Elsevier

Information Fusion

Volume 18, July 2014, Pages 161-174
Information Fusion

Integrated biometrics template protection technique based on fingerprint and palmprint feature-level fusion

https://doi.org/10.1016/j.inffus.2013.09.001Get rights and content

Abstract

Multi-biometric systems are known to be universal and more accurate in biometric recognition. However, the storage of multiple biometric templates as separate entities pose major threats to user privacy and system security. Therefore, we propose to fuse multiple biometric modalities at the feature level in order to obtain an integrated template and to secure the fused templates using a hybrid template protection method. The proposed method is made out of a feature transformation technique known as Random Tiling and an equal-probable 2N discretization scheme. The former enables the revocability of the template and the latter converts the feature elements into binary representations according to the area under the genuine interval curve, in order to offer better privacy protection to the template. Our experimental results show that the proposed multi-biometric template protection scheme demonstrates better verification results as compared to their unibiometric counterparts while preserving template security.

Introduction

In comparison with traditional personal recognition methods such as knowledge-based methods (i.e. passwords) and token-based methods, biometrics offer more usable security as users are required to physically interact with the system. Biometric systems are thought to be more convenient to use because users do not have to remember any passwords or to carry any token. These advantages have enabled biometrics to replace and strengthen the existing token and password-based technologies in diverse applications such as access control and electronic data security.

Normally, most biometric systems employ only one biometric modality for identity management, i.e. only a single distinguished biometric source is utilized during the recognition process. As a result, unibiometric systems are intolerant of noise arising from distorted input data acquired by the sensor, signal distortion caused by environmental factors and changes of physical traits over time. In contrast, a multi-biometric system offers the following benefits [1]: (i) lower error rate – an amalgam of the information acquired from various sources could possibly reduce error rate, (ii) improved availability – if one biometric trait is missing, this can be supported by other available traits, (iii) higher degree of freedom – a multi-biometric system is able to recognize a user even if he or she uses only a subset of the employed biometrics, (iv) less susceptible to spoof attacks – spoofing of multiple traits at the same time is not easy, and (v) higher robustness – a noisy biometric sample can be clarified by other samples which contain sufficient discriminative information.

In general, a biometric system comprises of four parts, namely, (i) sensor module – to acquire raw biometric impression(s), (ii) pre-processing and feature extraction module – to enhance the acquired impression(s), and to extract salient characteristics from them, (iii) matcher module – to compare the query features with the stored template in order to produce a match score, (iv) decision module – to authenticate or reject a user by comparing the match score against a predefined threshold.

In this context, the storage of digitized biometric data in a centralised database raises concerns about their protection and usage, especially since such information is vulnerable to be used for a variety of motives, without the awareness or permission of the owner. This causes an increase in anxiety among the public over privacy issues since the owners might worry that such information may be distributed for commercial purposes to unauthorized entities. Meanwhile, an adversary could jeopardize the integrity of the biometric database by utilizing stolen biometric data for illegal purposes. Once a biometric template is breached, it has to be considered permanently broken to avoid further damage to the system. To further complicate matters, a biometric cannot be substituted with another because of the strong connection between a user’s biometrics and its template. Thus, a secure template protection method is essential for biometrics systems.

This work primarily focuses on multi-biometric template protection method that encompasses biometric feature level fusion, feature transformation and bit-extraction. In the proposed solution, two different biometric modalities, i.e. fingerprint and palmprint are initially fused at the feature level. Subsequently, we recommend a revocable feature extraction method which is known as Random Tiling and a reliable bit extraction method that makes use of the fused features for the purpose of template protection.

After going through the literature, which will be presented in Section 2, we noticed that score and decision level fusions are useful to improve the accuracy of a biometric system. Unfortunately, the individual biometric modalities have to be kept as separate entities. This causes major overheads in securing a biometric system as the template protection measure has to be repeated multiple times in order to generate different biometric modalities. Meanwhile, the data level fusion should be capable of producing a unified biometric template. However, the raw biometric input is usually noisy and contains too much redundant data. This may result in performance degradation.

On the contrary, feature level fusion overcomes the weakness of data level fusion since pre-processing and feature extraction are performed prior to fusion in order to remove noise and to extract salient features. As compared with other fusion strategies, feature level fusion requires the template protection measure to be done only once by integrating the palmprint features with the fingerprint features, while still delivering similar or possibly better recognition performance.

