Multi-modal Biometric Protection System Using SURF Filter with BioHashing Algorithm

Multi-modal Biometric Protection System Using SURF Filter with BioHashing Algorithm

Bedad FatimaAdjoudj Réda 

EEDIS Laboratory, Djillali Liabes University of Sidi Bel Abbes, Sidi Bel Abbes 22000, Algeria

Corresponding Author Email:
1 March 2019
26 May 2019
1 September 2019
| Citation



Biometrics relates to the use of physiology and/or behaviour to determine or verify the identity of individuals. Despite the advantages of these biometric systems over traditional authentication systems that use passwords and ID cards, they are still vulnerable to specific limitations that can significantly degrade their functionality. The confidentiality of biometric data must be protected to ensure the privacy of users, while security mechanisms are necessary to ensure this protection. The system’s resistance to attacks can also be overcome by multimodal biometrics. Multibiometrics consists in concatenating multiple biometric data to improve authentication performance. Unfortunately, even well-established multi-biometric systems suffer from vulnerabilities, the most crucial elements being the biometric models that store the user's biometric data. If a user's biometrics is compromised, a person's identity and privacy are also compromised, since it is impossible to revoke or reissue the biometric template. Therefore, secure and revocable multi-biometric templates protection schemes are required to ensure multi-biometric system integrity and user privacy. This article presents a revocable multi-biometric scheme to secure multi-biometric data and focusses on the BioHashing method which is a recent technique capable of addressing the reversal of the confidentiality and security problem.


multi-biometric, security, fusion, BioHashing, revocable

1. Introduction

Biometrics, defined as the quantitative study of human biological, morphological or behavioral characteristics, are a very active field of research. Currently, biometric technologies are most often based on the fingerprint method, which is considered one of the most reliable in return for its intrusiveness [1].

There are two modes of operation in the application of a biometric system: Verification (comparison with a model) and Identification (search for the identity of one person among several).

Unimodal biometric systems suffer from several problems that lead to the use of a single biometric trait susceptible to noise, poor capture, poverty in terms of confident biometric points and in particular to the deterioration of the quality of biometric entry. The introduction of multimodal biometric systems is one solution to these problems.

Multi-biometric systems are based on the richness and availability of the individual's discriminating information. Multibiometrics offers enormous advantages to recognition systems by improving accuracy and reducing the risk of identity falsification [2].

There are 5 forms in multi-biometric system: multi sensor, multi algorithm, multi instance, multi samples and multi modal. However, multi-biometric systems are not limited only to these forms. The implementation of these systems must take into account their employment, also called fusion [3].

The fusion represents the integration of several pieces of information (homogeneous or heterogeneous) into a single usable piece of information. The combination of several biometric systems can be done at four different levels: at the data level, at the level of the characteristics extracted, at the level of the scores resulting from the comparison module or at the level of the decisions [4].

Furthermore, it is important to note that the study of security via multi-biometric systems faces a major problem against the remained challenge. This problem consists in the complexity of the algorithms for extracting and matching biometric models and protecting them.

Several solutions are proposed in the literature to ensure the protection of multi-biometric data including crypto-biometric algorithms [5-6] or transformation algorithms (revocable biometrics) [7-8].

This article aims to provide a contribution to the security of a multi-biometric system in order to improve the protection of user recognition.

As a result, multiple and repeated fingerprint instances were combined at the score level, taking advantage of the SURF algorithm's scaling and rotating invariability, as well as its speed [9].

In addition, a behavioural biometric method (the most commonly used) with a model protection scheme was also used. This solution has the advantage of being easy to use and very fast. The biometric reference is stored in the database as BioCode (binary code linked to the biometric data) that can be revoked in the event of an attack.

This paper is organized as follows. Section 2 briefly presents the state of art existing solutions for multi-biometric data protection. The proposed method is described in Section 3. Section 4 illustrates the performance of the proposed solution based on the experimental results. Finally, Section 5 concludes the paper and provides some perspectives.

2. Multi Biometric Model Protection

Security is a necessity for robust and user-friendly biometric systems. The main objective of research in the field of multi-biometric protection is to generate industrial projects that are presented in a generic framework. The system should be able to integrate n models, without the need to follow specific fusion levels for their representations, (k representation might be involved).

