Genetic programming for multibiometrics

Abstract

Biometric systems suffer from some drawbacks: a biometric system can provide in general good performances except with some individuals as its performance depends highly on the quality of the capture… One solution to solve some of these problems is to use multibiometrics where different biometric systems are combined together (multiple captures of the same biometric modality, multiple feature extraction algorithms, multiple biometric modalities…). In this paper, we are interested in score level fusion functions application (i.e., we use a multibiometric authentication scheme which accept or deny the claimant for using an application). In the state of the art, the weighted sum of scores (which is a linear classifier) and the use of an SVM (which is a non linear classifier) provided by different biometric systems provide one of the best performances. We present a new method based on the use of genetic programming giving similar or better performances (depending on the complexity of the database). We derive a score fusion function by assembling some classical primitives functions (+, ∗, −, … ). We have validated the proposed method on three significant biometric benchmark datasets from the state of the art.

Introduction

Every day, new evolutions are brought in the biometric field of research. These evolutions include the proposition of new algorithms with better performances, new approaches (cancelable biometrics, soft biometrics, …) and even new biometric modalities (like finger knuckle recognition (Kumar & Zhou, 2009), for example). There are many different biometric modalites, each classified among three main families (even if we can find a more precise topology in the literature):

• biological: recognition based on the analysis of biological data linked to an individual (e.g., DNA analysis Hashiyada, 2004, the odor Korotkaya, 2003, the analysis of the blood of different physiological signals, as well as heart beat or EEG Riera, Soria-Frisch, Caparrini, Grau, & Ruffini, 2008);

• behavioural: based on the analysis of an individual behaviour while he is performing a specific task (e.g., keystroke dynamics Gaines, Lisowski, Press, & Shapiro, 1980, online handwritten signature Fierrez & Ortega-Garcia, 2008, the way of using the mouse of the computer Weiss, Ramapanicker, Pranav, Noble, & Immohr, 2007, voice recognition Petrovska-Delacretaz, El Hannani, & Chollet, 2007, gait dynamics (way of walking) Nandini & Kumar, 2008 or way of driving Benli, Duzagac, & Eskil, 2008);

• morphological: based on the recognition of different particular physical patterns, which are, for most people, permanent and unique (e.g., face recognition Turk & Pentland, 1991, fingerprint recognition Maltoni, Jain, & Prabhakar, 2009, hand shape recognition Kumar & Zhang, 2006, or blood vessel Xu, Guo, Hu, & Cheng, 2005, …).

Nevertheless, there will always be some users for which a biometric modality (or method applied to this modality) gives bad results, whereas, they are better in average. These low performances can be implied by different facts: the quality of the capture, the instant of acquisition and the individual itself but they have the same implication (impostors can be accepted or user need to authenticate themselves several times on the system before being accepted). Multibiometrics allow to solve this problem while obtaining better performances (i.e., better security by accepting less impostors and better user acceptance by rejecting less genuine users) and by expecting that errors of the different modalities are not correlated. In this paper, we propose a generic approach for multibiometric systems.

We can find different types of biometric multimodalites (Ross, Nandakumar, & Jain, 2006). They use:

• different sensors of the same biometric modality (i.e., capacitive or resistive sensors for fingerprint acquisition);

• several different representations for the same capture (i.e., use of points of interest or texture for face or fingerprint recognition);

• different biometric modalities (i.e., face and fingerprint recognition);

• different instances of the same modality (i.e., left and right eye for iris recognition);

• multiple captures (i.e., 25 images per second in a video used for face recognition);

• an hybrid system composed of the association of the previous ones.

We are interested in the first four cases in this paper. Our objective is to automatically generate fusion functions which combine the scores provided by different biometric systems in order to obtain the most efficient multibiometrics authentication scheme.

In order to compare different multibiometrics systems, we need to present the how to evaluate them. Several works have already done on the evaluation of biometric systems (Theofanos et al., 2008, ISO, 2006). Evaluation is generally realized within three aspects:

• performance: it has for objective to measure various statistical criteria on the performance of the system (Capacity Bhatnagar & Kumar, 2009, EER, Failure To Enroll (FTE), Failure To Acquire (FTA), computation time, ROC curves, etc. ISO, 2006);

• acceptability: it gives some information on the individuals’ perception, opinions and acceptance regarding the system;

• security: it quantifies how well a biometric system (algorithms and devices) can resist to several types of logical and physical attacks such as Denial of Service (DoS) attack.

