Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
Vol
.
5
,
No
. 3,
J
une
2
0
1
5
,
pp
. 57
9~
58
5
I
S
SN
: 208
8-8
7
0
8
5
79
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
Fingerprint Authentication
Schemesfor Mobile Devices
Ji
nh
o H
a
n
Department o
f
Liberal Studies (C
omput
e
r),
Korean Bi
ble
Uni
v
ersi
ty
,
Korea
Article Info
A
B
STRAC
T
Article histo
r
y:
Received
Ja
n 17, 2015
Rev
i
sed
Ap
r
20
, 20
15
Accepte
d
May 5, 2015
In certa
in appli
c
ations, fing
erpri
n
t authent
i
c
a
tio
n sy
st
em
s require tem
p
lat
e
s
to be s
t
ored in datab
a
s
e
s
.
The po
s
s
i
ble leak
age of
thes
e fingerprin
t
tem
p
lat
e
s
can l
ead to s
e
r
i
o
u
s
s
ecurit
y
and
privac
y th
rea
t
s
.
Therefor
e, wi
th
the frequ
ent
use of fingerp
rint au
thentication on
mobile devices, it h
a
s become
increasingly
important to keep finge
rprint
data safe. Due to rapid
developments in
optical
e
quipm
ent, b
i
om
etri
c s
y
s
t
em
s are now able
to gain
the same biometric images r
e
peatedly
, so
strong f
eatur
es can b
e
selected with
precis
i
on
. S
t
ron
g
featu
r
es
ref
e
r
to hi
gh-quality
f
eatur
es
which can
be easily
distinguished
fro
m other f
eatur
es in b
i
om
etric raw images. In
this paper
,
w
e
introduce an algorithm that id
entifies
these strong featur
es with cer
tain
probability
from
a given fingerp
rint im
age. Onc
e
values
are ex
t
r
act
ed from
thes
e fe
atures
, t
h
e
y
are us
ed
as
the auth
enti
ca
tio
n data
. Us
ing the geom
etri
c
information of
these strong
features, a
can
cela
ble fingerprint
tem
p
late can be
produced, and
the process of ex
tracting
v
a
lu
es and geometri
c inf
o
rm
ation
is
further
explained. Because this
is a
light-weight auth
entication s
c
heme,
this
template has practical usag
e for low pe
rformance mobile
devi
ces
. F
i
nall
y,
w
e
demonstrate th
at our proposed
scheme
s are s
ecure and th
at the user’s
biometric raw data of
the fingerpr
i
nt ar
e
s
a
f
e
,
even
when th
e m
obil
e
dev
i
ce
i
s
lost or stole
n
.
Keyword:
Bio
m
e
t
rics
C
r
y
p
t
o
gra
p
hy
Fi
nge
r
p
ri
nt
Te
m
p
l
a
t
e
s
M
obi
l
e
Devi
ce
s
Copyright ©
201
5 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
Ji
nh
o Ha
n,
Depa
rt
m
e
nt
of
Li
beral
St
u
d
i
e
s (C
om
put
er
),
Ko
rea
n
Bible
Uni
v
ersity
,
32
D
o
ngi
l
-
ro
(s
t
)
2
1
4
-
g
i
l
,
N
o
wo
n-
g
u
,
Se
oul
,
K
o
rea
,
Em
a
il: h
j
in
ob@b
i
b
l
e.ac.k
r
1.
INTRODUCTION
The fi
nge
r
p
ri
nt
aut
h
e
n
t
i
cat
i
on sy
st
em
i
s
no
w wi
del
y
consi
d
ere
d
t
o
be a c
o
nve
ni
ent
h
u
m
a
n
identification
m
e
thod for m
obile de
vices [1]. In the ne
a
r
future,
one m
obile de
vice will have one or
m
o
re
b
i
o
m
etric te
m
p
lates su
ch
as fi
n
g
e
rp
rin
t
tem
p
lates an
d
th
is calls fo
r
h
i
gh
er
secu
rity m
easu
r
es to
k
e
ep
th
e d
a
ta
safe.