The generic system has been applied for the protection of biometric data. Figure 1 presents the representation

Figure 1. A generic protection framework for a generic multi-biometric template at the characteristics [10]

An ideal biometric model protection scheme should have the following four properties [5]:

  • Diversity
  • Revocability
  • Irreversibility
  • Performance

Biometric model protection schemes proposed in the literature can be classified into two categories (Figure 2), namely (i) biometric cryptosystems and (ii) transformation approaches [11].

The common feature to all these methods is that raw biometric data is not stored directly in the database; data is either stored on an external medium (smart card, token) or stored after transformation [11].

Figure 2. Categories protection of multi-biometric models

2.1 Multi-biometric cryptosystems

Multi-biometric cryptosystems are the combination between crypto-system and multi biometry. The principle of classical crypto systems [12] is to improve safety personal authentication systems based on biometrics. They have been associated with the principle of biometric recognition.

These approaches aim to minimize the rate of biometric data stored for protected models and system database, in general. In the majority of biometric encryption systems, the operation is as follows: when recording, an error correction code j is applied to the biometric model B and key K to extract the dataset H (the set H is called Helper Data). During the authentication process, an error correction code j is applied using data H and Q test pattern to find the K key the way the helper to extract the data is extracted (Figure 3).

In addition, biometric cryptosystems can be classified into two categories [13-14]: keybinding and key-generation.

Figure 3. General mechanism for authentication of key-binding and key generation biometric cryptosystems [13]

When the helper data is obtained using a key that is independent of the biometric characteristics, it is a key-binding crypto-system. If the data helper is derived only from the biometric model and the key is generated directly from the biometric features, it is a key-generation cryptosystem [15].

2.2 Characteristics transformation

The main idea of feature transformation approaches is to convert an unprotected biometric model into a protected model using a transformation function [16]. Based on the system and chosen method, the function of transformation can take several forms and can require the use of certain transformation parameters (for example, a user key). In the event of theft or compromise of the transformed biometric models, the transformation parameters are modified to update the protected biometric model [17].

To prevent impostors from tracking authorized users (legitimate users) registered in multiple systems, and thus protect privacy, different transformation settings or even different transformation functions must be applied to each application.

In general, these approaches are as follows: suppose that X will be transformed into T-coded data when a user using an F function. For verification purposes, the request must be biometric, Y will be transformed into T’, always using the F function. Successful authentication is achieved if T is close to T’ using a certain measure of similarity. To ensure the revocability of the system, a data S in the form of a key is assigned to each user U. The S key is then considered as an input parameter of the transformation function F. A revocation implies the direct replacement of this user key.

It will be a salting mechanism where transformation is an operation that combines X with random data generated from S. However, S is considered as the seed of a pseudo-random generator (Figure 4).

Figure 4. Generic functioning of revocable transformations [18]

The characteristics transformation schemes can be classified into two categories, the invertible transformation (known as Salting or also Biohashing) and the non-reversible transformation [13-14]. For Biohashing, it is a two-factor technique based on the use of a Random projection [19]. It is a technique that uses random orthogonal matrices to project biometric models into other areas or spaces where distances between pre- and post-transformation models are maintained.

As non-invertible transformation, an original model can be protected by using a non-reversable function that is in most tasks a one-way function [14].

One of the most significant forms of the function of transformation in the context of irreversible transformation and consists of using distortions or geometric transformations to protect biometric models [17].

3. Proposed Multi-Modal Biometric System Protection

This section presents the design and implementation of the application to ensure the recognition of individuals. The proposed system is based on fingerprint and 5 main modules were added: enrollment, feature extraction, protection module (Biohashing), fusion, matching module and decision. Figure 5 shows the general principle of the proposed method.

Figure 5. Diagram of the proposed multi-modal biometric system protection

(1) Enrollement: During the enrollment phase, the user must provide his/her fingerprints to the sensor to generate his/her reference templates. In this step, several instances are stored and the same biometric line trait is repeated to represent the variability of this l trait. The fingerprint recognition was chosen for several reasons:

- Fingerprint recognition is the oldest biometric technique.

- This is one of the methods that has been very successful. Its usefulness has been proven for the forensic service, as well as criminology.

- Nowadays, reliable biometric techniques such as fingerprint recognition are needed everywhere for the authentication of people and can therefore be used for very high security applications.

- Its simplicity of acquisition and public acceptance.

- The existing of a real fingerprint database.

