Ci a ion: Benlamoudi, A.;
Bekhouche, S.E.; Ko ichi, M.; Bensid,
K.; Ouahabi, A.; Hadid, A.;
Taleb-Ahmed, A. Face P esen a ion
A ack De ec ion Using Deep
Backg ound Sub ac ion. Senso s
2022,22, 3760. h ps://doi.o g/
10.3390/s22103760
Academic Edi o s: Michal Cho as,
Ra al Kozik and Ma ek Pawlicki
Recei ed: 3 Ap il 2022
Accep ed: 12 May 2022
Published: 15 May 2022
Publishe ’s No e: MDPI s ays neu al
wi h ega d o ju isdic ional claims in
published maps and ins i u ional a il-
ia ions.
Copy igh : © 2022 by he au ho s.
Licensee MDPI, Basel, Swi ze land.
This a icle is an open access a icle
dis ibu ed unde he e ms and
condi ions o he C ea i e Commons
A ibu ion (CC BY) license (h ps://
c ea i ecommons.o g/licenses/by/
4.0/).
senso s
A icle
Face P esen a ion A ack De ec ion Using Deep
Backg ound Sub ac ion
Azeddine Benlamoudi 1, Salah Eddine Bekhouche 2, Maa ou Ko ichi 1, Khaled Bensid 1,
Abdeldjalil Ouahabi 3,* , Abdenou Hadid 4and Abdelmalik Taleb-Ahmed 4
1
Labo a oi e de Génie Élec ique, Facul é des Nou elles Technologies de l’In o ma ion e de la Communica ion,
Uni e si é Kasdi Me bah Oua gla, Oua gla 30 000, Alge ia; benlamoudi.azeddine@uni -oua gla.dz (A.B.);
ko ichi.maa ou @uni -oua gla.dz (M.K.); bensid.khaled@uni -oua gla.dz (K.B.)
2
Depa men o Compu e Science and A i icial In elligence, Facul y o In o ma ics, Uni e si y o he Basque
Coun y UPV/EHU, 20018 San Sebas ian, Spain; [email p o ec ed]
3UMR 1253, iB ain, INSERM, Uni e si é de Tou s, 37000 Tou s, F ance
4Ins i u d’Elec onique de Mic oélec onique e de Nano echnologie (IEMN), UMR 8520,
Uni e si é Poly echnique Hau s de F ance, Uni e si é de Lille, CNRS, 59313 Valenciennes, F ance;
abdenou [email p o ec ed] (A.H.); [email p o ec ed] (A.T.-A.)
*Co espondence: [email p o ec ed]
Abs ac :
Cu en ly, ace ecogni ion echnology is he mos widely used me hod o e i ying
an indi idual’s iden i y. Ne e heless, i has inc eased in popula i y, aising conce ns abou ace
p esen a ion a acks, in which a pho o o ideo o an au ho ized pe son’s ace is used o ob ain
access o se ices. Based on a combina ion o backg ound sub ac ion (BS) and con olu ional neu al
ne wo k(s) (CNN), as well as an ensemble o classi ie s, we p opose an e icien and mo e obus
ace p esen a ion a ack de ec ion algo i hm. This algo i hm includes a ully connec ed (FC) classi ie
wi h a majo i y o e (MV) algo i hm, which uses di e en ace p esen a ion a ack ins umen s (e.g.,
p in ed pho o and eplayed ideo). By including a majo i y o e o de e mine whe he he inpu
ideo is genuine o no , he p oposed me hod signi ican ly enhances he pe o mance o he ace
an i-spoo ing (FAS) sys em. Fo e alua ion, we conside ed he MSU MFSD, REPLAY-ATTACK, and
CASIA-FASD da abases. The ob ained esul s a e e y in e es ing and a e much be e han hose
ob ained by s a e-o - he-a me hods. Fo ins ance, on he REPLAY-ATTACK da abase, we we e able
o a ain a hal - o al e o a e (HTER) o 0.62% and an equal e o a e (EER) o 0.58%. We a ained
an EER o 0% on bo h he CASIA-FASD and he MSU MFSD da abases.
Keywo ds: biome ics; ace p esen a ion a ack; deep lea ning
1. In oduc ion
Indi iduals can be success ully iden i ied and au hen ica ed using biome ic ea u es
and ai s. Hence, i is app op ia e o access con ol and global secu i y sys ems ha
depend on pe son ecogni ion, which is achie ed h ough he use o a a ie y o biome ic
modali ies, anging om he classic inge p in h ough he ace, i is, ea [
1
–
4
] and, mo e
ecen ly, ein and blood low. Fu he mo e, a numbe o spoo ing me hods ha e been
de eloped in o de o o e come such biome ic sys ems [
5
]. When someone ies o ge
a ound a ace biome ic sys em by placing a ake ace in on o he came a, his is known
as a p esen a ion a ack. Ne e heless, compa ed o o he modali ies, he abundance o s ill
ace images o ideo sequences on he in e ne has made i excep ionally easy o ob ain a
pe son’s acial da a.
The spoo ing de ec ion li e a u e discusses mul iple ypes o p esen a ion a ack in-
s umen s, such as p in , eplay, silicon masks, and makeup a acks. The ocus o ou wo k
is on he i s wo a acks, namely p in and eplay a acks. The p in a ack spoo s 2D
ace ecogni ion sys ems by using p in ed pho og aphs o a subjec , whe eas he eplay
Senso s 2022,22, 3760. h ps://doi.o g/10.3390/s22103760 h ps://www.mdpi.com/jou nal/senso s
Senso s 2022,22, 3760 2 o 17
a ack p esen s a ideo o a bona ide p esen a ion o a oid li eness de ec ion. Fu he mo e,
he low cos o launching a ace p esen a ion a ack ins umen has inc eased he p e alence
o he p oblem. Face ecogni ion sys em spoo ing media anges om low-quali y pape
p in s o high-quali y pho og aphs, as well as ideo s eams played in on o he biome ic
au hen ica ion sys em senso .
Fea u e ex ac ion is a c i ical componen o he ace p esen a ion a ack de ec ion
ask when using a classical machine lea ning classi ie . Con olu ional neu al ne wo k(s)
(CNN) can also be used o p edic sco es. This la e is a c ucial componen o deep lea ning
algo i hms, such as he ResNe -50 [
6
] p e- ained model, which has been s udied o a
ew yea s unde a a ie y o condi ions and scena ios. In ou wo k, we used backg ound
sub ac ion (BS) wi h CNN o p edic each ame in he inpu ideo and ank he sco e
using he MV algo i hm o de e mine whe he he inpu ideo is eal o ake.