A fi
nge
rp
ri
nt
aut
h
ent
i
cat
i
on sy
st
em
con
s
i
s
t
s
of a
n
a
ppl
i
cat
i
on an
d i
t
s
t
e
m
p
l
a
t
e
st
ored i
n
a
dat
a
base
. The
leak
ag
e
of th
ese fing
erprin
t tem
p
la
tes en
ta
ils seriou
s security
and privacy t
h
reats.
In this pa
pe
r, we propose secure a
nd efficient
fi
n
g
er
pri
n
t
aut
h
e
n
t
i
cat
i
on s
c
hem
e
s whi
c
h are desi
g
n
e
d
t
o
s
u
i
t
m
obi
l
e
devi
ces
, a
n
d
e
xpl
ai
n
t
h
e
p
r
oc
ess o
f
p
r
o
d
u
ci
ng
ca
ncel
abl
e
f
i
nge
rp
ri
nt
t
e
m
p
l
a
t
e
s t
h
at
pr
ot
ect
t
h
e
u
s
er’s
priv
acy. Ou
r su
gg
estion
is
b
a
sed
on
t
h
e assu
m
p
ti
o
n
th
at th
e
fing
erp
r
i
n
t sam
p
les are
o
f
ro
bu
st
qu
ality,
n
e
ith
er p
a
rtial
n
o
r wet, with
g
ood
co
nd
itio
n
i
m
ag
es.
Si
n
c
e th
e user
h
a
s en
oug
h ti
m
e
to
au
th
en
ticate h
i
s own
fing
erp
r
i
n
t th
ro
ugh
m
u
ltip
le trials, th
is assu
m
p
tio
n
is re
aso
n
ab
le.
Stro
ng
feat
u
r
es from
th
e fing
erprin
ts are
u
s
ed
fo
r th
is sch
e
m
e
. Stro
ng featu
r
es refer
to
h
i
gh-qu
ality featu
r
es wh
ich
can
b
e
d
i
sting
u
i
sh
ed
easily fro
m
othe
r features in biom
etric
raw im
ages [2]
.
Through our
(w-k
)
Select
a
l
g
o
rith
m
we will sh
o
w
th
at stron
g
featu
r
es
can
be
fo
und
with
certain
p
r
ob
abilit
y.
W
e
ex
tract v
a
lu
es from
th
e featu
r
es and
t
h
eir
g
e
ometric
inform
ation. T
h
e propose
d cancelable
fi
n
g
erprin
t tem
p
la
te is co
m
p
o
s
ed
of th
ese
v
a
lu
es and
g
e
ometric
inform
ation extracted
from these
feature
s
. The
propos
ed schem
e
s ar
e p
r
odu
ced
u
s
ing
th
is tem
p
la
te.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Fi
nger
p
ri
nt
Au
t
h
ent
i
c
at
i
o
n Sc
heme
sf
or Mo
bi
l
e
Devi
ces
(
J
i
n
ho
H
a
n)
58
0
Geom
et
ri
c has
h
i
n
g
[3]
i
s
a
m
e
t
hod
f
o
r
fi
ndi
ng
t
w
o-
di
m
e
nsi
o
nal
ob
jec
t
s.
W
e
use
t
h
i
s
m
e
t
hod t
o
e
x
t
r
act
coo
r
di
nat
e
s
(x,
y)
v
a
lu
es of
th
e str
ong
f
eatu
r
es.
Sect
i
on
2 c
o
nt
ai
ns a
di
scu
ssi
on
o
f
pre
v
i
o
us
w
o
r
k
rel
a
t
e
d
t
o
sec
u
re a
u
t
h
ent
i
cat
i
on sy
st
em
s usi
n
g
bi
om
et
ri
cs. Sect
i
on
3 des
c
ri
bes t
h
e
p
r
oce
ss o
f
sel
ect
i
n
g st
r
o
n
g
feat
ures
fr
om
fi
nger
p
ri
nt
i
m
ag
es an
d
g
e
n
e
r
a
ting
cancelab
le f
i
ng
erpr
in
t tem
p
lates.
I
n
section
4
,
w
e
th
en
pr
opose o
u
r
sch
e
m
e
s in
wh
ich
th
e h
a
sh
values
of
features a
r
e
used as
ve
rification inform
at
ion
by t
h
e a
u
the
n
tication system
. Be
cause
our
propose
d
schem
e
s are light-weight and efficient
,
they suit lo
w p
e
rf
orm
a
nce
m
obi
l
e
de
vi
ces.
