This content has subsequently be published as a white paper on the "European Biometrics Portal"
Biometrics is the science and technology of measuring and statistically analyzing biological data. In information technology, biometrics usually refers to automated technologies for measuring and analyzing an individual’s physical and behavioural characteristics; such as fingerprints, irises, voice patterns, facial patterns and gait. The analysis of such data is then used for identification or verification purposes, depending on need. Multi-modal biometrics, or biometric fusion, is the process of combining information from multiple biometric readings, either before, during or after a decision has been made regarding identification or authentication from a single biometric.
Before considering multi-modal biometrics it is important to understand the core features of a conventional (uni-modal) biometric system. Such a system can be decomposed into four components:
The ability of the system to perform well (within the limits of its design) is based almost solely upon the quality of the biometric captured. A well captured biometric is rich in distinguishing information, which in turn gives the feature extraction algorithms the best chance of finding a match with existing records.
However, the ability of the system to capture high quality biometric samples is reduced by many factors. Dirty fingerprint sensors, photographic light intensity or a voice altered by a cold, may all reduce quality of a reading to the extent that multiple readings of the same biometric (e.g., the same person’s index finger) can produce a range of widely-differing samples. Add to this the fact that a minority of individuals may not be able to provide a given biometrics (e.g., the fingerprints of manual workers are often too degraded to be captured well), and a conventional biometric system may soon become a burden on its operators.
The accuracy of a biometric system can be measured from two statistical measurements:
The FAR and the FRR are inversely proportional, forcing a trade-off between security and convenience when using biometric systems. A system that is easy to use, meaning convenient for both user and system administrator, may allow unauthorised access to a secure area, whilst a highly secure system may require continual human intervention to allow access to authorised users not recognised by the device. Multi-modal biometric systems may be able to improve this trade off.
Multi-modal biometric systems capture two or more biometric samples and use fusion to combine their analyses to produce a (hopefully) better match decision by simultaneously decreasing the FAR and FRR.
Multi-modal biometrics systems can be designed to work in five ways:<
A combination of uncorrelated modalities (e.g., fingerprint and face or two fingers) is usually expected to result in a better performance than a combination of correlated modalities [Jain], (e.g., multiple captures of the same finger or multiple matching algorithms).
Fusion may take place at any point during the processing, giving three possible fusion methodologies:
The data captured from each sensor is used to create a feature vector. In the case of fingerprints, the vector might include three dimensions: the location, shape and orientation of each minutia. In the ideal case, this vector uniquely identifies a given person in feature space; more likely is that the vector identifies a subset of the enrolled population. Combining the feature vector from each biometric creates a vector that has a higher dimensionality and higher probability of uniquely identifying a person in feature space.
Once the feature vectors from each biometric mode have been constructed, they are passed to their individual matching algorithms, which attempt to match them against previously captured templates. The individual matching probabilities are then combined (a variety of methods may be used, e.g., Neural Networks, Discriminate Analysis or Weighted Sums) to form a result from which a decision may be made.
Strictly speaking, Probabilistic-Decision level fusion is not ‘truly’ fusion, although it is generally termed as such. The principle behind this method is that each biometric system makes a match based solely upon the biometric it captures and then passes a binary “match/no-match” vector to a decision module, which forms a conclusion based on a majority vote scheme. Some systems also include a method to weight the decision towards more highly regarded biometrics (e.g., iris over retina scans).
Most observers accept that fusion produces better results when performed at the Feature-Extraction level rather than at the data matching or decision levels (see for example [Delac] & [Ross]), because the data is combined at its most information-rich stage and before external contamination from, e.g., matching algorithms. However, feature level fusion is difficult, as the relationship between individual biometrics’ feature spaces may not be straightforward, e.g., how are data from fingerprints and irises combined? Additionally, the higher the dimensionality of the feature space, the longer the analysis computation time, therefore a trade-off between speed and accuracy may need to be made at the application level. [Jain] explains that modality independence is important when improving biometric systems through fusion, e.g., uncorrelated biometrics, such as fingerprint and iris, are expected to result in a higher improvement rate than correlated biometrics, such as two independent readings of the same fingerprint.
The problems associated with feature-level fusion are noted in the literature, with most fusion research being carried out in the data matching and probabilistic decision level fusion areas (see for example [Hong]). However, it should be noted that [Daugman] presents arguments that, correctly, dismiss decision-level fusion and concludes that a strong biometric is better alone than in combination, at the decision level with a weaker one.
