Study: Evaluation of Fingerprint Recognition Technologies BioFinger. Public Final Report

Transcription

1 Study: Evaluation of Fingerprint Recognition Technologies BioFinger Public Final Report Version

2 Content Study: Evaluation of Fingerprint Recognition Technologies BioFinger 1 1 Summary Objective Results Structure of the Report 8 2 Biometric Authentication with Fingerprint Recognition Systems Introduction Requirements on a Biometric System Operative Capability of a Biometric System Fingerprint Recognition Prob lem Definition Fingerprint Scanning Pattern Classification Fingerprint Image Comparison Image of the Fingerprint Identification Procedure 20 3 Evaluation of Biometric Systems Description of the Evaluation Criteria Types of Errors Objective Comparison of Fingerprint Systems Experimental Determination of the ROC curves Determination of the Probability Density Functions Calculation of FNMR(T) and FMR(T) Determination of the ROC curves Police-Related Application Scenarios of Biometrics Systems and their Requirements regarding Error Rates 28 4 Investigations with Test Persons Inclusion of the Database Sensors and Algorithms Description of the Sensors Description of the Algorithms U1 Influence of the Sensors on Verification U2 Influence of Feature Extraction on Verification U3 Influence of the Algorithms (MSA) on Verification U4 Influence of the Sensors on Fingerprint Image Quality Contrast Average Value of Grayscales Separability 57 2 Federal Office for Information Security

3 4.6 U5 + U6 Influences on the Fingerprints U5 Influence of the Sensors on the Fingerprints Influence of Feature Extraction on the Fingerprints 60 5 Test Results of Various Systems Introduction Evaluation of the Fingerprint Quality Comparison of the Systems ROC Curve for Sensor ROC Curve for Sensor ROC Curve for Sensor ROC Curve for Sensor ROC Curve for Sensor ROC Curve for Sensor ROC Curve for Sensor ROC Curve for Sensor ROC Curve for Sensor ROC Curve for Sensor ROC Curve for Sensor ROC Curve for Sensor Description of the System ROC curve ROC Curve for Algorithm ROC Curve for Algorithm ROC Curve for Algorithm ROC Curve for Algorithm ROC Curve for Algorithm ROC Curve for Algorithm ROC Curve for Algorithm Investigations with the Fingerprint Database Description of the Databases Research on the Differentiability with Similar Fingerprints Description of the Examination Results Research on the Ageing Characteristics of Fingerprints Description of the Examination Results Examination of Ageing According to Age Groups 103 Federal Office for Information Security 3

4 6.4 Simulation of the Reduced Sensor Area Standards and Universal Fingerprints Feasibility and Algorithm Methods Feasibility Algorithm Procedures Documentation of Standards General Standards Fingerprint-Specific Standards Bibliography Table of Abbreviations Federal Office for Information Security

5 1 Summary 1.1 Objective As a biometric identification property, fingerprints have had a long tradition and are a synonym for the uniqueness (of man). Up until recently, it was only the resulting fingerprint image that was exclusively used as an identification feature; no further processing was carried out. Human fingerprints were almost solely used for forensic purposes in dactyloscopy. Dactyloscopists examine fingerprints with regard to details that can be used to identify people. Evidence of a fingerprint found at a scene of a crime can thus be allocated to a person as the one who left that trace. Since fingerprints can be classified, they can be categorized into various finger classes by making use of the fact that due to the ridge flow so-called patterns (loops, arches, whorls) are formed and that due to the interruptions of the ridges, anatomic characteristics (minutiae) are shaped. Thanks to the large dactyloscopic information content in individual prints, a dactyloscopic expert can determine, by comparison, whether individuals are identical or not. In the past, it took a lot of time to find one person in a hard copy database (identification) and then to prove that the fingerprints at the site of the crime and in the database were identical. The initial use of computers for identification purposes was limited by a quick searching of an electronic database. Dactyloscopic experts provided the details necessary for that searching process. Since computer performance capacities have increased, image processing of fingerprints and thus their electronic evaluation became possible. Initially, dactyloscopic systems analyzed and extracted all known details, i.e. patterns and the set of features. As far as their application in an access control system was concerned, the use of these comprehensive details resulting from fingerprints proved to be impractical. Processing time was too long and the amount of extracted details too large. As a result, the amount of data was reduced, i.e. certain patterns were treated as negligible. Additionally, the number of minutiae was reduced. Mostly, for today s access control systems, minutiae are simply defined as ridge endings or ridge bifurcations. More recent developments are aiming to use not just the minutiae but also the image information of the surroundings of a particular detail by covering it with a filter mask (e.g. by use of Gabor Filter). As part of the project called BioFinger Evaluation of Fingerprint Recognition Systems Fingerprint Technologies, the characteristics of fingerprint recognition systems are analyzed. The background of this project is the possible integration of fingerprints in German personal documents in order to improve the verification of the holder of the document (i.e. ID Cards, driver s licenses, passports). Hence, the very aim of the BioFinger Project is the verification, i.e. the examination of the identity claimed by the person (1 on 1 comparison). On the other hand, with regard to envisaged application, identification (one on x-comparison), with which a person is to be identified by comparing him/her with x number of people in a database, does not play any role. Within this context, a number of examinations are carried out in the BioFinger Project, which are to clarify the suitability of some chosen products. The question is this: Using today s systems or components, are there fingerprint recognition systems that have verification characteristics, or can they be assembled. Due to the special demands on personal documents, i.e. usable lifetime of ten years, the ageing of fingerprints with regard to their characteristic to identify people, is very significant. At the beginning, a market analysis, which includes all fingerprint technologies available on the German market, is carried out. Furthermore, a number of selected systems of foreign companies are included in the investigation. Promising systems are chosen from this survey. For this purpose, software algorithms and sensor hardware that is used are tested. This examination is meant to clearly show possible significant differences in those fingerprint recognition technologies. The associated ROC (Receiver Operating Characteristic) curves are set up in order to assess efficiency and comparability of the chosen fingerprint recognition technologies. Federal Office for Information Security 5

