Finger vein authentication technology and financial applications

Transcription

1 1 Finger vein authentication technology and financial applications Mitsutoshi Himaga 1 and Katsuhiro Kou R&D Division, Hitachi-Omron Terminal Solutions, Corp., Owari-asahi, Aichi, Japan. [email protected] Terminal Systems Group, Hitachi-Omron Terminal Solutions, Corp., Shinagawa-ku, Tokyo, Japan. [email protected] 1 Introduction 1.1 What is finger vein? Vascular network patterns are unique biometric feature that are robust against temporal and environmental changes. Researchers at Hitachi Central Research Laboratory started the fundamental research of finger vein biometrics for personal identification in mid 90 s and revealed that the biometric feature was extremely competitive (Kono, Ueki, and Umemura 2000). From medical point of view, the following features of finger vein are known: 1) Universality Arteries and veins are vital organ that supply and circulate sufficient oxygen and nutrition to finger and other part of human body. It is well-known anatomical and clinical fact that veins of mm diameter exist in every finger of healthy individuals. 2) Uniqueness Embryology proves that blood vessel paths are generated by probabilistic factors such as tissue hypoxia (low-oxygen condition). The influence of inherited factors are almost ignorable in the peripheral and it is reasonable to regard finger vein patterns as individually unique feature. 3) Permanency

2 2 Fundamental blood vessel network is formed before birth. The vascular patterns, whose diameter is mm, are then kept stable by tight interfaces of endothelial cells forming blood vessels and other cells exist in the neighbouring area. Blood vessel patterns of this thickness are supported by constant blood flow and do not disappear by age-related factor. Neovascularisation is, in principle, only caused by pathological conditions such as cancer or arteriosclerosis and does not occur to healthy tissue. 1.2 Why finger vein? Unlike other conventional biometric features such as fingerprint or hand geometry, finger vein patterns do not leave any traces or information that can be used to duplicate the biometric data. As finger veins exist beneath human being s skin, they are completely hidden and unexposed even during the authentication process. It is, therefore, impossible to steal or copy the biometric patterns by photography or video recording, which makes it extremely difficult to duplicate the biometric data. Yet another advantage of using finger vein is its flexibility. Most finger vein devices are designed to accept four to six fingers per person (i.e. index, middle and ring fingers of both hands), which means this biometric feature can be applied for far more people than other modalities that have less samples per person. This feature is even more important if the biometric data should be duplicated by some unknown method. Unlike PIN number, biometric features including finger vein cannot be changed for life. However, as long as finger vein biometrics concerns, the enrolled pattern can be replaced up to six times (!) The compactness of finger vein biometric features is also an important advantage. Finger vein biometric systems can be easily implemented and embedded into small spaces such as computer mice, telephone handsets or even door handles, by which users can obtain authorisation of the devices without paying any special attention. 2 System overview 2.1 Introduction In this section, a typical hardware architecture of finger vein authentication system and its related technologies are described by taking an example of one of the most successful commercial products. The finger vein authentication device illustrated here is Hitachi-Omron Terminal Solutions Model TS-E3F1 UB Reader series, the world s first finger vein authentication device widely applied for banking use.

3 3 Fig.1 shows the exterior of TS-E3F1. This compact device is designed to be used as a part of larger system such as ATMs or door access control. For that reason, it does not have a keypad or any other similar devices for ID number input. TS-E3F1 has instead an interface port on its back (USB1.1 or RS-232C, depending on the model), so that an external computer can control the device. The communication between TS-E3F1 and the controller and is fully encrypted and only one authorized computer is permitted to connect to the FV device at a time for even higher security. The grey and black part is the Fig. 1. Finger vein authentication device UB Reader optical unit on which users are to place an enrolled finger to be authenticated. A large World s first finger vein authentication device introduced for banking. multi-colour LED indicator is embedded on the optical unit to show the status of the device. The detail specifications of TS-E3F1 series can be found at (Hitachi-Omron 2006). 2.2 Hardware architecture TS-E3F1 is comprised of four major parts, namely an optical unit, CPU, SD-RAM, Flash memory and a communication interface (Fig. 2). The finger vein device is connected to a control unit such as PC by either RS-232C or USB interface and controlled by the commands sent through the interface. The commands are received by the communication LSI and processed by the firmware stored on the Flash memory and the SD-RAM. Up to 1,000 finger vein templates can be stored in the Flash memory. The optical unit has a pair of infrared LEDs aligned on both sides above the CCD camera as shown in Fig.3. The light intensity of LEDs is adaptively controlled in order to capture images optimised for finger vein pattern extraction. The hardware specifications of TS-E3F1 are shown in Table 1.