(2) Feature extraction: This step represents the core of the multi-biometric recognition system. The information that will be saved in the database is extracted from the image for later use in the A specific algorithms is used to extract the biometric characteristics. However, it is important to find algorithms in which the results of the characteristic vectors are stable in size, because the Biohashing Algorithm is applied to the biometric data represented by a real value vector of fixed length.

There are several recent techniques widely used in computer vision problems, such as Gabor filters [20], the LBP (Local Binary Pattern) [21] operator with a set of its variants and the SURF method. The SURF method was used in this research study.

The result of this process of feature extraction   is a vector of features its size is 64 features for each image.

(3) SURF Overview: The Speeded Up Robust Features (SURF) algorithm, which can be translated as robust accelerated features, developed by [9], is a feature detector and descriptor. It was presented by researchers at ETH Zurich and the Catholic University of Leuven for the first time in has the major advantage of being both invariant to rotations and changes. Moreover, it is simple and fast. SURF was partly inspired by the descriptor SIFT [22], that it surpasses in speed.

(4) The SURF algorithm: The implementation of the SURF algorithm requires several steps. The first is to detect points of interest on the image and the second is to describe these points of interest using a vector of 64 features (Figure 6).

Figure 6. Description of the SURF algorithm

The SURF algorithm uses fast-Hessian for the detection of points of interest and an approximation of the Haar wavelets to calculate the descriptors. The fast-Hessian is based on the study of the Hessian matrix:

$H(x, y, \sigma)=\left[\begin{array}{ll}{\operatorname{Lx} x(x, y, \sigma)} & {\operatorname{Lxy}(x, y, \sigma)} \\ {\operatorname{Lxy}(x, y, \sigma)} & {\operatorname{Lyy}(x, y, \sigma)}\end{array}\right]$    (1)

where, Lij (x, y, σ) is the second derivative along the directions at i and j of L with:

L(x , y , σ) = G(x , y , σ) * I (x , y )   (2)


$G(x, y, \sigma)=\frac{1}{\sigma \sqrt{2 \pi}} e^{\frac{\left(x^{2}+y^{2}\right)}{2 \sigma^{2}}}$   (3)

And I is the starting image.

The maximization of the determinant of this matrix makes it possible to obtain the coordinates of the points of interest on a given scale. This step brings an invariance of the points of interest compared to the scaling. The determinant is defined as:

$\operatorname{det}(H(x, y, \sigma))=\sigma^{2}\left(\operatorname{Lx} x(x, y, \sigma) L y y(x, y, \sigma)-L_{x y}^{2}(x, y, \sigma)\right)$    (4)

This step therefore makes it possible to detect the candidate points of interest. The algorithm then comprises intermediate steps intended to provide more precision in their location.

The descriptors are calculated using the Haar wavelets. They make it possible to estimate the local orientation of the gradient and thus to provide the invariance with respect to the rotation. The responses of the Haar wavelets are computed in x and y, in a circular window, whose radius depends on the scale factor of the point of interest detected. These specific responses contribute to the formation of the feature vector corresponding to the key point.

(5) Biohashing: The Biohashing method is an algorithm used for biometric data that is expressed by a fixed-length real value vector and generates a binary model named BioCode of length less than or equivalent to the original size.

For BioHashing, the goal is to generate a unique code, called biocode, using two data: the biometric model and a random number. A better protection, the data must be in the form of a token or a key USB.  same transformation scheme at the same time:

- During enrollment, where the biocode is stored instead of the biometric model.

- During the verification step, where a new biocode is generated from the random number assigned to the user during the enrollment.

Using different random numbers for different applications, this principle ensures revocability and diversity of the biocode. Figure 7 represents the BioHashing process.

It can be noticed that the model is a two-way system authentication factors. Therefore, the transformation function combines a random, grain-shaped number, stored in a token, whose biometric model is represented by a fixed-length vector X = (x1, ..., xn), XϵRn [7].

The result of this process is a vector of features. Its size is 64 features for each image contains the 0 and 1.

Figure 7. General principle of BioHashing with fingerprints [8]

(6) Matching: In this step, the correspondences between the biocode of the image request and the biocodes of the images of the base are searched.he objective is to form the couples of the nearest descriptors. Therefore, the method "NN-DR", meaning Nearest Neighbor Distance Ratio [23], is used. The image I is thus characterized by the formula:

Y(I) = {ki = (xi,yiii,vi) ǀ i = 1 : N(I)}

where, N(I)is the number of points of interest detected in I; (xi, yi) is the position of the point of interest i in I; (σi, Ɵi) is the scale and orientation of the point of interest i; vi is the descriptor vector of the point of interest i.

