Encryption and Data Management Architectureto Protect Biometric Security

Obaidul Maleka, Rabita Alamgira, Laila Alamgirb, and Mohammad Matinc

Center for Biometrics and Biomedical Research,VAa, Howard University, DCb, and University of Denver, COc

Abstract—In this paper, a novel symmetric encryptionalgorithm and its data management architecture to protectbiometric security and privacy is proposed. Unlike currentbiometric encryption, the proposed method uses crypto-graphic keys in conjunction with extracted MultiBiometricsto create cryptographic bonds. To further enhance thesecurity protection and to improve authentication accuracy,a data management architecture is being developed. Theproposed method is being tested on images from three pub-lic databases: the “Put Face Database”, the “Indian FaceDatabase”, and the “CASIA Fingerprint Image DatabaseVersion 5.1”. The performance of the proposed solutionhas been evaluated using the Equal Error Rate (EER) andCorrect Recognition Rate (CRR). The experimental resultsdemonstrate the effectiveness of the proposed method.

Index Terms—Biometric encryption, data management,MultiBiometrics, security, and unlinkability attack.

I. INTRODUCTION

With the unprecedented growth of biometric systems,concerns about biometric security are the crucial issuesfor the 21st century. Not only does the biometric template(i.e. features) contain the unique and sensitive physiologi-cal and behavioural traits of an individual, it is also unary,and cannot be revoked or reissued if compromised. Thefeatures extracted from the biometric traits are stored inthe database during enrollment in order to compare andauthenticate the legitimacy of the subject of interest. Thiscomparison also performs in the unencrypted domain,since the authentication accuracy can be largely influ-enced by a small variation in the feature properties if ittakes place in the encrypted domain. Therefore, concernsabout the security protection of biometric features areof paramount importance in the exploration of biometricsystems. Ideally, the security of the template can beaccomplished using mathematical algorithms that mustbe difficult to decrypt by the unintended recipients.In addition, a template protection algorithm should beirreversible, robust, and revokable [1-3].

In this paper, a novel symmetric biometric encryptionalgorithm and its Data Management Architecture (DMA)is proposed that protects the stored and dynamic bio-metric templates against security, privacy, and unlinka-bility attacks. In contrast to current biometric encryp-tion, this method uses cryptographic keys in conjunction

with extracted MultiBiometrics to create cryptographicbonds, called “BioCryptoBond”. To further enhancesecurity and privacy protection and to improve authen-tication accuracy, a multilayered DMA architecture isalso proposed. The theoretical foundation of the proposedmethod along with the model evaluation and experimen-tal results have also been presented in this paper.

The remainder of the paper is organized as follows:Section II presents the literature review and prerequi-sites; the detailed analysis and algorithmic formulationof the proposed biometric encryption and authentica-tion systems are presented in Section III; Section IV

presents the biometric Data Management Architecture(DMA); Section V studies the possible attacks; exper-imental results and discussions are given in Section V I;and finally, the conclusions are presented in Section V II .

II. LITERATURE REVIEW

The proposed method is based on the Biometric En-cryption (BE). In addition, a data management archi-tecture has been proposed to enhance the security ofbiometric features.

A. Cavoukian et al. [4] proposed a biometric en-cryption algorithm based on facial biometrics. In theirmethod, the system is composed of two distinct stages: i)Creation of a watch list consisting of a maximum of fivepatrons; and ii) Implementation of a biometric encryptionmodule and released keys for each of the top matchpatrons, as well as the generation of a match alert bythe system that is then reviewed by administrators. ThisBE method can achieve an optimal FAR at the costof FRR. In their self-exclusion model, K. Martin et al.[5] proposed a biometric encryption algorithm based ona small subset of the subject’s facial biometric database.The authors here used feature vectors for their key bind-ing process to secure the cryptographic key. It is a novelmodel, and they achieved low FAR at the cost of FRR.K. Nandakumar et al. [6] proposed a fuzzy vault schemewhere the authors derived a multibiometrics templatefrom multiple templates of a single user. They usedfingerprint minutiae points and iriscodes templates andtransformed them into a multibiometrics vault. Here, the

Int'l Conf. IP, Comp. Vision, and Pattern Recognition | IPCV'15 | 145

authors didn’t properly address the challenges associatedwith the unlinkability attacks.