4 4 RS-232C / USB Optical Unit (Camera + IR LEDs) Control unit (e.g. PCs or custom controllers) Communication LSI CPU (HITACHI SH-4) SD-RAM Flash Memory Fig. 2. Block diagram of TS-E3F1 Table 1 Hardware specifications of TS-E3F1 Interface USB1.1 or RS-232C OS Windows NT4.0/2000/XP (USB not supported by NT4.0) Dimensions 75 x 157 x 48 (mm) (WxDxH) Weight Approx. 0.2 (kg) Power DC12V Fig. 3. Infrared LED illumination LEDs are coloured in this picture for visualisation.

5 5 2.3 Image capturing technology Finger vein patterns are extracted by utilizing the optical characteristics of blood, or more precisely, haemoglobin. Blood flowing in veins contains deoxidised haemoglobin, which absorbs near infrared light. Since blood vessel walls are almost transparent in the range of infrared light wavelength, rays of infrared light incident on the blood vessel are not reflected but absorbed by the haemoglobin flowing inside, which results in dark network vein patterns. The optical unit of finger vein authentication device is equipped with a pair of aligned LEDs placed on both sides of the finger, which supplies infrared illuminations while capturing the finger vein pattern. The wavelength of infrared light is carefully selected by evaluation so that the contrast between the veins and the background is maximised. The infrared rays emitted from the LEDs penetrate through the finger and reach the CCD camera unit embedded beneath the tinted acrylic cover under the finger, where the raw infrared images are generated (Fig 4). LEDs Acrylic cover LEDs Veins Camera unit Fig. 4. Finger vein imaging The raw infrared images themselves are, however, not suitable for biometric base images as the contrast is not homogeneous across the field of view. Fig 5 shows a pair of raw infrared images of an identical finger. Image (a) was captured by using the illumination such that the lefthand side LEDs are brighter than the right-hand LEDs. Image (b) was captured by using the illumination of the opposite settings. As illustrated, some parts of the finger are too bright and saturated whilst (a) (b) Fig. 5. Infrared images of finger vein patterns The upper side of these images corresponds to the left-hand side of finger. Veins can be seen in the finger as a dark network patterns. others are too dark to locate veins. As the quality of finger vein patterns largely affect to the accuracy and the reliability of the entire authentication system, the LED

6 6 illumination control is very important and one of the key technologies used in the FV systems. It is noted that the near infrared illumination used by Hitachi s finger vein system is completely safe for human beings and animals as the intensity of the light is no more than one part per million of the natural near infrared rays contained in sunlight. Additionally, arteries are not imaged by this device as they are laying in deeper part of finger and have much less deoxidised haemoglobin. 2.4 Typical implementations Fig. 6 shows two of the most popular implementations of finger vein authentication system. The first one is a smart card system, in which user s biometric templates are encrypted and stored in a smart card that is owned by the user. In this architecture, neither a large-scale biometric database nor data security measures to protect the database are required and, consequently, the system maintenance costs are typically low. Since smart cards are equipped with a processor inside, it is possible to use it to execute matching processes on the card. This approach is very safe because biometric templates are always in the secured area on the smart card and never loaded out to the system. This method is known as match-on-card technology and regarded as one of the most protected Fig. 6. Two typical implementations of FVsystem for financial use