In sect
i
on
5
,
we
d
e
m
o
n
s
tr
ate th
at o
u
r
pr
oposed
sch
e
m
e
s
b
a
sed
o
n
o
n
e
-ti
m
e
p
a
sswo
r
d (
O
TP)
ar
e secu
r
e
, an
d
th
e
u
s
er’
s
bi
om
et
ri
c dat
a
of
fi
n
g
er
p
r
i
n
t
s
are al
so sa
fe, e
v
en
w
h
e
n
t
h
e m
obi
l
e
devi
ce i
s
l
o
s
t
or st
ol
e
n
.
F
i
nal
l
y
,
concl
u
si
o
n
s a
r
e di
sc
usse
d i
n
sect
i
on
6.
2.
RELATED WORK
Si
nce t
h
e i
n
t
r
o
duct
i
o
n o
f
o
u
t
s
t
a
ndi
ng c
once
p
t
s
suc
h
as fu
zzy
co
m
m
i
tm
ent
[4]
an
d fuz
z
y
vaul
t
[5]
schem
e
s, w
h
i
c
h l
o
cks
bi
om
et
ri
c dat
a
f
o
r
safet
y
, t
h
ere
ha
ve
bee
n
st
udi
es
o
n
h
o
w
t
o
p
r
ot
ect
bi
om
et
ri
c
t
e
m
p
l
a
t
e
s by
usi
ng r
o
bu
st
ha
sh f
unct
i
ons
[
6
, 7]
. T
o
ge
ne
rat
e
cancel
abl
e
fi
nge
r
p
ri
nt
t
e
m
p
l
a
t
e
s R
a
t
h
a et
al.
pr
o
pose
d
a har
d
-t
o-i
n
ve
rt
t
r
ansf
orm
a
t
i
on [8
, 9]
. B
y
caref
ully se
lectin
g
strong
feat
ure
s
that are easier for a
specific
user to re
plicate, Ra
ndom
i
zed B
i
om
et
ri
c Tem
p
l
a
t
e
s (R
B
T
s)
[
2
]
ha
ve al
s
o
be
e
n
pr
op
ose
d
,
cr
eat
i
ng a
di
ffi
c
u
l
t
envi
r
o
nm
ent
for at
t
ackers t
o
m
a
ke g
u
esses. B
i
om
et
ri
c key
ext
r
act
i
on i
s
a m
e
t
hod
t
o
get
fi
xe
d bi
nary
fr
om
bi
om
et
ri
c t
e
m
p
l
a
t
e
s. M
a
ny
st
u
d
i
e
s
have
al
so s
u
g
g
est
e
d m
e
t
h
o
d
s
o
f
m
a
ki
ng c
r
y
p
t
o
gra
p
hi
c k
e
y
s
fr
o
m
vari
ous
bi
om
etri
cs.
Iri
s
[1
0,
1
1
]
,
face
[
1
2]
an
d
fi
n
g
er
vei
n
[
13]
a
r
e s
o
m
e
exam
pl
es.
Fi
gu
re
1.
Sel
e
c
t
i
ng st
r
o
ng
feat
ures
an
d
ge
om
et
ri
c has
h
i
n
g
o
f
t
h
e
feat
u
r
es
Algo
rith
m
1
.
Sp
ecificatio
n of
th
e
(w-k)
S
elect
alg
o
rith
m
Inp
u
t:
fingerprint samples
(
β
1
,…
,
β
w
),
w,
k
Output:
the strong features
F
1:
k
i
= 0 (
i=1,…,
N)
//
N
is
the
nu
m
b
er of a
ll f
e
a
t
u
r
esof origin
al
fin
g
erprint
β
2:
for
j
←
1
to
w
do
{
3:
(
f
1
,.
..
,f
N
)
←
Find(
β
j
) //Find fingerprint features
4:
for
i
←
1
to
N
do
{
5:
if
f
i
ex
ists
then
k
i
++
6:
}
7:
}
8:
F
←
f
i
ha
s
k
i
s
u
ch th
at
k
i
≥
k
9:
re
tur
n
F
3.