In short, the multi-biometric system produces superior recognition accuracy as compared to that of its unibiometric counterparts. However, if the fused template is stored unprotected in the database, it will be vulnerable to leakage and compromise of personal information as a result of database attacks. As more than one of the biometric modalities is involved, a single security breach may result in the downfall of multiple biometric modalities. Therefore, we assert that protecting the biometric template is a prerequisite in order to strengthen the biometric system.

While biometric salting approaches such as [2], [3] are known for their strength in generating diversified and revocable templates, they are susceptible to invertibility attacks. Similarly, key generation methods [4], [5], [6], [7] may produce non-invertible bit-string templates. However, the produced bit-strings are usually irrevocable. To address these issues, we propose a 3-stage hybrid template protection method as listed below for a secure biometric system:

  • (a)

    Feature level fusion – to integrate the fingerprint and palmprint biometrics at feature level in order to obtain a unified template.

  • (b)

    Random Tiling (RT) – to extract unique and random characteristics from the fused feature using a user-specific key for revocability and diversity.

  • (c)

    Equal-probable 2N discretization (eq2N) – to produce the ultimate bit-string template from the unified feature vector according to the assigned interval index.

The proposed 3-stage hybrid template protection method inherits the advantages of both biometric salting and key generation approaches, while overcoming the weaknesses of each one. The RT method is applied to achieve the revocability property of the template protection method, while preserving recognition performance. On the other hand, the eq2N equal-probable discretization scheme serves as the key generation approach to derive a template bit-string for each user from the ordered set of extracted feature vectors to alleviate the stolen-token problem in the RT. In comparison with the hybrid method proposed in [8], the proposed method (a combination of RT and eq2N) exhibits a better performance in both the plain verification and the stolen token scenario. In addition to that, the suggested approach is also able to further widen the gap between the imposter and genuine distributions as compared to the combination of BioPhasor with equal-width 2N quantization approach [8]. Furthermore, eq2N is less susceptible to the possible revelation of the exact location of each genuine user’s probability density function (PDF), and thus offers better privacy protection of biometric data. We will justify the above claims in the experimental analysis section.

A preliminary study of the proposed hybrid template protection method has been presented in [9]. However, the viability of the proposed method is not fully justified without a rigorous analysis and substantial empirical evaluation with respect to four template protection properties. To supplement this, we have introduced several new items in this paper as stated below:

  • A rigorous experiment setup with a new fingerprint dataset [10] has been added to provide two additional pairs of fusion combinations in addition to the initial four pairs.

  • A new work detailing the different dimension selection of fingerprint and palmprint images is added to explain how the two biometric modalities could be fused and represented into a unified feature vector for the use of the subsequent template protection method. A set of experimental analysis is included to evaluate the performance results of fusion methods against its unibiometric counterparts in Section 4.2.

  • Performance validation of unibiometrics with template protection method is given in Section 4.3.

  • Separability property is examined: the resulting distances between genuine and impostor distributions are analyzed against equal-width discretization scheme [8]. The wider the distance between the two distributions, the better the recognition result is. Essentially, a clear separation could produce 0% equal error rate.

  • Revocability and diversity properties are examined. Firstly, pseudo-impostor distributions are generated to simulate the effects of producing refreshed template bit-strings based on a new user-specific token. Revocability is verified if the pseudo-impostor distribution closely resembles the original impostor distribution. Whereas, if the pseudo-impostor distribution does not overlap or only minimally overlaps with the genuine distribution, this phenomenon could verify the diversity property.

The structure of this paper is organized as follows: Section 1 presents an overview of this research in addition to its motivations and contributions, Section 2 conveys the literature survey of biometric fusion and template protection, while Section 3 covers the proposed feature-level fusion based hybrid template protection method – feature level fusion for fingerprint and palmprint, Random Tiling, and equal-probable 2N discretization, In Section 4, experimental results validating the effectiveness of the proposed approach are discussed. Section 5 presents the summary of finding, and finally, conclusion remarks are provided in Section 6.

Section snippets

Biometric fusion

Typically, a biometric fusion scheme can be categorized into feature, sensor, decision or score level, subject to the kind of information being merged [11]. The amount of information content generally decreases towards the later modules of a biometric system. For instance, raw data obtained by the sensor module holds the highest information content, while the decision module produces only an accept/reject output.