7 7 authentication procedure to date. The match-on-card technology is increasingly popular and employed for high security systems such as banks. The second one is a server system, which is relatively complicated than the smart card system. Users biometric templates are stored in a data storage system, typically located in a remote processing centre. Users are normally required to input an ID code by keypad or swiping a magnetic card so that the system can identify the user and retrieve the user s biometric templates from the server. The matching processes can be executed at the remote server, terminal computers or anywhere depending on the system requirement. In the case of TS-E3F1, however, all matching and image processing algorithms are executed inside the hardware itself for higher security. Although the data maintenance costs tend to be higher than the smart card systems, the initial cost can be suppressed by employing this framework especially when the biometrics is introduced as an additional security for an existing magnetic card system. This framework is also preferred in case 1:N matching is required for a relatively small system of a limited number of users. In both systems, typically two or more fingers per person are enrolled in preparation for troubles such as injury. In order to exclude frauds and to guarantee the quality of template data, it is very important to have a trusted administrator attended during the enrolment process. The features of both systems are compared and summarised in Table 2. Matching process Biometric data management Security measures Templates Table 2 System comparison Smartcard system match-on-card or intradevice matching End-users are responsible for their own biometric template stored on smartcard. Security measures for lost cards are necessary: Bio template encryption Card R/W validation Template must be re-written when issuing a new card. Server system Intra-device matching System administrator is responsible for the users biometric templates stored on server. Security measures for bio database at server are necessary: Network security Database encryption No template needs to be written on card.

8 8 3 Matching algorithm 3.1 Introduction In general, the preferable characteristics for a matching algorithm are as follows. - Extremely low FAR - Reasonable FRR - Practical processing time In the real world, however, the performance requirements are largely dependent on the system applied and the following features may also need to be considered in practical use. - Configurable FAR/FRR - Robust security measures such as encryption The FAR/FRR are typically configured by the threshold value to determine the acceptable similarity scores. It is well known that these two performance indices have a trade-off relationship and the threshold value configuration is one of the most sensitive and critical factors to characterise the identification system s behaviour. 3.2 Pattern Extraction The infrared images captured by the optical unit (see section 2.3) are firstly given to the pattern extraction algorithm before matching. The pattern extraction algorithm locates finger veins appear in the infrared images and converts the infrared images from 256-level greyscale to three-level image. In the conversion process, the intensity values of all pixels in the input 256-level images are re-assigned by the following rules. 0, if the pixel of interest belongs to the background I(x,y) = 2, if the pixel of interest belongs to finger vein (1) 1, otherwise where I(x,y) is the intensity value of the pixel at (x,y). More intuitively, pixels most likely to be a part of finger veins are assigned to be white and those likely to be the background are assigned to be black. Grey is assigned to all other pixels, which are literally classified as grey area. These three-level images are tested by the feasibility checking logic and only those passed the check-

9 9 ing are used as valid templates or samples in the next matching algorithm. In the actual implementation, the enrolment process is typically repeated several times in order to reduce false enrolment due to poor image contrast and to obtain the most representative vein pattern of the finger. Fig.7 shows an input greyscale image and its corresponding threelevel image. Fig. 7. Finger vein pattern extraction 3.3 Matching algorithm The matching algorithm is rather simple. The similarity of the template and the input pattern is evaluated by the distance. The distance is calculated based on the number of pixels of the same intensity value at the corresponding coordinate on each of the three-level images. In order to calculate the distance precisely, image segmentation and registration techniques are carefully applied in both software and hardware approach. The details of the matching algorithm employed for TS-E3F1 is, however, not to be disclosed for security reason. For relevant matching technology, please refer (Miura, Nagasaka, and Miyatake 2003) and (Miura, Nagasaka, and Miyatake 2004) 4 Performance evaluations 4.1 Introduction In order to evaluate and compare the performances of biometric systems, it is extremely important to employ a widely recognised, objective methodology. Hitachi- Omron s TS-E3F1 was tested and evaluated based on two different evaluation standards, namely, JIS TR X0079 and INCITS JIS TR X0079 is a performance evaluation guideline titled Evaluation Method for Accuracy of Vein Authentication Systems defined by Japan Industrial Standard in 2003, which is especially designed for vascular biometric systems. JIS TR X0079 is mainly focuses on the accuracy of the identification systems and the performances are evaluated in terms of FAR and FRR at a fixed threshold configuration. INCITS is a scenario testing and reporting standard for biometric performance evaluations defined by the American

