Research Article Face Spoof Attack Recognition Using...

Research ArticleFace Spoof Attack Recognition UsingDiscriminative Image Patches

Zahid Akhtar and Gian Luca Foresti

Department of Mathematics and Computer Science University of Udine Via delle Scienze 206 33100 Udine Italy

Correspondence should be addressed to Zahid Akhtar zahidakhtaruniudit

Received 3 December 2015 Revised 12 March 2016 Accepted 14 April 2016

Academic Editor Kwok-Wai Cheung

Copyright copy 2016 Z Akhtar and G L Foresti This is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work is properlycited

Face recognition systems are now being used in many applications such as border crossings banks and mobile payments Thewide scale deployment of facial recognition systems has attracted intensive attention to the reliability of face biometrics againstspoof attacks where a photo a video or a 3D mask of a genuine userrsquos face can be used to gain illegitimate access to facilities orservices Though several face antispoofing or liveness detection methods (which determine at the time of capture whether a faceis live or spoof) have been proposed the issue is still unsolved due to difficulty in finding discriminative and computationallyinexpensive features and methods for spoof attacks In addition existing techniques use whole face image or complete video forliveness detection However often certain face regions (video frames) are redundant or correspond to the clutter in the image(video) thus leading generally to low performances Therefore we propose seven novel methods to find discriminative imagepatches which we define as regions that are salient instrumental and class-specific Four well-known classifiers namely supportvectormachine (SVM) Naive-Bayes Quadratic Discriminant Analysis (QDA) and Ensemble are then used to distinguish betweengenuine and spoof faces using a voting based scheme Experimental analysis on two publicly available databases (Idiap REPLAY-ATTACK and CASIA-FASD) shows promising results compared to existing works

1 Introduction

In the last years face recognition systems have gained interestdue to facersquos rich features that offer a strong biometric cueto recognize individuals for a wide variety of applications inboth law and nonlaw enforcements [1] In fact facial recog-nition systems are already in operation worldwide includingUSVISIT which is a US Customs and Border Protection(CBP) management system UIDAI that provides identity toall persons resident in India andMicrosoftKinect which usesface recognition to access dashboard and automatic sign-in toXbox Live profile Similarly face biometrics is also nowadaysbeing used ubiquitously as an alternative to passwords onmobile devices such as Android KitKat mobile OS LenovoVeriFace Asus SmartLogon and Toshiba SmartFace

Despite the great deal of progress in facial recognitionsystems vulnerabilities to face spoof attacks are mainlyoverlooked [2] Facial spoof attack is a process in which afraudulent user can subvert or attack a face recognition sys-tem by masquerading as registered user and thereby gaining

illegitimate access and advantages [1 3ndash5] Face spoofingattack is a major issue for companies selling face biometric-based identity management solutions [6] For instance atNew York City nonwhite robbers disguised themselves aswhite cops using life-like latex masks and were caughtrobbing a cash-checking store in 2014 (see Figure 1 also forother recent face spoof attacks)

Recent study reported in [1] suggests that the successrate of face spoof attacks could be up to 70 even whena state-of-the-art Commercial Off-The-Shelf (COTS) facerecognition system is used Therefore we could infer thateven COTS face recognition systems are mainly not devisedto effectively distinguish spoof faces from genuine live facesAs a matter of fact this vulnerability of face spoofing toface recognition systems is now enlisted in the NationalVulnerability Database of the National Institute of Standardsand Technology (NIST) in the US

Typical countermeasure to face spoof attacks is livenessdetection method which aims at disambiguating humanlive face samples from spoof artifacts [2 7] There exist

Hindawi Publishing CorporationJournal of Electrical and Computer EngineeringVolume 2016 Article ID 4721849 14 pageshttpdxdoiorg10115520164721849

2 Journal of Electrical and Computer Engineering

(a) (b)

Figure 1 (a) In New York a robber disguised as white cop using latex masks robbing a cash-checking store (b) In 2010 a passenger boardeda plane in Hong Kong with an old man mask and arrived in Canada to claim asylum

several face antispoofing or liveness detection techniques [7ndash15] However face spoofing attacks remain a problem dueto difficulties in finding discriminative and computationallyinexpensive features and techniques for spoof recognitionMoreover publishedmethods are limited in their scope sincethey mainly use whole face image or complete video forliveness detection Nevertheless often certain face imageregions (video frames) are redundant or correspond to theclutter in the image (video) leading thus generally to lowperformances

It is thus essential to develop robust efficient andcompact face antispoofing (or liveness detection) methodswhich are capable of generalizing well to discriminativeclass-specific information and imaging conditions To thisaim in this paper we propose a simple and effective solutionbased on discriminative image patches In particular wepropose seven novel fully automated algorithms to highlightregions of interest in face images We define these regions(or image patches) to be discriminative (ie specific toa particular class live or spoof) consistent (ie reliablyappearing in different face images or video frames) salient(ie conspicuous regions) and repetitive (ie frequentlyappearing in the image set of specific class) The basic notionis ldquointeresting patches are those that are specific to a faceimage (or video frame) and should contain features thatgive assistance to discriminate a given live face image fromspoofed onerdquo Based on this definition two of the sevenproposed image patch selectionmethods (ieMAXDIST andDEND-CLUSTER) donot employ any training or prior learn-ing However the remaining techniques use simple clustering(ie CP and CS) image intensity (ie IPI) image quality(ie IQA) or diversity filter (ie DF) to obtain discriminativepatches For final classification we exploited four well-known classifiers namely support vector machine (SVM)Naive-Bayes Quadratic Discriminant Analysis (QDA) andEnsemble using voting based scheme Experimental analysison two publicly available databases (Idiap Replay-Attackand CASIA-FASD) shows good results compared to existingworks

The added advantages of the proposed framework are(i) being cheap (ii) very low complexity (iii) needing oneface image (ie the same face sample acquired for facerecognition) to detect whether it is genuine or spoof attack(iv) being nonintrusive (v) being user-friendly and (vi) beingeasy to embed in already functional face recognition systemsplus no requirement of new piece of hardware

The reminder of the paper is organized as follows Exitingliterature works on face liveness detection are discussed inSection 2 The proposed approaches to determine the dis-criminative image patches and spoof classification schemesare described in Section 3 Experimental datasets protocolsand results are presented in Section 4 A conclusion is drawnin Section 5

2 Related Work

Despite great deal of advancements in face recognitionsystems face spoofing still poses a serious threat Most of theexisting academic and commercial facial recognition systemsmay be spoofed by (see Figure 2) (i) a photo of a genuine user(ii) a video of a genuine user (iii) a 3D face model (mask) ofa genuine user (iv) a reverse-engineered face image from thetemplate of a genuine user (v) a sketch of a genuine user (vi)an impostor wearing specific make-up to look like a genuineuser (vii) an impostor who underwent plastic surgery to looklike a genuine user The most easiest cheapest and commonface spoofing attack is to submit a photograph of a legitimateuser to the face recognition systems which is also known asldquophoto attackrdquo

Typical countermeasure (ie face liveness detection orantispoofing) techniques can be coarsely classified in threecategories based on clues used for spoof attack detection(i) motion analysis based methods (ii) texture analysis basedmethods and (iii) hardware-based methods In what followswe provide a brief literature overview of published face spoofrecognition techniques along with their pros and cons

(i) Motion Analysis Based Methods These methods broadlytry to detect spontaneous movement clues generated whentwo dimensional counterfeits are presented to the camera ofthe system such as photographs or videos Therefore Pan etal [7] exploited the fact that human eye-blink occurs onceevery 2ndash4 seconds and proposed eye-blink based livenessdetection for photo-spoofing using (spontaneous) eye-blinksThis method uses an undirected conditional random fieldframework to model the eye-blinking which relaxes theindependence assumption of generative modelling and statesdependence limitations from hidden Markov modelling Itis evident that real human faces (which are 3D objects)will move significantly differently from planer objects andsuch deformation patterns can be employed for livenessdetection For example Tan et al [8] considered Lambertian

Journal of Electrical and Computer Engineering 3

(a) (b)

(c) (d)

(e)

(f) (g)

Target Reconstructed

Figure 2 Examples of face spoofing using (a) photograph (b) video (c) 3Dmask (d) sketch (e) reverse-engineered face image (f) make-up(skillful application of make-up to look like Michel Jackson) and (g) plastic surgery (this boy underwent excessive plastic surgery to look likeJustin Bieber)

reflectance model with difference-of-Gaussians (DoG) toderive differences of motion deformation patterns between2D face photos presented during spoofing attacks and 3D livefaces It does so by extracting the features in both scenariosusing a variational retinex-based method and difference-of-Gaussians (DoG) [9] based approach Then the features areused for live or spoof classification Reported experimentsshowed promising results on a dataset consisting of realaccesses and spoofing attacks to 15 clients using photo-qualityand laser-quality prints Kollreider et al [10] proposed a live-ness detection approach based on a short sequence of imagesusing a binary detector which captures and tracts the subtlemovements of different selected facial parts using a simplifiedoptical flow analysis followed by a heuristic classifier Thesame authors also presented a method to combine scoresfrom different experts systems which concurrently observethe 3D face motion approach introduced in the former workas liveness attributes like eye-blinks or mouth movements Inthe similar fashion Bao et al [11] also used optical flow toestimate motion for detecting attacks produced with planarmedia such as prints or screens