GENERATION OF TEMPLATE
3.
1. Str
o
n
g
Fe
atu
res of
Fi
ng
erpri
nt
It is
widely known that
bi
omet
ric data ca
nnot
be
precis
e
ly reproduced each tim
e
it is m
easured.
Whe
n
a
bi
om
et
ri
c sy
st
em
use
s
t
h
ese
dat
a
,
i
t
has t
o
s
o
l
v
e the prob
lem
o
f
erro
r to
leran
ce.
Howev
e
r,
d
u
e
to
th
e
devel
opm
ent
of opt
i
cal
eq
ui
p
m
ent
,
t
h
e bi
om
et
ri
c sy
st
em
s
a
r
e now able to repeatedly
ob
tain
al
m
o
st th
e
sam
e
b
i
o
m
etric i
m
a
g
es. Stron
g
featu
r
es are
h
i
gh
-qu
a
lity feat
u
r
es wh
ich
can b
e
d
i
stin
gu
ish
e
d
easily from o
t
h
e
r
feat
ure
s
i
n
bi
o
m
et
ri
c raw i
m
ages.
Usi
ng a
n
err
o
r c
o
rrect
i
o
n m
e
t
hod l
i
k
e
qua
nt
i
zat
i
on
[
2
]
,
we ca
n ar
ri
v
e
at
t
h
e
sam
e
and re
pe
atable val
u
es
from
the strong
feature
s
.
An exam
ple of s
e
lecting fi
ve s
t
rong features
in a
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
5, No
. 3,
J
u
ne 2
0
1
5
:
57
9 – 5
8
5
58
1
fing
erp
r
i
n
t i
m
a
g
e is sh
own
in Fig
u
r
e
1
.
Stron
g
feat
u
r
es
o
f
certain
p
r
ob
ab
i
lity
are resu
lts o
f
th
e pro
p
o
s
ed
(w-
k)
Select
a
l
g
o
r
i
t
h
m
wh
ich
is sh
own
in
Algo
rith
m
1
.
If so
me stro
ng
feat
u
r
es
of 10
0
p
e
rcen
t
p
r
o
b
a
b
ility are
fo
u
n
d
,
t
h
ey
ca
n be
use
d
as cr
y
p
t
o
g
r
a
phi
c
ke
y
s
. Here
,
w
is th
e nu
m
b
er o
f
in
pu
t fing
erprin
t i
m
ag
es and
k
is th
e
m
i
nim
u
m
num
ber
o
f
t
r
i
a
l
s
t
o
fi
n
d
t
h
e
feat
ures
.
We say
F
is th
e
set of
(w-
k
)
st
rong
feat
ure
s
. Wh
e
n
ϕ
=
(k/w
)
*100
, we can
say
F
is the set of
(
ϕ
%)
strong feat
ures
.
3
.
2
.
G
e
n
e
ra
t
i
ng
F
i
ng
erp
rint
T
emplate
with Str
o
ng
Features
Aft
e
r e
n
r
o
l
l
m
ent
pre
-
pr
oces
si
ng
pha
se w
h
i
c
h c
onsi
s
t
s
of
one
-
p
i
x
el
-
w
i
d
t
h
t
h
i
n
ni
n
g
st
age a
n
d
p
o
s
ition
i
ng
stag
e wit
h
an
enro
lled
fi
n
g
e
rp
ri
n
t
im
ag
e, we
can
select features, called
as
min
u
tiae, as sho
w
n
in
Fi
gu
re
1 an
d e
x
t
r
act
val
i
d
m
i
nut
i
ae val
u
es.
A m
i
nut
i
a
can be sp
eci
fi
ed
by
i
t
s
coor
di
nat
e
s
(x,
y
)
, angl
e
(
θ
)
, and
its typ
e
(t)
whi
c
h i
s
endi
ng
or
bi
fu
rcat
i
o
n, su
ch as
m
i
=
(x
i
, y
i
,
θ
i
, t
i
)
[14]
. T
o
ap
pl
y
err
o
r c
o
r
r
ect
i
on t
o
m
i
nut
i
a
e
m
i
, eac
h m
i
nut
i
a
sh
oul
d
be
q
u
a
nt
i
zed t
o
t
h
e
r
a
nge
s
of
val
u
e
s
, w
h
e
r
e
x
i
and
y
i
m
a
y h
a
v
e
16
rang
es
(
0
-1
5),
θ
i
32
ran
g
es (
0
-
3
1),
and
t
i
4 ran
g
e
s
(0-
3
), res
p
ec
t
i
v
el
y
.