The verification between the two images I1 and I2 corresponds to the calculation of the associations number between the two sets Y (I1) and Y (I2).

An association is defined by a double mapping between two points of interest. For the point of interest x, the point of interest y of the nearest among the set of points d interest of I2.

In addition, the second nearest point of interest is verified to be far enough away from x using a threshold value C:

 d(xy) = min {Zϵ Y (I2)} d(x,y)    (5)

d(x,y) = C * d(x,y’)     (6)

d(x,y’)= min{Z ϵ Y(I2) ,d(x,z) >d(x,y)}d(x,z)     (7)

The correspondence between x and y is accepted if the ratio $\frac{d\left(x, y^{\prime}\right)}{d(x, y)}$ is less than a threshold C. This highlights that « must look a lot like «y» but not as all the other descriptors in the database.

The distance d(.,.) is the  calculated Euclidean distance between the two normalized descriptors corresponding to the points of interest. If these two conditions are not met, then the point x is not mapped to the point y [23]. This method is called "NN-DR" for Nearest Neighbor Distance Ratio.

(7) Fusion: Fusion considers the problem of multi-biometric data protection applications or transactions between multi-biometric system architecture and protection algorithms that take place across the fusion levels.

This work is based on two-tier coupling: multi-methods and Biohashing. The Biohashing algorithm is applied to the fusion level. Therefore, the fusion at the score level gives the best compromise between the wealth of information and the ease of implementation.

The two methods of fusion: Multi sample (to take into account the variations of the pose and the image quality acquired) and multi instance of 3 fingers of the left hand are applied to add more security and reduce the error rate. 

During the test, the four scores of the samples of the same instance are fused, after combining the three scores of three instances, having a global score and using a simple method. The weighted sum that is described in next section a balanced weight equal to one.

weighted sum: The weighted sum makes it possible to give different weights to each subsystem according to their individual performance or their interest in the multimodal system [24]. Strategies used in the combination, such as Sum Rule, Decision Tree, Linear Discriminant Analysis [25].

(8) Decision: This step represents the entry of a similarity matrix that contains all the merged scores, the system accepts the client if it has a maximum score (maximum number couple of the point of interest).

4. Experimental Results

4.1 Database

The experiments were conducted on the SDUMLA-HMT fingerprint sub database [40]. It includes images acquired by 5 different sensors on 1 person. For each person, 6 fingers are recorded: the thumb, the index and the middle finger of each hand; each of the fingers is recorded 8 times (Table 1). The 3 fingers of the left hand have been selected. The images captured by the AuthenTec AES2501 sensor, which is a capacitive type sensor using the scanning capture method, allowed to obtain better results.

Table 1. Description of the fingerprint sub-database




















4.2 Distribution of the database

Develop a recognition application, it is necessary to have two databases: one to perform the learning and the other to test the techniques and determine their performance:

- Learning Images: The first, third, fifth, and seventh images of each person are used for the learning phase (4/8).

- Test Images: The 4 remaining images of each person were used for different tests.

5. Performance Evaluation

This section presents the results of experiments conducted on the SDUMLA HMT database.

In the identification scenario, the performance of the proposed approach is presented as recognition rates and CMC curves. In the authentication scenario, the performance of the proposed approach is presented as Verification Rates (VR) and ROC Curves.

5.1 Classification rate

The SURF algorithm detects a variable number of points of interest for gallery fingerprints and test fingerprints, forming a dictionary. Its size has a direct impact on the recognition performance. A small dictionary speeds up the calculation, but this results in a significant degradation of the recognition performance.

There can be between 800 and 1500 points of interest returned by the SURF detector on each footprint. This makes the dictionary large and thus increases the cost of calculation of e proposed approach.

The effect of dictionary size (number of points) on the performance of the proposed approach for fingerprint recognition is studied. This number was experimentally evaluated in the classification task by increasing the number of points progressively and calculating the classification rate correct.

Recognition accuracy increases dramatically with increasing dictionary size, as shown in Figure 8. Thus, it can be noticed the improvement of the recognition rate of the proposed system without the application of Biohashing algorithm versus uni-modal system.