Since the frequency of facial motion is restricted by thehuman physiological rhythm thus motion based methodstake a relatively long time (usually gt 3 s) to accumulate stablevitality features for face spoof detection Moreover they maybe circumvented or confused by other motions for examplebackground motion in the video attacks

(ii) Texture Analysis Based Methods This kind of methodsexamines the skin properties such as skin texture and skinreflectance under the assumption that surface properties ofreal faces and prints for example pigments are differentExamples of detectable texture patterns due to artifacts areprinting failures or blurring Li et al [13] described a methodfor print-attack face spoofing by exploiting differences in the2D Fourier spectra of live and spoof images The methodassumes that photographs are normally smaller in size andcontain fewer high frequency components compared to real

faces The method only works well for downsampled photosof the attacked identity but likely fails for higher-qualitysamples In [14 16 17] authors developed microtextureanalysis based methods to detect printed photo attacksOne limitation of presented methods is the requirementof reasonably sharp input image Recently Galbally et al[3] designed a face spoof detection scheme based on 25different image quality measures 21 full-reference measuresand 4 nonreference measures However all 25 image qualitymeasures are required to get good results and no face-specificinformation has been considered in designing informativefeatures for face spoof detection

Compared to other techniques texture analysis basedalgorithms are generally faster to classify a spoof attackNevertheless they could be easily overfitted to one particularillumination and imagery condition and hence do not gener-alize well to different spoofing conditions

(iii) Hardware-Based Methods Few interesting hardware-based face antispoofing techniques have been proposed so farbased on imaging technology outside the visual spectrumsuch as 3D depth [18] complementary infrared (CIR) ornear infrared (NIR) images [15] by comparing the reflectanceinformation of real faces and spoof materials using a specificset-up of LEDs and photodiodes at two different wavelengthsPreliminary efforts on thermal imaging for face livenessdetection have also been exploited including the acquisitionof large database of thermal face images for real and spoofedaccess attempts [19] Besides numbers of researchers haveexplored multimodality as antispoofing techniques for facespoofing attacks They have mainly considered the combina-tion of face and voice by utilizing the correlation betweenthe lips movement and the speech being produced [20]where a microphone and a speech analyzer were requiredSimilarly challenge-response strategy considering voluntaryeye-blinking and mouth movement following a request fromthe system has been studied in [10] Though hardware-based methods provide better results and performances they


Table 1 Summary of different face spoof detection techniques

Method Main features used Pros and cons

Motion analysis based methods

Motion detection [10]Eye-blink detection using conditional random fields(CRF) [7]Face motion detection using Optical Flow Lines (OFL) [10]Context-based using correlation between face motion andbackground motion [17]

(i) Good generalization capability(ii) High computational cost(iii) Easily circumvented by fake motions

Texture analysis based methodsFace texture using Lambertian model [8]Face texture using LBPs [17]Texture + shape combining LBPs + Gabor Wavelets +HOG [5]

(i) Fast response(ii) Low computational cost(iii) Poor generalization capability

Hardware-based methodsMultimodality face and voice [20]Thermal images [19]Reflectance in 3D [21]

(i) Better generalization capability(ii) Extra hardware requirement(iii) High cost of the system

Face detectionand

normalization

Patch gridcreation and

selection of top Kdiscriminative

patches

Classifier

Rest of the facerecognition

system

An input video

Randomly selected frame

Normalized face image

Selected patches

Fusion

Selectedpatch 1

Selectedpatch K

Live

Spoof

Finaldecision

Labelselected patch 1

Labelselected patchK

Figure 3 The proposed face spoof recognition algorithm based on discriminative image patches

require extra piece of hardware which increases the cost ofthe system A summary with relevant features of the mostrepresentative works in face antispoofing is presented inTable 1

Though there exist several face antispoofing or livenessdetection techniques face spoof attacks remain an issuebecause of difficulty in finding discriminative and compu-tationally inexpensive features and mechanisms for spoofrecognition Reported methods are limited in their scopesince theymainly use full image or complete video for livenessdetection In particular there is a lack of investigation onhow specific image patches rather than full image performin spoof detection As often image regions (video frames)are redundant or correspond to the clutter in the image(video) leading thus generally to low performances and highcomputational cost towards this direction we propose sevennovel methods to find discriminative image patches whichwe define as regions that are salient instrumental and class-specific Experimental results show that the proposed meth-ods obtain comparable performances to existing techniques

3 The Proposed Method forFace Spoof Recognition

Face spoof detection can be seen as a two-class classificationproblem where the input face image (or video) has to be

flagged as either live or spoof The keynote of the processis attaining a discriminant feature set together with anappropriate classification scheme that gives the probability ofthe image (or video) realism Practical face spoof detectionrequires that a decision be made based on single imageor a limited number of frames in the video-based systemIn this work thus we aim to design simple but effectivesolution based on discriminative image patches using a singleface frameimage We define these image patches to bediscriminative consistent salient and repetitive The notionis that the interesting patches are those that are specific to aface image (or video frame) and should contain features thathelp discriminate a given live face image from spoofed one

Figure 3 shows the schematic diagram of the proposedface spoof recognition algorithm based on discriminativeimage patches The proposed framework first randomlyselects a single frame from a given face video (in caseof image-based system the given single image is used)Then face is detected using Local SMQT Features andSplit-Up Snow Classifier [30] Subsequently the detectedface is densely divided into a grid of nonoverlapping localpatches These patches are ranked based on their discrimi-native power The top 119870 patches are selected using specificdiscriminative image patch selection method among theproposed techniques (explained below in detail) For eachselected image patch features are extracted that are then fedinto particular classifier (ie SVM Naive-Bayes QDA or


Discriminative imagepatches selection

DEND-CLUSTERING

CP(Cluster Pairing)

CS(cluster space) MAXDIST

IQA(image quality

assessment)

DF(diversity filter)

Nonclustering basedmethodsClustering based methods

IPI(intensity-based patch of interest)

Figure 4 Classification of the seven discriminative image patches selection methods proposed in this work

Ensemble classifier) The classification results of individualpatches are combined by a majority-voting based scheme toobtain the final binary decision genuine or spoof face

31 Discriminative Image Patches Selection Methods In whatfollows we give the details of the proposed sevenmethods fordiscriminative image patches selection The proposed patchselectionmethods are grouped into two categories clusteringbasedmethods and nonclustering basedmethods For clarityin Figure 4 we show a diagram with the patch selectionmethods classification followed in this section

311 Clustering BasedMethods Thepatch selectionmethodsin this category rely on a clustering algorithm at any specificstage of the procedure Three clustering based methodsproposed in this work are as follows

(1) DEND-CLUSTERING In this discriminative imagepatches selection technique the 119899 patches in the givenface image are grouped into T clusters such that patcheswithin a cluster are more similar to each other than patchesbelonging to different clusters Then for each cluster aprototype (representative) patch that typifies the members ofthat cluster is chosen resulting in T discriminative patchesSince this method uses the dendrogram [31] to choose thediscriminative patches thus we call it DEND-CLUSTERING

For each normalized face image (I) first a set of densepatches P

119894(I) isin R119872times119873

119899

119894=1is taken where 119899 is the total

number of dense patches Computation of the dissimilarityscores between patches is needed to perform clusteringTherefore first the dissimilarity between successive patches iscomputed by comparing the respective featuresThen hierar-chical clustering [31] is exploited because the representationof the 119899 patches is in the form of an 119899 times 119899 dissimilarity matrixinstead of an 119899times119901 patternmatrix where119901 is the dimension offeature vector In particular an agglomerative complete linkclustering algorithm [31] is used in this work The outcomeof this algorithm is a dendrogram (a binary tree) where eachterminal node corresponds to a patch and the intermediatenodes indicate the formation of clusters The discriminative119870 patches are selected as follows

(i) Find the pairwise distance scores between the 119899

patches to form the dissimilarity matrixD(ii) Apply the complete link clustering algorithm on D

and generate the dendrogram L Use the dendrogramL to identify T clusters The method in [31] automat-ically determines the threshold distance to cut thedendrogram and identify exactly T clusters

(iii) In each of the clusters identified in step (ii) selecta patch whose average distance from the rest of thepatches in the cluster is minimum If a cluster hasonly 2 patches choose any one of the two patches atrandom

(iv) The patches selected in step (iii) are arranged indescending order on the basis of their ideal selectionmeasure (ISM) value which is computed as

ISM (P) =

119875minus2

sum

119909=1

119876minus2

sum

119910=1

119866 (119909 119910) (1)

where P is a patch of size 119875 times 119876 and 119866(119909 119910) is theimage gradient at location (119909 119910)

(v) The top 119870 patches are selected as discriminativepatches

It is worth mentioning that steps (i)ndash(iii) in DEND-CLUSTERING method have close resemblance with thetechnique in [31] for fingerprint template selection Here weextended the technique by proposing step (iv) to be utilizedfor ranking and selection of discriminative patches

