You are on page 1of 9


Signal Processing for Biometric Systems
(DSP Forum)

Anil K. Jain, Rama Chellappa, Stark C. Draper, Nasir Memon, P. Jonathan Phillips and
Anthony Vetro

TR2007-071 November 2007

This IEEE Signal Processing Magazine (SPM) forum discuses signal processing applications,
technologies, requirements and standardization of biometric systems. The forum members bring
their expert insights into issues such as biometric security, privacy, and multi biometric and fusion
techniques. The invited forum members are Prof. Anil K. Jain (Michigan State University),
Prof. Rama Chellappa (University of Maryland), Dr. Stark C. Draper (University of Wisconsin-
Madison), Prof. Nasir Memon (Polytechnic University), and Dr. P. Jonathan Phillips (National
Institute of Standards and Technology). The moderator of the forum is Dr. Anthony Vetro
(Mitsubishi Electric Research Labs and associate editor of SPM).

IEEE Signal Processing Magazine

This work may not be copied or reproduced in whole or in part for any commercial purpose. Permission to copy in whole or in part
without payment of fee is granted for nonprofit educational and research purposes provided that all such whole or partial copies include
the following: a notice that such copying is by permission of Mitsubishi Electric Research Laboratories, Inc.; an acknowledgment of
the authors and individual contributions to the work; and all applicable portions of the copyright notice. Copying, reproduction, or
republishing for any other purpose shall require a license with payment of fee to Mitsubishi Electric Research Laboratories, Inc. All
rights reserved.

c Mitsubishi Electric Research Laboratories, Inc., 2007
201 Broadway, Cambridge, Massachusetts 02139

MERLCoverPageSide2 .

ized or peer-to-peer applications? Are yahoo. Jonathon Phillips. For instance. Jain: In some applications. Anil P. Therefore.905886 Evaluation (ICE) 2006 show that face tion methods with biometrics-based IEEE SIGNAL PROCESSING MAGAZINE [146] NOVEMBER 2007 . to enter many metric applications will be an integral government facilities a person must show Moderator: What are the advantages of part of our everyday life (for instance. bio- nant applications continue to be in the ble of outperforming humans on recog. Stark C. and law enforcement domains.J. ment applications and for common non. In others. it is not clear if biometrics will comparable to iris recognition for verifica- discuses signal processing have a big market outside the security. In either deployed at border control points in the cryptography? Are biometrics good for case. Recent results show can/will they replace other technologies? transactions online). tity system.2007. Prof. For algorithms were used in state offices for Dr. one of the and authentication purposes. are looking for wider applications. A. I feel that biometric applications tions currently performed by humans. vidual for obtaining government future. categories. Phillips: Historically. and been used in the US-VISIT system for biometrics fit into broader technological associate editor of SPM). years are compared for face matching The invited forum members are Prof. a security guard an identification card or using biometric technologies and to access ATMs or conduct sensitive badge with a photo. applications. United Arab Emirates and has been pro. fingerprints were detecting potential abusers of welfare Wisconsin-Madison). tion applications such as access control. Since the events of biometrics will diffuse into a wider range tor of the forum is Dr. Nasir Memon. but I do not know if these appli- (Polytechnic University).C. face recognition Rama Chellappa (University of Maryland).[dsp FORUM] Anil or SPM_columns_forums@ in the United Kingdom’s national iden. 2001. low. I also recall K. fingerprints have of applications. Jain (Michigan State University). incoming visitors to the United States. it is the obtain entry access in buildings and illumination changes. we getting close to the point where bio. most well-known biometric technologies that in the mid-1990s. Rama Chellappa. privacy. such as peer-to-peer worm attacks? All seemed to mature in recent years. Jain. The forum younger generation catches on to the related to what you just mentioned members bring their expert insights into idea of biometrics. do biometrics Our readers may agree or disagree Iris recognition technology has been complement or compete with public-key with the ideas discussed next. Are ular. Draper: To understand whether Standards and Technology). technologies. Draper. benefits. Prof. and standardi. metrics can replace or supplement the area of access control (for instance. been used. to nizing unfamiliar faces across existing technology. it is useful to decide how (Mitsubishi Electric Research Labs. Jonathon Phillips (National Institute of law-enforcement applications such as S. will be employed in workplace access Results from the Face Recognition Vendor I think that ultimately the decision to Test (FRVT) 2006 and Iris Challenge add or replace existing personal recogni- Digital Object Identifier 10. and Dr. we invite you to share your com. pre- passport control)? recognition algorithms now have the venting multiple enrollments by an indi- R.K. Nasir Memon recorded and examined for law-enforce. and Anthony Vetro Signal Processing for Biometric Systems T his IEEE Signal Processing control and airport access control. Chellappa: In the immediate potential to be incorporated into applica. military. Anthony Vetro September 11. phase where we have the technology but facial images of an adult separated by ten and multibiometric and fusion techniques. Draper (University of most of the last century. that face recognition algorithms are driver licenses. face only viable approach (for instance. cations are still in effect. recognition is capable of performance Magazine (SPM) forum Overall. R. P.or high-security applications? How ments with us by e-mailing avetro@ posed for maintaining identity integrity appropriate are biometrics for central- merl. The modera. has been fingerprint recognition. I feel that we are in a involves biometrics for passport renewal: issues such as biometric security. This depends very much on whether the application of face recognition that is zation of biometric systems. systems. Chellappa: An interesting new requirements. Stark C. background checks. P. For instance. human face recognition has long are important questions to be answered.1109/MSP. and passports). or will the domi. biometric technologies more or less sus- While face recognition technologies in ceptible to important classes of attacks Moderator: Biometric technology has biometric applications have been less pop.