An exa
m
pl
e of geom
et
ri
c hashi
n
g
wi
t
h
fi
ve m
i
n
u
t
i
ae i
s
sho
w
n i
n
Fi
g
u
r
e 1.
W
i
t
h
t
h
i
s
m
e
t
hod,
we can ext
r
act
m
i
nut
i
ae i
n
f
o
rm
at
i
on fr
om
st
rong
feat
ure
s
. T
a
bl
e 1
sh
ow
s h
a
sh
tab
l
e of
v
ector
41
PP
whi
c
h
has
x, y
co
or
di
nat
e
s
o
f
Fi
g
u
r
e
1’s
fi
ve m
i
nut
i
ae w
i
t
h
basi
s
(4
,1
).
If
ot
he
r has
h
t
a
bl
es fo
r t
h
e st
ro
ng
feat
u
r
es ar
e
neede
d
,
has
h
t
a
bl
e of
vect
o
r
42
PP
or
vect
o
r
43
PP
can be m
a
de,
respectively.
The ge
om
et
ri
c i
n
fo
rm
at
i
on o
f
fi
ve m
i
nut
i
ae i
s
sho
w
n i
n
F
i
gu
re 2,
w
h
i
c
h
are di
st
ances
(d
1
,d
2
,d
3
,d
4
)
and angles
(a
1
,a
2
,a
3
,a
4
)
bet
w
een st
r
o
ng
fea
t
ures
(F
1
,F
2
,F
3
,F
4
,F
5
)
. Our fi
n
g
e
rp
rin
t
tem
p
late is co
m
p
o
s
ed
of
geom
et
ri
c i
n
form
at
i
on an
d
m
i
nut
i
ae val
u
es o
f
st
r
o
ng features
. In the verificati
on
phase,
ge
om
etric
i
n
f
o
rm
at
i
on i
s
pr
ovi
de
d f
o
r
t
h
e al
i
gnm
ent
of i
n
p
u
t
fi
n
g
e
r
pri
n
t
i
n
searc
h
i
ng f
o
r st
r
o
n
g
feat
ure
s
. Let
n
be a
num
ber
of
st
r
o
ng
feat
ures
.
Ge
om
et
ri
c i
n
form
at
i
on
G
i
s
as
s
h
ow
n
bel
o
w
G =
(
d
1
,…,
d
n-1
, a
1
,…,
a
n-1
).
Min
u
tiae v
a
lues are actu
a
l au
th
en
ticatio
n
data th
at is
u
s
ed
in
u
s
er v
e
ri
ficatio
n
.
Orig
inal
min
u
tiae
val
u
es
M
i
s
as
sho
w
n
bel
o
w
M = (
m
1
,…,m
n
).
Fin
a
lly, ou
r fing
erp
r
i
n
t tem
p
la
te is as fo
llows
FingTemp
=
(G
|| M)
.(
||
:
bi
nary
c
o
ncat
ena
t
i
on)
In
ord
e
r to
check
security fu
n
c
tion
s
, we will ran
d
o
m
iz
e
m
i
n
u
tiae v
a
l
u
es and
in
sert
th
em in
to
th
e
te
m
p
late. Ran
d
o
m
iz
in
g
m
i
n
u
tiae in
fo
rm
atio
n
will b
e
ex
p
l
ai
ned
in section
4
.
Tabl
e
1.
Has
h
t
a
bl
e o
f
st
ro
n
g
f
eat
ures
wi
t
h
ba
si
s (4
,1
)
x y
basis
point
0.
5
0
(
4
,
1
)
1
1.
4
-
0
.
25
(
4
,
1
)
2
0.
4
-0
.5
0.
4
0.
8
0
-1
.3
(4
,1
)
(4
,1
)
(4
,1
)
3
4
5
Fi
gu
re
2.