(2) CP (Cluster Pairing) Apparently the discriminationpower of patches (features) decidesmaximumpossible classi-fication accuracy and thus prior knowledge of ldquohow clutteredthe features (patches) may berdquo and ldquotheir contribution toclasses separability in the feature spacerdquo can help to designand accomplish better classification scheme and accuracyTo this aim in this method first two independent setsof clusters are generated using genuine and spoof attacksamples respectively Since overlapping of interclass clustershas great effect on classification accuracy therefore pairs


of overlapped corresponding clusters of two independentsets are formed using minimum distance between themFinally patches which do not belong to both clusters of agiven interclass highly overlapped cluster pair are selected asdiscriminative patches In other words if a patch belongs toboth clusters of a given interclass cluster pair it means thatits features cause overlapping different classes in the featurespace which might thus lead to lower classification accuracyThe steps followed to obtain top 119870 discriminative patchesusing this method are as follows

(i) Two independent sets of clusters are generated usinglive and spoof attack training samples respectivelyEach class is represented by 119883 number of clustersscattered in the feature space 119870-means clusteringtechnique is exploited in thiswork for cluster creation

(ii) All possible pairs of corresponding clusters of twoindependent sets are formed using

119863mean (119862119894 119862119895) =

10038171003817100381710038171003817119898119894

minus 119898119895

10038171003817100381710038171003817

119883

119895=1le 120578 (2)

where 120578 is threshold 119898119909is center of 119862

119909 119862119894is a

given cluster from live class clusters set and 119862119895is

a given cluster from spoof class clusters set 119863meanis appropriate for detecting spherical and compactcluster pairs since each cluster is represented only byits center point

(iii) For a given face image 119870 patches are chosen asdiscriminative patches which do not belong to bothelements of the interclass clusters

(3) CS (Cluster Space) Principally information classes cannotbe described efficaciously by a single well-defined groupingin a spectral space Thus it is better to represent them by agroup of spectral classes (clusters) which is prime inferenceof thismethod It is worth noting that thismethod is identicalto the above-mentioned Cluster Pairing (CP) method Inthis method 119883 number of clusters are generated using bothlive and fake labeled training patches together we nameresulting spectral space as cluster space For each clusteran individual classifier (IC) is trained hence resulting in119883 number of individual classifiers Given an image patchits discriminative value (DV) is computed as an averageof the probabilities given by all ICs Later the patches aresorted based on their DV with respect to other patchesFinally patches corresponding to 119870 largest DV values areselectedThismethod (cluster space + IC) provides ameans ofoptimizing the variance and correlation present in all classesand samples Following are the steps executed to designatetop discriminative patches

(i) Using training datasetrsquos labeled patches 119883 number ofclusters are generated using both live and fake samplestogether119870-means clustering algorithm is exploited inthis work for cluster creation

(ii) For each cluster an individual classifier (IC) (in thiswork SVM) is trained using ground truth

(iii) The patches of a given face image are arranged indescending order on the basis of their respective DV

DV = exp(

1

119883

119883

sum

119894=1

119875119894 (P)) (3)

where 119875119894is the probability given by 119894th classifier

trained on 119894th cluster and P is the candidate patch(iv) The top 119870 patches are selected as discriminative

patches

312 Nonclustering Based Methods Unlike the clusteringbasedmethods techniques in this category do not require anyclustering algorithm Following are the four nonclusteringbased discriminative patch selection methods

(1) MAXDIST This method of discriminative patches selec-tion is based on the assumption that candidate discriminativepatches are maximally dissimilar from the other patchesin the given face image Therefore first the dissimilaritybetween successive patches is computed The method thensorts the patches based on their average dissimilarity scorewith respect to other patches and selects those patches (iediscriminative patches) that correspond to the 119870 largestaverage dissimilarity scores We refer to this method asMAXDIST since discriminative patches are selected using amaximum dissimilarity criterion

Following steps are followed to select top discriminativepatches

(i) An 119899 times 119899 dissimilarity matrix (D) is generated whereeach element D(119894 119895) 119894 119895 isin 1 2 119899 is the distancescore between features of patches 119894 and 119895

(ii) For the 119895th patch the average dissimilarity score (119889119895)

with respect to the remaining (119899 minus 1) patches iscomputed by finding the average of the elements in119895th row ofD

(iii) The average values obtained in step (ii) are orderedin descending order and the top 119870 patches that havethe largest average dissimilarity scores are selectedas discriminative patches since they are the mostldquodissimilarrdquo patches in the image and hence they arerepresenting typical data measurements

For classification performance point of view smaller 119870

values might not be able to sufficiently seize the inter- andintraclass variability whichmay lead to inferior performanceHowever larger 119870 values on the other hand would becomputationally demanding Thus a rational value of 119870by taking above-mentioned factors into account has to bespecified A similar method has been proposed in [32] forpeak frame selection in a given facial expression videoUnlike [32] in this work we employ the technique to selectdiscriminative patches in an imageframe

(2) IQA (Image Quality Assessment) This algorithm usesimage quality measures to select discriminative patches thusnamed as image quality assessment (IQA) The method


assumes that the discriminative patches will have differentquality fromnondiscriminative patchesThe expected qualitydifferences between discriminative and nondiscriminativepatches may be local artifacts color levels luminance levelsdegree of sharpness entropy structural distortions or naturalappearance

This framework exploits 4 general reference-based imagequalitymeasures thus having a very lowdegree of complexityFirst four distinct label-sets for all patches are obtained usingfour different image qualitymeasuresThe labels are ldquodiscrim-inativerdquo and ldquonondiscriminativerdquo The method selects onlythose patches which are flagged as discriminative by all four-image quality assessment techniques

In particular reference-based IQAmethods are employedin this scheme that rely on the availability of a clean undis-torted reference image to estimate the quality of the testimage However in typical spoof recognition such a referenceimage is unknown because only the input sample is availableTherefore to circumvent this limitation the same technique(filtering the image with a low-pass Gaussian kernel) thatis successfully being used for image manipulation detection[33] and for steganalysis [34] is utilized Following steps areexecuted to attain top 119870 discriminative patches

(i) The normalized face image (I) is filtered with a low-pass Gaussian kernel in order to generate a smoothedversion I

(ii) Two corresponding sets of dense patches P119894(I) isin

R119872times119873119899

119894=1and P

119895(I) isin R119872times119873

119899

119895=1are taken where

119899 is the total number of patches

(iii) Four label matrices (LPSNR LNCC LTED and LGMSD)using following four-image quality measures (IQM)are generated The patches are flagged as ldquodiscrimi-nativerdquo if their IMQ is greater than or equal to thethreshold The image quality measures are as follows

(a) Peak Signal to Noise Ratio (PSNR) It com-putes the distortion between two correspondingpatches (of size119875times119876) on the basis of their pixel-wise differences as follows

PSNR (P P) = 10 log[

max (P2)

MSE (P P)

] (4)

where

MSE (Mean Squared Error)

=

1

119875119876

119875

sum

119909=1

119876

sum

119910=1

(P119909119910

minus P119909119910

)

2

(5)

(b) Normalized Cross-Correlation (NCC) The cor-relation function can also be used to quantifythe similarity between two digital image patches[3] Here a variant of correlation basedmeasureis obtained by considering the statistics of the

angles between the pixel vectors of the originaland distorted patches as

NCC (P P) =

sum119875

119909=1sum119876

119910=1(P119909119910

sdot P119909119910

)

sum119875

119909=1sum119876

119910=1(P119909119910

)

2 (6)

(c) Total Edge Difference (TED) Edge features aregenerally considered as one of the most infor-mative datasets in computer vision algorithmsThus we considered edge-related quality mea-sure since the structural distortion of an imageis deeply linked with its edge degradation TheTED measure is calculated as follows

TED (P P) =

1

119875119876

119875

sum

119909=1

119876

sum

119910=1

100381610038161003816100381610038161003816PE119909119910

minus PE119909119910

100381610038161003816100381610038161003816 (7)

In this work we use Sobel operator to build thebinary edge maps PE and PE

(d) Gradient Magnitude Similarity Deviation(GMSD) GMSD finds the pixel-wise GradientMagnitude Similarity (GMS) between thereference and distorted patches It uses apooling strategy based on standard deviation ofthe GMS map to predict accurately perceptualimage qualityTheGMSD is obtained as follows

GMSD (P P) = radic1

119880

119880

sum

119894=1

(GMS (119894) minus GMSM)2 (8)

where 119880 is the total number of pixels in thepatch GMS is Gradient Magnitude Similar-ity map and GMSM is Gradient MagnitudeSimilarity Mean calculated by applying averagepooling to the GMSmapTheGMS is calculatedas

GMS (119894) =

2m119903 (119894)m119889 (119894) + 119888

m2119903

(119894) + m2119889

(119894) + 119888

(9)

where 119888 is a positive constant that suppliesnumerical stability while m

119903and m

119889are gra-

dient magnitude images obtained by convolu-tion using Prewitt filters along horizontal andvertical directions respectively The GMSM isobtained as

GMSM =

1

119880

119880

sum

119894=1

GMS (119894) (10)

Clearly a higher GMSM score means higherimage qualityWe refer reader to [35] for furtherdetails of GMSD technique

(iv) The patches flagged as discriminative by all abovefour-image quality assessment techniques areselected

LPSNR cap LNCC cap LTED cap LGMSD (11)


(v) The patches selected in step (iv) are arranged indescending order on the basis of their average valuesof (4) (6) (7) and (8)