10 10 National Standards Institute (ANSI) in The ANSI compliant scenario testing for TS-E3F1 was conducted by a third-party testing organisation the International Biometric Group, LLC [ as a part of its public biometric system testing programme called Comparative Biometric Testing Evaluation based on JIS-TR X0079 (a) Evaluation method In compliant with JIS TR X0079, FAR and FRR were calculated at a fixed threshold value using an adequately large number of samples. FRR was calculated based on on-line genuine attempts. Subjects are to present each finger until either the finger is accepted as a genuine attempt or rejected by time-out. The attempt duration was set to five seconds, which is a typical configuration for the commercial use of this device. The FV device captures images at a frame rate of approximately 100 ms and evaluates the distance (or similarity, in other words) between each frame and its enrolled template at real time. The device stops image capturing as soon as the distance becomes lower than the preset threshold value and gives a message authenticated to the test subject. If the distance does not become lower than the threshold value within the attempt duration of five seconds, the attempt is regarded as rejection, i.e. false rejection. FAR was, on the other hand, calculated by an off-line batch process. The templates and samples obtained in the genuine attempts were used for the batch process. FAR was computed by matching all combinations of the templates and samples except one genuine combination and counting the number of falsely accepted cases. (b) Evaluation samples Samples are collected from a population of office workers in Japan from October to December in Four fingers per subject are used for the evaluation, e.g. index and middle fingers of both hands. Each test subject presents a finger for 10 times. The number of test subjects was 2,673 and the number of tested genuine attempts was 2,673 subjects * 4 fingers * 10 times = 106,920. (2) The number of impostor attempts was (2,473 subjects * 4 fingers) * (2,673 subjects * 4 fingers -1) * 10 = 1,143,081,720. (3) (c) Results

11 11 There was no false rejection in 106,920 genuine attempts and a very small number of false acceptances in 1,143,081,720 impostor attempts at the device s default threshold value. FAR was less than % and FRR was %. 4.3 Third-party evaluation by International Biometric Group (a) Testing organisation International Biometric Group, LLC (New York, USA) conducts a third-party objective evaluation tests called Comparative Biometric Testing approximately once a year since CBT is one of the most reliable and influential third-party testing programmes specialised in biometric systems. The evaluated modalities in the past CBTs include finger print, iris, face, voice, hand geometry, key strokes and signature biometrics and more than 50 biometric systems from the world were tested so far. The CBT reports are distributed to thousands of commercial and governmental organisations throughout the world and provide them useful information. The CBT methodology is compliant with ANSI INCITS and the performance tests are conducted independently from participating biometrics vendors. CBT is a scenario testing that is designed in compliant with ISO/IEC JTC1 SC : Testing Methodologies for technology and Scenario Evaluation. (b) Evaluation method In CBT programme, biometric systems are evaluated in both usability and accuracy. The usability was assessed in terms of failure to enrol rates (FTE), failure to acquire rates (FTA), and transaction duration. The FTE and FTA represent the system s ability to successfully enrol or acquire biometric features from test subjects, which is important especially when the system is to be applied to a large population. Transaction duration was measured for both enrolment and matching process and includes not only the hardware s processing time but also the time test subjects spent to operate the system. The accuracy was assessed in terms of match rates, namely, False Match Rate (FMR) and False Non-Match Rate (FNMR). Since FMR and FNMR have a trade-off relationship that are determined by the given threshold value, the system s overall characteristics were evaluated by Detection Error Trade-off curves, or DET (Martin, Doddington, Kamm, Ordowski, and Przybocki 1997).

12 12 Unlike FAR or FRR, samples failed to capture were excluded from the FMR and FNMR calculation. FAR and FRR are given by the equations below: FAR = FMR * (1-FTA) (4) FRR = FTA + FNMR * (1-FTA) (5) (c) Evaluation samples Two templates were created per test subject by scanning the right index finger and the right middle finger during the enrolment process. The matching samples were collected in a form of streaming image volume. One streaming image volume contains 50 frames of biometric feature, which corresponds to a five-second presentation of finger in the real-time implementation. In order to emulate real world applications such as ATMs, the streaming image volumes are captured on transaction basis. Three streaming volumes are created per transaction and three transactions per finger were executed. Total of nine streaming image volumes were collected per finger as matching sample data. 650 test subjects are publicly recruited from the residents of New York City and its neighbouring area. The test subject population has a variety of ethnicity and wide range of age. The statistics profile of the test subject population is shown in Fig. 8 (a) (c). (a) (b) (d) Usability results Table XX shows the usability results. Levels are rated by IBG according to its original rating criteria, where level 4 is the best and level 1 is the poorest. (c) Fig. 8. Statistics profile of test subject population