In this paper, we are only interested in performance evaluation (because the fusion approach is not modality dependant and perception and security depend on the used modalities). The main performance metrics are the following ones:

• FAR (False Acceptance Rate) which represents the ratio of impostors accepted by the system;

• FRR (False Rejection Rate) which represents the ratio of genuine users rejected by the system;

• EER (Error Equal Rate) which is the error rate when the system is configured in order to obtain a FAR equal to the FRR;

• ROC (Receiver Operating Characteristic) curve which plots the FRR depending on the FAR and gives an overall overview of system performance;

• AUC (Area Under the Curve) which gives the area under the ROC curve. In our case, smaller is better. It is a way to globally compare performance of different biometric systems.

We can also present the HTER (Half Total Error Rate) which is the mean between the FAR and FRR for a given threshold (this error rate is interesting when we cannot get the EER).

There are several studies on multibiometrics. The fusion can be operated on different points of the mechanism:

• template fusion: the templates captured by different biometric systems are merged together, then the learning process is realized on these new templates (Raghavendra et al., 2009, Rattani and Tistarelli, 2009). Fig. 1(a) presents this type of fusion. The fusion process is related to a feature selection in order to determine the most significant patterns to minimize errors.

• decision fusion: the decision is taken for each of the biometric authentication system, then the final decision is done by fusing the previous ones (Ross & Jain, 2004).

• rank fusion: the decision is done with the help of different ranks of biometric identification systems. The main method is the majority vote (Zuev & Ivanov, 1999).

• score fusion: the fusion is realized considering the output of the classifiers. The Fig. 1(b) presents this type of fusion.

Buyssens, Revenu, and Lepetit (2009) showed the interest of biometric fusion for face recognition combining the image in visible and infrared color spaces with convolutional neural networks. In Montalvao Filho and Freire (2006), Mantalvao and Freire have combined keystroke dynamics with voice recognition, it seems it is the first time that multibiometrics has been done with keystroke dynamics and another biometric modality. In Hocquet (2007), Hocquet et al. demonstrated the interest of fusion in keystroke dynamics in order to improve the recognition rates: three different keystroke dynamics functions are used on the same capture. The sum operator (consisting in summing the different scores) seems to be the most powerful approach in the literature.

These fusion architectures are quite simple but powerful. Results can yet be improved (in term of error rate or computation time) by using different architectures. A cascade fusion (Allano, 2009) is another interesting approach. A first test is done, if the user is correctly verified as the attended client or if it is detected as an impostor, the algorithm stops. Otherwise, another biometric authentication (with another capture from another modality) is proceeded until obtaining a decision of acceptance or rejection, or reaching the end of the cascade. So, instead of using one decision threshold, each test (except the last one) needs two thresholds: one for rejection and one for acceptance. All scores between these thresholds are considered in an indecision zone. This mechanism is presented in Fig. 1(c). Another advantage of this method is to decrease the verification time by not using all the modalities, they are used only if necessary. This method has been successfully applied on a multibiometric system using face and fingerprint recognition in a mobile environment (where acquisition and computation times are important) (Allano, 2009).

Another kind of architecture has been proposed: it is a hierarchical fusion scheme (Teh, Teoh, Tee, & Ong, 2009) (called multiple layers by their authors). Shen et al. have presented this method with two different keystroke dynamics methods. The fusion is done at different steps, and involves different mathematical operations on scores (sum, weighted sum, product, min, max) and logical operations decision (comparison to a threshold, or, and) on differents templates extracted from the same capture. An extended version to any multibiometric system is presented in Fig. 1(d). We think our work can be seen as a generalization of this paper.

It is also possible to model the distribution of the genuine and impostor matching scores, we talk about Density-based score fusion. In Nandakumar, Chen, Dass, and Jain (2008), scores are modelled with a Gaussian Mixture Model and have been tested on three multibiometric databases involving face, fingerprint, iris and speech modalities.

Concerning non linear algorithms, Support Vector Machine (SVM) can also be used in a fusion process. Each score to combine is arranged in a vector and a training set is used to learn the SVM model. In Czyz, Sadeghi, Kittler, and Vandendorpe (xxxx), the SVM fusion to improve face recognition gives slightly better performances than weighted sum. Voice and online signature have been fused with SVM in Garcia-Salicetti, Mellakh, Allano, and Dorizzi (2005). In this experiment, arithmetic mean gives best results with noise free data, while SVM gives equivalent results with noisy data.

In this paper, we are interested in biometric modality independent transformation-based score fusion (Nandakumar et al., 2008) where the matching scores are first normalized and second combined. We have previously seen that in this case, arbitrary functions are often used. Our work is based on these various fusion architectures based on score fusion in order to produce a score fusion function automatically generated with genetic programming (Koza & Rice, 1992).

By the way, the definition of a fusion architecture is still an open issue in the multibiometrics research field (Ross & Poh, 2009), because the range of possible fusion configurations is very large. We think that using automatically generated fusion functions can bring a new solution to solve this kind of problems.