Inspi ed by he wo k o ame di e ence and mul ile el ep esen a ion
(FDML) [7],
we p opose an e ec i e sys em o ace p esen a ion a ack de ec ion. To do his, we
sugges using he backg ound subs uc ion me hod in he p ep ocessing s ep o adjus he
ace’s mo ion. The MV algo i hm is used o imp o e he pe o mance a e as well as he
decision o he inpu ideo a e p edic ing he sco e o each ame by ResNe -50. To es ou
sys em, we used ideos om nume ous public ace spoo da abases wi h a ying quali y,
esolu ions, and dynamic anges. We also compa ed ou esul s o hose o a numbe o
cu en s a e-o - he-a app oaches. The ollowing a e he main con ibu ions o his wo k:
•
Imp o ing ace p esen a ion a ack de ec ion using BS ha disc imina es he mo ion
o eal ace om a ake one.
•
Fine- uning he ResNe -50 model o he ace p esen a ion a ack de ec ion ask o
ex ac meaning ul deep acial ea u es.
•
Using he MV algo i hm o inc ease he classi ica ion a e o he sys em, which is
clea ly obse ed when he me hodology ou pe o med p e ious me hodologies in he
li e a u e, acco ding o he esul s o ou expe imen s.
•
Tackling he senso in e ope abili y p oblem by including he expe imen s o in e -
da abase and in a-da abase es s.
The emainde o he pape is s uc u ed as ollows. Sec ion 2desc ibes ela ed
wo k on ace p esen a ion a ack de ec ion. Then, ou app oach is desc ibed in de ail in
Sec ion 3. Sec ion 4summa izes he expe imen al esul s and p o ides a compa a i e
analysis. The sec ion also desc ibes he da abases ha we used in ou es s. Sec ion 5d aws
some conclusions and highligh s some u u e di ec ions. Abb e ia ions de ine he main
ac onyms used in his pape .
2. Rela ed Wo k
P esen a ion a ack can be de ec ed in a a ie y o ways. In his pape , we ocus on
wo ypes o ace p esen a ion a ack de ec ion me hods: handc a ed and deep lea ning-
based me hods. In his sec ion, we p esen mos p e ious wo k in ace p esen a ion a ack
de ec ion. Howe e , we only ocus on hose ha a e hema ically close o ou goals
and con ibu ions.
2.1. Handc a -Based Techniques
Tex u e ea u es, which can desc ibe he con en s and de ails o a speci ic egion in an
image, a e an impo an low-le el ea u e in ace p esen a ion a ack de ec ion me hods.
The e o e, he analysis o image ex u e in o ma ion is used in many echniques, such
as comp essed sensing, which p ese es ex u e in o ma ion and denoising a he same
ime [
8
,
9
]. These echniques based on handc a ed ea u es p o ide accu a e ea u es ha
inc ease he de ec ion a e o a spoo ing sys em. Smi h e al. [
10
] p oposed a me hod o
coun e ing a acks on ace ecogni ion sys ems by using he colo e lec ed om he use ’s
ace as displayed on mobile de ices. The p esence o absence o hese e lec ions can be
u ilized o es ablish whe he o no he images we e cap u ed in eal ime. The algo i hms
use simple RGB images o de ec p esen a ion a ack. These s a egies can be classi ied in o
Senso s 2022,22, 3760 3 o 17
wo ca ego ies: s a ic and dynamic app oaches. The s a ic is used on a single image, whils
dynamic is used on he ideo.
The majo i y o app oaches o dis inguishing be ween eal and syn he ic aces a e
ocused on ex u e analysis. A ashloo e al. [
11
] combined wo spa ial– empo al desc ip o s
using ke nel disc iminan analysis usion. They a e mul iscale bina ized s a is ical image
ea u es on h ee o hogonal planes (MBSIF-TOP) and mul iscale local phase quan iza ion
on h ee o hogonal planes (MLPQ-TOP). To dis inguish be ween eal and ake indi iduals,
Pe ei a e al. [
12
] also expe imen ed wi h a dynamic ex u e ha was based on local
bina y pa e n on h ee o hogonal planes (LBP-TOP). The good esul s o LBP-TOP a e
due o he ac ha empo al in o ma ion is c ucial in ace p esen a ion a ack de ec ion.
Ti unaga i e al. [
13
] used local bina y pa e n(s) (LBP) o dynamic pa e ns and dynamic
mode decomposi ion (DMD) o isual dynamics. Wen e al. [
14
] p oposed an image
dis o ion analysis-based me hod (IDA). To ep esen he ace images, ou di e en ea u es
we e used: blu iness, colo di e si y, specula e lec ion, and ch oma ic momen s, also
elying on he ea u es ha can de ec di e ences be ween eal image and ake one wi hou
cap u ing any in o ma ion abou he use ’s iden i y. Pa el e al. [
15
] in es iga ed he impac
o di e en RGB colo channels (R, G, B, and g ay scale) and di e en acial egions on he
pe o mance o LBP and dense scale in a ian ea u e ans o m (DSIFT) based algo i hms.
Thei in es iga ions e ealed ha ex ac ing he ex u e om he ed channel p oduces he
bes esul s. Boulkena e e al. [
16
] p oposed a colo ex u e analysis-based ace p esen a ion
a ack de ec ion app oach. They employed he LBP desc ip o o ex ac ex u e ea u es
om each channel a e encoding he RGB images in wo colo spaces: HSV and YCbC ,
and hen conca ena ed hese ea u es o dis inguish be ween eal and ake aces.
Some me hods, such as [
17
], ha e ecen ly used use -speci ic in o ma ion o imp o e
he pe o mance o ex u e-based FAS echniques. Ga cia e al. [
18
] p oposed ace p esen a-
ion a ack de ec ion by looking o Moi é pa e ns caused by digi al g id o e lap whe e
hei de ec ion is based on equency domain peak de ec ion. Fo classi ica ion, hey used
suppo ec o machines (SVM) wi h an adial basis unc ion ke nel. They s a ed o un
hei es s on he Replay A ack Co pus and Moi é da abases. O he ace p esen a ion
a ack de ec ion solu ions a e based on ex u es on 3D models, such as hose used in [
19
].