Di
st
a
n
ces a
n
d a
ngl
e
s
bet
w
een
st
r
o
ng
feat
ures
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Fi
nger
p
ri
nt
Au
t
h
ent
i
c
at
i
o
n Sc
heme
sf
or Mo
bi
l
e
Devi
ces
(
J
i
n
ho
H
a
n)
58
2
4.
OU
R CO
NST
R
U
C
TIO
N
We
have
p
r
o
p
o
se
d o
n
e
basi
c schem
e
and
t
w
o
ext
e
nde
d
schem
e
s. In
schem
e
1, we
sim
p
l
y
use
FingTemp
an
d
rand
om
num
ber
r
. I
n
sche
m
e
2, we
use Fi
ngTe
m
p
,
r
,
and use
r’s
pa
ssw
or
d
π
to increase
security and in schem
e
3, we co
nsider t
h
e case of losi
ng
one of
n
st
rong features.
We as
sum
e
that with error
co
rrectio
n
su
ch
as sim
p
le q
u
a
n
tization
,
we can
ob
tain
t
h
e
(1
00
%)
str
o
ng feat
ure
s
of the sam
e
user in our
schem
e
s.
4.
1. Scheme
1
As e
xpl
ai
ne
d a
b
o
v
e,
l
e
t
FingT
e
mp = (G || M)
b
e
th
e orig
i
n
al fing
erp
r
i
n
t tem
p
la
te g
i
v
e
n fro
m
th
e raw
im
age.
Enrollment:
Gi
ve
n a
sec
u
ri
t
y
param
e
t
e
r
k
, let
p
be
a prime
num
b
er.
T
h
e enrollm
ent
algorithm
ch
oo
ses a
r
a
ndo
m
n
u
m
b
e
r
r
∈
Z
p
*
an
d
has
h
f
u
nct
i
o
n
s
H
1
: {0,1}
*
→
Z
p
*
, a
n
d
H
2
:
Z
p
*
→
{0
,1}
k
.
1)
C
o
m
put
e
M
of
FingTemp
,
M
h1
=
H
1
(M)
.
2)
C
o
m
put
e
M’
=
M
h1
⊕
r
.(
⊕
: bitwise exclusi
v
e-or,
XOR)
3)
C
o
m
put
e
r
h2
=
H
2
(r)
.
Later,
r
h2
will b
e
u
s
ed as
verificatio
n inform
at
io
n
.
Rando
m
i
zed
m
i
n
u
tiae
M’
is tak
e
n to
m
a
k
e
a
cancelable
fingerprint tem
p
late. Our ca
ncelab
le fi
n
g
e
rprin
t
te
m
p
late is as fo
llo
ws
C
anFi
n
g
T
em
p =
(
G
||
M
’
)
.
If a u
s
er
’s
ID
is n
ecessary for au
th
en
ticatio
n in
th
e d
e
v
i
ce,
ID
,
C
a
nFi
n
gT
emp
and
r
h2
are store
d
in a
dat
a
base
as
fol
l
ows
DB rec
o
rd: {
ID
,
(G ||
M’
)
,
r
h2
}.
Verific
a
ti
on
:
W
i
t
h
th
e u
s
e of th
e inp
u
t
b
i
o
m
etric tem
p
la
te wh
ich
is ob
tain
ed b
y
a fi
n
g
e
rprin
t
scan
ner,
s
o
m
e
one
’s m
i
nut
i
ae val
u
e
s
β
is
pro
v
i
d
e
d to
t
h
e
v
e
rification
al
g
o
rith
m
.
Geometric in
form
a
tio
n
G
is
use
d
t
o
fi
n
d
e
x
act
po
si
t
i
on
o
f
t
h
e m
i
nut
i
ae.
1)
C
o
m
put
e
β
h1
=
H
1
(
β
)
.
2)
C
o
m
put
e
β
’
=
β
h1
⊕
M’
.
3)
C
o
m
put
e
(
β
’)
h2
=
H
2
(
β
’ )
.
4)
If
(
β
’)
h2
=
r
h2
then output “
Y
es” which indicates that
th
e
u
s
er is au
th
en
tic, or “No
”
which
means that the
user is
not.
4.