A. Ross et al. [7] proposed a visual cryptographymethod to protect the privacy of the biometric templates.In their method, an image is decomposed into two hostimages and stored in the two central databases. Theoriginal image can only be revealed when two imagesare available simultaneously. C. Lee et al. [8], introduceda two factor method for generating cancelable fingerprinttemplates using local minutia information. The transfor-mation function is associated with the randomly gener-ated PIN number, which is used to change the biometrictemplate. The major drawback of this method is that ithas a tradeoff between performance and changeability. D.Maio et al. [9], implemented a multihashing algorithm,where the scores of selected fingerprint matchers andthose obtained by a face authenticator are combined. Fur-thermore, to enhance the performance of this system, arandom subspace based method is further combined withthe similarity matching scores. However, this method iscomputationally expensive. A. Teoh et al. [10] proposeda multispace random projection method. The distance-preserving property of multispace random projection isanalyzed based on a normalized inner product, and anapproximately zero EER is achieved; however privacyand changeability are the main concerns in this paper.

A. PrerequisitesThis section introduces some fundamental concepts

related to the proposed method before getting into itsdetailed analysis.

1) Unlinkability Attack: True anonymity requires un-linkability, which is the ability of the system to performmultiple operations anonymously. Unlinkability also im-plies the incapability of retrieving the information ofone individual based on the information of another.Additionally, it is the measurement of the strength of asystem (or object) to be unlinkable. Unlinkability is thecore property for any authentication process, making itdifficult for a third party recipient to be associated withthe unauthorized information.

2) Data Segmentation and Foreign Key: Data seg-mentation is known as data grouping, and is a branchof the data mining operation. It is the process of ex-tracting and segmenting data in such a way that thesystem would be able to factorize the data, reduce itsvolume, and classify it. It is also capable of storingdata in different locations of the database system withthe intention of increasing overall system performanceand security. However, prior to carrying out a datasegmentation analysis, appropriate care should be takento decide which key parameters could be used forthe segmentation process. This is especially importantbecause the failure of biometric segmentation means

that the system was not able to detect useful biometricfeatures. Indexing is another technique that can be usedin the data segmentation process to put segmented datain order. In addition, a foreign key is used in conjunctionwith the indexing process to create a link and establisha relationship amongst segmented data within databasesystem.

III. Biometric encryption and AuthenticationThe biometric templates h(t) created from the images

received from the output of the Sequential SubspaceEstimator (SSE) studied in [11],[12] are the desiredtemplates. These biometric templates along with refer-ence pointers will be stored (enrollment-Fig. 4) in thedatabases for the authentication process. The securityand confidentiality of the stored and dynamic biometricfeatures are dependent on their level of protection fromsecurity, privacy, and unlinkability attacks. Therefore, theobjective of this section is to present a secure, robust, andreliable encryption and authentication algorithm for theseprotections.

A. EncryptionThe cryptographic architecture of this method is de-

signed to deal with two categories of people: the au-thorized user and the subject (target). A detailed systemdiagram and processing method for generating user andsubject biometric encrypted bonds BioCryptoBond ispresented in Fig. 1.

The algorithmic architecture and formulation for cre-ating BioCryptoBond bonds have been stated below.

1) BioCryptoBondu: The steps that are involved increating the user cryptographic bond BioCryptoBondu,are stated below:(i) Extract and compute orientation angle (θ) (Fig. 2)

from received user fingerprint features.(ii) The tensor operation is performed on the user

filtered fingerprint biometric template h(t) (i.e. minutiaepoints) as a function of orientation angle.(iii) Output from the tensor operation is converted into

an orthogonal matrix Π.(iv) A digital random key Ku for the user is gen-

erated (Fig. 1) and fused with an orthogonal matrixof vectors Π, creating the user cryptographic bond,BioCryptoBondu. This cryptographic bond bindingprocess can be formulated as follows:

T = θ × �(s)

Π = Tor

BioCryptoBondu = Π×Ku (1)

where T is the output from the tensor operation; thesubscript or is the orthogonal operator; and �(s) is thefourier transformation of h(t).