Because he a acke in 3D models u ilizes a mask o spoo he sys em, he in oduc ion o
w inkles migh be ex emely help ul in de ec ing he a ack. The p esen ed wo k in [
19
]
examines he iabili y o pe o ming low-cos assaul s on 2.5D and 3D ace ecogni ion
sys ems using sel -manu ac u ed h ee-dimensional (3D) p in ed models.
2.2. Deep Lea ning-Based Techniques
Ac ually, deep lea ning is used in a a ie y o sys ems and applica ions o biome ic
au hen ica ion [
20
], whe e he deep ne wo k can be ained using a numbe o pa e ns.
A e lea ning all o he da ase ’s unique ea u es, he ne wo k can be used o iden i y simila
pa e ns. Deep lea ning app oaches ha e mos ly been used o lea n ace p esen a ion a ack
de ec ion ea u es. Mo eo e , deep lea ning is e icien a classi ica ion (supe ised lea ning)
and clus e ing asks (unsupe ised lea ning). Thus, he sys em assigns class labels o he
inpu ins ances in a classi ica ion ask, bu he ins ances in clus e ing app oaches a e
clus e ed based on hei simila i y wi hou he usage o class labels.
To ain models wi h signi ican disc imina i e abili ies, Yang e al. [
21
] used a deep
CNN a he han manually cons uc ing ea u es om sc a ch. Quan e al. p oposed a semi-
supe ised lea ning-based a chi ec u e o igh ace p esen a ion a ack h ea s using only a
ew agged da a, a he han depending on ime-consuming da a anno a ions. They assess
he eliabili y o selec ed da a pseudo labels using a empo al consis ency equi emen . As a
esul , ne wo k aining is subs an ially acili a ed. Mo eo e , by p og essi ely inc easing
he con ibu ion o unlabeled a ge domain da a o he aining da a, an adap i e ans e
mechanism can be implemen ed o elimina e domain bias. Acco ding o he au ho s in [
22
],
hey use a ype o g ound h ough (GT) e med app -GT in conjunc ion wi h he iden i y
in o ma ion o he spoo image o gene a e a genuine image o he app op ia e subjec in
Senso s 2022,22, 3760 4 o 17
he aining se . A me ic lea ning module cons ains he gene a ed genuine images om
he spoo images o be nea he app -GT and a om he inpu images. This educes he
e ec o changes in he imaging en i onmen on he app -GT and GT o a spoo image.
Jia e al. [
23
] p oposed a uni ied unsupe ised and semi-supe ised domain adap a ion
ne wo k (USDAN) o c oss-scena io ace p esen a ion a ack de ec ion, wi h he pu pose o
educing he dis ibu ion misma ch be ween he sou ce and a ge domains. The ma ginal
dis ibu ion alignmen module (MDA) and he condi ional dis ibu ion alignmen module
(CDA) a e wo modules ha use ad e sa ial lea ning o ind a domain-in a ian ea u e
space and condense ea u es o he same class.
Raw op ical low da a om he clipped ace egion and he comple e scene we e used
o ain a neu al ne wo k by Feng’s eam e al. [
24
]. Mo ion-based p esen a ion a ack
de ec ion does no need a scenic model o mo ion assump ion o gene alize. They p esen
an image quali y-based and mo ion-based li eness amewo k ha can be used oge he
using a hie a chical neu al ne wo k.
In hei wo k [
25
], Liu e al. p oposed a deep ee ne wo k (DTN) ha lea ns cha ac e -
is ics in a hie a chical o m and may de ec unan icipa ed p esen a ion a ack ins umen
by iden i ying he ea u es ha a e lea ned.
Yu e al. [
26
] in oduces wo new con olu ion and pooling ope a o s o encoding ine-
g ained in a ian in o ma ion: cen al di e ence con olu ion (CDC) and cen al di e ence
pooling (CDP). CDC ou pe o ms anilla con olu ion in ex ac ing in insic spoo ing
pa e ns in a numbe o si ua ions.
As desc ibed in Qin e al. [
27
], adap i e inne -upda e (AIU) is a no el me a lea ning
app oach ha uses a me a-lea ne o ain on ze o- and ew-sho FAS asks u ilizing a newly
cons uc ed Adap i e Inne upda e Me a Face An i spoo ing (AIM-FAS).
Acco ding o Yu e al. [
28
], he mul i-le el ea u e e inemen module (MFRM) and
ma e ial-based mul i-head supe ision can help inc ease BCN’s pe o mance. In he i s
app oach, local neighbo hood weigh s a e eassembled o c ea e mul i-scale ea u es, while
in he second, he ne wo k is o ced o acqui e s ong sha ed ea u es in o de o pe o m
asks wi h mul iple heads.
CDC-based ame le el FAS app oaches, p oposed by he au ho s in [
29
], ha e been
de eloped. These pa e ns can be cap u ed by agg ega ing in o ma ion abou in ensi y
and g adien . In compa ison o a anilla con olu ional ne wo k, he cen al di e ence
con olu ional ne wo k (CDCN) buil wi h CDC has a mo e obus modeling capabili y.
CDCN++ is an imp o ed e sion o CDCN ha inco po a es he sea ch backbone ne -
wo k wi h he mul iscale a en ion usion module (MAFM) o collec ing mul i-le el CDC
ea u es e ec i ely.
Spa io empo al an i-spoo ne wo k (STASN) is a new a en ion mechanism in en ed
by Yang e al. [
30
] ha combines global empo al and local spa ial in o ma ion, allowing
hem o examine he model’s unde s andable beha io s.
To imp o e CNN gene aliza ion, Liu e al. [
31
] p oposed o use inno a i e auxilia y
in o ma ion o supe ise CNN aining. A new CNN-RNN a chi ec u e o lea ning he
dep h map and PPG signal om end- o-end is also p oposed.
Wang e al. [32] p oposed a dep h-supe ised a chi ec u e ha can e icien ly encode
spa io empo al in o ma ion o p esen a ion a ack de ec ion and de elops a new app oach
o es ima ing dep h in o ma ion om se e al RGB ames. Sho - e m ex ac ion is ac-
complished h ough he use o wo unique modules: he op ical low-guided ea u e block
(OFFB) and he con olu ion ga ed ecu en uni s (Con GRU). Jou abloo e al. [
33
] p o-
posed a new CNN a chi ec u e o ace p esen a ion a ack de ec ion, wi h app op ia e
cons ain s and supplemen a y supe isions, o disce n be ween li ing and ake aces,
as well as long- e m mo ion. In o de o de ec p esen a ion a acks e ec i ely and e -
icien ly, Kim e al. [
34
] in oduced he bipa i e auxilia y supe ision ne wo k (BASN),
an a chi ec u e ha lea ns o ex ac and agg ega e auxilia y in o ma ion.
