVM.4 5.6 3 631631

Table I presents the performance of the proposed scheme on artefact 1 samples. It can be observed that the proposed scheme has demonstrated the best performance with the lowest ACER of 3.6% while the second best performance is noted with the LPQ-SVM with an ACER of 6.8%. Table II indicate the quantitative performance of the proposed scheme on artefact 2 samples. Here also it can be observed that the proposed method have emerged as the best method with the lowest ACER of 3%. A similar observation is noted with the artefact 3 as indicated in the Table III in which the proposed scheme has demonstrated the best performance with ACER of 2.4%. Thus based on the obtained results obtained the proposed method is emerged as the best PAD scheme for finger vein biometric systems. Table III: Performance results for the proposed PAD algorithm on the Artefact 3 Ensemble of BSIF SVM 1.8.8 5.8 GLCM-SVM 77.2 16.4 46.8 LBP(3X3) u2 SVM 5.2 8.8 7 LPQ SVM 4.4 2.8 12.6 LBP Variance SVM 9.6 16.4 13 2D FFT SVM 79.2 7.2 43.2 2D Cepstrum SVM 8.8 35.6 22.2 Local Entropy Map SVM 6.8 8.8 7.8 Quality based Techniques Block-wise Sharpness - SVM 6 4.8 5.4 Block-wise standard deviation - SVM 7.6 18.8 13.2 Proposed Method Steerable Pyramids - SVM 4.4.4 2.4 V. CONCLUSION The goal of the PAD algorithm research is to improve the reliability of a finger vein biometric system by automatically detecting and mitigating the direct attacks. In this work, we present a novel algorithm based on the steerable pyramid and SVM. The proposed PAD algorithm was evaluated on three different kinds of finger vein artefact samples that shows the consistent high performance in detecting the finger vein artefact. A comparative evaluation is presented by evaluating the performance of the proposed method with ten different PAD algorithms used widely in biometric spoof detection. The experimental results demonstrate the efficacy of the proposed scheme with the lowest ACER of 3.6% on artefact 1, 3.% on artefact 2 and 2.4% on artefact 3. [3] S. Marcel, M. S. Nixon, and S. Z. Li. Handbook of Biometric Anti-Spoofing: Trusted Biometrics Under Spoofing Attacks. Springer Publishing Company, Incorporated, 214. [4] N. Miura, A. Nagasaka, and T. Miyatake. Extraction of finger-vein patterns using maximum curvature points in image profiles. IEICE Transactions on Information and Systems, 9(8):1185 1194, 27. [5] D. T. Nguyen, Y. H. Park, K. Y. Shin, S. Y. Kwon, H. C. Lee, and K. R. Park. Fake finger-vein image detection based on fourier and wavelet transforms. Digital Signal Processing, 23(5):141 1413, 213. [6] R. Raghavendra, M. Avinas, C. Busch, and S. Marcel. Finger vein liveness detection using motion magnification. In IEEE International Conference on Biometrics: Theory, Applications and Systems, Sept. 215. [7] R. Raghavendra, K. Raja, J. Surbiryala, and C. Busch. Finger vascular pattern imaging; a comprehensive evaluation. In Asia-Pacific Signal and Information Processing Association, 214 Annual Summit and Conference (APSIPA), pages 1 5, Dec 214. [8] R. Raghavendra, K. B. Raja, J. Surbiryala, and C. Busch. A low-cost multimodal biometric sensor to capture finger vein and fingerprint. In International Joint Conference on Biometrics (IJCB), pages 1 7, Sep 214. [9] R. Raghavendra, J. Surbiryala, and C. Busch. An efficient finger vein indexing scheme based on unsupervised clustering. In 1st IEEE International Conference on Identity, Security and Behaviour Analysis (ISBA 215). IEEE, 215. [1] R. Raghavendra, J. Surbiryala, K. B. Raja, and C. Busch. Novel finger vascular pattern imaging device for robust biometric verification. In Imaging Systems and Techniques (IST), 214 IEEE International Conference on, pages 148 152. IEEE, 214. [11] P. Tome, R. Raghavendra, C. Busch, S. Tirunagari, N. Poh, B. H. Shekar, D. Gragnaniello, C. Sansone, L. Verdoliva, and S. Marcel. The 1st competition on counter measures to finger vein spoofing attacks. In The 8th IAPR International Conference on Biometrics (ICB), May 215. [12] P. Tome, M. Vanoni, and S. Marcel. On the vulnerability of finger vein recognition to spoofing. In International Conference of the Biometrics Special Interest Group (BIOSIG), pages 1 1, Sept 214. VI. ACKNOWLEDGMENT This work is funded by the EU 7th Framework Program (FP7/27-213) under grant agreement n o 284862 for the large-scale integrated project FIDELITY. REFERENCES [1] W. T. Freeman and E. H. Adelson. The design and use of steerable filters. IEEE Transactions on Pattern Analysis & Machine Intelligence, (9):891 96, 1991. [2] ISO/IEC JTC1 SC37 Biometrics. ISO/IEC WD 317-3:214 Information Technology - presentation attack detection - Part 3: testing and reporting and classification of attacks. International Organization for Standardization, 214. 632632