146 Int'l Conf. IP, Comp. Vision, and Pattern Recognition | IPCV'15 |

Fig. 1: System Architecture -BioCryptoBond

2) BioCryptoBondFP : The steps that are in-volved in creating the subject cryptographic bondBioCryptoBondFP using (subject) h(t) fingerprint fea-tures (i.e. minutiae points (Figs. 1 and 2)) are statedbelow:(i) Randomly generated key Ks is transformed into

the orthogonal matrix Πfp.(ii) Matrix Πfp is fused with the filtered fin-

gerprint output h(t), creating the cryptographic bondBioCryptoBondFP .

This bond binding process can be formulated as fol-lows:

Πfp = Ksor

BioCryptoBondFP = Πfp ×�(s) (2)

3) BioCryptoBondF : The steps that are in-volved in creating the subject cryptographic bondBioCryptoBondF using subject facial biometrics (i.e.facial area; size and relative positions of eyes and lips(Figs. 1 and 3)) are stated below:(i) An arbitrary interface pointer β is received. This

interface pointer is generated by the system upon suc-cessful user authentication.(ii) Tensor operation is performed as a function of β

on received filtered facial biometric features h(t).(iii) Output of tensor operation is converted into the

orthogonal matrix Πf .(iv) Matrix Πf is fused with the same randomly

generated digital key Ks used to create the subject’sBioCryptoBondFP bond. This bond binding processcan be formulated as follows:

T = β ×�(s)

Πf = Tor

BioCryptoBondF = Πf ×Ks (3)

4) BioCryptoBondFF : The steps that are in-volved in creating the subject cryptographic bondBioCryptoBondFF using MultiBiometrics (the fusionof facial and fingerprint biometrics) are stated below:(i) Subject filtered fingerprint and facial biometrics are

concatenated (or fused), and a MultiBiometrics templateis created.(ii) Concatenated matrix or MultiBiometrics is con-

verted to orthogonal matrix Πff .(iii) Orthogonal matrix is fused with a randomly

generated digital secret key Ks′ (Fig. 1) and aBioCryptoBondFF bond is created.

This bond binding process can be formulated as fol-lows:

c(t) = c[h1(t) + h2(t)]

C(s) = �(c(t))

Πff = Cor(s)

BioCryptoBondFF = Πff ×Ks′ (4)

where h1(t) and h2(t) represent filtered outputs forfingerprint and facial biometrics, respectively; c(t) rep-resents the concatenate operation; and C(s) representsthe fourier transform of the concatenate operation.

B. AuthenticationIn the case of the user authentication process, fin-

gerprint biometric features received from the autho-rized user are combined with the cryptographic bond,BioCryptoBondu, and the digital secret key is released.In this stage, authentication is performed to ensure thelegitimacy of the user and to release the user secret keyKu. The user authentication process is shown in Fig. 5.

During the user authentication cycle, the same algo-rithmic operation stated in Eq. (1) is performed on thelive user fingerprint features, generating the matrix Π.

Int'l Conf. IP, Comp. Vision, and Pattern Recognition | IPCV'15 | 147

(a) Original Image

θ

(b) Ridge Orientation

Core point (185, 178)

(c) Core Point

50 100 150 200 250 300

50

100

150

200

250

300

(d) Minutiae Points

Fig. 2: Fingerprint Biometrics –Features Extraction

Fig. 3: Facial Biometrics –Features Extraction

Afterwards, Π is combined with the previously storedBioCryptoBondu to release the key Ku. This processcan be stated as follows:

T = θ ×�(s)

Π = Tor

Ku = Π×BioCryptoBondu (5)

Once the secret key is activated and released, it ishashed with the user biometric features and generates thereference pointers required to complete the final level ofauthenticity of the user. Afterwards, this reference pointeralong with the secret key allows the user to access thesystem. This process can be formulated as follows:

Ru = H[Ku

× Ip]

Required Info = Ru [dBu] (6)

where Ru is the user reference pointer.

Finally, a triggering signal is processed to initialize aninterface between user and subject, if the user authen-ticity is found positive. This interface allows the user toprepare a system platform for receiving inputted subjectbiometric features. The system also releases an interfacepointer β, which is required to ensure that the systemis ready to enroll, authenticate, and release the subject

information in the presence of the legitimate user andthe subject of interest.

IV. Biometric Data Management ArchitectureThe main objective of the biometric DMA architecture

is to enhance the security protection of the stored and dy-namic biometric features. In this case, a multilayered andMultiBiometrics data management architecture has beenproposed to protect the users’ and the subjects’ biometricfeatures. The cryptographic bonding architecture and itsprocess have already been presented in previous sections.The hash function, Hot-Key, and segmentation processesare integral parts of this management architecture, andare presented in the following subsections.