Huszá e al. [35] p oposed a deep lea ning (DL) app oach o add ess he p oblem o
p esen a ion a ack ins umen s occu ing om ideo. The app oach was es ed in a new
Senso s 2022,22, 3760 5 o 17
da abase made up o se e al ideos o use s juggling a oo ball. Thei algo i hm is capable
o unning in pa allel wi h he human ac i i y ecogni ion (HAR) in eal- ime.
Roy e al. [36]
p oposed an app oach called he bi-di ec ional ea u e py amid ne wo k (BiFPN) o de ec
p esen a ion a acks because he app oach con aining high-le el in o ma ion demons a es
negligible imp o emen s. Ali e al. [
37
]—based on s imula ing eye mo emen s by using he
use o isual s imuli wi h andomized ajec o ies o de ec p esen a ion a ack ins umen .
Ali e al. [
38
]—by he combina ion o wo me hods, which a e head-de ec ion algo i hm
and deep neu al ne wo k-based classi ie s. The es in ol ed a ious ace p esen a ion
a acks in he mal in a ed in a ious condi ions.
I appea s ha mos o he exis ing handc a and deep lea ning-based ea u es may
no be op imal o he FAS ask due o he limi ed ep esen a ion capaci y o in insic
spoo ing ea u es. In o de o lea n mo e obus ea u es o he domain shi as well as mo e
disc imina i e pa e ns o li eness de ec ion, we p opose deep backg ound sub ac ion
and majo i y o e algo i hm o ake in o accoun bo h dynamic and s a ic in o ma ion.
3. P oposed App oach
Figu e 1desc ibes he o e all s uc u e o ou p oposed app oach, which is di ided
in o h ee modules: backg ound sub ac ion, ea u e lea ning, and da a classi ica ion.
To begin, we used he backg ound sub ac ion be ween consecu i e ames o ex ac
mo ion, we can also call his echnique BS. Then, he ea u es we e ex ac ed using he
ResNe -50 ans e lea ning model on he o eg ound o BS. Finally, o dis inguish be ween
eal and ake aces o each ame, we used a classi ica ion laye ha employed a ully
connec ed laye . A e ha , we used MV o p edic whe he he inpu ideo was eal o no .
In he subsec ions ha ollow, all subsys ems (modules) will be discussed.
Inpu s eam Ou pu masks
Classi ica ion Module
Fea u es lea ning using ans e
lea ning
Resne 50
D opou + Relu
Classi ica ion Laye (Wi h me ics upda es)
Laye 1 x 2
Real
Fake
Majo i y Vo es
Real
Fake
FD
FD
FD
FD
FD
ResNe 50
Ou pu
FC Laye 1x1024
Real
Fake
Real
Fake
Real
Fake
Real
Fake
Backg ound Sub ac ion Module Fea u e Lea ning Module
Figu e 1. F amewo k o ou p oposed app oach.
3.1. Backg ound Sub ac ion Module
In ou esea ch wo k, we used a ace p esen a ion a ack de ec ion sys em based
on an ex ended BS algo i hm. The backg ound sub ac ion app oach, which is based on
he p emise o ob aining he pixels in he image sequence di e ence ope a ion o do wo
o h ee con inuous ames, is he mos commonly used ac ion a ge de ec ion measu e.
Using an image pixel alue ob ained by sub ac ing he di e ence image and he bina ized
di e ence image, i he pixel alue change h eshold is less han a p ede ined one, we can
eel his as a backg ound pixel in he adjacen ame. I he pixel alue o an image a ea
changes d ama ically, i is possible o deduce ha his is due o he ac ion o de ec ing
Senso s 2022,22, 3760 6 o 17
spoo in he image caused by hese symbols as o eg ound pixel egions. While aking
dynamic in o ma ion in o accoun , a pixel egion based on symbolic ac ions can de e mine
he posi ion o he a ge in he image.
Backg ound sub ac ion is applied o images and he h eshold esul is displayed
as a o eg ound image. Figu e 2shows an example o he ou pu . This is a low-cos and
ine ec i e me hod o de ec ing mo ion in a ideo s eam. The image
P
is ans o med
in o a g ey-scale (in ensi y) image
I
. Then, gi en he image
I
and he p e ious image
I −1
,
he cu en ou pu is R , whe e:
R (x,y) = I (x,y)i |I (x,y)−I −1(x,y)|>T
0 o he wise (1)
T
is he alue o he h eshold pa ame e . In ou si ua ion, we jus u ilized a h eshold
o emo e he pixels wi h he same alues ac oss he wo ames. The o eg ound pixels
ake he alue o he cu en ame i he e is mo ion. The o eg ound pixel is se o ze o i
he e is no mo ion.
We apply a mo ing window mechanism ha akes he i s ame wi hin a window o
5 ames. This means ha we ake a ame and d op 4 ames in each cycle. Fo example,
in a ideo o 150 ames, we use 30 ames and d ops 120 ames.
Real P in ed Replay
RGB G ay BS
Figu e 2.
Example o a genuine ace and co esponding p in and eplay a acks in g ey-scale and BS.
3.2. Fea u e Lea ning Module
Fea u e lea ning (FL) is a se o app oaches in machine lea ning ha allow a sys em o
disco e he ep esen a ion needed o ea u e de ec ion, p edic ion, o classi ica ion om a
p ep ocessed da ase au oma ically. This enables a machine o lea n he ea u es and apply
hem o a speci ic ask-like classi ica ion and p edic ion. Fea u e lea ning can be achie ed in
deep lea ning by ei he c ea ing a comple e CNN o ain and es he collec ion o images o
adap ing a p e- ained CNN o classi ica ion o p edic ion o he new images-se . T ans e
lea ning is he la e s a egy used in he deep lea ning domain. T ans e lea ning is a
machine lea ning echnique in which a model c ea ed o one ask is u ilized as he basis
o a model on a di e en ask.
T ans e lea ning (TL) is commonly used in deep lea ning (DL) applica ions o allow
you o use a p e- ained ne wo k o sol ing new classi ica ion asks. To mee he new
lea ning asks, he lea ning pa ame e s o he p e- ained ne wo k wi h andomly ini ialized
Senso s 2022,22, 3760 7 o 17
weigh s mus be ine- uned. T ans e lea ning is ypically conside ably as e and easie
o lea n/ ain han building a ne wo k om he ini ial concep . T ans e lea ning is an
op imiza ion and a quick way ha can sa e ime o imp o e e iciency.