(vi) The top 119870 patches are selected as discriminativepatches

(3) DF (Diversity Filter) This method considers that thediscriminative patches are discernibly diverse from the otherpatches in the given face image The discriminative patchesare obtained using the combination of a trained classifier(we refer to such classifier as diversity filter) and a selectionprocedure that selects the patches based on their incrementalcontribution to the discriminative patch detection perfor-mance This method is close to object detection methods[36 37] where similar procedure is carried out to capturevisually varied parts of objects at a fixed pose or viewpointUnlike the proposed method techniques in [36 37] usebefore diversity filter a preprocessing step that is creatingvarious clusters corresponding to different parts appearancesin images Also distinct diversity filter is employed for eachsuch cluster while our method single diversity filter is usedfor all samples and image patches without any clusteringMoreover the diversity tradeoff parameter in our method iscomputed dynamically for each image whereas in [36 37] astatic value is utilized for all images

In particular we learn the diversity model of patchesbased on their properties that can be computed from the filteritself The intuition is that across image categories good filterexhibits common traits such as low clutter and gradients thatare spatially correlatedTherefore we train a ranking functionwith the objective to produce the order of diversity qualityof patches The function uses a weight to control tradeoffbetween estimated rank of a patch and the diversity it addswhich discourages adding patches similar to the ones alreadyselected even if this patch is highly ranked Following are thesteps required to select 119870 discriminative patches

(i) Using training dataset a classifier (diversity filter hereSVM) is trained

(ii) The patches (119899) of a given face image are arranged indescending order on the basis of following equation

argmax119894

119894

minus 120572max1119905

119878119894119895

(12)

where 119894is the probability given by diversity filter 119878

119894119895

denotes similarity between patches 119894 and 119895 and 120572 =

max(eigenvalues (cov(119878119894119895

))) is the diversity tradeoffparameter It is worth mentioning that in (12) duringthe first iteration 119905 = 119899 (total number of patches)and then in each successive iteration 119905 is reduced by1 such that the patch selected in foregoing iteration isremoved

(iii) The top 119870 patches are selected as discriminativepatches

(4) IPI (Intensity-Based Patch of Interest) Local intensityinhomogeneity can be exploited to find the regions shapes

and edges of similar kind in an image [38] However our aimhere is to disregard the image patches (regions) with similarfeatures in order to avoid redundancy Therefore in thismethod to determine the discriminative patches we applyan approach identical to standard statistical background-subtraction approach (which is most commonly used toaccost intensity inhomogeneity) [39] The proposed methoddoes not use any preprocessing step that is foregroundand background models based on recursive or nonrecursivetechniques like in [39] Following steps are executed to attain119870 discriminative patches

(i) A set of dense patches P119894(I) isin R119872times119873

119899

119894=1are taken

where 119899 is the total number of patches (of size 119875 times 119876)

(ii) A label matrix (FIPI) is generated using a standardstatistical background-subtraction approach

FIPI

=

Discriminative ifsum119875

119909=1sum119876

119910=1

1003816100381610038161003816P119894(119909 119910) minus P

119894

1003816100381610038161003816

120590 (P119894)

gt 120578

Nondiscriminative otherwise

(13)

where 120578 is threshold which is estimated using similarprocedure as explained above in IQA method

(iii) The patches flagged as discriminative in step (ii) arearranged in descending order on the basis of theirvalues using (13)

(iv) The top 119870 patches are selected as final discriminativepatches

32 Classification Method For final classification whetherthe face is genuine or spoof we used majority-voting basedscheme that exploits four well-known classifiers support vec-tor machine (SVM) Naive-Bayes (NB) Quadratic Discrim-inant Analysis (QDA) and Ensemble based on AdaBoostalgorithm

4 Experiments

In this section we evaluate the proposed approach on twodifferent publicly available databases REPLAY-ATTACK [4]and CASIA-Face Antispoofing Database (FASD) [22]

41 Datasets

411 REPLAY-ATTACK [4] This dataset is composed ofshort videos of both real-access and spoofing attempts of50 different subjects acquired with a 320 times 240 resolutioncamera The datasets were collected under controlled (witha uniform background and artificial lighting) and adverse(with natural illumination and nonuniform background)conditions The face spoof attacks were created by forginggenuine verification attempts of the respective subjects viaprinted photos displayed photosvideos on mobile phonersquosscreen and displayed photosvideos on HD screen


Table 2 Summary of two databases used in this study

Database Number of subjects Number of videos Resolution Attack type

REPLAY-ATTACK [4] 50 (i) 200 live(ii) 1000 spoof 320 times 240

(i) Printed photo(ii) Displayed photo (mobileHD)(iii) Replayed video (mobileHD)

CASIA-FASD [22] 50 (i) 150 live(ii) 450 spoof

640 times 480Dagger

480 times 640oplus

1280 times 720⋆

(i) Printed photo(ii) Cut photo(iii) Replayed video

Dagger oplus and ⋆ indicate low- normal- and high-quality camera

412 CASIA-FASD [22] This database contains videorecordings of real and fake faces for 50 different identitiesBoth real-access and spoof attacks were captured usingthree camera resolutions low resolution normal resolutionand high resolution Three kinds of attack attempts wereconsidered warped photo attacks cut photo attacks andvideo attacks The dataset is divided into two subsets fortraining and testing 20 and 30 identities respectively Table 2provides a summary of the above two databases

42 Evaluation Protocols For REPLAY-ATTACK dataset wefollowed the same standard protocols specified in [4] for theexperiments The dataset contains three totally independentdatasets in terms of users The train and development setsare used for training and parameter tuning respectively Thefinal results are computed on test The performance of theproposed liveness detection system was evaluated as in [3 417] usingHalf Total Error Rate (HTER) computed asHTER=(FAR+FRR)2 where FRR and FAR stand for False RejectionRate and False Acceptance Rate respectively

For CASIA-FASD database we followed the benchmarkprotocols specified in [22]The test protocol consists of sevenscenarios The first three scenarios are to study the effect ofimaging quality (1) low quality (2) normal quality and (3)

high quality The next three scenarios are (4) warped photoattacks (5) cut photo attacks and (6) video attacks Finally(7) is the overall scenario (here all data are combined togetherto give a general and overall evaluation) The classifiertraining and parameters tuning were performed on trainingset while the results are reported in terms of Equal Error Rate(EER) on the test set

In a given video frame first the face is detected Thedetected face image is then normalized to 128 times 128 pixels anddensely divided into a grid of nonoverlapping local patchesof size 16 times 16 Out of total 119899 number of patches only 40peculiar patches are selected as discriminative patches TheLBP (Local Binary Patterns) technique is utilized to extractthe features both for final classification and for discriminativepatch selection methods (to obtain dis(similarity) scoreand clustering) Figure 5 shows examples from REPLAY-ATTACK and CASIA-FASD database of a given face imageand corresponding selected discriminative patches usingproposed seven image patches selection methods

43 Experimental Results The experimental results onREPLAY-ATTACK and CASIA-FASD databases are reported

in Tables 3 and 4 respectivelyWe compared the performanceof proposedmethodwithmost eminent techniques publishedin the literature such as methodologies in [4] (based on localbinary pattern features with SVM classifier) [22] (groundedon multiple difference-of-Gaussian (DoG) filters to extractthe high frequency information) [23] (usingMultiscale LocalBinary Patterns with SVM) [3] (which makes use of generalfull-reference and nonreference image quality measures)[16] (exploiting correlation between head motion and back-ground that is estimated using optical flow) [28] (encodinginformation with a Histogram of Oriented Optical Flow(HOOF)) [23] (utilizing both texture and motion estimationalong with preprocessing for motion magnification) and[1] (based on image distortion analysis features which iscombination of specular reflection blurriness chromaticmoment and color diversity properties)

The results in both Tables 3 and 4 show that theproposed method in general achieves better accuracy thanexisting techniques under specific combination of discrim-inative patch selection method and classification schemeFor instance in the experiment using the REPLAY-ATTACKdatabase it is easy to see in Table 3 that when the proposedframework is implemented using DEND-CLUSTERING-Ensemble or MAXDIST-Ensemble combinations the HalfTotal Error Rate (HTER) is 500 which is much lower thanmethod in [25] (ie PCA + LBP + SVM (2050)) Similarlyin Table 4 we can see that the proposed system achieves errorrate better than or similar to the state-of-the-art methodsunder overall scenario

The MAXDIST patch selection method achieves betterperformances on average with the four classification tech-niques used in this study Additionally MAXDIST CS andDF patch selection algorithms demonstrate good general-ization capability not only for disparate datasets but alsofor spoofing attacks with varying qualities and fabricationmethods Quite to the contrary CP and IQA methods fail toattain proficient generalization aptitude Beside patch selec-tion algorithm choice of feature classification scheme alsoplays vital role in accomplishing preferable performancesTo this end it is easy to see in Tables 3 and 4 that amongSVM Naive-Bayes (NB) QDA and Ensemble based onAdaBoost classifiers Ensemble performs best under varyingfeatures datasets attack types and amount of training andtesting samples owing to its ability of reducing the variancesaveraging out the biases and most unlikeliness of overfittingThe NB and QDA classifiers in this study are quite sensitive


(a) (b) (c) (d)

(e) (f) (g) (h)