albeit slower than anticipated. Gait analysis may be building. signature. simultaneous palm security breaches are still possible. While government. To begin with. a similar policy needed to prove its effectiveness for per. and matching biometric data. lenges. which in turn depends systems. But.g. So I agree that it is important to based on fingerprint. on the move.g. keystroke Despite that. These concerns The main advantages of a biometric mandated biometric requirements mostly center on the security of biomet- system are as follows: 1) it offers greater increase the market and need for bio. and unchecked manner. Memon: Although biometrics pro- gram and the Hong Kong national ID and gain public support. metric solutions. it is hard to conclude which still in its infancy. For example. alarms). they do nothing to whose biometric data is captured. technologies are and the privacy impli. However.g. New Jersey. why are we con- many mandated government applica. since biometrics are usu- To do that. revoked or replaced. but rized people without using up staff time enforced such that no one should write further testing on larger data sets is to check identities. bio- find out where and how biometrics should print. palm A. Among the various biomet.K.C. I believe recognition systems received comparably less attention. ric data and the privacy of individuals security and convenience than tradi. iris. personal identi. access to Disney Parks in other methods or cannot be enforced dangers of using them in a ubiquitous Orlando). we need to understand nologies (e. the associated signal processing chal- far. multispectral and touch. less for fingerprints. Jain: Human biometric traits can can be used to track the activity and of recognition. whole hand imager that the introduction of biometrics the existing tion credentials. must be one that is important to the for authentication and/or identification.. it the user is present at the point and time A. allay widespread concerns about privacy N. hand geometry. But the point I am Moderator: Effective protection of the the personal behavior and habits of the making is that the real question may be biometric data and handling of privacy users than any weakness in the biomet- not whether they can be easily stolen. them. worth the price. Let us discuss each and outline efit analysis or return on investment. of attacks. dynamics). this has been easier to justify in S. after they were but how much we are willing to invest usability are typically listed as require. and speech have metric systems are vulnerable to a variety be used. Privacy concerns arise from the passwords) while maintaining suffi. working with which is authorized individuals into the school keyboard). with other methods. If a company policy can be effective in surveillance should be based on a cost-ben. Memon: I tend to believe that no have both anatomical and behavioral on developing techniques for capturing matter how secure biometrics can be. If we print and hand shape) will improve the Consider the iris recognition compare a system where everyone uses performance as well as usability of the Teacher-Parent Authorization Security fingerprints (which can be left on car biometric systems. what other technolo. iris. individual and either is not possible with there is widespread concern about the tions (e. Both appear to be ty issues came up that had to do with quite ridiculous. increase the acceptability of biometrics. years back to let parents and other written on a piece of paper beside the able is gait-based. tion scenario that proves the unique cerned with securing biometric systems tions in the United States and overseas capabilities of biometrics is needed. To and biometric data? (the already-mentioned US-VISIT pro. it remains compromised convenience (no need to remember face recognition. Draper: I believe that an applica. Hence. face. it is being increasingly adopt. the application vide a simple and effective mechanism card). New sensor tech. the New Egypt. Some users. palmprint) and unacceptable in most open societies. to protect them and whether they are ments for biometric technologies and held the door open for another person IEEE SIGNAL PROCESSING MAGAZINE [147] NOVEMBER 2007 . components. ed in a number of commercial applica. recognized based on their iris scans.. face. issues. Although the iris down their passwords. gait. ally a component (embedded) in a security that biometrics have some intrinsic dif. and ric systems. even with ferences from other types of identifica. tional methods of personal recognition assuage the concerns of the public or Security concerns stem from the fact [credentials—ID cards. a number of securi- should wear gloves.g. behavioral (e. users were satisfied. while they are more likely to complement exist. biometric systems worked well and for fingerprints could be that everyone son authentication.K. and pass. there has been a lot of cations of using biometrics. iris infrastructure (ID cards. fact that biometric data is so tightly ciently high accuracy and ensuring that tions to date? bound to a person’s identity. fingerprint. ear.. In fact. while keeping out unautho- is better. iris at a distance. forever. System (T-PASS) that was installed in doors) and a system where everyone R. widespread be classified into two main categories: behavior of individuals and violate their adoption of biometric technology will anatomical (e. been most effective. privacy in ways that would be entirely really depend on how secure biometric hand geometry. Voice or speech biometrics research done over the past few decades N. that biometric data cannot be easily fication numbers (PINs). Chellappa: On the other hand. ric traits that have been proposed in the security and privacy issues have ing security systems instead of replacing literature. So on the application scenario. a couple of uses strong passwords (which can be only true remote biometric that is avail.. gies are available for biometric applica. high recognition accuracy. Jain: Like any security system. captures fingerprint. biometrics cannot be forgotten. Once a biometric is words] and 2) it gives users greater Moderator: In addition to fingerprint and compromised.