6 In addition, the algorithms were examined with regard to their capability of differentiating between socalled biometric twins (persons whose fingerprints were classified as being similar by the fingerprint recognition systems). This was under-girded by examining the influence of the ageing process on the algorithm performance. The Federal Office of Criminal Investigation (BKA) provided the specially selected databases. 1.2 Results Achievable Recognition Performance The examination has shown what kind of performance today s technology can achieve. The result was that half of the tested systems 1 had an EER 2 [Equal Error Rate) of less than 5%. One third achieved EERs below 3%. In the top range (EER 1%), there are 8% of the tested systems. As far as the verification of passport or identity card holders is concerned, the recognition system will probably be run in such a way as to have an FAR 3 [False Acceptance Rate] that is better than the EER, e.g. FAR = 1. Even though this leads to a worse FRR 4 [False Reject Rate], half of the tested systems still generates FRRs below 10% for this operational mode. About 23% of the tested systems can still reach FRRs of 3% or less. 1 Combination of scanner and algorithm 2 EER: Equal Error Rate; see for definition 3 FAR: False Acceptance Rate; see for definition 4 FRR: False Reject Rate; see for definition 6 Federal Office for Information Security

7 Number of test pieces absolute % 80% 60% 40% 20% 0% cumulated 70% 60% 50% 40% 30% 20% 10% 0% FFR with FAR=1/1000 This means that, if mutually compatible components (scanner and algorithm) are carefully chosen, only one out of 1,000 persons with a false ID card would, despite his/her false identity, be accepted by the biometric system. However, the probability of wrongly rejecting a person with a correct ID card would be about 1:50. Thus, this technology shows an effective improvement to people comparing faces with ID card pictures. Influence of Components A few combinations of sensors and algorithms led either to a notably higher error rate or were not compatible at all. Comparing results of individual sensors showed significant differences. For instance, the best sensor achieved an error rate that was ten times lower than the worst one. Optical sensors operating with the method of frustrated total internal reflection achieved the best results. Differences between algorithms were notably less pronounced. The best algorithm achieved error rates that were three times lower than those of the worst algorithms. Influence of the Age of Reference Data The possible use of fingerprints in personal documents raises the question of whether recognition ability stays the same if reference and verification data were not recorded within a certain period of time but rather at large intervals. In principle, the wider the time frame, the worse the FRR that is to be expected. Based on the examinations that were carried out, it can be estimated that the FRR doubles if the time period reaches ten years. Standards and Universal Fingerprints The different templates of the various algorithms for recognizing fingerprints showed a great variety in design despite the fact that they had some features in common. Some systems extract only minutiae; others, however, additionally extract patterns or else they use image information either exclusively or in addition. As a consequence, one system would, under normal circumstances, not generate the details, which are needed by the other system for its verification purposes at its usual level of performance. Although the smallest possible common amount of data with which all algorithms can work guarantees the inter-operability of various individual systems, it would, nevertheless, have negative effects on the biometric recognition performance and increase error rates. Highest interoperability without a decrease in performance can be reached if fingerprint images are stored instead of features. However, this requires additional feature extraction for every single verification procedure. Federal Office for Information Security 7