Figure 8. Evaluation of the classification rate according to the number of points used in the three systems: Uni-modal (S.UNI), multi-sample without protection (S.M.E) and multi-instance multi-sample with protection (S.M.E.M.I)

5.2 The CMC curve

Figure 9. Performance of the proposed approach versus the number of fingers

In order to evaluate the performance of the proposed approach, the CMC (Cumulative Match Characteristic) curve was chosen. This curve displays the cumulative identification rate according to the rank of the distribution. This provides an indication of the degree of proximity to obtain the correct match of the correspondence at rank-1 is incorrect.

The Figure 10 indicates that the proposed approach gives better results compared to the other two systems (uni-modal system and unprotected multi-sample system).

Therefore, recognition rate was improved at rank1of 91 % in the unimodal system; while 96 % using several samples and finally 100 % using repeated and multiple instances with Biohashing as shown in Figure 10.

Figure10. Comparison between CMC curves: The proposed. Uni-modal system and multi-sample system

5.3 The ROC curve

An effective approach should ideally have high verification rates (VR) associated with low false acceptance rates (FAR). Figure 11 shows that the FAR=100 % the proposed approach compared to 84 % for a multi biometric system without protection.

Thus, Figure 9 justifies the choice of this research to use three fingers.

Figure 11. Comparison between ROC curves:

The proposed approach (S.M.E.M.I). Uni-modal system (S.UNI) and multi-sample system without protection (S.M.E)

Table 2 provides a comparison of the recognition accuracy of the proposed Biohashing [26] and related work.

Table 2. Comparison between the proposed approach and previous studies (recognition accuracy)






Performance (%)






Jeong and al., 2006 [27]

Combination of two methods of extraction of PCA and ICA characteristics, and vector tronsformation by Biohashing






Maiorana and al., 2011 [28]

Non-invertible transformations






Paul and al., 2012 [29]

Random projection and transformation based on extraction and selection of features







Canuto and al., 2013 [30]

Fusion in the context of the recognition of revocable multi biometrics

Voice data






Rathgeb and al., 2014 [31]

Multi biometrics revocable based on bloom filters






Rathgeb and al., 2015 [32]

Multi-biometric revocation based on Bloom filters and fusion-level features.







Damasceno and al., 2015 [33]

Four revocable transformations (Interpolation, BioHashing, BioConvolving and DoubleSum)

Touch Analytics





Stokkenes and al., 2016 [34]

Protection of a multi-biometric system based on the Bloom filter


two peri-ocular regions





Yildiz and al., 2017 [35]

Security and confidentiality of a multi-biometric model by superimposing several biometric data and fusion of multi-biometric models







Bringer and al., 2017 [36]

Security of protected biometric models based on Bloom filters







Jegede and al., 2018[37]

Matrix Transformation for revocable multi biometric model protection







Bedad and Adjoudj, 2018 [38]

BioHashing and fusion at feature level







Dwivedi and al., 2019 [39]

Hybrid fusion of score and decision levels for cancelable multimodal biometric verification







The proposed approach

BioHashing and fusion at the score level

Fingerprint multi instance





The results of the previous work show that the proposed approach gives a performance equal to the performance of Bedad and Adjoudj [38] and less compared to other research studies. The use of BioHashing is to modify the original model corresponding to the method used. It also gives a high protection for data security.

The common point between the two approaches is the use of Biohashing, but there are several differences in the details between the two approaches. However, Biohashing proves its effectiveness despite changes in methods (feature extraction, fusion and matching).

Table 3. Comparison of the security of the proposed approach and related work by Bedad & Adjoudj [38]



Feature extraction



Bedad & Adjoudj, 2018 [38]


LBP filter

Feature level

Hamming distance

The proposed approach



Score level

Nearest Neighbor Distance Ratio

6. Conclusion

This article studied the problem of protecting multi-biometric systems using invertible reversible biometry: Biohashing.

Firstly, the biometric system and their disadvantages were presented and then multi-biometric system was introduced as the solution to these disadvantages.

Then the application of revocable biometrics to diversify and secure biometric and multi-biometric data so that they do not directly use the original data and thus ensure the protection of users' privacy was studied.