transformed template. Another ers can help address this problem. finger. two problems. legislation already tion-theoretic sense. himself/herself could provide the trans- form’s parameters in terms of a password Moderator: What signal processing Moderator: What are the main risks or PIN. stolen ment based only on reported security Moderator: How would you define the biometric attacks are more severe from breaches). larity functions.S. secure sketches. Examples include fuzzy Although producing a gummy clone of Privacy Magazine. software (finger deformation) have hacker might use a very specific target This continues to be an active area of been implemented. while very unlikely. As the incidence of identi.[dsp FORUM] continued entering the building behind them (sce. instead of storing the original with security versus privacy tradeoff. assurance of solutions addressed by fuzzy hashes.K. Integrating cryptographic self-regulation. and covert recognition. Jain. techniques along with biometric match. recon. 1. may sniff it. ty fraud increases. show that the current approaches lead to ger imaging from Lumidigm. the system could store only associated with biometric systems will open a door behind them so they could its noninvertible transformed version not be the bottleneck in adopting this get back into the building easily without (for example. in the end. unlike ID less repudiable than other types of access ficiently rigorous security formulation cards and passwords. the techniques promised (e. In my opinion. Mar. are designed for attack remains quite complicated. which is related to the privacy of than real threats to system security. The types of that’s when.. a hash). independent regulatory organizations. A. many of the tech- structing a fake finger from a latent ingly come to play for positively niques proposed security in an informa- fingerprint to perform a fake biometric recognizing people. Published works based on hardware (multispectral fin- tion to the security of biometric systems. fuzzy vault. 2. vol. It has whether we will find the perfect nonin. biometrics does raise three needs before they can be safely used in security issue. a duplicate [such as a three-dimensional anonymous and to control access to your N. biometric-based accesses are proof of security. there is a lot of work that still needs to be encrypted and sometimes doubly cation scope. the template must be decrypted and use of biometrics can be addressed by S. personal information” (from A. real-world applications. Jain: In applications where rics and the risk of compromising a bio. information. U. In outside the building during their breaks biometric signal in the database during my opinion. and use simple similarity The latter refers to the case when a as biometrics to limit access to sensitive measures. rity of biometric systems and privacy of systems? the system can issue a new one using a biometric data? A. Privacy Concerns. Biometric templates are systematic privacy concerns: unintend. stolen). a latent fingerprint). to remain based system. requires strong recognition schemes such discrete data.. Multimodal bio- and present the system with a copy of a research in biometrics.g. we must come to terms problem involved some users who went example. These possible abuses or unacceptable biometric data. and pres. A likely knows an authorized user person. This requires a tradeoff between fake biometric attacks remain a serious metric template are some of the most invertibility and discriminability of the concern. “Biometric Recognition: Security & past few years for secure storage of bio- is not an easy process in practice. it is metric system would allow system admin. hash. A number of solu. and enforcement by fuzzy vaults. that would provide the guarantees one tions have been proposed to address this However. vitality detection mechanisms publicized and frequently cited in rela. If a biometric is com. they are ed functional scope. et very clever techniques proposed in the ent it to the biometric systems.” IEEE Security & metric data. nario referred to as tailgating). medical records. etc. The user technology in the long run. using the iris system again. security and privacy issues to take a walk or smoke and propped enrollment. are IEEE SIGNAL PROCESSING MAGAZINE [148] NOVEMBER 2007 . can make your life free of intrusions. and been reported that insiders commit vertible transform. these attacks. iris) have also been proposed to thwart about 80% of all cyber-crimes (an assess. Automated database continuous and requires complex simi- metric template. Still. Inc. this al. different traits (e. Memon: There have been some (3-D) mold of the fingerprint]. for matching purposes. 2003). However. never stored in their raw form.) and The former refers to the case when a some loss in the matching performance. However. unintended appli. control mechanisms. Second. pp. can acquire a sample biometric (for A. For But. real biometric data is system deals with a compromised bio. In my opinion. If a hacker ever compromises techniques are available to ensure secu- related to the security of biometric such a (transformed) biometric template.K. It is not clear metric systems that incorporate several known person’s biometric sample. 33–42. Jain: The risk of stolen biomet. ters. different transform or different parame. that are designed for real-world biometric compromised forever. But.K.g. Draper: I agree. biometrics will increas. Jain: “Privacy is the ability to lead er risk than an attack on a password- example. no.C. they suffer from one of ing user) is relatively simple. First. the individual second requirement for biometric sys.K.-Apr. a hacker government regulation.. niques. the point of user perception/acceptance breaching the system’s security very tems. Since a user has istrators to track all accesses to privileged data are either ad hoc and without formal only a limited number of biometrics. access mechanisms through a secure bio. biometric data? fake biometric attack presents a small- ally. face. done towards the security and privacy of encrypted. or do not provide a suf- they are not easy to replace. In such cases. and secure sketch tech- an available real finger (from a consent.