8 Possibilities for Improvement Since this examination was carried out with all fingers except for the small finger, further improvement can be expected if only fingers with a large area (thumb, index finger) are used. Since not only the best fingerprints were used but rather all images were analyzed, an improved recognition performance can be expected, if for example in case of a wrong rejection further verification attempts are allowed for or if a quality control is carried out at enrollment. The follow-up study, BioFinger2, shall show what kind of improvement can be reached if several fingerprints are used for verification purposes. 1.3 Structure of the Report The individual chapters of the report are briefly described below. The second chapter describes biometric systems in general. At the beginning, there is a short introduction into the topic of biometrics and a description of a biometric system is given. The requirements for biometric systems are explained and performance parameters of a biometric system are defined. In the second half of the chapter, the fingerprint recognition procedure is discussed in detail. The third chapter elaborates on the evaluation criteria of biometric systems as well as the definition of such criteria. Furthermore, evaluation metrics are linked to concrete police-related application scenarios. Thus, the concrete feasibility of the tested system can be evaluated with regard to its intended purposes. The fourth chapter describes the examination of fingerprints from test persons. It is introduced with the description of the recording of the database and contains a table of sensors, the set-up of the database from the fingerprints of various persons as well as an explanation of the database analysis in order to filter out errors. Within the framework of the study, a number of examinations (E1 through to E6) are carried out, which are also mentioned in this chapter. U 1 Influence of the sensor quality on the verification quality (FAR, FRR) U 2 Influence of the quality of feature extraction methods (PE) on the verification quality U 3 Influence of the various matching systems (MSA) on the verification quality U 4 Influence of the various sensors on the quality of the fingerprint images (resolution, etc.) U 5 Influence of the sensor quality on the quality of generated feature vectors U 6 Influence of the various PEs on the generated feature vectors. Which features are extracted by an algorithm (e.g. "Only coordinates of the minutiae" or "coordinates of minutiae and directions", number of features, data quality of features, etc.)? The fifth chapter summarizes the results of the above-mentioned examinations for various sensors and algorithms. For the purpose of this test, eleven sensors and seven algorithms were used. The sixth chapter contains a description of the investigations carried out using the fingerprint image database provided by the Federal Office of Criminal Investigation (BKA). The seventh chapter describes today's standards (with regard to fingerprint recognition technologies); it also discusses the possibility of a universal fingerprint standard. Darmstadt, 20 May Federal Office for Information Security

9 2 Biometric Authentication with Fingerprint Recognition Systems 2.1 Introduction In order to classify fingerprint recognition as a biometric procedure, the following terms need to be defined: [BRO02], [TTT02]: Static features are anatomic characteristics of the body, which either change very little or not at all in the course of life (fingerprints, eye color, iris, genetic data, etc.). Dynamic features are behavioral characteristics of man (handwriting, walk, voice, etc.). Passive acquisition is "pass by" acquisition (e.g. of a face by a camera). Active acquisition describes an acquisition process involving the person (i.e. fingerprint). Identification: Establishing of identity (1 on x-comparison; who is this person?). For identification purposes, the biometric feature is compared with all reference details stored in the biometric system. If any characteristics match, the identification process was successful and the corresponding name (e.g. User ID) belonging to this reference feature can be processed further. Verification: Confirmation of identity (1 on 1 comparison; is this person who she/he claims to be?). For verification purposes, the user states his/her identity to the biometric system in advance (e.g. his/her User ID is entered via a keyboard or chip card). Then the system only has to compare the biometric feature with one reference feature that matches the User ID. If they are identical, the verification process was successful. Verification is done significantly faster than identification if the number of reference features / users is very high. At the same time, verification is much more reliable than identification, especially if the number of reference features is very high. Authentication: Attestation of genuineness (proof of identity, e.g. by identification or verification). Authorization: Authorization means "empowerment. Following a successful authentication (identification or verification) using a biometric system, a person is given permission to carry out certain actions or to use certain services. Biometric system: Biometric recognition systems process biometric features of a person with the aim of confirming or rejecting that person's identity by using previously gathered reference data. In general, all biometric systems are made up of the following components: data input, preprocessing, feature extraction, classification, and calculation of reference data. For adapting to changes in the biometric pattern, an adaptive procedure can be used. Figure 2.1 demonstrates the basic set-up of a biometric system. Figure 2.2 shows the verification process. Data input is carried out via a sensor. The data is pre-processed and normalized prior to and during the comparison of patterns. For classification purposes (i.e. for categorizing fingerprint image types into given finger classes) both pre-processed data or extracted features can be used. The initial input data or features are compared with respective reference data. In order to choose reference data in the reference database, the user may, for example, indicate his personal identification number. As an alternative, reference data may also be stored on a storage medium, such as a chip card, which the user holds. As far as adaptive procedures are concerned, if the classification was positive, the results thus achieved can be used for updating reference data. Nowadays, the demand for reliable identification procedures is increasing. Currently, we encounter the issue of personal identification e.g. in e-commerce, access control facilities, in the fight against terrorism etc. Even though identification by means of an object, e.g. an identity card, is still fulfilling its purpose, it is continually loosing its importance in our modern, electronically communicating world of more than 6 billion people. For this reason, biometry has, especially in recent times, been getting Federal Office for Information Security 9