Subsequently, the impact of Biohashing in the multi-biometric system was studied by presenting the different classification rates or comparing them between the unprotected multi-biometric system and the proposed multi-biometric system. As a result, the effectiveness of security in multi-biometric systems is observed.

Based on the promising results obtained in this study, some perspectives include advanced hybridization methods between revocable biometrics and cryptography to improve confidentiality in the protection of multi-biometric data.


[1] Adjoudj, R., Gafour, A.K., Boukelif, A., Lehireche, A. (2008). La reconnaissance des visages: une comparaison entre les réseaux des neurones compétitifs et les réseaux des neurones à spike. Traitement du Signal, 25(6): 449-458.

[2] Ross, A.A., Nandakumar, K., Jain, A.K. (2006). Handbook of Multibiometrics, Springer, 193, ISBN: 978-0-387-33123-2. 

[3] Oloyede, M.O., Hancke, G.P. (2016). Unimodal and multimodal biometric sensing systems: A review. IEEE Access, 4: 7532-7555.

[4] Nanni, L., Lumini, A. (2006). Empirical tests on biohashing. Neurocomputing, 69(16): 2390-2395.

[5] Maltoni, D., Maio, D., Jain, A., Prabhakar, S. (2009). Handbook of fingerprint recognition. Springer Science & Business Media. 

[6] Barni, M., Bianchi, T., Catalano, D., Di Raimondo, M., DonidaLabati, R., Failla, P., Fiore, D., Lazzeretti, R., Piuri, V., Piva, A., Scotti, F. (2010). A privacy-compliant fingerprint recognition system based on homomorphic encryption and finger code templates. BTAS.

[7] Teoh, A.B.J., Ngo, D., Goh, A. (2004). Biohashing: two factor authentication featuring fingerprint data and tokenised random number. Pattern Recognition, 37(11): 2245-2255.

[8] Nagar, A., Nandakumar, K., Jain, A.K. (2006). Biometric template transformation: A security analysis. Proceedings of SPIE, Electronic Imaging, Media Forensics and Security XII.

[9] Bay, H., Ess, A., Tuytelaars, T., Gool, L.V. (2008). Speeded-up robust features (surf). Computer Vision and Image Understanding, 110(3): 346–359.

[10] Rathgeb, C., Busch, C. (2012). Multi-biometric template protection: Issues and challenges. In New trends and developments in biometrics. In Tech.

[11] Belguechi, R., Le-Goff, T., Cherrier, E., Rosenberger, C. (2011). Study of the robustness of a cancelable biometric system. In Network and Information Systems Security (SARSSI), 2011 Conference on, pp. 1-7.

[12] Degabriele, J.P., Paterson, K.G. (2010). On the (in) security of IPsec in MAC-then-encrypt configurations. In Proceedings of the 17th ACM Conference on Computer and Communications Security, 493-504.

[13] Anil, K.J., Karthik, N., Nagar, A. (2001). Featured Article: A machine-learning phase classification scheme for anomaly detection in signals with periodic characteristics. EURASIP Journal on Advances in Signal Processing, 40(3): 614–634.

[14] Rathgeb, C., Uhl, A. (2011). A survey on biometric cryptosystems and cancelable biometrics. EURASIP Journal on Information Security, 2011(1), 3.

[15] Moujahdi, C., Ghouzali, S., Mikram, M., Rziza, M., Bebis, G. (2012). Spiral cube for biometric template protection book chapter. Lecture Notes in Computer Science, 235-244.

[16] Ratha, N.K., Connell, J.H., Bolle, R.M. (2001). Enhancing security and privacy in biometrics-based authentication systems. IBM systems Journal, 40(3): 614-634.  

[17] Ratha, N., Connell, J., Bolle, R.M., Chikkerur, S. (2006). Cancelable biometrics: A case study in fingerprints. In Pattern Recognition, 2006. ICPR 2006. 18th International Conference on, 4: 370-373.

[18] Belguechi, R.O. (2015). Sécurité des systèmes biométriques: Révocabilité et protection de la vie privée. PhD thesis, Ecole nationale Supérieure en Informatique Alger. 

[19] Jaishanker, K.P., Vishal, M.P., Chellappa, R., Ratha, N.K. (2010). Sectored random projections for cancelable iris biometrics. In Acoustics Speech and Signal Processing (ICASSP), 2010 IEEE International Conference on, Dallas, TX, USA, pp. 1838-1841.