It’s important to understand how cryptographic hash functions to them. even in a localized area pressed. they have a long way changes are the most challenging.” These models would vary with to the “biometric channel” relating the processing challenges to achieving this? biometric modality. the full impact of this during enrollment so that they can be Good solutions cannot strengthen one approach has not been established. settings. but reason.very interesting variations on traditional R. tication). in the case of security tech- natural pair of follow-on questions is changes in the underlying biometric or niques such as fuzzy hashes or fuzzy what modalities present the highest bar. Fuzzy hashes. tion.g. and the presence of dust. between the source data and the deci. As noted by security will be far from optimum. N. around an ATM machine) that the extra sweat. etc. ment biometric) and the relationship tions. depending on the angle at which the been discussed? ceivably collect enough people’s iris finger was placed. better understanding of the Even if the signal processing chal. Opportunities might biometric system. A. from the measurement process. have been effectively handled using 3-D biometrics. recognition accuracy and attacker will be able to break in regard.C. to provide a secure end-to-end solution.. used to protect information in different sion data (the data presented at authen. to in such remote authentication scenarios be exactly reproduced. and he places it on the plate center vaults. variations due to pose formance. tions required to deal robustly with the ric and measurement processes. how hard it is S. effort needs be devoted to refining the sta- security afforded by the fuzzy hash S. recognition. The incentive to ture of the source data (say the enroll. effectively matched with partial prints link to the detriment of another. A Nasir. for multimodal for deployment. View-invariant having cooperative and habituated Different protocols and technologies signatures for gait-based identification users. will also improve the recogni- enhance security in different parts of the have received some attention. biometric matching and have great variations due to pose and illumination Refinements in either could lead to potential. sensors that capture 3-D the transmission of biometric-derived efforts are underway for superesolving fingerprints have been developed that information over the network. in a magnetic Moderator: A third requirement for bio. One may draw an analogy with raw fingerprint data. Over major improvements in system per- to go before they mature and are ready the last five years. e. fuzzy even a single bit of error. on a CD. aim to protect biometric two fingerprint scans of the same indi. If a user wants A problem posed by using biometrics ficulty is that biometric samples cannot to be recognized—for instance.K. components that underlie many signal oping codes matched to the particular er is high enough to prevent the processing techniques consist of a good source and channel under considera- development of an illegal repository of understanding of the statistical struc. oil. information. come from exploiting recent modeling one link in the chain. gerprints. an attacker could con. during As far as resolution (or lack thereof). Existing methods are effec. good models. tistical models of the underlying biomet- would be obviated. Jain: I would add that As an example. Again. Many questions In gait-based human recognition the advances in machine learning or uni- remain on how to integrate such a piece variations due to pose create even more versal compression. an variability intrinsic to biometric meas. the usefulness of that vidual will not be identical due to sen- extra protection is based on the assump. soned manner. scars. less of how the biometric is stored. the performance of biometric systems in “leakable” different biometric modalities Even the minutiae extracted will vary terms of the requirements that have are. but the tion accuracy. Similar to work on 3-D face biometric information is stored. Take the example of fin- network: in a central repository where problem is yet unsolved. with the right amount of pressure. Without tionary of collected iris information. Different classes of codes are metrics become widespread. the main dif. given a dic. The face images in a video for improved face provide 3-D models of fingerprints weakest link would be the one attacked. consider biometric tive when the human is walking in a advances in sensor technology. Furthermore. effort needs be devoted to devel- rier to collection and whether that barri. The first could be termed the recording. morphable models. The key vaults. In metric systems is that they yield high “source model” and the latter the “noise the same vein. etc. fronto-parallel direction. However. For instance. But. over a wireless channel. the sensing system and environment. Memon: In my view. such variations may result from Second. First. For example. and authentication over a public network. as well as with the biometric as measured at enrollment IEEE SIGNAL PROCESSING MAGAZINE [149] NOVEMBER 2007 . and traditional access a laptop—the user is more like- is that the data communicated over the cryptographic primitives do not allow ly to ensure that his finger is clean network is just bits. What are the signal model. Draper: I believe that progress scans (say. this strengthens only able progress has been made. sor acquisition noise. urements need more work. Draper: The algorithmic solu. Chellappa: In face recognition. That is. Variations due to source and noise models would allow us lenges can be surmounted to produce a illumination changes have not been to fuse likelihood ratios in a more rea- provably secure and operationally robust addressed as effectively yet. during authentication. challenges. requires a two-pronged approach. we need codes matched recognition accuracy. etc. recognition. So one cannot Moderator: How would you summarize tion that the raw biometric itself is a apply digital signatures or traditional the main directions of R&D to improve secret.C. storage and communication applica- compile such databases will grow as bio.