Figure 3 shows an improvement in the results of a biometrics system achieved through the use of a multi-modal fusion biometric system. The system used to produce these results fused the data at the data matching stage. The graph shows the improvement of using a fusion system over a single biometric, with the fused data improving the best single biometric, fingerprint in this case, by 20%. Usually, an improvement in false acceptance rate is obtained at the cost of lowering the genuine acceptance rate, but it can be seen that this is not the case here, with the fused system matching the best FAR of the individual biometric systems.
Many data matching fusing methods exist, three of the most popular being: The ‘Sum Rule’, Decision Tree Determination and Linear Discriminate Analysis
Where: 
and 
are the scores for the fingerprint and iris biometrics respectively; and 
is a weight that can be computed using training data or determined dependent upon the quality of captured data. To ensure the combined score is meaningful the component scores must be normalised.
In its simplest form a decision tree takes the form of a pyramidal flow chart representing a predictive or deterministic model, i.e., a mapping of observational values to conclusions about the observed item's target value. Therefore in the case of a biometric system the captured biometric is compared against a stored template and is accepted as genuine if its determined value is above a set threshold, or target value. Both the threshold and decision tree structure maybe determined from a training data set.
Linear Discriminate Analysis (LDA) is used to determine, from a set of vectors, a binary result. Thus LDA aims to transform the original matching score space to a lower dimensional space, maximising the separation between the genuine acceptance rates and false acceptance rate classes with as little loss in discrimination as possible.
[Ross] have investigated each of these methods for combining 3 biometrics; facial images, hand geometry and finger prints. They find that the ‘Sum Rule’ method returns the best improvement over uni-modal systems. Their results, approximately 15% improvement in genuine acceptance rates and a 10% reduction in FAR, show that significant improvements can be made over uni-modal biometric systems.
Whilst the uni-modal biometrics systems used by Jain did not produce high ‘genuine accept rates’, ([Mansfield] produce similar genuine acceptance rates as those achieved after [Jain]’s Fusion procedure), the process of multi-modal fusion is demonstrated to be successful.
However, whilst using multiple biometrics produces advantages, it also introduces disadvantages, amongst which are:
| Advatntages | Disadvantages |
|---|---|
| Improved population coverage is achieved by reducing the failure to enrol rate; | Multiple biometrics means multiple enrolment difficulties; |
| Lower False Acceptance Rates (FAR) and False Reject Rates (FRR) can be achieved through the combination; | The process of determining the correct feature vector combinations required to fuse, e.g., iris and fingerprint data, is not simplistic and adds to a systems R&D overheads; |
| The ability to spoof the system is reduced as replicating multiple biometrics traits is more difficult than replicating one; | System processing times are increased due to the complex computations required; |
| A challenge-response authentication procedure may be facilitated that increases the probability that a live user is interfacing with the system; users could be challenged to enter a random and different sequence of biometrics each time. |
A key feature of fusion technology is the ability to choose the degree of reliance on each of the biometric traits in order to maintain an acceptable balance between confidence and convenience for its users and operators. An additional benefit is that the weighting can be applied at population level, an individual level or indeed anywhere in between. [Ross] suggest that user specific weighting could enhance fusion further; The weights, used to associate different emphasises to the different biometrics, could be used to ensure that a user with poor fingerprints had more importance placed on the other biometrics in a given system, increasing the likelihood of genuine acceptance. The biometric ‘weightings’ may be hard coded, e.g., an operator may more heavily weight the facial recognition component of a particular subset of workers who perform manual labour and therefore have poorly recognisable fingerprints; or organically learnt over time by examining the stored template of the user, e.g., a system that required consistent clarification over the identity of an individual, perhaps due to an eigenface3 that is easily matched to a large number of the enrolled population, may be set to learn that the higher score achieved from the voice print is to be more heavily weighted, removing the burden of intervention from its operator.
No Biometric is truly universal; the National Institute of Standards and Technology (NIST) report that fingerprints cannot be obtained from 2% of the American population [NIST], whilst [Golfarelli] show that hand geometry and Facial biometrics have a limited number of distinguishable patterns, 105 and 103 respectively, indicating that they have a limited ‘uniqueness’ factor. Whilst Multi-Modal biometrics systems overcome these problems there are two overall limiting factors that affect them:
The long term vision for multi-modal biometrics is to take capture units from different vendors and manufacturers and combine them in a simplistic manner. However, there are many barriers to this ideal. Primarily, the combination of individual biometric vectors into a coherent multi-modal vector is not a simplistic task. It is dependent on, the oft proprietary, extraction algorithms. In addition to this, there is, currently, no standard biometric data exchange format that manufacturers can reliably assume they will be given data in, although this is under address by, amongst others, the American ‘M1 biometrics’ body.