[20] Fogel, I., Sagi, D. (1989). Gabor filters as texture discriminator. Biological Cybernetics, 61(2): 103-113. 

[21] Ojala, T., Pietikainen, M., Harwood, D. (1994). Performance evaluation of texture measures with classification based on kullback discrimination of distributions. Proceedings of the 12th IAPR International Conference on, Jerusalem, Israel, 1: 582-585.

[22] David, G.L. (2004). Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2): 91-110.

[23] Oyallon, E., Rabin, J. (2015). An analysis of the surf method. Image Processing on Line, 5: 176-218. 

[24] Hafs, T. (2016). reconnaissance biométrique multimodale basée sur la fusion en score de deux modalités biométriques: l’empreinte digitale et la signature manuscrite cursive en thesis, universite badji mokhtar-annaba. 

[25] Sander, K. (2003). Multibiometric authentication. Term Project CS, 104: 40-41.

[26] Bedad, F., Adjoudj, R. (2018). Multi-biometric template protection: An overview. In International Conference in Artificial Intelligence in Renewable Energetic Systems, pp. 70-79. 

[27] Jeong, M., Lee, C., Kim, J., Choi, J.Y., Toh, K.A., Kim, J. (2006). Changeable biometrics for appearance based face recognition. In Biometric Consortium Conference, 2006 Biometrics Symposium: Special Session, pp. 1-5.

[28] Maiorana, E., Campisi, P., Neri, A. (2011). Cancellable biometrics for on-line signature recognition. In New Technologies for Digital Crime and Forensics: Devices, Applications, and Software, pp. 290-315.

[29] Paul, P.P., Gavrilova, M. (2012, August). Multimodal cancelable biometrics. In Cognitive Informatics and Cognitive Computing (ICCI* CC), 2012 IEEE 11th International Conference, pp. 43-49.

[30] Canuto, A.M., Pintro, F., Xavier-Junior, J.C. (2013). Investigating fusion approaches in multi-biometric cancellable recognition. Expert Systems with Applications, 40(6): 1971-1980. 

[31] Rathgeb, C., Busch, C. (2014). Cancelable multi-biometrics: Mixing iris-codes based on adaptive bloom filters. Computers and Security, 42: 1-12.

[32] Rathgeb, C., Gomez-Barrero, M., Busch, C., Galbally, J., Fierrez, J. (2015). Towards cancelable multi-biometrics based on bloom filters: A case study on feature level fusion of face and iris. In Biometrics and Forensics (IWBF), 2015 International Workshop, pp. 1-6.

[33] Damasceno, M., Canuto, A.M., Poh, N. (2015). Multi-privacy biometric protection scheme using ensemble systems. In 2015 International Joint Conference on Neural Networks (IJCNN), pp. 1-8.

[34] Stokkenes, M., Ramachandra, R., Sigaard, M.K., Raja, K., Gomez-Barrero, M., Busch, C. (2016). Multi-biometric template protection-A security analysis of binarized statistical features for bloom filters on smartphones. In Image Processing Theory Tools and Applications (IPTA), 2016 6th International Conference, pp. 1-6.

[35] Yildiz, M., Yanikoğlu, B., Kholmatov, A., Kanak, A., Uludağ, U., Erdoğan, H. (2017). Biometric layering with fingerprints: template security and privacy through multi-biometric template fusion. The Computer Journal, 60(4): 573-587. 

[36] Bringer, J., Morel, C., Rathgeb, C. (2017). Security analysis and improvement of some biometric protected templates based on Bloom filters. Image and Vision Computing, 58: 239-253. 

[37] Jegede, A., Udzir, N.I., Abdullah, A., Mahmod, R. (2018). Revocable and non-invertible multibiometric template protection based on matrix transformation. Pertanika Journal of Science & Technology, 26(1): 341-352. 

[38] Bedad, F., Adjoudj, R. (2018). Secured multimodal biometric system. Journal of Multimedia Processing and Technologies, 9(3): 77.

[39] Dwivedi, R., Dey, S. (2019). A novel hybrid score level and decision level fusion scheme for cancelable multi-biometric verification. Applied Intelligence, 49(3): 1016-1035. 

[40] Yin, Y.L., Liu, L.L., Sun, X.W. ( 2011). Sdumla-hmt: A multimodal biometric database. In Chinese Conference on Biometric Recognition, pp. 260-268.