One of the problems that provide good-quality fingerprint images associated with fusion techniques for bio. and I one method of increasing security.. ric for more than one trait. One can also design a hybrid multibio. finger. views) to account for intra-user varia. These sources are: (for templates). Chellappa: Although biometric 1) to improve the matching accuracy. Jain: There are three main rea. and processing time need a single type of biometric. A more principled biometric data can be sniffed by a worm. Unfortunately. I believe that fusion the population coverage. Many sources of information can be critical as it decreases the throughput of tical models and the corresponding considered in designing a multibiometric the system). systems with minimum increase in from still images. despite the possible ric datasets. visual and infrared face cameras). metric domain? of multibiometrics is still a largely unex- some individuals who may not be able to A. have proven to be useful for many fusion techniques. 2) system cost for authentica- matched codes will vary from one type system. reduce false accept rates (FARs) and cessing. more work is needed in the area (feature extraction and matching) small—this limits the evaluation of vari- of multifactor (multimodal) biometrics algorithms. and semantic retrieval. A large number of ad hoc techniques S. Using multimodal biometrics is identity. ciently protected. Draper: I agree. and fusion methods in partic. A further challenge for the detailed reasons for using a multibio. data mining. 3) while there are large from a minutiae-based feature set of a is imaged using multiple sensors (e. R.. face and P. and distributed and embed. vidual (the verification time is particularly One should note that both the statis.. 2006 and ICE 2006 measures perform- improve the accuracy of biometric sys. the trouble with trait (e. Memon: This is also a nice way to structure for determining which fusion include sensor design and acquisition combine the strengths of biometrics and methods are appropriate for different algorithms. the biometric research communities is metric system and what are the data Moderator: Multibiometric methods in to develop experimental and data collec- sources of such a system? general.e. Jain: Some of the major problems plored area. munications. such FRVT and ICE aim to help answer ded computing. The chal. Phillips: Anil did an excellent job ernment secrets with single-modal bio. vidual biometric traits (face. What are inconvenience. and 3) metrics are as follows: 1) collection from for the correlation among the different to minimize incidence of spoof attacks. one should not protect important gov. ■ Multimodal: Evidence provided by tion and evidence weighting. it multiple sources of information takes single biometric algorithm. tor applications. Fusion of IEEE SIGNAL PROCESSING MAGAZINE [150] NOVEMBER 2007 . public domain databases available for indi- finger to a Fourier-based feature set. and ease of sales and sensor cost. have been tried. two face approach using likelihood ratio is being fingerprints can be lifted off a car door. still needs to be addressed is accounting can be identified using their iris. disadvantage in terms of cost of the sys. What for fingerprint and faces. Each has its own advantage and tion in commercial applications of bio- of biometric to another. recommended as the optimal approach and iris data can be extracted from tions in the trait. lenge here is how to combine different ance of iris recognition.[dsp FORUM] continued to the biometric as measured at would be difficult to present fake biomet. P. other security schemes while perhaps applications. 4) which fusion since single-modality biometrics ■ Multiinstance: This uses multiple technique is the best has yet to be decid- schemes are limited. Phillips: Multibiometrics and metric system by integrating a subset Numerous fusion methods exist in the improved techniques for matching bio. for example.g. ■ Multisensor: A single biometric trait to be considered. finger) are used for establishing of articulating the potential of multibio- metrics. human-computer inter. instances or units of the same body ed. The combination of FRVT Moderator: Among the solutions to increase in the inconvenience. iris). image pro.. amount of storage further vary between representations of use. the overall system can this question using specific multibiomet- be quite secure. and N.K. faces and gait. Other areas of importance N. and audio and faces. for instance. sons for using multibiometric systems: application domains including com. that avoids the need for score normaliza- high-resolution photographs. Memon: On the other hand. and it is generally agreed single-modality biometric systems is ■ Multisample: This uses a single that the simple score level fusion works that many biometrics lack secrecy. the same biometric (e. fusion techniques have been developed i. does not exist the experimental infra- to pursue. time to enroll and verify. metric and fusion techniques. false reject rates (FRRs). in ■ Multialgorithm: The same biomet. They will tem.g. If each factor is suffi.J.J. for example. of the five scenarios mentioned above biometric literature.C. biometrics in multibiometric challenges encryption. Hence. the size of public domain multibio- addition to understanding data and noise ric data is processed using different metric databases available is rather models.K. Raw sensor to acquire multiple samples of well most of the time. left and right irises). more time to enroll and to verify an indi- authentication. different body traits (e. ous fusion algorithms. face recognition tems you have mentioned multibiomet.g. secure communications and keeping the costs low by using multifac. emphasize the word potential. from 3-D scans. there metric samples are interesting avenues via fusion.g. 2) to increase are the specific problems in the bio. Comparisons of different faces.. tion protocols for further assessing A. and face recognition ric (or multifactor) techniques. metrics such as verification at point of from iris to fingerprint to face. ular.