However, extracting higher level information such as raw biometric data (e.g., raw

The proposed hybrid template protection method

A general outline of the proposed hybrid template protection approach is provided in Fig. 1. In the beginning, a user registers in the system by submitting his fingerprint and palmprint. The captured biometric data will then be fused at the feature level. Based on the user-specified key, a random feature set is generated by applying Random Tiling to the fused feature. Finally, the random features are discretized to generate the template bit-string. During verification, the same procedure is

Experimental settings

There are all together three fingerprint and two palmprint databases. In every database, 100 subjects and 8 images for each participant were randomly chosen for the experiments. First and foremost, the ROI of each fingerprint and palmprint pair was extracted and standardized to 150 pixels × 150 pixels.

The details and denotations of the three fingerprint databases are as follow:

  • (a)

    F1: Database from [29] a.k.a. FVC2004 DB1 Set A database. The optical sensor “V300” by CrossMatch was used for collecting

Summary of findings

In general, our findings can be summarized as follows:

  • Our approach integrates a robust fusion scheme, hybrid feature transformation and discretization (key generation) approach to address various criteria of the template protection method. From the comparison of the two hybrid approaches (the proposed method and equal-width 2N discretization scheme), the proposed method exhibits promising performance in both the plain and the stolen token scenario.

  • Due to the unique random rectangles generated

Conclusion

In this work, we have presented a 3-stage hybrid biometric template protection method for a multi-biometrics system. The first stage is a fusion method that integrates palmprint and fingerprint at feature level for the production of unified templates which reduces the possible overheads of the template protection process. The second stage – RT and third stage – equal-probable 2N discretization make up the hybrid template protection method. Empirical evaluations indicate that this method could

Acknowledgements

This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Science, ICT and Future Planning (2013006574).

References (34)

  • A. Ross et al.

    Information fusion in biometrics

    Pattern Recognition Letters

    (2003)
  • Y.-F. Yao et al.

    Face and palmprint feature level fusion for single sample biometrics recognition

    Neurocomputing

    (2007)
  • J. Yang et al.

    Feature-level fusion of fingerprint and finger-vein for personal identification

    Pattern Recognition Letters

    (2012)
  • A. Lumini et al.

    An improved BioHashing for human authentication

    Pattern Recognition

    (2007)
  • A.B.J. Teoh et al.

    Random multispace quantization as an analytic mechanism for BioHashing of biometric and random identity inputs

    IEEE Transactions on Pattern Analysis and Machine Intelligence

    (2006)
  • A.B.J. Teoh et al.

    BioHashing: two factor authentication featuring fingerprint data and tokenised random number

    Pattern Recognition

    (2004)
  • C. Soutar, D. Roberge, A. Stoianov, R. Gilroy, B.V.K. Vijaya Kumar, Biometric encryption using image processing, in:...
  • U. Uludag, S. Pankanti, S. Prabhakar, and A.K. Jain, Biometric cryptosystems: issues and challenges, in: Proceedings of...
  • Y.J. Chang, W. Zhang, T. Chen, Biometrics-based cryptographic key generation, in: IEEE International Conference on...
  • C. Chen, R.N.J. Beldhuis, T.A.M. Kevenaar, A.H.M. Akkermans, Biometric binary string generation with detection rate...
  • A.B.J. Teoh, K.-A. Toh, W.K. Yip, 2^N discretization of biophasor in cancellable biometrics, in: Proceedings of the...
  • Y.J. Chin, T.S. Ong, A.B.J. Teoh, M.K.O. Goh, Multimodal biometrics based bit extraction method for template security,...
  • FVC2002 fingerprint database....
  • A. Ross, An introduction to multibiometrics, in: Proceedings of the 15th European Signal Processing Conference, 2007,...
  • S.C. Dass, K. Nandakumar, A.K. Jain, A principled approach to score level fusion in multimodal biometric systems, in:...
  • K. Nandakumar et al.

    Likelihood ratio-based biometric score fusion

    IEEE Transactions on Pattern Analysis and Machine Intelligence

    (2008)
  • S. Ribaric et al.

    A biometric identification system based on eigenpalm and eigenfinger features

    IEEE Transactions on Pattern Analysis and Machine Intelligence

    (2005)
  • Cited by (0)

    View full text