10 more important since it combines personal identification with unambiguous and unchangeable characteristics of man. With ever-increasing and evermore complex technologies, exact personal identification is imperative. By using identification processes, it is for example possible to regulate access to certain objects by granting certain rights. Everyone who was positively identified and thus accepted is given preestablished privileges. In the police, identification (e.g. dactyloscopy) plays an important role. These are only two of many cases where "biometric" identification is used. Man has certain unambiguous features (in the sense of physical characteristics) which are formed in the earliest phases of human life as part of a random process (randotypical) and which are different for each individual. One of the first biometric features that was discovered and scientifically investigated was the fingerprint. The surface of the inguinal skin of man and of most mammals shows patterns and their variety seems to be endless. For example, the ridges of the inguinal skin on the fingers of humans are different. Ridges form various patterns (loops, arches, whorls) which in connection with interruptions of the ridges (minutiae) differ from finger to finger. For forensic purposes, fingerprints were used as early as at the end of the 19 th century in order to identify people (dactyloscopy) [HEI27]. With the advancement of technology, the issue of safety has become more important. For access controls, analyzing fingerprints biometrically has been playing an increasingly important role. Registration module Template database Biometric sensor Characteristics IDENTIFICATION MODULE Biometric sensor Features Comparison Figure 2.1: General biometric system 10 Federal Office for Information Security

11 Data acquisition Reference data Reference data adaptation Pre-processing Reference formation Classification Data Feature extraction Feature vector Decision original / falsification Figure 2.2: Sequence of biometric verification 2.2 Requirements on a Biometric System Each physiological or behavioral feature can be used as a biometric characteristic for personal identification processes as long as they fulfill the following requirements: Universality: Every person has to have this feature, Uniqueness: No two or more people with the same feature must exist, Constancy: The feature does not change significantly in the course of time, Collectability: The feature must be measurable or collectable. However, biometric features that are currently being used mostly do not fulfill all of the abovementioned requirements. Hence, they are only partly suitable for a practical application in biometric systems. In addition, further practical aspects have to be taken into consideration: Performance of the system which allows for quantitative statements with regard to identification accuracy and speed as well as the required robustness in the face of system-related factors, Acceptability of the system in its practical application, Fake resistance of the system, i.e. robustness against direct methods of tricking the system. Hence, for most applications, practical biometric systems have to perform with and at an acceptable identification accuracy and speed, to have reasonable requirements with regard to biometric features, to be non-invasive, to be accepted by the users, to be sufficiently robust against misuse. Federal Office for Information Security 11

12 2.3 Operative Capability of a Biometric System Although digitisation and the ensuing reduction of biometric data, which is necessary for its electronic processing, result in a very large classification of various features, it, nevertheless, shows a very rough granularity. For this theoretical reason, the system s answer whether someone is an authorized person or not, will not be an absolute yes or no ; the answer will be expressed with a certain quantitative index or matching score instead. An identification result may read as follows: the index, according to which the biometric patterns of person A are identical with the stored data of person A, is In this case, the index is the result of the matching algorithm and can, for example, show the degree to which the biometric characteristics between the actual data and the stored reference features are identical. Each matching process, will report an other numerical value which is a reflection of the statistical and system-related variations. By choosing a threshold value for these results, the right identifications are separated from the false ones. An analogous distribution can be made if an attacker (person B) is wrongly identified as person A. In this case, a threshold value also separates "false" and "right" results. These two distributions and a common threshold result in four different cases: (Person A is correctly identified as A (correct identification), (Person A is rejected as A (false rejection), (Person B is rejected as A (correct rejection), (Person B is accepted as A (false acceptance). Frequency No correspondence Acceptance threshold Correspondence FRR FAR 0 Correspondence value 100 (a) and (c) are correct, (b) and (d) are erroneous cases. If the above-mentioned frequency distributions for (b) and (d) are integrated, i.e. by fixing the threshold value in such a way as to define the upper or bottom integration limit, there will only be two results, which, once they are standardized, show the so-called False Acceptance Rate (FAR) and the False Rejection Rate (FRR). FAR is defined as the probability that a person B is accepted as person A. FRR is defined as the probability that person A is rejected. Due to the overlapping of both distributions, compromises have to be made when fixing the threshold values for the system; a lower FRR usually leads to a higher FAR and vice versa (Figure 2.3). Usually, the performance of a biometric system for automated personal identification is defined by the FAR and FRR factors. For example, if FAR is zero, it means that no attacker was accepted. Figure 2.3: FAR and FRR There are additional parameters such as verification and identification speed, which are used to determine the performance of the systems. Due to the one-on-one comparison method in verification processes, the speed is mainly limited by the time the computer needs to carry out the verification 12 Federal Office for Information Security