A. Hot-Key FunctionThe Hot-Key function is the compound function key

generated from a combination of the reference pointerand foreign key. The foreign key (F) is a 32-bit digitalkey generated from the primary (indexed) biometricfeatures.

The first step of this process is to create a refer-ence pointer for the user (or subject) from the systemgenerated 32 − bit digital key hashed with the primary

148 Int'l Conf. IP, Comp. Vision, and Pattern Recognition | IPCV'15 |

(a) User Enrollment

(b) Subject Enrollment

Fig. 4: Enrollment and Possible Attacks

biometric features. This reference pointer is used to storethe encrypted features (enrollment) in the user databases.In this case, the biographical information is stored in theuser database dBu and the encrypted biometric featuresare stored in the Encryptionu database. This referencepointer is used to establish a relationship between userdatabases.

In the case of a subject database, the reference key isgenerated in the same way as the user. This referencekey is hashed with the indexed foreign key generatedfrom the subject biometric features. The output of this

hash function is called the Hot-Key (Φhk) function, andits main objective is to create an extra-layer of securityfor the stored and dynamic biometric features of thesubject. A description of the subject multilayered andMultiBiometrics authentication process is not includedhere, but successful user authentication in the presenceof the subject is required. The experimental result ofthis process has been included in Section V I . Thesubject’s biographical information and biometric featuresare stored in the subject databases (dBs, EncryptionF ,EncryptionFP , and EncryptionFF ), and the generated

Int'l Conf. IP, Comp. Vision, and Pattern Recognition | IPCV'15 | 149

Fig. 5: User Authentication Process and Possible Attacks

reference pointers are used to create a link betweenthe subject databases using the reference table and datasegmentation process. The system architecture of thismethodology is presented in Figs. 4(b) and 6.

B. Segmentation ProcessThe main purpose of the segmentation process is to

cluster (or group) the subject biometric features andbiographical information based on the address pointerscreated as shown in Figs. 4(b) and 6. This clustering pro-cess uses the index biometric features of face, fingerprint,and MultiBiometrics (fusion of face and fingerprint). Inthis process, a hash key function in conjunction withthe composite foreign key and reference pointer areimplemented to construct the Hot-Key algorithm. Thedata segmentation technique along with the Hot-Keyalgorithm are employed in order to develop a securebiometric DMA architecture. A reference table is created,which serves as a link list (or address pointer) forkeeping reference addresses and locating records storedin the subject databases. The relationship between subjectdatabases is also maintained by the reference table asshown in Figs. 4(b) and 6.

V. Possible Attacks -AuthenticationThe possible attacks on the user authentication process

are shown in Fig. 5 (attacks on enrollment process are not

included here). The subject authentication is dependenton a successful user authentication process, and thetypes of attacks on it are the same. The experimentalresults of the subject (and user) authentication process arepresented in Section V I . User fingerprint biometrics isbeing used during the authentication process. If attackersare able to intervene at the sensor or communicationchannel, they still won’t be able to access the system,since the biometric features need to be transformed as afunction of the user fingerprint orientation angle beforethe authentication process occurs. Even if the attackersare able to obtain access to the system through a singlepoint, they won’t have the right to access other users’or subjects’ information, since the biometric systems areunlinkable and the physical presence of the subject isrequired along with user in order to retrieve the biometricand biographic information. The databases are protectedby multilayered encryption, hence single point accessability won’t allow the attacker to retrieve unauthorizedinformation or distinguish the identity of the subject (oruser) from the received information.

Furthermore, in this DMA architecture, the biometricinformation is segmented, and reference pointers are usedto establish a link between these segmented biometrics.In this method, it is not possible to obtain the originalbiometrics from these reference pointers and vice versa.As well, it is not possible to know the individual’s iden-

150 Int'l Conf. IP, Comp. Vision, and Pattern Recognition | IPCV'15 |

Fig. 6: Biometric Data Management Architecture

tity or construct (or guess) the original biometric featuresof an individual from the segmented biometrics storedin the databases. Databases (or information) are seg-mented and transformed, complete authorized processingis required in order to access the system. Therefore, thissystem is invincible to unlinkable attacks, and imposterscannot retrieve data based on information found in otherparts of the system.