and also ing better than humans for recognition bits would be on the attackers who try engender much discussion at committee of frontal faces across illumination to guess the key. If retransmissions are inde. commercial. secrecy can be ensured by keeping the not a good measure for biometric sys. Biometric Profiles gerprint sensor can affect performance. We can expect real progress in the need to be addressed in this area for next. the probability that all trans. liveliness testing may recognition performance of real biomet.. experimental results show that correlations that are hard to understand Reporting (M1. biometric acquisition system is to reli- processing contexts. for biometrics to play across a range of either machine or human. In are examples of topics that are active Moderator: Usability is important for bio- communications.C. such touched on in this forum are AHGEMS human operator’s decision-making. with the same latter has produced a report (available on candidate face images. the bur. that computers are capable of perform- den to analyze the correlation of the rics and E-authentication. indication of next-generation issues. and 5) whether S. INCITS-M1. both biometric methods interoperability between systems by speci. as well as its list of participants. However. in face recognition. biometrics. ic organizations. usually the best practice current groups whose foci have been in a manner that helps facilitate a is to choose the bits uniformly. In such cases. Testing the effi- entropy. U. 2002. 3) quantifying the inher- the data is good.J. we can actually create much the INCITS-M1 Web site) detailing many cacy of this method requires a compari- longer keys. understanding of how the data is corre. the much more time compare with the uni. IEEE SIGNAL PROCESSING MAGAZINE [151] NOVEMBER 2007 . where the bits are correlat. Memon: Correlations among bio. INCITS-M1 members include biometric systems can be divided into two but fails almost always for the most chal. Two metric matching needs to be presented graphic cipher. Performance Testing and Also. instructions on a poster. metric systems to gain user acceptance. On one hand. The design of the sensor. positions for international stan. such a fingerprint or metric authentication is much more to provide comprehensive and well-found. not control the sensor and so the sensor rules are used. a system with the same key length.S. What we are seeing is that.g. A sampling of issues performance. figuration of the sensor. Draper: The main domestic bio.6). even if the entropy of security community that there is a role yield performance superior to that of the key is actually the same. The height of the fin- with better performance and/or security.correlated classifiers is a difficult prob. Its task groups include Technical the sensor may use a single finger or lated would allow us to design systems Interfaces (M1. Multibiometric than the biometric data. generation biometric standards? not be possible as the authenticator may ric systems when more sophisticated P. I agree that the 1/SC-37. The goal of a important than in many other signal ed U. INCITS-M1 serves as the ably acquire a high-quality biometric N. Phillips: The biometric standard. face image. will not be acceptable. On the tee. demonstrate the growing importance of P.4). scale biometric systems to areas that are (e. con- metric data are indeed a difficult issue to international organization ISO/IEC JTC. What efforts are needed pendent. applications? other hand. cases: acquisition and decision-making. The could present an operator with a list of can also argue that. when biometric standards community to ensure requirements. include: 1) methods to ensure changes (humans are asked to recog- force) attacker would have to spend the privacy of biometric data. dards activities. in a biometric system. Let us consider a simple example. to increase the user acceptance of bio- mission attempts fail drops exponentially The timing and activities of this commit. if a packet research areas. and when challenging fying common data exchange formats and its) of biometric data. Recent results have found hard to exploit. ent entropy (and therefore security lim- perform well. a sys. context were. The ad hoc groups within M1 give an In the second case. that are raised in the report on biomet.3). for example. Technical Advisory Group for the sample. and a naive (brute. transmission fails. and Societal Issues (M1. metric standardization committee is considerations. issues of security and privacy in net. Phillips: Human interaction with tem that works well for the average user. a user provides a Robust worst-case performance in bio. lenging users.g.. biometrics. and user inter- deal with. metric systems for a broader range of with the number of attempts. son between operator and computer ed in some complex manner that is worked contexts. meetings. and academ. which was established in June face type are important. Correctly. e. ization activities range from the nuts can be fed synthetic data.C. In the first case. to a system. Formats (M1. but it differentiates them systems and biometric quality metrics somewhat from other applications. one can always S. What critical issues requirements of remote authentication fail. and there seem to exist promising design retransmit. Moderator: There is ongoing work in the applications with differing security lem. 4) the additional examples are presented. governmental.2). Draper: Average performance is and bolts of interoperability of large. This is not necessarily unique to active research topics. both methods application interfaces.J. matching) algorithm secret rather tems. we (Biometrics and E-Authentication). One the other hand.5) and Cross Jurisdictional video instructions are better than can also work to our advantage. I tend to believe that (M1. demonstrate to the cryptography-focused human and machine recognition results form-key case. founded in November 2001. For that the entropy of the key is maximum (multibiometric systems) and AHGBEA example. the result of bio- When we choose a key for a crypto. A prime objective of the committee is biometric sample. For instance. 2) how to nize unfamiliar faces). Data Interchange four-finger slap.S.