13 algorithm. Usually, it is quite easy to meet the speed requirements in these cases. As for identification processes, however, and especially in systems which have millions of entries the number of required comparisons does limit the overall speed of the system. 2.4 Fingerprint Recognition Prob lem Definition n the context of the term "identification through fingerprint images", fingerprints are generally accepted as human fingerprint images. Identification can functionally be split up into the following three basic tasks: (Fingerprint scanning, (Fingerprint classification, and (Fingerprint comparison. Fingerprints can be aquired as color prints or via sensors which store the ridges on a medium (glass, paper, sensor surface, etc.). During the classification process, fingerprint images are optionally allocated to a certain category based on the global orientation of the ridges while the location of the minutiae is marked as well. The comparison determines whether two fingerprint images are identical, i.e. whether they belong to the same person (finger). The complete process of a fingerprint image analysis (comparison of patterns) can be divided into six steps (Confer Figure 2.4). Scanning of the fingerprint image Image quality improvement Image processing Pattern classification Feature extraction Verification phase Figure 2.4: Process of fingerprint analysis 1. Scanning of a fingerprint image. The quality of the scanned image is the decisive factor for automatic identification purposes. It is desirable to use a high-definition fingerprint scanner which is able to tolerate different skin types, damages, dryness, as well as the humidity of the finger surface. 2. Image quality improvement. By using image quality improvement, an optical improvement of the structures (ridges) on the scanned image can be achieved. 3. Image processing. Image processing means the preparatory phase for feature extraction and classification purposes. 4. Feature classification. Fact is that all fingerprints show certain global similarities, which allow for rough classification into three principal finger classes. However, classification is a rather difficult process both for algorithm-based decisions as well as for man-made decisions since some Federal Office for Information Security 13

14 fingerprints cannot be clearly allocated to a concrete finger class. Nowadays, pattern classification is only used in dactyloscopic systems, e.g. AFIS (Automated Fingerprint Identification System) of the Federal Office of Criminal Investigation (BKA). This method is not feasible for access systems. 5. Feature extraction. In this phase, the location of the minutiae (ridge bifurcations and ridge endings) in the fingerprint is detected and extracted. In practice, scanned fingerprint images show differing qualities. The algorithm performance is negatively influenced by a poor image quality. 6. Verification phase. In the verification phase two feature vectors are being compared. The algorithm performance strongly depends on the quality (significance) of the extracted minutiae and on the comparison process. Below, we will describe in more detail the fingerprint scanning, feature classification, and fingerprint comparison processes Fingerprint Scanning Depending on whether the scanning process is carried out on- or off-line, the fingerprint image can either be a color image, e.g. on paper or an image of a life finger obtained through a sensor In case of a color print, rolling the finger on a surface generates the image of the ridges, e.g. on paper after that the finger is moistened with ink. An example of such rolled fingerprint images can be seen in Figure 2.5. By evenly rolling the finger from one side of the nail to the other, all line-related data is reliably recorded in the image. Afterwards, these images can be scanned or electronically photographed. In dactyloscopy, this method has already been used for well over 100 years. Thus, since a complete "overall imaging" of the finger is done, in addition to a higher number of ridges and minutiae, "macro features" (i.e. delta and nucleus) are recorded as well. Even though they are part of each and every ridge pattern (with the exception of the "arch" pattern which does not have a delta), they are not always printed. A disadvantage of this method is a possible distortion, which may occur through pressing and rolling the finger while taking the fingerprint. Furthermore, a quality feedback is not possible which may lead to a decrease in the quality of the fingerprints. From the user s point of view, this is an unpleasant and slow method. It is also unsuitable for partially automated access controls. Figure 2.5: Color image The term life image of a finger is a comprehensive term for images that are obtained directly by placing the finger on a suitable sensor. There are a vast number of various methods, which can be used for scanning ridges. They include: Optical sensors, Electrical field sensors, Polymer TFT sensors (TFT Thin Film Transistor), Thermal sensors, 14 Federal Office for Information Security

15 Capacitive sensors, Contactless 3D-sensors, and Ultrasound sensors. A biometric sensor is the hardware component of a biometric system, which initially supplies biometric measurements. Depending on the biometric method in use, there are different kinds of sensors. Optical sensors use light for obtaining fingerprint images confer Figure 2.6. Electrical field sensors measure local variations of the electrical field, which is generated on the finger surface relief upon the emission of a small electrical signal. Polymer TFT sensors measure the light, which is emitted upon contact when the finger is laid on the polymer substrate. Thermal sensors register the thermal finger image. In capacitive sensors, the sensor and the finger surfaces together form a capacitor. The capacity thereof changes based on the skin relief (skin ridges and grooves) confer Figure 2.6. These local changes are measured and thus represent the fingerprint. The above-mentioned sensors are used in connection with the data-processing module as on-line systems. They substitute the off-line method in which fingerprints are, for example, taken on paper before they are digitised later on. Image quality strongly depends on the "contrasts" that were achieved between the ridges and the adjacent grooves. Since there is a feedback to image-processing algorithms for on-line methods, it is relatively easy to immediately check the quality of fingerprint images that were just obtained. The life image is usually recorded by lightly placing the finger on the surface of the sensor. Since it is not so user-friendly, the finger's surface is only rolled in the context of AFIS-systems (as in the dactyloscopic method). Of course, in this case, only such ridges can be captured that are recorded as a result of being directly in contact with the sensor surface. Hence, compared to rolled fingerprint images, the life image generates the image of a smaller section of the finger's surface but, in addition, it might also have smaller distortions of the image. Figure 2.6: Capacitive sensor, optical sensor Currently, the most frequently used life image technology is the optical method. Upon placing the finger on the sensor's glass pane (prism), the elevations of the papillary lines are in contact with the glass; the grooves, on the other hand, are not in contact with it. Basically, the recording device consists of a light source (LED) and a CCD camera, both of which are located within the device on the other side of the glass pane. The light of the LED illuminates the glass at a certain angle and the photoelement receives the reflected light. The course of the beam runs in such a way that the incoming light on the contact ridges is scattered as if on a mirror surface and then reflected back on the CCD camera. There where the grooves are behind the glass pane, the light passes through; these spots remain dark. An example for such a fingerprint image can be seen in Figure 2.7. Figure 2.8 shows the design of some types of sensors. Federal Office for Information Security 15