In his sec ion, a ans e lea ning echnique is applied by ine- uning a p e ained
ResNe -50 model on ImageNe da ase using mul iple spoo ing da ase s whe e he ou pu
o he las FC laye is changed o ou pu wo classes ( eal/ ake). The ne wo k is called
ResNe -50, due o he ac ha i has 48 con olu ion laye s, along wi h 1 MaxPool and 1
A e age Pool laye , and i in oduced he use o esidual blocks.
3.3. Classi ica ion Module
Da a classi ica ion is a i al p ocess o sepa a ing la ge da ase s in o classes o
decision-making, pa e n de ec ion, and o he pu poses. Fo mul i-class classi ica ion
p oblems wi h mu ually exclusi e classes, a classi ica ion laye uses a ully connec ed laye
o compu e he c oss-en opy loss.
The ea u es om ResNe -50 a e passed ia a FC laye made o 1024 neu ons wi h a
40% d opou o p e en o e - i ing in he classi ica ion module. Ha ing ollowed ha ,
he uni s we e ac i a ed wi h a ec i ica ion mechanism called ReLU. MAX (X, 0) is he
ReLu unc ion, which se s all nega i e alues in he ma ix
X
o ze o while keeping all o he
alues cons an . The eason o choosing ReLU is ha deep ne wo k aining wi h ReLU
ended o con e ge conside ably as e and mo e eliably han deep ne wo k aining wi h
sigmoid ac i a ion. Finally, he ou pu laye comp ised o one neu on uni p og ammed
o compu e p obabili ies o he classes using he sigmoid unc ion (bina y classi ie ). A
sigmoid is a ma hema ical unc ion ha akes a ec o o k eal alues and changes i o a
wo-p obabili y p obabili y dis ibu ion.
We employed o ing ensemble in ou es s o classi y each subjec ( ideo) as eal o
ake. A o ing ensemble (some imes known as a "majo i y o ing ensemble") is a machine
lea ning model ha inco po a es p edic ions om se e al o he models, such as mul iple
p edic ions in each ame a e he inpu ideo’s las laye (classi ica ion laye ). The p edic-
ions o each label a e combined, and he label wi h he majo i y o e is o ecas ed (see
Figu e 1, classi ica ion module) o de e mine i he inpu ideo belongs o a eal o ake
one. The majo i y o ing ensemble c ea es o ecas s based on he mos common one. I is a
s a egy ha can be u ilized o boos pe o mance, wi h he goal o ou pe o ming e e y
ame used independen ly in he ensemble.
4. Expe imen al Resul s and Analysis
In his sec ion, he employed benchma k da ase s will be in oduced i s , ollowed by
a b ie desc ip ion o he e alua ion c i e ia. A e ha , we p esen and analyze a se ies o
expe imen s ha we assume demons a e he e icacy o he p oposed BS-CNN+MV based
ace p esen a ion a ack de ec ion echnique.
4.1. Da abase and P o ocol
In o de o assess o he e ec i eness o ou p oposed p esen a ion a ack de ec-
ion echnique, we pe o med a se o expe imen s on well known da abases whe e he
op h ee challenging da abases we e used: he CASIA-FASD (h p://www.cbs .ia.ac.
cn/english/FaceAn iSpoo Da abases.asp (accessed on 4 July 2014)); ace an i-spoo ing
da abase, Replay-A ack (h ps://www.idiap.ch/da ase / eplaya ack (accessed on 6 Au-
gus 2014)) da abase, and MSU (h ps://d i e.google.com/d i e/ olde s/1nJCPdJ7R6
7xOiklF1omk z4yHeJwhQsz (accessed on 8 May 2014)) mobile ace spoo ing da abases.
Those da abases con ain ideo eco dings o eal and ake a acks. A b ie desc ip ion o
hese da abases is gi en as ellow:
The CASIA-FASD da abase [
39
] is a da ase o ace p esen a ion a ack de ec ion.
This da abase con ains 50 genuine subjec s in o al and he co esponding ake aces a e
cap u ed wi h high quali y om he o iginal ones. The e o e each subjec con ains 12 ideos
(3 genuine and 9 ake) unde h ee di e en esolu ions and ligh condi ions, namely low
Senso s 2022,22, 3760 8 o 17
quali y, no mal quali y, and high quali y. Mo eo e , h ee ake ace a acks a e designed,
which include wa ped pho o a ack, cu pho o a ack, and ideo a ack. The o e all
da abase con ains 600 ideo clips and he subjec s a e di ided in o subse s o pe o ming
aining and es s in which 240 ideos o 20 subjec s a e used o aining and 360 ideos
o
30 subjec s
o es ing. Tes p o ocol is p o ided, which consis s o 7 scena ios o a
ho ough e alua ion om all possible aspec s (see Figu e 3).
Low
Quali y
No mal
Quali y
High
Quali y
Real
Face Wa ped Pho o
A ack
Cu Pho o
A ack Video
A ack
Figu e 3. Samples om he CASIA FASD da abase.
Among he popula da abases designed o he p esen a ion a ack de ec ion applica-
ion, one can ind he Replay-A ack da abase [
40
]. This da abase consis s o 1300 ideo o
eal-access and a ack a emp s o 50 subjec s, (See Figu e 4). Howe e , hese ideos we e
aken using a buil -in webcam on a MacBook lap op unde wo sepa a e scena ios (con-
olled and ad e se). In addi ion, wo came as we e used o c ea e he aked acial a ack
o each pe son in high- esolu ion images and ideos: a Canon Powe Sho SX150 IS and an
iPhone 3GS came a. Mo eo e , ixed a acks and hand a acks a e he wo ypes o a acks.
The e a e en ideos in each subse : ou mobile a acks wi h a esolu ion o 480
×
320 pixels
on an iPhone 3GS sc een, hen, by using a i s gene a ion iPad wi h a sc een esolu ion
o 1024
×
768 pixels, ou high- esolu ion sc een a acks we e pe o med. On A4 pape ,
wo ha d-copy p in a acks (p in ed on a T iumph-Adle DCC 2520 colou lase p in e )
occupied he whole a ailable p in ing su ace. I should be no ed ha he comple e se o
ideos is di ided in o h ee non-o e lapping subse s o aining, de elopmen , and es ing
in o de o e alua e hem.