security. interests include signal processing. activity modeling and recognition. that be? Engineering. He is a Senior an extensive business contact list. can recognition potential provided by biomet. College Park. com- improve performance for users with rics. However. “Bootstrap white Medical Image Computing and Computer-Assisted 20th ed edition. 965–978. P. Jonathon Phillips (jonathon@nist. 2006. Björnemo. facial images or fingerprints team. Draper: In my view. M.g. also be impacted by the selection of algo. Online: http://bartleby. 2006.” tial importance sampling. M. pp. mind. New York: Springer-Verlag. Doucet. Rennes–Saint Malo. stream- more challenging biometrics (e.J. munications. the success and acceptance of Nasir Memon (memon@poly. pp. being able to seg. He is a is with Mitsubishi Electric Research Web to a personal biometric information Fellow of the IEEE. tors. professor of electrical and interests include information coding. “White matter Westin. Brun.E. [11] M. tography. 1918. media streaming. and Statistics and Probability His current research interests include development of personal biometric departments. no. computer and network security. biometric security. O’Donnell. information whose finger prints do not register well crime. keyless access to office rooms. Behrens. Conf.. Draper (sdraper@ece.. and that do not threaten basic Brooklyn. M. 7th Int. evaluation methodologies. probe data (i. and eliminating fraud. 3-D modeling. queuing. Bartleby. 8th Int. 24. usable. cars. including Handbook Anthony Vetro (avetro@merl. and other devices. Brun. O’Donnell and C. [10] T. H. using normalized cuts. to adjust car seats. “Characterization and Dreusicke. research interests are face and gait recogni- most affected by sensor design and can biometric systems may be employed for tion. is a sion. no. images. vol. 140–147. professor in the Computer Science databases for a match with an unlabeled ty to create systems that are cost effective. [12] O. Jain (jain@cse. Oct. usability is R. Cambridge. Johansen-Berg.” IEEE Trans. sensors) could be sent via the mobile of Biometrics. positions of mirrors. [SP] IEEE SIGNAL PROCESSING MAGAZINE [152] NOVEMBER 2007 . 8.-F.. Medical Image Computing and Meeting (ISMRM’02). e. He is the author of a number of Senior Member of the IEEE. Clare. ISMRM Annu. Shenton. 2005. G. R. Lazar and A. J. “Clustering fiber tracts Amer. W. ment and extract features of an iris at used to personalize settings in a given anomaly detection in still and video long range (one to a few meters) and in space.” in Proc. Anatomy of the Human Body. Electrical and Computer gov) is an electronic engineer with NIST. Honolulu. N. R. 2005. M. Westin.M. what would ments in the Computer Science and P.-F. securing our theory.-F. 1077–1088. He is a on many devices).J.e.K. markerless motion capture for analysis of rithms.” NeuroImage. where system’s provider for matching against a Rama Chellappa (rama@cfar. tempera. and N. comes next for biometrics. M. 2004. and an affiliate pro. in Proc. “White matter tractography using sequen. [16] L. and multidimension- provide a capability to identify people on fessor of computer science at University of al signal processing. and video indexing real time would improve the acceptability ture control. 