2. Scheme
2
Let
π
be t
h
e u
s
er’s
pass
wo
r
d
. Fo
r a m
o
re secure
veri
fi
cat
i
on
pr
ocess
we
use t
h
e us
er
’s
fi
nge
r
p
ri
nt
te
m
p
late an
d
passwo
r
d
,
t
h
at i
s
to
say a two
-facto
r
au
th
en
ticatio
n
.
Enrollment:
Gi
ven a secu
r
i
t
y
param
e
t
e
r
k
, let
p
be a pri
m
e num
ber. The en
rol
l
m
ent
al
gori
t
h
m
ch
oo
ses a
r
a
ndo
m
n
u
m
b
e
r
r
∈
Z
p
*
an
d
has
h
f
u
nct
i
o
n
s
H
1
: {0,1}
*
→
Z
p
*
, a
n
d
H
2
:
Z
p
*
→
{0
,1}
k
.
1)
C
o
m
put
e
M
of
FingTemp
,
M
h1
=
H
1
(M)
2)
C
o
m
put
e
π
h1
=
H
1
(
π
)
.
3)
C
o
m
put
e
M’
=
M
h1
⊕
r
.
4)
C
o
m
put
e
M’’
=
M’
⊕
π
h1
5)
C
o
m
put
e
r
h2
=
H
2
(r)
.
Later,
r
h2
w
ill b
e
u
s
ed as
v
e
rificatio
n
informatio
n
.
Randomized
m
i
n
u
tiae
M’’
is tak
e
n to
m
a
k
e
a
can
celab
le
fingerprin
t
tem
p
lat
e
. Now,
ou
r fing
erp
r
i
n
t tem
p
la
te is as fo
llows
C
anFi
n
g
T
em
p =
(
G
||
M
’’)
.
If t
h
e
user
’s
ID
is
n
ecessary
for au
th
en
ticatio
n in
t
h
e
d
e
v
i
ce,
ID
,
Can
F
ingTemp
an
d
r
h2
ar
e
s
t
or
ed
in
a d
a
tab
a
se as
fo
llo
ws
DB rec
o
rd: {
I
D
, (
G
||
M’
’)
, r
h2
}.
Verific
a
ti
on
:
W
i
t
h
th
e u
s
e of th
e inp
u
t
b
i
o
m
etric tem
p
la
te wh
ich
is ob
tain
ed b
y
a fi
n
g
e
rprin
t
scan
ner,
som
e
one
’s m
i
nut
i
ae val
u
es
β
a
n
d
al
so hi
s
pas
s
w
o
r
d
are provided t
o
the
ve
ri
fi
cat
i
on al
go
r
i
t
h
m
.
Geom
etric inform
ation
G
is
u
s
ed
to fi
n
d
th
e ex
act po
sition
of th
e m
i
n
u
tiae.
1)
C
o
m
put
e
β
h1
= H
1
(
β
)
.
2)
C
o
m
put
e
h1
= H
1
(
)
.
3)
C
o
m
put
e
β
’ =
β
h1
⊕
M’’
.
4)
C
o
m
put
e
β
’’
=
β
’
⊕
h1
.
5)
C
o
m
put
e
(
β
’’)
h2
=
H
2
(
β
’’)
.
6)
If
(
β
’’)
h2
=
r
h2
then
out
put “Yes” whic
h indicates that
th
e u
s
er is au
t
h
en
tic, o
r
“No
”
which
means that the
user is
not.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
5, No
. 3,
J
u
ne 2
0
1
5
:
57
9 – 5
8
5
58
3
4.
3. Scheme
3
As e
xpl
ai
ne
d a
b
o
v
e,
l
e
t
FingT
e
mp = (G || M)
b
e
th
e orig
i
n
al fing
erp
r
i
n
t tem
p
la
te g
i
v
e
n fro
m
th
e raw
im
age. He
re
we conside
r
the
case of losi
ng
one
o
f
n
st
r
o
ng
feat
u
r
es. We m
a
ke
FingTe
mp
(-
1
)
, … ,
Fi
ngTemp
(-
n
)
suc
h
that
Fi
ngTemp
(-
i
)
=
(G ||
M
(-
i
)
)
and
M
(-
i
)
= M
-
m
i
.