16 Figure 2.7: Life image of a finger An ever re-occurring question is that of life recognition of a finger (temperature, fluorescence, pulse), the question being whether a scanner or sensor would accept an artificial finger. This question was already addressed in a study entitled BioIS [ZWI00]. The conclusions reached in this study have not changed significantly until today [CT02]. Figure 2.8: Types of sensors Pattern Classification The global pattern of papillary lines occurring in the central area of the tip of the finger constitutes a specific configuration, which is sufficient for a rough systematic classification. For fingerprint classification purposes, only a part of the entire image, called Pattern Area, is used. The Pattern Area is defined as the inner area, which is limited by two lines, so-called Type Lines. Two singular points are part of this central area of the fingerprint image (Confer Figure 2.9): (a) the delta (several of which may exist; only sample arches do not have deltas) and (b) the nucleus. Delta, which is sometimes also called "outer border" is usually located at the fringe of the fingerprint image. An image of papillary lines is called a delta if it is similar to the Greek capital letter delta. It is formed by two parting ridges or by a ridge bifurcation and a third ridge that is convex and coming from another direction. Some examples of a delta configuration are shown in Figure It is rather hard to define the nucleus of an individual fingerprint due to vast variations in the curving of the inner lines. Therefore, a specific point is simply chosen as the nucleus as though it was the center of the corresponding pattern. Figure 2.11 shows some examples of a nucleus configuration. Another important quantitative factor in classifying images is the number of lines. This means the number of lines that touch or cross the imaginary connection between the nucleus and the delta. Due to the great complexity of various line configurations, it is often difficult to clearly determine the number of lines. Figure 2.12 shows three simple examples for the number of lines. 16 Federal Office for Information Security

17 Figure 2.9: Type lines Figure 2.10: Delta configurations Figure 2.11: Nucleus configurations Figure 2.12: Examples of the number of lines According to the definitions given above, fingerprint categories can be described as follows (pursuant to the Henry classification system [HEN03]): In loops, one or more ridges enter into the central area, they form a curve, touch or cross the imaginary lines between the delta and the nucleus and return to the same side from which they came. There are three decisive characteristics for classifying lines as a loop: (a) at least one suitably curved papillary line, (b) a delta, and (c) a number of lines other than zero. Depending on the orientation of the line's curve, a differentiation is made between right (clockwise) and left (anticlockwise) loops. Approximately 60 to 65 % of human fingerprints belong into this category. Whorls have at least two deltas. In their nucleus, ridges form a twist. Even though this definition is very general, it expresses the main characteristic of this category. Whorls can be split up into further Federal Office for Information Security 17

18 categories: (a) flat whorls, (b) whorls with a medium slant, (c) double whorls, and (d) random whorls. About 30 to 35 % of all fingerprints belong into this category. Arches are a rather special type of fingerprint. Less than 5 % of all fingerprints belong into this category. Arches can be split up into two categories: (a) flat arches and (b) high arches. In flat arches, the ridges enter at the side, form moderate and nearly parallel waves in the center and exit on the opposite side. In high arches, the wave is stronger in the middle. The route of all lines is no longer parallel and part of the lines seemingly exerts pressure from below. Figure 2.13: Flat arch, left loop, right loop, high arch, and whorl Obviously, due to the vast variation in the spectrum of fingerprints, the classification is always a big problem both for experts as well as for automatic systems. The allocation into categories is a very complex task. Dactyloscopy experts need a lot of experience in order to do their work reliably. Figure 2.13 shows some examples for individual categories. Figure 2.14 demonstrates examples of fingerprint images, which are very difficult to classify. Figure 2.14: Left loop, high arch Fingerprint Image Comparison Data about the fingerprint category and further global characteristics, such as the number and position of the centers, deltas, and ridges, does provide enough information for a certain differentiation of fingerprints. However, the true individuality of fingerprints is determined by the anatomic characteristics of the ridges (minutiae) and their respective orientation. Whether they can be recorded in their entirety depends on the conditions when the fingerprint was taken as well as on its quality. The most frequently occurring minutiae are Ridge ending and Ridge bifurcation. Ridge ending defines the end of a line, while ridge bifurcation is defined as a point in the ridge where the line is separated into two branches. Minutiae are usually stable and robust with regard to conditions occurring during the scanning process. Figure 2.15 shows some examples. Minutiae can be characterized by their type, by x- and y-coordinates in a coordinate system, and by their direction. Figure 2.16 shows the directions. 18 Federal Office for Information Security