Figure 5 Examples of selected discriminative patches using proposed patches selection methods Top row of (a)ndash(h) face from REPLAY-ATTACK database Bottom row of (a)ndash(h) face from CASIA-FASD database A normalized face image (a) of a subject and the discriminativeselected patches using (b) CS (c) DEND-CLUSTERING (d) IPI (e) IQA (f) CP (g) MAXDIST and (h) DF methods

to patch selection approaches Specifically though NB enjoyssimplicity and computational efficiency it substantially per-forms poorly under diverse attack conditions This maybe occurring due to its assumption that all attributes areindependent (ie no correlation between variables) sinceit was pointed out in [20 40] that correlation mappingis beneficial to procure better accuracy and generalization

capability in biometric liveness detection Moreover NBalso assumes that the samples follow Gaussian distributionHowever Gaussian distribution assumption is generally truefor small biometric datasets But spoofing databases areheterogeneous that contain different spoof attack types andsizes and thereby NB either gets overfitted or fails to addressthe problem of concept-drift


Table 3 Comparison of the proposed method (with SVM QDANaive-Bayes (NB) and Ensemble based classifiers) on REPLAY-ATTACK database with existing methods

Method HTER ()Multi-LBP [23] 2025IQA [3] 1520GLCM (Unicamp) [24] 1562LBP119906281

[4] 1610LBP119906282

+ LBP1199062162

+ LBP119906281

+ SVM [4] 1387PCA + LBP + SVM [25] 2050Motion [16] 1170DoG-LBP + SVM [1] 1110LBP-TOP [26] 851IDA [1] 741Proposed DF-SVM 687Proposed DF-NB 801Proposed DF-QDA 730Proposed DF-Ensemble 623Proposed CS-SVM 625Proposed CS-NB 744Proposed CS-QDA 687Proposed CS-Ensemble 600Proposed DEND-CLUSTERING-SVM 598Proposed DEND-CLUSTERING-NB 887Proposed DEND-CLUSTERING-QDA 611Proposed DEND-CLUSTERING-Ensemble 500Proposed IQA-SVM 623Proposed IQA-NB 1105Proposed IQA-QDA 775Proposed IQA-Ensemble 562Proposed IPI-SVM 750Proposed IPI-NB 830Proposed IPI-QDA 619Proposed IPI-Ensemble 600Proposed CP-SVM 837Proposed CP-NB 918Proposed CP-QDA 712Proposed CP-Ensemble 680Proposed MAXDIST-SVM 587Proposed MAXDIST-NB 801Proposed MAXDIST-QDA 612Proposed MAXDIST-Ensemble 500

By metaknowledge analysis it was found that spoofattacks with higher resolutionquality are comparativelyharder to be recognized as also pointed out in [3 22]for instance high-quality eye cut-off attacks in which thecropped portions are filled by real eyes of the attackers leadingthus to the high quality spoofing attacks that are having acombination of real and spoofed face features Furthermore

between REPLAY-ATTACK and CASIA-FASD databasesCASIA-FASD database is more challenging as it incorporatesmore practical states such as variant of spoof attacks (eg cutphoto attack simulating eye-blinking) and samples with highquality (resolutions) All in all results also suggest that forsystems or datasets based on low- or normal-quality samplesit is advisable to adopt CS method with Ensemble classifierto reach desirable performance accuracies while MAXDISTwith Ensemble is better choice for systemsdatasets based onhigh-quality samples especially videos

On the whole it can be stated that use of only certainimage locations can significantly influence the face anti-spoofing accuracy Namely the proposed method uses onlyselected discriminative patches and attains higher-rankingprecision unlike the state-of-the-art methods which exploitwhole face imageframevideo leading hence generally tothe clutter in the feature representations and to their lowperformances

In many face recognition applications there is no accessto the video or image sequences of the user However a largenumber of existing face antispoofing solutions need video orsequences of images (ie either for motion or for temporalinformation) to attain high accuracy Accordingly they haveless usability since they are not devised to work on a singlestatic face image Conversely the proposed method is single-image algorithm (ie the method requires just one inputimage and not a sequence of them) Therefore the proposedmethod is more useful in various applications Further itis evident from the experimental results that the proposedframework is robust and performs well across diverse typesof spoof attacks materials and techniques (such as printedimage attack video-replayed attack cut photo attack andimage or video resolutions) although only specific face imagearea is considered Consequently the proposed method ismuch realistic and useful in real-world where a priori attack(artifact) types (paper mobile or resolution) which attackermight utilize are unpredictable

To sum up the performances shown by proposed algo-rithm confirm that contemplating the inherent differencesof discriminant abilities among various face image locationsis useful for consistently recognizing well the facial spoofattacks In other words we show that it seems feasible to useonly certain face image patches instead of whole face imageto reduce significantly the error rates

5 Conclusion

The vulnerability of face recognition systems to spoof-ing attacks is a largely accepted reality which has led togreat advances in face antispoofing (especially face livenessdetection) technologies Despite the remarkable advancescounteracting face spoof attacks has yet proven to be achallenging task Moreover existing face liveness detectionmethods use whole face image or complete video for livenessdetection However often image regions (video frames) areredundant or correspond to the clutter in the image (video)thus leading generally to low performancesTherefore in thispaper we propose using just discriminative image patchesfor face liveness detection In particular we present seven


Table 4 Comparison of the proposed method (with SVM QDA Naive-Bayes (NB) and Ensemble based classifiers) on CASIA-FASDdatabase with existing methods in terms of EER ()

Method Low quality Normal quality High quality Warpedphoto attack

Cut photoattack Video attack Overall

DoG [22] 1300 1300 2600 1600 600 2400 1700IQA [27] 3170 2220 569 2610 1831 3441 3245LBP + SVM baseline 1650 1720 2300 2470 1670 2700 2480Multi-LBP baseline 1277 1666 2666 1555 2555 1777 1777HOOF [28] 1666 3000 2611 1555 1777 3888 2111Mag-HOOF 1722 3333 2277 1222 2000 3660 2222HOOF + Multi-LBP 944 2055 1666 1000 1666 2444 1555Motion-MLBP [23] 722 1333 2944 1444 2222 1333 1574Motion magnification [23] 611 2333 1388 1000 1444 2000 1444Color texture [29] 780 1010 640 750 540 810 620Proposed DF-SVM 753 665 628 694 821 797 671Proposed DF-NB 777 779 666 700 766 816 900Proposed DF-QDA 578 701 565 697 788 715 781Proposed DF-Ensemble 465 599 657 594 649 600 611Proposed CS-SVM 646 643 597 656 872 727 854Proposed CS-NB 669 630 599 749 801 768 869Proposed CS-QDA 733 889 609 743 821 768 797Proposed CS-Ensemble 651 559 530 531 613 712 759Proposed DEND-CLUSTERING-SVM 739 709 593 735 822 842 807Proposed DEND-CLUSTERING-NB 698 672 740 765 823 800 845Proposed DEND-CLUSTERING-QDA 711 689 790 703 699 889 867Proposed DEND-CLUSTERING-Ensemble 589 606 558 533 542 602 516Proposed IQA-SVM 632 700 597 773 771 769 733Proposed IQA-NB 666 766 632 733 845 800 822Proposed IQA-QDA 635 778 831 878 787 806 869Proposed IQA-Ensemble 602 657 576 727 687 727 727Proposed IPI-SVM 834 799 872 789 856 874 836Proposed IPI-NB 888 745 835 722 890 812 883Proposed IPI-QDA 601 733 688 733 878 914 857Proposed IPI-Ensemble 600 678 610 627 724 800 722Proposed CP-SVM 887 823 908 764 937 890 950Proposed CP-NB 910 907 839 974 985 945 934Proposed CP-QDA 867 834 900 878 851 834 859Proposed CP-Ensemble 801 727 734 682 713 823 760Proposed MAXDIST-SVM 725 576 668 928 828 842 856Proposed MAXDIST-NB 737 698 733 733 800 831 800Proposed MAXDIST-QDA 711 678 732 884 802 842 840Proposed MAXDIST-Ensemble 526 600 530 578 549 502 507

novel methods to obtain discriminative patches in a faceimage (or randomly selected lone video frame) The featuresof selected discriminative image patches are fed to a specificclassifier (ie SVM Naive-Bayes QDA or Ensemble) Theclassification results of these patches are combined by amajority-voting based scheme for the final classificationof genuine and spoof faces Experimental results on twopublicly available databases show comparative performances

compared to the existing works The future works includedevising more novel techniques for attaining discriminativeimage patches and inclusion of temporal information in theproposed method for higher security applications

Competing Interests

The authors declare that they have no competing interests


References

[1] D Wen H Han and A K Jain ldquoFace spoof detection withimage distortion analysisrdquo IEEE Transactions on InformationForensics and Security vol 10 no 4 pp 746ndash761 2015

[2] Z Akhtar C Micheloni and G L Foresti ldquoBiometric livenessdetection challenges and research opportunitiesrdquo IEEE Securityamp Privacy vol 13 no 5 pp 63ndash72 2015

[3] J Galbally S Marcel and J Fierrez ldquoImage quality assessmentfor fake biometric detection application to iris fingerprint andface recognitionrdquo IEEE Transactions on Image Processing vol23 no 2 pp 710ndash724 2014

[4] I Chingovska A Anjos and S Marcel ldquoOn the effectiveness oflocal binary patterns in face anti-spoofingrdquo in Proceedings of theInternational Conference of the Biometrics Special Interest Group(BIOSIG rsquo12) pp 1ndash7 Darmstadt Germany September 2012