The pa e n ecogni ion and image p ocessing (PRIP) g oup a Michigan S a e Uni e -
si y de eloped a publicly a ailable MSU-MFSD [
14
] da abase o ace p esen a ion a acks.
The da abase con ains 280 ideo clips o a emp ed pho o and ideo a acks on 35 clien s. I
was c ea ed using a mobile phone o cap u e bo h genuine and p esen a ion a ack. This
was accomplished using wo ypes o came as: (1) he buil -in came a in he MacBook Ai
13 inch (640
×
480) and (2) he on - acing came a on he Google
Nexus 5 And oid
phone
(720
×
480). Each subjec ecei ed wo ideo eco dings, he i s o which was aken using
a lap op came a and he second wi h an And oid came a (See Figu e 5). High- esolu ion
ideo was eco ded o each subjec u ilizing wo de ices o c ea e he a acks: (1) Canon
Powe Sho 550D SLR came a, which cap u es 18.0 megapixel pho os and 1080p high-
de ini ion ideo clips; (2) iPhone 5S back- acing came a, which cap u es 1080p ideo clips.
The e a e h ee ypes o p esen a ion a ack ins umen s, he i s one (1) high- esolu ion
Senso s 2022,22, 3760 9 o 17
eplay ideo, he i s ype o p esen a ion a ack ins umen is a high- esolu ion eplay
ideo a ack using an iPad Ai sc een, wi h a esolu ion o 2048
×
1536, he second is a
mobile phone eplay ideo a ack using an iPhone 5S sc een, wi h a esolu ion o 1136
×
640,
and he hi d is a p in ed pho o a ack using an A3 pape wi h a ully-occupied p in ed
pho o o he clien ’s biome y, wi h a pape size o : 11
×
17 (279 mm
×
432 mm), p in ed
wi h an HP Lase Je CP6015xh p in e a a esolu ion o 1200
×
600 dpi. Finally, o assess
pe o mance, he 35 subjec s in he MSU-MFSD da abase we e di ided in o wo subse s: 15
o aining and 20 o es ing.
Ad e se
Scena io
Con olled
Scena io
Real Access Pho o A ack
Fixed
Pho o A ack
Hand
Video A ack
Hand
Video A ack
Fixed
Figu e 4. Examples o eal accesses and a acks in di e en scena ios.
Google Nexus 5
sma phone
came a
Mac Book Ai
13" lap op
came a
Genuine aces Spoo aces by
p in ed pho o
Spoo aces by
iPhone
Spoo aces by
iPad
Figu e 5.
Example images o genuine and a ack p esen a ion o one o he subjec s in he MSU-
MFSD da abase.
Senso s 2022,22, 3760 16 o 17
4.
Khaldi, Y.; Benzaoui, A.; Ouahabi, A.; Jacques, S.; Taleb-Ahmed, A. Ea Recogni ion Based on Deep Unsupe ised Ac i e
Lea ning. IEEE Sens. J. 2021,21, 20704–20713. [C ossRe ]
5.
Benlamoudi, A. Mul i-Modal and An i-Spoo ing Pe son Iden i ica ion. Ph.D. Disse a ion, Depa men o Elec onics and
Telecommunica ions, Facul y o New Technologies o In o ma ion and Communica ion (FNTIC), Uni e si y o Kasdi Me bah,
Oua gla, Alge ia, 2018.
6.
He, K.; Zhang, X.; Ren, S.; Sun, J. Deep esidual lea ning o image ecogni ion. In P oceedings o he IEEE Con e ence on
Compu e Vision and Pa e n Recogni ion, Las Vegas, NV, USA, 27–30 June 2016; pp. 770–778.
7.
Benlamoudi, A.; Aiadi, K.E.; Oua i, A.; Samai, D.; Oussalah, M. Face an ispoo ing based on ame di e ence and mul ile el
ep esen a ion. J. Elec on. Imaging 2017,26, 043007. [C ossRe ]
8.
Mahdaoui, A.E.; Ouahabi, A.; Moulay, M.S. Image Denoising Using a Comp essi e Sensing App oach Based on Regula iza ion
Cons ain s. Senso s 2022,22, 2199. [C ossRe ]
9.
Haneche, H.; Boud aa, B.; Ouahabi, A. A new way o enhance speech signal based on comp essed sensing. Measu emen
2020
,
151, 107117. [C ossRe ]
10.
Smi h, D.F.; Wiliem, A.; Lo ell, B.C. Face ecogni ion on consume de ices: Re lec ions on eplay a acks. In . Fo ensics Secu .
IEEE T ans. 2015,10, 736–745. [C ossRe ]
11.
A ashloo, S.R.; Ki le , J.; Ch is mas, W. Face spoo ing de ec ion based on mul iple desc ip o usion using mul iscale dynamic
bina ized s a is ical image ea u es. In . Fo ensics Secu . IEEE T ans. 2015,10, 2396–2407. [C ossRe ]
12.
de F ei as Pe ei a, T.; Komulainen, J.; Anjos, A.; De Ma ino, J.M.; Hadid, A.; Pie ikäinen, M.; Ma cel, S. Face li eness de ec ion
using dynamic ex u e. EURASIP J. Image Video P ocess. 2014,2014, 1–15. [C ossRe ]
13.
Ti unaga i, S.; Poh, N.; Wind idge, D.; Io liam, A.; Suki, N.; Ho, A.T. De ec ion o ace spoo ing using isual dynamics. In .
Fo ensics Secu . IEEE T ans. 2015,10, 762–777. [C ossRe ]
14.
Wen, D.; Han, H.; Jain, A.K. Face spoo de ec ion wi h image dis o ion analysis. In . Fo ensics Secu . IEEE T ans.
2015
,10, 746–761.
[C ossRe ]
15.
Pa el, K.; Han, H.; Jain, A.K.; O , G. Li e ace ideo s. spoo ace ideo: Use o moi é pa e ns o de ec eplay ideo a acks. In
P oceedings o he Biome ics (ICB), 2015 In e na ional Con e ence on IEEE, Phuke , Thailand, 19–22 May 2015; pp. 98–105.
16.
Boulkena e , Z.; Komulainen, J.; Hadid, A. Face an i-spoo ing based on colo ex u e analysis. In P oceedings o he Image
P ocessing (ICIP), 2015 IEEE In e na ional Con e ence on IEEE, Quebec Ci y, QC, Canada, 27–30 Sep embe 2015; pp. 2636–2640.
17.