Biometrics is fashionable; and in the news, because of the national identity card debates underway in some countries, and in people’s pockets because of the transition to electronic passports around the world. Biometrics has been a favourite in the movies for many years and you can even go to the Microsoft web site and buy a keyboard with a fingerprint reader built-in.
So biometrics is here and here to stay; and it is of course the solution to all of our problems: crime, terrorism, housing benefit fraud and the hole in the ozone layer. Well, no. The real world is of course more complicated (and more interesting). To understand how to develop strategies to make the most of the rapidly advancing technology of biometrics, biometrics must be viewed in the right context.
Biometrics is set to make major improvements to as many areas that require secure identification as can be imagined; fingerprint entry to homes and cars, travel and maternity units, for example. It is the single most convenient means of identification that we all possess and cannot leave behind or forget. However, biometric technology has a long and rough journey ahead of it to gain acceptance. It is not a panacea, as is often portrayed in movies and any high profile mistakes are likely to be quickly vilified by those who oppose its use and the propagation of what is seen as a “1984” Orwellian state.
Fused Biometrics has the possibility to make identification more secure and more accurate than single biometric systems. How would one replace or overcome the problem of a compromised fingerprint? Biometric Fusion could help – by not relying on a single biometric, as could weighting – by more heavily relying on those biometrics not compromised. Building such systems is not, however, straightforward. How does one combine the multi-dimensional feature vectors of iris, fingerprint and facial recognition systems?
It is more likely that a hybrid system of traditional digital security and biometrics (uni-modal in the short term, progressing to multi-modal in the future) will prevail. Digitally signed biometric tokens and/or databases, crossed checked with the user require a fraudster to not only spoof a biometric but also break the encryption or tamper-resistant enclosure used to store the biometric template.
So is fusion required if current security techniques and uni-modal biometrics can overcome security barriers? Simply put, yes. Even very good identification biometrics are imperfect at large scale. A good system will deliver a “false reject” rate of 0.4% - though this is enough to annoy a regular user, or more importantly, to encourage customers to flood call centers with complaints. Biometric fusion is slowly but surely showing that these FRR can be driven down; providing, in turn, a more user friendly experience.
Whilst biometrics will endure and improve verification procedures, a thought should be spared for a world where other areas of a security system, rather than the identification process, become the weak link and, therefore, are more likely to become the target of attack. One consequence to be considered is that since biometrics is the one ‘verification’ method that cannot be forgotten or left behind and multi-modal biometric systems are decreasing the likelihood of spoofing, the individual becomes the weakest link. Whilst biometrics are touted as being able to vastly increase authentication procedures, e.g., entry to secure areas, will they correspondingly increase attacks on ones person, e.g., kidnapping to enable forceful use of ones iris pattern?
Biometrics will succeed only if:
| [ARAJ] | Arun Ross and Anil Jain “Information Fusion in Biometrics”, Dept Computer Science, Michigan State University, 2002. |
| [Daugman] | John Daugman, Combining Multiple Biometrics, The Computer Laboratory, Cambridge University, Daugman, 2004. |
| [Delac] | Delac and Grgic, “A Survey of Biometric Recognition Methods”, 46th International Symposium Electronics in Maine ELMAR-2004. |
| [Golfarelli] | Golfarelli, Maio and Maltoni, “On the Error-Reject Trade-Off in Biometric Verification Systems”, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 19, No. 7 July 1997. |
| [Hong] | Hong et al., “Can Biometrics Improve performance?”, Proc. AutoID’99 Summit, NJ, p59-64. |
| [IAfB] | International Association for Biometrics (IAfB) and International Computer Security Association (ICSA), 1999 Glossary of Biometric Terms. |
| [Jain] | Jain, Ross and Prabhakar, “An Introduction to Biometric Recognition”, IEEE Transactions on Circuits and Systems for Video Technology, Vol. 14, No. 1 January 2004. |
| [Mansfield] | Mansfield et al., “Biometric Product testing Final Report”, UK Biometrics Working Group, 2001. |
| [NIST] | Report to US Congress “Summary Of NIST Standards For Biometric Accuracy, Tamper Resistance And Interoperability”, Nov., 2002. |
| [Ross] | Ross and Jain, “Information Fusion in Biometrics”, Pattern Recognition letters 24 p2115-2125, Sep. 2003. |
Print
Email
News!