[5] J Maatta A Hadid and M Pietikainen ldquoFace spoofing detec-tion from single images using texture and local shape analysisrdquoIET Biometrics vol 1 no 1 pp 3ndash10 2012

[6] Z Akhtar C Micheloni C Piciarelli and G L ForestildquoMoBio LivDet mobile biometric liveness detectionrdquo in Pro-ceedings of the 11th IEEE International Conference on AdvancedVideo and Signal Based Surveillance (AVSS rsquo14) pp 187ndash192Seoul Republic of Korea August 2014

[7] G Pan L Sun ZWu and S Lao ldquoEyeblink-based anti-spoofingin face recognition from a generic webcamerardquo in Proceedingsof the IEEE 11th International Conference on Computer Vision(ICCV rsquo07) pp 1ndash8 Rio de Janeiro Brazil October 2007

[8] X Tan Y Li J Liu and L Jiang ldquoFace liveness detection froma single image with sparse low rank bilinear discriminativemodelrdquo in Proceedings of the 11th European Conference onComputer Vision (ECCV rsquo10) K Daniilidis P Maragos and NParagios Eds vol 6316 of Lecture Notes in Computer Sciencepp 504ndash517 Crete Greece September 2010

[9] Y Li and X Tan ldquoAn anti-photo spoof method in face recogni-tion based on the analysis of fourier spectra with sparse logisticregressionrdquo in Proceedings of the Chinese Conference on PatternRecognition (CCPR rsquo09) Nanjing China November 2009

[10] K Kollreider H Fronthaler and J Bigun ldquoNon-intrusiveliveness detection by face imagesrdquo Image and Vision Computingvol 27 no 3 pp 233ndash244 2009

[11] W Bao H Li N Li andW Jiang ldquoA liveness detection methodfor face recognition based on optical flow fieldrdquo in Proceedingsof the International Conference on Image Analysis and SignalProcessing (IASP rsquo09) pp 233ndash236 IEEE Taizhou China April2009

[12] G Pan L Sun Z Wu and Y Wang ldquoMonocular camera-based face liveness detection by combining eyeblink and scenecontextrdquo Telecommunication Systems vol 47 no 3 pp 215ndash2252011

[13] J Li Y Wang T Tan and A K Jain ldquoLive face detectionbased on the analysis of Fourier spectrardquo in Proceedings ofthe Biometric Technology for Human Identification vol 5404of Proceedings of SPIE pp 296ndash303 Orlando Fla USA April2004

[14] J Bai T Ng X Gao and Y Shi ldquoIs physics-based livenessdetection truly possible with a single imagerdquo in Proceedingsof IEEE International Symposium on Circuits and Systems pp3425ndash3428 Paris France May-June 2010

[15] Z Zhang D Yi Z Lei and S Z Li ldquoFace liveness detection bylearning multispectral reflectance distributionsrdquo in Proceedingsof the IEEE International Conference on Automatic Face and

Gesture Recognition andWorkshops (FG rsquo11) pp 436ndash441 SantaBarbara Calif USA March 2011

[16] J Komulainen A Hadid M Pietikainen A Anjos and S Mar-cel ldquoComplementary countermeasures for detecting scenic facespoofing attacksrdquo in Proceedings of the International Conferenceon Biometrics (ICB rsquo13) pp 1ndash7 Madrid Spain June 2013

[17] A Anjos and S Marcel ldquoCounter-measures to photo attacksin face recognition a public database and a baselinerdquo inProceedings of the International Joint Conference on Biometrics(IJCB rsquo11) pp 1ndash7 Washington DC USA October 2011

[18] T Wang and S Z Li ldquoFace liveness detection using 3dstructure recovered from a single camerardquo in Proceedings of theInternational Conference on Biometrics (ICB rsquo13) pp 1ndash6 IEEEMadrid Spain 2013

[19] T I Dhamecha A Nigam R Singh and M Vatsa ldquoDisguisedetection and face recognition in visible and thermal spec-trumsrdquo in Proceedings of the 6th IAPR International Conferenceon Biometrics (ICB rsquo13) pp 1ndash6 Madrid Spain June 2013

[20] G Chetty and M Wagner ldquoLiveness detection using cross-modal correlations in face-voice person authenticationrdquo inProceedings of the 9th European Conference on Speech Commu-nication and Technology (Interspeech rsquo05) pp 2181ndash2184 LisbonPortugal September 2005

[21] N Kose and J-L Dugelay ldquoReflectance analysis based counter-measure technique to detect face mask attacksrdquo in Proceedingsof the 18th International Conference on Digital Signal Processing(DSP rsquo13) pp 1ndash6 IEEE Fira Greece July 2013

[22] Z Zhang J Yan S Liu Z Lei D Yi and S Z Li ldquoA faceantispoofing database with diverse attacksrdquo in Proceedings of the5th IAPR International Conference on Biometrics (ICB rsquo12) pp26ndash31 IEEE New Delhi India April 2012

[23] S Bharadwaj T I Dhamecha M Vatsa and R Singh ldquoFaceanti-spoofing via motion magnification and multifeature vide-olet aggregationrdquo Tech Rep IIITD-TR-2014-002 2014

[24] I Chingovska J Yang Z Lei and D Yi ldquoThe 2nd competitionon countermeasures to 2D face spoofing attacksrdquo in Proceedingsof the IEEE International Conference on Biometrics (ICB rsquo13) pp1ndash6 Madrid Spain June 2013

[25] S Tirunagari N Poh D Windridge A Iorliam N Suki andA T S Ho ldquoDetection of face spoofing using visual dynamicsrdquoIEEE Transactions on Information Forensics and Security vol 10no 4 pp 762ndash777 2015

[26] T de Freitas Pereira A Anjos J M De Martino and SMarcel ldquoCan face anti-spoofing countermeasures work in a realworld scenariordquo in Proceedings of the 6th IAPR InternationalConference on Biometrics (ICB rsquo13) pp 1ndash8 Madrid Spain June2013

[27] J Galbally and S Marcel ldquoFace anti-spoofing based on generalimage quality assessmentrdquo in Proceedings of the 22nd Interna-tional Conference on Pattern Recognition (ICPR rsquo14) pp 1173ndash1178 Stockholm Sweden August 2014

[28] R Chaudhry A Ravichandran G Hager and R Vidal ldquoHis-tograms of oriented optical flow and Binet-Cauchy kernels onnonlinear dynamical systems for the recognition of humanactionsrdquo in Proceedings of the IEEE Conference on ComputerVision and Pattern Recognition (CVPR rsquo09) pp 1932ndash1939Miami Fla USA June 2009

[29] Z Boulkenafet J Komulainen and A Hadid ldquoFace anti-spoofing based on color texture analysisrdquo in Proceedings of theIEEE International Conference on Image Processing (ICIP rsquo15)pp 2636ndash2640 Quebec City Canada September 2015


[30] M Nilsson J Nordberg and I Claesson ldquoFace detection usinglocal SMQT features and split up snow classifierrdquo in Proceedingsof the IEEE International Conference on Acoustics Speech andSignal Processing (ICASSP rsquo07) pp II-589ndashII-592 HonoluluHawaii USA April 2007

[31] U Uludag A Ross and A K Jain ldquoBiometric templateselection and update a case study in fingerprintsrdquo PatternRecognition vol 37 no 7 pp 1533ndash1542 2004

[32] S Zhalehpour Z Akhtar and C Eroglu Erdem ldquoMultimodalemotion recognition based on peak frame selection fromvideordquoSignal Image and Video Processing 2015

[33] S Bayram I Avcibas B Sankur and N Memon ldquoImagemanipulation detectionrdquo Journal of Electronic Imaging vol 15no 4 Article ID 041102 17 pages 2006

[34] I Avcibas N Memon and B Sankur ldquoSteganalysis using imagequality metricsrdquo IEEE Transactions on Image Processing vol 12no 2 pp 221ndash229 2003

[35] W Xue L Zhang X Mou and A C Bovik ldquoGradient mag-nitude similarity deviation a highly efficient perceptual imagequality indexrdquo IEEE Transactions on Image Processing vol 23no 2 pp 684ndash695 2014

[36] L Bourdev and J Malik ldquoPoselets body part detectors trainedusing 3D human pose annotationsrdquo in Proceedings of the IEEE12th International Conference on Computer Vision (ICCV rsquo09)pp 1365ndash1372 Kyoto Japan September 2009

[37] P F Felzenszwalb R B Girshick D McAllester and DRamanan ldquoObject detection with discriminatively trained part-based modelsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 32 no 9 pp 1627ndash1645 2010

[38] U Vovk F Pernus and B Likar ldquoA review of methods for cor-rection of intensity inhomogeneity in MRIrdquo IEEE Transactionson Medical Imaging vol 26 no 3 pp 405ndash421 2007

[39] S Cheung and C Kamath ldquoRobust techniques for backgroundsubtraction in urban traffic videordquo in Proceedings of the IEEEConference on Visual Communications and Image Processing(VCIP rsquo07) pp 1ndash12 2007

[40] Z Akhtar C Micheloni and G L Foresti ldquoCorrelation basedfingerprint liveness detectionrdquo in Proceedings of the Interna-tional Conference on Biometrics (ICB rsquo15) pp 305ndash310 PhuketCity Thailand May 2015