Chingo ska, I.; dos Anjos, A.R. On he use o clien iden i y in o ma ion o ace an ispoo ing. IEEE T ans. In . Fo ensics Secu .
2015,10, 787–796. [C ossRe ]
18.
Ga cia, D.C.; de Quei oz, R.L. Face-spoo ing 2D-de ec ion based on Moi é-pa e n analysis. IEEE T ans. In . Fo ensics Secu .
2015
,
10, 778–786. [C ossRe ]
19.
Galbally, J.; Sa a, R. Th ee-dimensional and wo-and-a-hal -dimensional ace ecogni ion spoo ing using h ee-dimensional
p in ed models. IET Biom. 2016,5, 83–91. [C ossRe ]
20.
Bekhouche, S.E.; Do naika, F.; Benlamoudi, A.; Oua i, A.; Taleb-Ahmed, A. A compa a i e s udy o human acial age es ima ion:
Handc a ed ea u es s. deep ea u es. Mul imed. Tools Appl. 2020,79, 26605–26622. [C ossRe ]
21. Yang, J.; Lei, Z.; Li, S.Z. Lea n con olu ional neu al ne wo k o ace an i-spoo ing. a Xi 2014, a Xi :1408.5601.
22.
Xu, Y.; Wu, L.; Jian, M.; Zheng, W.S.; Ma, Y.; Wang, Z. Iden i y-cons ained noise modeling wi h me ic lea ning o ace
an i-spoo ing. Neu ocompu ing 2021,434, 149–164. [C ossRe ]
23.
Jia, Y.; Zhang, J.; Shan, S.; Chen, X. Uni ied unsupe ised and semi-supe ised domain adap a ion ne wo k o c oss-scena io
ace an i-spoo ing. Pa e n Recogni . 2021,115, 107888. [C ossRe ]
24.
Feng, L.; Po, L.M.; Li, Y.; Xu, X.; Yuan, F.; Cheung, T.C.H.; Cheung, K.W. In eg a ion o image quali y and mo ion cues o ace
an i-spoo ing: A neu al ne wo k app oach. J. Vis. Commun. Image Rep esen . 2016,38, 451–460. [C ossRe ]
25.
Liu, Y.; S ehouwe , J.; Jou abloo, A.; Liu, X. Deep ee lea ning o ze o-sho ace an i-spoo ing. In P oceedings o he IEEE/CVF
Con e ence on Compu e Vision and Pa e n Recogni ion, Long Beach, CA, USA, 16–20 June 2019; pp. 4680–4689.
26.
Yu, Z.; Wan, J.; Qin, Y.; Li, X.; Li, S.Z.; Zhao, G. Nas- as: S a ic-dynamic cen al di e ence ne wo k sea ch o ace an i-spoo ing.
a Xi 2020, a Xi :2011.02062.
27.
Qin, Y.; Zhao, C.; Zhu, X.; Wang, Z.; Yu, Z.; Fu, T.; Zhou, F.; Shi, J.; Lei, Z. Lea ning me a model o ze o-and ew-sho ace
an i-spoo ing. In P oceedings o he AAAI Con e ence on A i icial In elligence, New Yo k, NY, USA, 7–12 Feb ua y 2020;
Volume 34, pp. 11916–11923.
28.
Yu, Z.; Li, X.; Niu, X.; Shi, J.; Zhao, G. Face an i-spoo ing wi h human ma e ial pe cep ion. In Eu opean Con e ence on Compu e
Vision; Sp inge : Be lin/Heidelbe g, Ge many, 2020; pp. 557–575.
29.
Yu, Z.; Zhao, C.; Wang, Z.; Qin, Y.; Su, Z.; Li, X.; Zhou, F.; Zhao, G. Sea ching cen al di e ence con olu ional ne wo ks o ace
an i-spoo ing. In P oceedings o he IEEE/CVF Con e ence on Compu e Vision and Pa e n Recogni ion, Sea le, WA, USA,
13–19 June 2020; pp. 5295–5305.
30.
Yang, X.; Luo, W.; Bao, L.; Gao, Y.; Gong, D.; Zheng, S.; Li, Z.; Liu, W. Face an i-spoo ing: Model ma e s, so does da a. In
P oceedings o he IEEE/CVF Con e ence on Compu e Vision and Pa e n Recogni ion, Long Beach, CA, USA, 15–20 June 2019;
pp. 3507–3516.
31.
Liu, Y.; Jou abloo, A.; Liu, X. Lea ning Deep Models o Face An i-Spoo ing: Bina y o Auxilia y Supe ision. In P oceedings o
he IEEE Con e ence on Compu e Vision and Pa e n Recogni ion, Sal Lake Ci y, UT, USA, 18–23 June 2018; pp. 389–398.
Senso s 2022,22, 3760 17 o 17
32.
Wang, Z.; Zhao, C.; Qin, Y.; Zhou, Q.; Qi, G.; Wan, J.; Lei, Z. Exploi ing empo al and dep h in o ma ion o mul i- ame ace
an i-spoo ing. a Xi 2018, a Xi :1811.05118.
33.
Jou abloo, A.; Liu, Y.; Liu, X. Face de-spoo ing: An i-spoo ing ia noise modeling. In P oceedings o he Eu opean con e ence on
compu e ision (ECCV), Munich, Ge many, 8–14 Sep embe 2018, pp. 290–306.
34.
Kim, T.; Kim, Y.; Kim, I.; Kim, D. Basn: En iching ea u e ep esen a ion using bipa i e auxilia y supe isions o ace an i-spoo ing.
In P oceedings o he IEEE/CVF In e na ional Con e ence on Compu e Vision Wo kshops, Seoul, Ko ea, 27–28 Oc obe 2019.
35.
Huszá , V.D.; Adhika la, V.K. Li e Spoo ing De ec ion o Au oma ic Human Ac i i y Recogni ion Applica ions. Senso s
2021
,21,
7339. [C ossRe ] [PubMed]
36.
Roy, K.; Hasan, M.; Rup y, L.; Hossain, M.S.; Sengup a, S.; Taus, S.N.; Mohammed, N. Bi-FPNFAS: Bi-Di ec ional Fea u e Py amid
Ne wo k o Pixel-Wise Face An i-Spoo ing by Le e aging Fou ie Spec a. Senso s 2021,21, 2799. [C ossRe ] [PubMed]
37.
Ali, A.; Hoque, S.; De a i, F. Di ec ed Gaze T ajec o ies o Biome ic P esen a ion A ack De ec ion. Senso s
2021
,21, 1394.