VI. Experimental Results and DiscussionsIn this experiment, two types of authentication pro-

cesses have been performed: i) user authentication, andii) authentication and retrieval of the subject’s informa-tion. Therefore, the experimental results and resultantanalysis presented here are based on these two processes.

A. User AuthenticationIn this experiment, two user encrypted databases were

created for 30 users, then 10 users with fingerprintbiometrics from the public database “CASIA FingerprintImage Database Version 5.1”. The encrypted database setcomprised of 30 users has been used for authorized userfingerprints, and the encrypted database set comprisedof 10 users has been used for imposter fingerprints.The main objective of this process is to authenticate thelegitimacy of a user. An evaluation of the verification

performance of the encryption method is also presentedin this paper. In this case, each of the 40 users havebeen tested against the encrypted users’ biometrics storedin the databases. The performance of the verificationprocess has been evaluated based on the False Accep-tance Rate (FAR), False Rejection Rate (FRR), and EqualError Rate (EER). The experimentation results of thisverification process have been recorded in Table I , andthe graphical outcome of the FAR, FRR, and ROC arepresented in Fig. 7.

TABLE I: Performance Evaluation in (%) - FAR, FRR,and EER

Database Person FAR FRR EER

CASIA Fingerprint 40 users 1.20 3.50 2.40

Put Face 20 Subjects 1.45 8.50 4.70

Put Face 40 Subjects 1.75 9.30 5.10

Indian Face 10 Subjects 1.50 4.60 3.10

Indian Face 20 Subjects 1.86 5.40 3.45

Int'l Conf. IP, Comp. Vision, and Pattern Recognition | IPCV'15 | 151

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

False Rejection Rate

Fals

e A

ccepta

nce R

ate

Performance Evaluation −−FAR and FRR

Surveillance Zone

To Level Security

(a) FAR and FRR

0 10 20 30 40 50 60 70 80 90 1000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

ThresholdF

als

e E

rro

r R

ate

ROC Curve −Performance Evaluation

FARFRR

EER 2.4%

(b) ROC Curve

Fig. 7: User Fingerprint Biometrics -Verification Process

B. Authentication and Retrieval of the Subject’s In-formation

The performance of the proposed method has beenevaluated based on the images of these public databases:“Put Face Database” [13], “Indian Face Database” [14],and “CASIA Fingerprint Image Database Version 5.1”.The experimental results presented here are based on theauthorized users’ authentication processes using finger-print biometric features in the presence of the respectivesubjects. In this experiment, two sets of encrypted userdatabases and four sets of encrypted subject databaseshave been created from the original image databases.This experiment tested whether the subject’s informationcould be retrieved from their biometric database by legit-imate and illegitimate users, with or without the presenceof the subject. The percentages of Correct RecognitionRate (CRR), False Acceptance Rate (FAR), False Re-jection Rate (FRR), and Equal Error Rate (EER) havebeen determined, and experimental results have beenrecorded. The experimental results of this authentication(verification) process have been recorded in Table I .Simulation results of the legitimate (and illegitimate) userverification process for retrieving subject biometrics inthe presence (and without the presence) of the respectivesubjects are shown in Figs. 8. As well, the performanceof the identification process (CRR) has been recorded inTable II .

VII. ConclusionsA biometric system contains attributes that exclusively

represent an individual’s identity. These properties don’t

TABLE II: Performance Evaluation in (%) - CRR

Database 10-Subject 20-Subject 40-Subject Average

Put Face – 91.68 88.35 90.02

Indian Face 96.20 95.55 – 95.87

change and are difficult to lose or fake. The mainconcern for the exploration of the biometric system isto protect the security and privacy of these biometricfeatures. This cannot be neglected, otherwise it can revertthe overall process in the opposite direction, since thedamage to this system is irreversible and may cost morethan the system it is used for. The proposed MultiBio-metrics BioCryptoBond is secure and efficient, sincea 1.5% FAR has been achieved at the cost of 4.6%FRR. According to the experimental results, the proposedmethod is also found to be robust with a promisingEER of 3.1%. As well, BE along with the DMA archi-tecture provide multilayered protection against security,privacy, and unlinkability attacks for the dynamic andstored biometric features in the databases. It can beconcluded that the encryption method presented in thispaper is heuristic, robust, and reliable in comparisonto its counterparts. This is because, unlike other keybinding encryption systems, BE along with the biometricDMA architecture are implemented to enhance securityprotection and improve authentication accuracy. Without