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Active and Passive Electronic Components

Control Scienceand Engineering

Journal of



RotatingMachinery


Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics


Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation



DistributedSensor Networks



(a) (b)

Figure 1 (a) In New York a robber disguised as white cop using latex masks robbing a cash-checking store (b) In 2010 a passenger boardeda plane in Hong Kong with an old man mask and arrived in Canada to claim asylum

several face antispoofing or liveness detection techniques [7ndash15] However face spoofing attacks remain a problem dueto difficulties in finding discriminative and computationallyinexpensive features and techniques for spoof recognitionMoreover publishedmethods are limited in their scope sincethey mainly use whole face image or complete video forliveness detection Nevertheless often certain face imageregions (video frames) are redundant or correspond to theclutter in the image (video) leading thus generally to lowperformances

It is thus essential to develop robust efficient andcompact face antispoofing (or liveness detection) methodswhich are capable of generalizing well to discriminativeclass-specific information and imaging conditions To thisaim in this paper we propose a simple and effective solutionbased on discriminative image patches In particular wepropose seven novel fully automated algorithms to highlightregions of interest in face images We define these regions(or image patches) to be discriminative (ie specific toa particular class live or spoof) consistent (ie reliablyappearing in different face images or video frames) salient(ie conspicuous regions) and repetitive (ie frequentlyappearing in the image set of specific class) The basic notionis ldquointeresting patches are those that are specific to a faceimage (or video frame) and should contain features thatgive assistance to discriminate a given live face image fromspoofed onerdquo Based on this definition two of the sevenproposed image patch selectionmethods (ieMAXDIST andDEND-CLUSTER) donot employ any training or prior learn-ing However the remaining techniques use simple clustering(ie CP and CS) image intensity (ie IPI) image quality(ie IQA) or diversity filter (ie DF) to obtain discriminativepatches For final classification we exploited four well-known classifiers namely support vector machine (SVM)Naive-Bayes Quadratic Discriminant Analysis (QDA) andEnsemble using voting based scheme Experimental analysison two publicly available databases (Idiap Replay-Attackand CASIA-FASD) shows good results compared to existingworks

The added advantages of the proposed framework are(i) being cheap (ii) very low complexity (iii) needing oneface image (ie the same face sample acquired for facerecognition) to detect whether it is genuine or spoof attack(iv) being nonintrusive (v) being user-friendly and (vi) beingeasy to embed in already functional face recognition systemsplus no requirement of new piece of hardware

The reminder of the paper is organized as follows Exitingliterature works on face liveness detection are discussed inSection 2 The proposed approaches to determine the dis-criminative image patches and spoof classification schemesare described in Section 3 Experimental datasets protocolsand results are presented in Section 4 A conclusion is drawnin Section 5

2 Related Work

Despite great deal of advancements in face recognitionsystems face spoofing still poses a serious threat Most of theexisting academic and commercial facial recognition systemsmay be spoofed by (see Figure 2) (i) a photo of a genuine user(ii) a video of a genuine user (iii) a 3D face model (mask) ofa genuine user (iv) a reverse-engineered face image from thetemplate of a genuine user (v) a sketch of a genuine user (vi)an impostor wearing specific make-up to look like a genuineuser (vii) an impostor who underwent plastic surgery to looklike a genuine user The most easiest cheapest and commonface spoofing attack is to submit a photograph of a legitimateuser to the face recognition systems which is also known asldquophoto attackrdquo

Typical countermeasure (ie face liveness detection orantispoofing) techniques can be coarsely classified in threecategories based on clues used for spoof attack detection(i) motion analysis based methods (ii) texture analysis basedmethods and (iii) hardware-based methods In what followswe provide a brief literature overview of published face spoofrecognition techniques along with their pros and cons

(i) Motion Analysis Based Methods These methods broadlytry to detect spontaneous movement clues generated whentwo dimensional counterfeits are presented to the camera ofthe system such as photographs or videos Therefore Pan etal [7] exploited the fact that human eye-blink occurs onceevery 2ndash4 seconds and proposed eye-blink based livenessdetection for photo-spoofing using (spontaneous) eye-blinksThis method uses an undirected conditional random fieldframework to model the eye-blinking which relaxes theindependence assumption of generative modelling and statesdependence limitations from hidden Markov modelling Itis evident that real human faces (which are 3D objects)will move significantly differently from planer objects andsuch deformation patterns can be employed for livenessdetection For example Tan et al [8] considered Lambertian


(a) (b)

(c) (d)

(e)

(f) (g)



















Face detectionand

normalization



patches

Classifier


system

An input video



Selected patches

Fusion

Selectedpatch 1

Selectedpatch K

Live

Spoof

Finaldecision












DEND-CLUSTERING

CP(Cluster Pairing)


IQA(image quality

assessment)










119894(I) isin R119872times119873

119899








ISM (P) =

119875minus2

sum

119909=1

119876minus2

sum

119910=1

119866 (119909 119910) (1)









119863mean (119862119894 119862119895) =

10038171003817100381710038171003817119898119894

minus 119898119895

10038171003817100381710038171003817

119883

119895=1le 120578 (2)


119909 119862119894is a








DV = exp(

1

119883

119883

sum

119894=1

119875119894 (P)) (3)



patches

















R119872times119873119899

119894=1and P

119895(I) isin R119872times119873

119899






max (P2)

MSE (P P)

] (4)

where


=

1

119875119876

119875

sum

119909=1

119876

sum

119910=1

(P119909119910

minus P119909119910

)

2

(5)



NCC (P P) =

sum119875

119909=1sum119876

119910=1(P119909119910

sdot P119909119910

)

sum119875

119909=1sum119876

119910=1(P119909119910

)

2 (6)


TED (P P) =

1

119875119876

119875

sum

119909=1

119876

sum

119910=1

100381610038161003816100381610038161003816PE119909119910

minus PE119909119910

100381610038161003816100381610038161003816 (7)



GMSD (P P) = radic1

119880

119880

sum

119894=1



GMS (119894) =

2m119903 (119894)m119889 (119894) + 119888

m2119903

(119894) + m2119889

(119894) + 119888

(9)


119903and m

119889are gra-


GMSM =

1

119880

119880

sum

119894=1

GMS (119894) (10)











argmax119894

119894

minus 120572max1119905

119878119894119895

(12)


119894119895








119899

119894=1are taken



FIPI

=


119909=1sum119876

119910=1

1003816100381610038161003816P119894(119909 119910) minus P

119894

1003816100381610038161003816

120590 (P119894)

gt 120578


(13)





4 Experiments


41 Datasets








640 times 480Dagger

480 times 640oplus

1280 times 720⋆













(a) (b) (c) (d)

(e) (f) (g) (h)







[4] 1610LBP119906282

+ LBP1199062162

+ LBP119906281







5 Conclusion









Competing Interests



References













































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b)

(c) (d)

(e)

(f) (g)



















Face detectionand

normalization



patches

Classifier


system

An input video



Selected patches

Fusion

Selectedpatch 1

Selectedpatch K

Live

Spoof

Finaldecision












DEND-CLUSTERING

CP(Cluster Pairing)


IQA(image quality

assessment)










119894(I) isin R119872times119873

119899








ISM (P) =

119875minus2

sum

119909=1

119876minus2

sum

119910=1

119866 (119909 119910) (1)









119863mean (119862119894 119862119895) =

10038171003817100381710038171003817119898119894

minus 119898119895

10038171003817100381710038171003817

119883

119895=1le 120578 (2)


119909 119862119894is a








DV = exp(

1

119883

119883

sum

119894=1

119875119894 (P)) (3)



patches

















R119872times119873119899

119894=1and P

119895(I) isin R119872times119873

119899






max (P2)

MSE (P P)

] (4)

where


=

1

119875119876

119875

sum

119909=1

119876

sum

119910=1

(P119909119910

minus P119909119910

)

2

(5)



NCC (P P) =

sum119875

119909=1sum119876

119910=1(P119909119910

sdot P119909119910

)

sum119875

119909=1sum119876

119910=1(P119909119910

)

2 (6)


TED (P P) =

1

119875119876

119875

sum

119909=1

119876

sum

119910=1

100381610038161003816100381610038161003816PE119909119910

minus PE119909119910

100381610038161003816100381610038161003816 (7)



GMSD (P P) = radic1

119880

119880

sum

119894=1



GMS (119894) =

2m119903 (119894)m119889 (119894) + 119888

m2119903

(119894) + m2119889

(119894) + 119888

(9)


119903and m

119889are gra-


GMSM =

1

119880

119880

sum

119894=1

GMS (119894) (10)











argmax119894

119894

minus 120572max1119905

119878119894119895

(12)


119894119895








119899

119894=1are taken



FIPI

=


119909=1sum119876

119910=1

1003816100381610038161003816P119894(119909 119910) minus P

119894

1003816100381610038161003816

120590 (P119894)

gt 120578


(13)





4 Experiments


41 Datasets








640 times 480Dagger

480 times 640oplus

1280 times 720⋆













(a) (b) (c) (d)

(e) (f) (g) (h)







[4] 1610LBP119906282

+ LBP1199062162

+ LBP119906281







5 Conclusion









Competing Interests



References













































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation



















Face detectionand

normalization



patches

Classifier


system

An input video



Selected patches

Fusion

Selectedpatch 1

Selectedpatch K

Live

Spoof

Finaldecision












DEND-CLUSTERING

CP(Cluster Pairing)