[C ossRe ] [PubMed]
38. Kowalski, M. A S udy on P esen a ion A ack De ec ion in The mal In a ed. Senso s 2020,20, 3988. . [C ossRe ]
39.
Zhang, Z.; Yan, J.; Liu, S.; Lei, Z.; Yi, D.; Li, S.Z. A ace an ispoo ing da abase wi h di e se a acks. In P oceedings o he
Biome ics (ICB), 2012 5 h IAPR In e na ional Con e ence on IEEE, New Delhi, India, 29 Ma ch–1 Ap il 2012; pp. 26–31.
40.
Chingo ska, I.; Anjos, A.; Ma cel, S. On he e ec i eness o local bina y pa e ns in ace an i-spoo ing. In P oceedings o he 2012
BIOSIG—In e na ional Con e ence o Biome ics Special In e es G oup (BIOSIG), Da ms ad , Ge many, 6–7 Sep embe 2012;
pp. 1–7.
41.
Galbally, J.; Ma cel, S. Face an i-spoo ing based on gene al image quali y assessmen . In P oceedings o he 2014 22nd In e na ional
Con e ence on Pa e n Recogni ion (ICPR), S ockholm, Sweden, 24–28 Augus 2014; pp. 1173–1178.
42.
Pin o, A.; Ped ini, H.; Robson Schwa z, W.; Rocha, A. Face Spoo ing De ec ion Th ough Visual Codebooks o Spec al Tempo al
Cubes. Image P ocess. IEEE T ans. 2015,24, 4726–4740. [C ossRe ]
43.
Benlamoudi, A.; Samai, D.; Oua i, A.; Bekhouche, S.E.; Taleb-Ahmed, A.; Hadid, A. Face spoo ing de ec ion using Local bina y
pa e ns and Fishe Sco e. In P oceedings o he Con ol, Enginee ing & In o ma ion Technology (CEIT), 2015 3 d In e na ional
Con e ence on IEEE, Tlemcen, Alge ia, 25–27 May 2015; pp. 1–5.
44.
Yang, J.; Lei, Z.; Liao, S.; Li, S.Z. Face li eness de ec ion wi h componen dependen desc ip o . In P oceedings o he Biome ics
(ICB), 2013 In e na ional Con e ence on IEEE, Mad id, Spain, 4–7 June 2013; pp. 1–6.
45.
Benlamoudi, A.; Samai, D.; Oua i, A.; Bekhouche, S.; Taleb-Ahmed, A.; Hadid, A. Face spoo ing de ec ion using Mul i-Le el Local
Phase Quan iza ion (ML-LPQ). In P oceedings o he Fi s In e na ional Con e ence on Au oma ic Con ol, Telecommunica ion
and signals ICATS’15, El-Oued, Alge ia, 16–18 No embe 2015. [C ossRe ]
46.
Benlamoudi, A.; Bougou zi, F.; Zighem, M.; Bekhouche, S.; Oua i, A.; Taleb-Ahmed, A. Face An i-Spoo ing Combining MLLBP
and MLBSIF. In P oceedings o he 10ème Con é ence su le Génie Elec ique, Alge , Alge ie, 30 June 2017.
47.
Quan, R.; Wu, Y.; Yu, X.; Yang, Y. P og essi e T ans e Lea ning o Face An i-Spoo ing. IEEE T ans. Image P ocess.
2021
,
30, 3946–3955. [C ossRe ]
48.
Anjos, A.; Ma cel, S. Coun e -Measu es o pho o a acks in ace ecogni ion: A public da abase and a baseline. In P oceedings o
he Biome ics (IJCB), 2011 In e na ional Join Con e ence on IEEE, Washing on, DC, USA, 11–13 Oc obe 2011; pp. 1–7.
49.
Komulainen, J.; Hadid, A.; Pie ikainen, M.; Anjos, A.; Ma cel, S. Complemen a y coun e measu es o de ec ing scenic ace
spoo ing a acks. In P oceedings o he Biome ics (ICB), 2013 In e na ional Con e ence on IEEE, Mad id, Spain, 4–7 June 2013;
pp. 1–7.
50.
A ashloo, S.R.; Ki le , J.; Ch is mas, W. An anomaly de ec ion app oach o ace spoo ing de ec ion: A new o mula ion and
e alua ion p o ocol. IEEE Access 2017,5, 13868–13882. [C ossRe ]
51.
Xiong, F.; AbdAlmageed, W. Unknown p esen a ion a ack de ec ion wi h ace gb images. In P oceedings o he 2018 IEEE 9 h
In e na ional Con e ence on Biome ics Theo y, Applica ions and Sys ems (BTAS) IEEE, Redondo Beach, CA, USA, 22–25 Oc obe
2018; pp. 1–9.
52.
Boulkena e , Z.; Komulainen, J.; Li, L.; Feng, X.; Hadid, A. Oulu-npu: A mobile ace p esen a ion a ack da abase wi h eal-wo ld
a ia ions. In P oceedings o he 2017 12 h IEEE In e na ional Con e ence on Au oma ic Face & Ges u e Recogni ion (FG 2017)
IEEE, Washing on, DC, USA, 30 May–3 June 2017; pp. 612–618.
53.
Bha adwaj, S.; Dhamecha, T.I.; Va sa, M.; Singh, R. Compu a ionally e icien ace spoo ing de ec ion wi h mo ion magni ica ion.
In P oceedings o he IEEE Con e ence on Compu e Vision and Pa e n Recogni ion Wo kshops, Po land, OR, USA, 23–28 June
2013; pp. 105–110.
54.
F ei as Pe ei a, T.; Anjos, A.; De Ma ino, J.M.; Ma cel, S. Can ace an i-spoo ing coun e measu es wo k in a eal wo ld scena io?
In P oceedings o he Biome ics (ICB), 2013 In e na ional Con e ence on IEEE, Mad id, Spain, 4–7 June 2013; pp. 1–8.
55.
Boulkena e , Z.; Komulainen, J.; Hadid, A. Face spoo ing de ec ion using colou ex u e analysis. IEEE T ans. In . Fo ensics Secu .
2016,11, 1818–1830. [C ossRe ]
56.
Bian, Y.; Zhang, P.; Wang, J.; Wang, C.; Pu, S. Lea ning Mul iple Explainable and Gene alizable Cues o Face An i-spoo ing.
a Xi 2022, a Xi :2202.10187.