1032–1036. etc. Gray. Phillips: I would mention the Engineering.-F. pp. Maryland. May 2002. Computer-Assisted Intervention (MICCAI’05). intelligent search techniques can yield same time.C. S. and This would computer engineering. “A [15] L. Farneback. director of the Information Systems and developments would increase the accept. Knutsson. He is a Fellow of the IEEE. method for clustering white matter fiber tracts. vol. Great Books matter tractography (boot-trac). [9] A. Westin. Kikinis.[dsp FORUM] continued S. 2003. Woolrich. and retrieval.” MR imaging. estimation. and C. any system for reliable person is with the Department of Electrical and Improved matching algorithms can recognition must contain a biometric Computer Engineering at the University lead to better overall matching accuracy component. of iris biometrics. CA. Neuroradiology. Westin. Maryland. 27. Palm [13] A. often more significantly. and C. vol.E. vol. 2. P.-F. They may also be movement disorders. bio- information systems (PBIS) for mobile Academies panel on biometrics and is a metrics. [SP] [life SCIENCES] continued from page 126 [8] H. 5. Alexander. 524–532.L. he is a group manager responsible for personal biometric database. Bayesian approach for stochastic white matter trac. 50. Chellappa: My thought is that next. Med. Such rights to privacy and anonymity. Springs. and S. Shenton. and networking. Massachusetts. to the originating cell phone. Smith. Grimson. identifying criminals. Department at Polytechnic University. Sept. Kubicki. where he is the ciated user identification number).msu.umd. The results edu) is a Minta Martin Professor of multimedia technology. “A propagation of uncertainty in diffusion-weighted [14] A. homes. M. those vide useful societal value by deterring ing media. H. For example. He is a Web PBIS. Nunes.E. face recognition. no.html. Sequential Monte Carlo Methods in Practice. Because of the unique person of Wisconsin in Madison. In a mobile member of biometrics defense support computational psychophysics. tract clustering and correspondence in is a massive speed-ups when searching large their deployment will depend on our abili. and multimedia data one last thought or outlook on what Michigan State University with appoint. edi. acquired by a cell phone (using included books on biometrics. Park. His Member of the IEEE. and Web-enabled cell phones. New York. His ability of a biometric ments would also raise privacy concerns. Moderator: If you were to summarize University Distinguished Professor at digital forensics. PANELISTS research interests include data compres- Anil K.” Magnetic Resonance Med. Jain: With a wider perspective in Stark C. His research of the search could then be transmitted Engineering. Internet Security (ISIS) Laboratory.wisc. Imaging. Matthews. 2007. J. and C. Intervention (MICCAI’04). pp. He serves on the National computer vision. Springer. a probe without an asso. Conf. they have and will continue to pro.” in Proc. Westin. 2001. Jenkinson. pp. such develop. Faster algorithms and borders. 368–375. Brady. France. Gordon.g. Labs. pp. de Freitas. He is a Member of the IEEE. At the Member of the IEEE. A. Friman. M. His research and. multisensor fusion. Hawaii.