13 13 Table 3 Usability evaluation results Usability index Result Level Failure to Enrol Rate (FTE) 0.08% Level 4 Median Enrolment Transaction Duration (seconds) 33.3 Level 3 Transactional Failure to Acquire Rate (T-FTA) 0.06% Level 4 Median Recognition Attempt Duration (seconds) 1.23 Level 4 (e) Accuracy results The attempt-level and transactional DET curves are shown in Fig. 9. FMR and FNMR values are plotted over various threshold values. The closer a DET plot is to the origin (i.e. the bottom-left corner), the more preferable configuration of the threshold. Please note that both the horizontal and vertical axes are in logarithmic scale, where 1.00E+00 corresponds to 100%. Detection Error Tradeoff Curves 1.00E+00 Attempt level Transactional False Non-Match Rate (FNMR) 1.00E E E E E E E E E E E+00 False Match Rate (FMR) Fig. 9. Detection Error Tradeoff Curves of TS-E3F1

14 Evaluation summary The usability performance of TS-E3F1 was one of the best results ever marked in the history of CBT; three out of four indices marked the highest ratings Level 4 and one index rated as Level 3 was very close to Level 4. This proves that finger vein systems are quicker and more widely applicable than other modalities. As for the accuracy, TS-E3F1 once again demonstrated very competitive performance. FAR/FRR results in the JIS compliant evaluation was one of the best performances among many biometric systems. The CBT results were, on the other hand, not as good as JIS results despite of its stable performance over various threshold values. The possible reasons for this difference are as follows: (a) Non-interactive sample data collection: In CBT, sample streaming images are simply recorded without matching, which may have caused for some fingers to be kept in a wrong position. This may result in higher False Non-Match Rate. (b) Differences in user instruction: A printed instruction and minimum guidance by IBG staff were provided during CBT data collection, whilst skilled engineers gave detailed instructions for test subjects in JIS evaluation. Both of the performance evaluations are carefully designed and strictly compliant with their relevant standards as described earlier in this section. Although the results of these two evaluations are not directly comparable, evaluations based on these publicly-recognized standards are extremely informative and useful when choosing a biometric system from a variety of modalities. For further information of CBT round 6, refer (IBG 2006).

15 15 5 Applications in financial sector 5.1 Background of Biometrics Implication (a) Social problem of illegal withdrawal using counterfeit or stolen cash cards In Japan, the number of illegal withdrawal cases from ATM using counterfeit or stolen cash cards are rapidly increasing and it has been an object of public concern. The total amount of withdrawals using counterfeit or stolen cash cards from ATMs in Japan was reported to be as much as 19million US dollars in Those who lost a fortune by illegal withdrawals claim that banks are to blame for not taking sufficient precautions against this sort of frauds and the scope of legal responsibility has been actively discussed. (b) Government policy In reaction to these trends of public opinion, Japanese government embarked on a tougher new policy against illegal withdrawal. On 10th of February 2006, the Law concerning the Protection of Depositors from Illicit Deposit Withdrawals Using Counterfeit/Stolen Cash Cards through ATMs (so-called the Depositor Protection Law) went into effect aiming to prevent this sort of crimes. This law clearly defines the scope of legal responsibilities of financial institutions and account holders and requests financial institutions to compensate for the damage if the account holder is proven free from fault and innocent for the illegal withdrawal. (c) Financial institutions responses Financial institutions responses may be categorised into the following three measures. PIN code renewal (normally four digits) Introduction of smart cards Introduction of biometrics It is well known that some people are using their birthday or telephone numbers as PIN, which makes swindlers job easy. If a cash card was stolen together with a driving license for example, it is easy for the fraudster to obtain the genuine PIN code within a small number of trials and errors. For this reason, almost all financial institutions encourage their account holders to change PIN codes frequently.