19 Figure 2.15: Ridge ending, simple ridge bifurcation, twofold ridge bifurcation, threefold ridge bifurcation, simple whorl, twofold whorl, and side contact; Hook, point, interval, X-line, simple bridge, twofold bridge, and continuous line If two fingerprints belong into the same category and have a certain number of identical minutiae, it is quite safe to say that they come from the same finger. Y (x,y α (x,y α Ridge ending Bifurcation X Figure 2.16: Ridge ending and ridge bifurcation The general definition for the identicalness of any two fingerprint images consists of four criteria and says: The general pattern configuration has to be identical, The minutiae have to be qualitatively identical (qualitative factor), The quantitative factor says that a certain number of minutiae must be found (in Germany it is 12), and There has to be a mutual minutiae relationship specifying that corresponding minutiae must have a mutual relationship. In practice, a large number of complex identification protocols for fingerprint image comparisons have been proposed. These protocols are derived from the traditional dactyloscopic methodology and prescribe an exact procedure for trained specialists. Even though various protocols differ in the process flow of the comparison procedure and the definition of the decision, the basic steps remain the same. Typically, comparison is done in an iterative three-phase-process. It is hardest to compare two fingerprints that have similar feature configurations. If, however, both fingerprints are totally different as far as their feature configuration is concerned, it is impossible that these images are from the same finger. In the next step, significant minutiae are examined, the central area is located, and the minutiae are compared with each other. Afterwards, the decisive comparison of the minutiae is carried out where all minutiae of the fingerprints are compared with each other. A decision is made based on identified pairs and their configuration. Due to variations in fingerprint qualities, not all points are always clear or defined with the same quality. In such cases, experts use their discretion and experience in deciding whether images are identical or not. For instance, ridge bifurcations could be identified as ridge endings if little pressure was exerted in taking the fingerprint. Obviously, the experience of the experts always plays a certain key role when comparing fingerprints. As an example, Figure 2.17 shows the comparison of 18 such minutiae. Federal Office for Information Security 19

20 Figure 2.17: Dactyloscopic comparison with 18 corresponding minutiae Image of the Fingerprint Identification Procedure In this section, the individual steps of the application "Fingerprint Image Recognition by Comparing Minutiae" are documented in pictures (Confer Figure 2.18). 20 Federal Office for Information Security

21 Fingeprint scanning Direction field determination Image processing Image quality improvement Thinning Feature extraction Figure 2.18: Minutiae-based fingerprint identification procedure Federal Office for Information Security 21

22 3 Evaluation of Biometric Systems 3.1 Description of the Evaluation Criteria Different types of error rates are used as metrics for the operative capability of biometric authentication systems in general and for fingerprint image recognition systems in particular. The result of a comparison in the feature matcher within a fingerprint image recognition system is called Matching Score "s". It measures the similarity between the fingerprint image and the stored template. The closer s approaches 1 (if normalized between in the range [0,1]), the more likely it is that both fingerprints originate from the same finger. On the other hand, if s is near 0, it will be quite probable that both fingerprints are from different fingers. The decision of the system is determined by threshold T, i.e. if s passed the threshold, the fingerprints are regarded as being of the same finger (Matching Pair). If s is below the threshold, the fingerprints are regarded as being different (Non-Matching Pair). In connection with this, two erroneous decisions, i.e. two kinds of mistakes, can be made by biometric systems. False Match Two fingerprint images of different fingers are categorized as being identical. False Non-Match Two fingerprints of the same finger are categorized as being different. These two mistakes are often referred to as False Acceptance and False Rejection. In order to provide a clear understanding of the different kinds of errors that can occur, they shall be defined below Types of Errors False Acceptance Rate (FAR) The FAR is the probability that a biometric system falsely recognizes different characteristics as identical, thus failing to reject, for example, a potential intruder. Definition: FAR= Number of comparisons of different fingers resulting in a match Total number of comparisons of different fingers False Rejection Rate (FRR) The False Rejection Rate (FRR) is the probability that a biometric system falsely recognizes identical characteristics as being different, thus, for example refusing to accept an authorized person. Definition: FRR= Number of comparisons of the same fingers resulting in a non-match Total number of comparisons of the same fingers False Match Rate (FMR) The False Match Rate (FMR) indicates the proportion of persons who, in the characteristics comparison, were falsely accepted. Those attempts that were previously rejected (Failure To Acquire, FTA) due to a low quality (e.g. of the image) are, in contrast to FAR, not taken into consideration. Please note that it depends on the application whether a falsely accepted characteristic contributes to increasing the FAR or FRR. 22 Federal Office for Information Security