IQA(image quality

assessment)










119894(I) isin R119872times119873

119899








ISM (P) =

119875minus2

sum

119909=1

119876minus2

sum

119910=1

119866 (119909 119910) (1)









119863mean (119862119894 119862119895) =

10038171003817100381710038171003817119898119894

minus 119898119895

10038171003817100381710038171003817

119883

119895=1le 120578 (2)


119909 119862119894is a








DV = exp(

1

119883

119883

sum

119894=1

119875119894 (P)) (3)



patches

















R119872times119873119899

119894=1and P

119895(I) isin R119872times119873

119899






max (P2)

MSE (P P)

] (4)

where


=

1

119875119876

119875

sum

119909=1

119876

sum

119910=1

(P119909119910

minus P119909119910

)

2

(5)



NCC (P P) =

sum119875

119909=1sum119876

119910=1(P119909119910

sdot P119909119910

)

sum119875

119909=1sum119876

119910=1(P119909119910

)

2 (6)


TED (P P) =

1

119875119876

119875

sum

119909=1

119876

sum

119910=1

100381610038161003816100381610038161003816PE119909119910

minus PE119909119910

100381610038161003816100381610038161003816 (7)



GMSD (P P) = radic1

119880

119880

sum

119894=1



GMS (119894) =

2m119903 (119894)m119889 (119894) + 119888

m2119903

(119894) + m2119889

(119894) + 119888

(9)


119903and m

119889are gra-


GMSM =

1

119880

119880

sum

119894=1

GMS (119894) (10)











argmax119894

119894

minus 120572max1119905

119878119894119895

(12)


119894119895








119899

119894=1are taken



FIPI

=


119909=1sum119876

119910=1

1003816100381610038161003816P119894(119909 119910) minus P

119894

1003816100381610038161003816

120590 (P119894)

gt 120578


(13)





4 Experiments


41 Datasets








640 times 480Dagger

480 times 640oplus

1280 times 720⋆













(a) (b) (c) (d)

(e) (f) (g) (h)







[4] 1610LBP119906282

+ LBP1199062162

+ LBP119906281







5 Conclusion









Competing Interests



References













































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











DEND-CLUSTERING

CP(Cluster Pairing)


IQA(image quality

assessment)










119894(I) isin R119872times119873

119899








ISM (P) =

119875minus2

sum

119909=1

119876minus2

sum

119910=1

119866 (119909 119910) (1)









119863mean (119862119894 119862119895) =

10038171003817100381710038171003817119898119894

minus 119898119895

10038171003817100381710038171003817

119883

119895=1le 120578 (2)


119909 119862119894is a








DV = exp(

1

119883

119883

sum

119894=1

119875119894 (P)) (3)



patches

















R119872times119873119899

119894=1and P

119895(I) isin R119872times119873

119899






max (P2)

MSE (P P)

] (4)

where


=

1

119875119876

119875

sum

119909=1

119876

sum

119910=1

(P119909119910

minus P119909119910

)

2

(5)



NCC (P P) =

sum119875

119909=1sum119876

119910=1(P119909119910

sdot P119909119910

)

sum119875

119909=1sum119876

119910=1(P119909119910

)

2 (6)


TED (P P) =

1

119875119876

119875

sum

119909=1

119876

sum

119910=1

100381610038161003816100381610038161003816PE119909119910

minus PE119909119910

100381610038161003816100381610038161003816 (7)



GMSD (P P) = radic1

119880

119880

sum

119894=1



GMS (119894) =

2m119903 (119894)m119889 (119894) + 119888

m2119903

(119894) + m2119889

(119894) + 119888

(9)


119903and m

119889are gra-


GMSM =

1

119880

119880

sum

119894=1

GMS (119894) (10)











argmax119894

119894

minus 120572max1119905

119878119894119895

(12)


119894119895








119899

119894=1are taken



FIPI

=


119909=1sum119876

119910=1

1003816100381610038161003816P119894(119909 119910) minus P

119894

1003816100381610038161003816

120590 (P119894)

gt 120578


(13)





4 Experiments


41 Datasets








640 times 480Dagger

480 times 640oplus

1280 times 720⋆













(a) (b) (c) (d)

(e) (f) (g) (h)







[4] 1610LBP119906282

+ LBP1199062162

+ LBP119906281







5 Conclusion









Competing Interests



References













































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation













119863mean (119862119894 119862119895) =

10038171003817100381710038171003817119898119894

minus 119898119895

10038171003817100381710038171003817

119883

119895=1le 120578 (2)


119909 119862119894is a








DV = exp(

1

119883

119883

sum

119894=1

119875119894 (P)) (3)



patches

















R119872times119873119899

119894=1and P

119895(I) isin R119872times119873

119899






max (P2)

MSE (P P)

] (4)

where


=

1

119875119876

119875

sum

119909=1

119876

sum

119910=1

(P119909119910

minus P119909119910

)

2

(5)



NCC (P P) =

sum119875

119909=1sum119876

119910=1(P119909119910

sdot P119909119910

)

sum119875

119909=1sum119876

119910=1(P119909119910

)

2 (6)


TED (P P) =

1

119875119876

119875

sum

119909=1

119876

sum

119910=1

100381610038161003816100381610038161003816PE119909119910

minus PE119909119910

100381610038161003816100381610038161003816 (7)



GMSD (P P) = radic1

119880

119880

sum

119894=1



GMS (119894) =

2m119903 (119894)m119889 (119894) + 119888

m2119903

(119894) + m2119889

(119894) + 119888

(9)


119903and m

119889are gra-


GMSM =

1

119880

119880

sum

119894=1

GMS (119894) (10)











argmax119894

119894

minus 120572max1119905

119878119894119895

(12)


119894119895








119899

119894=1are taken



FIPI

=


119909=1sum119876

119910=1

1003816100381610038161003816P119894(119909 119910) minus P

119894

1003816100381610038161003816

120590 (P119894)

gt 120578


(13)





4 Experiments


41 Datasets








640 times 480Dagger

480 times 640oplus

1280 times 720⋆













(a) (b) (c) (d)

(e) (f) (g) (h)







[4] 1610LBP119906282

+ LBP1199062162

+ LBP119906281







5 Conclusion









Competing Interests



References













































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation















R119872times119873119899

119894=1and P

119895(I) isin R119872times119873

119899






max (P2)

MSE (P P)

] (4)

where


=

1

119875119876

119875

sum

119909=1

119876

sum

119910=1

(P119909119910

minus P119909119910

)

2

(5)



NCC (P P) =

sum119875

119909=1sum119876

119910=1(P119909119910

sdot P119909119910

)

sum119875

119909=1sum119876

119910=1(P119909119910

)

2 (6)


TED (P P) =

1

119875119876

119875

sum

119909=1

119876

sum

119910=1

100381610038161003816100381610038161003816PE119909119910

minus PE119909119910

100381610038161003816100381610038161003816 (7)



GMSD (P P) = radic1

119880

119880

sum

119894=1



GMS (119894) =

2m119903 (119894)m119889 (119894) + 119888

m2119903

(119894) + m2119889

(119894) + 119888

(9)


119903and m

119889are gra-


GMSM =

1

119880

119880

sum

119894=1

GMS (119894) (10)











argmax119894

119894

minus 120572max1119905

119878119894119895

(12)


119894119895








119899

119894=1are taken



FIPI

=


119909=1sum119876

119910=1

1003816100381610038161003816P119894(119909 119910) minus P

119894

1003816100381610038161003816

120590 (P119894)

gt 120578


(13)





4 Experiments


41 Datasets








640 times 480Dagger

480 times 640oplus

1280 times 720⋆













(a) (b) (c) (d)

(e) (f) (g) (h)







[4] 1610LBP119906282

+ LBP1199062162

+ LBP119906281







5 Conclusion









Competing Interests



References













































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation
















argmax119894

119894

minus 120572max1119905

119878119894119895

(12)


119894119895








119899

119894=1are taken



FIPI

=


119909=1sum119876

119910=1

1003816100381610038161003816P119894(119909 119910) minus P

119894

1003816100381610038161003816

120590 (P119894)

gt 120578


(13)





4 Experiments


41 Datasets








640 times 480Dagger

480 times 640oplus

1280 times 720⋆













(a) (b) (c) (d)

(e) (f) (g) (h)







[4] 1610LBP119906282

+ LBP1199062162

+ LBP119906281







5 Conclusion









Competing Interests



References













































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation















640 times 480Dagger

480 times 640oplus

1280 times 720⋆













(a) (b) (c) (d)

(e) (f) (g) (h)







[4] 1610LBP119906282

+ LBP1199062162

+ LBP119906281







5 Conclusion









Competing Interests



References













































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b) (c) (d)

(e) (f) (g) (h)







[4] 1610LBP119906282

+ LBP1199062162

+ LBP119906281







5 Conclusion









Competing Interests



References













































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation












[4] 1610LBP119906282

+ LBP1199062162

+ LBP119906281







5 Conclusion









Competing Interests



References













































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation
















Competing Interests



References













































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










References













































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation























RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation









Research Article Face Spoof Attack Recognition Using...

Documents

Transcript of Research Article Face Spoof Attack Recognition Using...