16 16 Introduction of smart card and biometrics is relatively a new movement among financial institutions in Japan. The trend is, however, rapidly spreading nationwide and forming the mainstream of financial security measure. 5.2 Introduction of biometric systems (a) Biometric modalities Currently, there are two biometrics modalities adopted for ATM application in Japan. Finger vein authentication systems Palm vein authentication systems (b) Current situation of finger vein authentication in Japan As of the end of September 2006, 42 users in financial sector adopted finger vein authentication system in Japan. Among them are Japan Post and three out of four Japanese mega banks. More than 10,000 finger vein authentication equipments had been installed at ATMs (Fig.10), and over 30,000 finger vein enrolment equipments (Fig.11) had been set up at teller s windows of financial institutions by the end of September Fig. 11. Finger vein enrolment device Fig. 10. Automated teller machine with FV device The world s first ATM equipped with finger vein biometrics AK-1 was released in 2004 and approximately 30,000 units are working in Japan as of November A smartcard reader/writer is attached to an FV device. Encrypted finger vein biometric data is securely recorded onto a smart card on site so that it can be issued to end users straightforwardly.

17 17 5 Conclusion and future plan Hitachi-Omron s TS-E3F1 is the first finger vein identification device tested by an internationally recognised third-party organization. The objective evaluations proved that finger vein identification technology is extremely competitive and superior in many aspects to other conventional biometric modalities in terms of both usability and accuracy. By taking advantage of the compactness of finger vein biometrics, even smaller hardware units are being developed. The compact FV module shown in Fig. 12 is as small as mobile phone (69mm(W)*85mm(D)*43mm(H)) and equipped with everything necessary for personal identification. Since all identification processes are executed in this small module, no powerful CPU or large memory is required for the control unit; virtually any processor can drive this module. Fig. 12. Finger vein authentication module One of the unique applications of this module is a key management system shown in Fig.13. As shown in Fig.13, the key management system has a keypad by which users input IDs before presenting their enrolled finger to the compact FV device installed below. The system allows the authenticated users to loan keys and records the status of key along with the user s information. Even smaller units are now being developed and some epoch-making prototype systems are proposed (Fig.14). This miniature implementation of FV system further enhances the scope of application and brings new concepts to conventional systems. Fig. 13. Key management system (a) (b) Fig. 14. Applications for automobiles (a) Keyless engine starter (b) Door handle with FV system. Only enrolled people are entitled to use the car. Users are required to type an ID number and place a finger to obtain a key to a restricted area / resources.

18 18 6 Acknowledgments Authors thank Prof Yoichi Seto at AIIT for his support, Dr Takafumi Miyatake, Dr Akio Nagasaka and Mr Naoto Miura at Hitachi Central Research Laboratory for providing technical information, Hitachi-Omron Terminal Solutions, Corp. for providing product information, photographs and evaluation data. References Kono, M., Ueki, H., and Umemura, S. (2000) A new method for the identification of individuals by using of vein pattern matching of a finger, Proceedings of the Fifth Symposium on Pattern Measurement, pp.9-12 (in Japanese) Miura, N., Nagasaka, A., and Miyatake, T. (2004) Feature Extraction of finger-vein patterns based on repeated line tracking and its Application to Personal Identification, Machine Vision and Applications, IAPR, Vol. 15, No. 4, pp Miura, N., Nagasaka, A., and Miyatake, T. (2003) Feature Extraction of Finger Vein Patterns Based on Iterative Line Tracking and Its Application to Personal Identification, IEICE transactions Vol.J86-D-II, No.5, pp (in Japanese) Hitachi-Omron Terminal Solutions, Corp.(2006) URL. Martin, A., Doddington G., Kamm T., Ordowski M. and Przybocki M.(1997) The DET Curve in Assessment of Detection Task Performance, Proceedings of the 5th European Conference on Speech Communication and Technology, vol. 4, pp IBG (2006) Comparative Biometric Testing Round 6 Public Report. International Biometric Group, LLC. New York.