23 False Non-Match Rate (FNMR) The False Non-Match Rate (FNMR) indicates the proportion of persons who, when comparing characteristics, were falsely not accepted. Those attempts that were previously rejected (Failure to Acquire, FTA) due to a low quality (e.g. of the image) are, in contrast to FRR, not taken into consideration. Again, it depends on the application whether a falsely non-accepted characteristic contributes to increasing the FRR or FAR. In contrast to the FAR and FRR, which are often used metrics in literature, the FMR and FNMR are calculated by the enrolled template through a number of comparisons. In contrast to it, the FAR and FRR are calculated via transactions and include, for example, the Failure to Acquire (FTA, confer below) rates as well. Figure outlines how the FMR and FNMR are calculated. The definitions of the error rates result from the probability densities for the comparison of different and identical fingerprints with regard to the threshold T: 1 = FMR ( T ) p ( s H ) ds FNMR( T ) p ( s H ) T u u with the Decision threshold T, Statement "different" H u (enrolled fingerprint and template come from different fingers), Statement "identical" H g (enrolled fingerprint and template come from the same finger), Probability density p which fulfils the hypothesis in brackets, and Matching Score s. p Different features Person Same feature Person T = 0 g g ds H u H g FNMR ( T) FMR ( T) 0 Threshold T 1 s Figure 3.1.1: FMR and FNMR Thus, the two error rates depend both on the probability densities p u (s H u ) and p g (s H g ), which characterize the system. They are a function of threshold T. Equal Error Rate (EER). The EER is defined by the condition FNMR (T) = FMR (T). In practice, the probability densities are discrete functions. Hence, the EER cannot be determined exactly. In contrast, an EER range can be established in which error rates match. As a result, if threshold T of the system is set accordingly, the same number of people are falsely accepted and falsely rejected. Furthermore, depending on the application, fixing threshold T in such a way that different error rates are generated, can be useful. Federal Office for Information Security 23

24 FMR EER FNMR Threshold T Figure 3.1.2: FAR, FRR, and EER in dependence on threshold T ZeroFNMR is the lowest limit of FMR, i.e. FNMR = 0. ZeroFMR is the lowest limit of FNMR, i.e. FMR = 0. Error FMR FNMR ZeroFNMR ZeroFMR EER 0 1 Threshold T Figure 3.1.3: ZeroFMR, ZeroFNMR, and EER The Failure To Acquire Rate (FTA) reflects the frequency at which a fingerprint image cannot be acquired by the sensor in automatic mode. This means the scanning of the fingerprint was rejected even though the finger was placed on the sensor. The higher the measured value, the less the sensor is suited for acquiring the fingerprint. In this sense, the error rate is a parameter for evaluating the sensor. The Failure To Enroll Rate (FTE) indicates the percentage of identities which cannot be enrolled by the biometric recognition system. FTE rates occur often in connection with systems which, by checking the fingerprint image quality, decide whether or not a template will be generated. This means that low quality fingerprint images will not be enrolled in the system. In this sense, FTE is a parameter evaluating the capability of the algorithm to process low quality fingerprint images. The Failure To Match Rate (FTM) indicates the percentage of enrolled fingerprints which can neither be matched nor generally processed with the stored biometric templates. This shows the incapability of the system to make a decision, i.e. unlike in case of false matches there is no result in which a wrong decision could be made. 24 Federal Office for Information Security

25 3.1.2 Objective Comparison of Fingerprint Systems Both FMR and FNMR depend on threshold T that was set and thus they are a function of T. For example, if the threshold is moved to the left (i.e. reduced), in order to increase the tolerance of the whole system, FNMR decreases and FMR rises accordingly. The system's performance in different operating points (threshold 7) can be shown in a ROC (Receiver Operating Characteristic) curve. This curve plots the FRR versus FAR thus eliminating the graph's dependence on threshold T. ROC curves are the standard approach for evaluating the performance of pattern recognition systems. These curves provide for objective comparisons in decision systems. Hence, they can be applied when comparing biometric systems in general and fingerprint recognition systems in particular. ROC curves either show the detecting rate (1 - FRR) or they show FRR as an FAR function. The graph chosen here shows FRR = f(far) FRR T falling EER FAR Figure 3.1.4: ROC -curve ROC curves offer the possibility of determining different operating points. A possible operating point is, e.g. the operating line for identical errors. The operating point of EER is determined by the (point of) intersection of the ROC curve and the straight line FRR = FAR. Linking Various Error Rates for the Purpose of Objective Comparisons If we look at a biometric recognition system from the outside as a Black Box 5, it does not matter where the FAR and FRR error rates come from. They consist of (1) errors resulting from the acquisition of images (FTA), (2) errors from enrolling fingerprints (FTE), and (3) FNMR and FMR errors resulting from the actual comparison of fingerprints: The FTA rate describes the percentage of fingerprints that could not be aquired. A higher FTA increases FRR and, on the other hand, decreases FAR. Hence, the portion of fingerprints that could be aquired is 1 FTA. ( ) The FTE rate describes the percentage of fingerprints, which could not be enrolled by their respective algorithms. Higher FTEs increase FRR and, consequently, reduce FAR. Hence, the portion of fingerprints that could be enrolled is ( 1 FTE). Consequently, this results in the following combined error rates: ( 1 FTA ) FTE : This is the proportion of fingerprints, which could be acquired but not enrolled. 5 The result is tested via the user interface without knowledge of Interna. Federal Office for Information Security 25