152 Int'l Conf. IP, Comp. Vision, and Pattern Recognition | IPCV'15 |

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

False Rejection Rate

Fals

e A

ccepta

nce R

ate

Performance Evaluation −−FAR and FRR

Surveillance Zone

Top Level Security

(a) FAR and FRR

0 10 20 30 40 50 60 70 80 90 1000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

ThresholdF

als

e E

rror

Rate

ROC Curve −Performance Evaluation

FARFRR

EER 5.1%

(b) ROC Curve

Fig. 8: MultiBiometrics Encryption –Put Face Database(40 subjects)

a successful authentication process, neither the secret keynor the biometric features can be retrieved independentlyfrom the encrypted bonds. In addition, even if the secretkey or the transformed biometric features are interceptedat any point of operation by the imposter, the originalbiometric features are not obtainable. Finally, top levelsecurity has also been maintained for subject biometrictemplates, since the retrieval of the subject’s biometricfeatures would also require the physical presence ofthe subject along with a successful user authenticationprocess.

REFERENCES

[1] A. Jain and A. Kumar, “Biometric of next generation: Anoverview”, To Appear in Second Generation Biometrics Springer,Aug. 2010.

[2] A. Menezes, P. Oorschot, and S. Stone, “Handbook of appliedcryptography”, CRC press, Jun. 1996.

[3] A. Cavoukian and A. Stoianov, “Biometric encryption: A positive-sum technology that achieves strong authentication, security andprivacy”, Information and Privacy Commissioner Ontario, Mar.2007.

[4] A. Cavoukian and T. Marinelli, “Privacy-protective facial recog-nition: biometric en- cryption proof of concept”, Information andPrivacy Commissioner, Ontario, Canada. Nov. 2010.

[5] K. Martin, H. Lu, F. Bui, K. Plataniotis, and D. Hatzinakos,“Abiometric encryption system for the selfexclusion scenario of facerecognition”, IEEE Systems Journal: Special Issue on BiometricsSystems, vol. 3, no. 4, pp. 440-450, Mar. 2009.

[6] K. Nandakumar and A. Jain, “Multibiometric template security us-ing fuzzy vault, Proceedings of 2nd IEEE International Conferenceon Biometrics: Theory, Applications, and Systems”, pp. 1-6, Jun.2008.

[7] A. Ross, K. Nandakumar, and A. Jain, “Handbook of multibiomet-rics”, Springer, Chapter 2, Mar. 2006.

[8] C. Lee, J.Choi, K. Toh, S. Lee, and J. Kim, “Alignment-free can-celable fingerprint templates based on local minutiae information”,IEEE Transactions on Systems, Man and Cybernetics, Part B, vol.37, no. 4, pp. 980-992, Oct. 2007.

[9] D. Maio and L. Nanni, “Multihashing, human authenticationfeaturing biometrics data and tokenised random number: a casestudy”, Elsevier Neurocomputing, vol. 69, no. 1, pp. 242-249, Jun.2006.

[10] A. Teoh, T. Connie, O. Ngo, and C. Ling, Remarks on biohashand its mathematical foundation, Information Processing Letter,no. 4, pp. 145-150, Sep. 2006.

[11] O. Malek, A. Venetsonoupoulous, D. Androutsos, and L. Zhao,“Sequential subspace estimator for biometric authentication”,ELSEVIER-Neurocomputing, vol. 148, pp. 294-309, Jan. 2015.

[12] O. Malek, A. Venetsonoupoulous, D. Androutsos, and L. Zhao,“Subspace state estimator for facial biometric verification”, IEEEProceedings of The International Conference on ComputationalScience and Computational Intelligence, Las Vegas, USA, vol. 1,pp. 137-143, Mar. 2014.

[13] A. Kasiski, A. Florek, and A. Schmidt, “The PUT Face Database”,Image Processing and Communications, vol. 13, no. 3-4, pp. 5964,Aug. 2008.

[14] V. Jain and A. Mukherjee, “The Indian Face Database”,http://www.cs.umass.edu/ vidit/IndianFaceDatabase/, Jun. 2002.

Int'l Conf. IP, Comp. Vision, and Pattern Recognition | IPCV'15 | 153