Enrollment:
Gi
ven a secu
r
i
t
y
param
e
t
e
r
k
, let
p
be a pri
m
e num
ber. The en
rol
l
m
ent
al
gori
t
h
m
ch
oo
ses r
a
nd
om
n
u
m
b
er
s
r
∈
Z
p
*
and
(r
(-
1
)
, … r
(-
n
)
)
∈
Z
p
*
, t
w
o
ha
sh
f
u
n
c
t
i
ons
H
1
: {0,1
}
*
→
Z
p
*
, a
n
d
H
2
:
Z
p
*
→
{0,1}
k
.
1)
C
o
m
put
e
M
of
FingTemp
,
M
h1
=H
1
(M)
.
C
o
m
put
e
M
(
-
1)
h1
=H
1
(M
(-
1
))
.
…
C
o
m
put
e
M
(
-
n)
h1
=H
1
(M
(-
n
)
)
.
2)
C
o
m
put
e
M’ =
M
h1
⊕
r
.
C
o
m
put
e
M’
(-
1
)
=
M
(-
1
)
h
1
⊕
r
(-
1
)
.
…
C
o
m
put
e
M’
(-
n
)
=
M
(-
n
)
h
1
⊕
r
(-
n
)
.
3)
C
o
m
put
e
r
h2
=
H
2
(r)
.
C
o
m
put
e
r
(-
1
)
h
2
=
H
2
(r
(-
1
)
)
.
…
C
o
m
put
e
r
(-
n
)
h
2
=
H
2
(r
(-
n
)
)
.
Later
, r
h2
, r
(
-
1)
h2
, …
,
r
(-
n
)
h
2
will b
e
u
s
ed
as
v
e
rification
i
n
fo
rm
atio
n
.
Rand
o
m
ized
min
u
tiae
M’,M
’
(-
1)
,…,
M’
(-
n
)
are use
d
in order t
o
m
a
ke cancel
able fi
nge
rprint
te
m
p
lates. Ou
r can
celab
le fin
g
e
rprin
t
temp
lates
are as
follows
C
anFi
n
g
T
em
p =
(
G
||
M
’
)
.
C
anFi
n
g
T
em
p
(-
1
)
= (G || M
’
(-
1
)
)
.
…
C
anFi
n
g
T
em
p
(-
n
)
= (G || M
’
(-
n
)
)
.
If a
u
s
er
’s
ID
is neces
sary
for authe
n
tication in the
de
vice,
ID,
r
h2
, r
(
-
1)
h2
,…
,
r
(-
n
)
h
2
, and
our cancela
b
le
fing
erp
r
i
n
t temp
lates are
sto
r
ed
in a
d
a
tab
a
se as fo
llows
DB rec
o
rd: {
ID
,
(G ||
M’
)
,
r
h2
}.
{
ID,
(
G
|
|
M
’
(-
1
)
),
r
(-
1
)
h
2
}.
…
{
ID,
(
G
|
|
M
’
(-
n
)
),
r
(-
n
)
h
2
}.
Verific
a
ti
on
:
W
i
t
h
th
e u
s
e of th
e inp
u
t
b
i
o
m
etric tem
p
la
te wh
ich
is ob
tain
ed b
y
a fi
n
g
e
rprin
t
scan
ner,
s
o
m
e
one
’s m
i
nut
i
ae val
u
e
s
β
is
pro
v
i
d
e
d to
t
h
e
v
e
rification
al
g
o
rith
m
.
Geometric in
form
a
tio
n
G
is
use
d
t
o
fi
n
d
t
h
e exact
po
si
t
i
on
of
the m
i
nutiae.
W
e
ass
u
m
e
that
β
has
n
or
n-
1
m
i
nutiae values.
1)
C
o
m
put
e
β
h1
= H
1
(
β
)
.
2)
C
o
m
put
e
β
’ =
β
h1
⊕
M’
.
C
o
m
put
e
β
’
(-
1
)
=
β
h1
⊕
M’
(-
1
)
.
…
C
o
m
put
e
β
’
(-
n
)
=
β
h1
⊕
M’
(-
n
)
.
3)
C
o
m
put
e
(
β
’)
h2
=
H
2
(
β
’)
.
C
o
m
put
e
(
β
’
(-
1
)
)
h2
=
H
2
(
β
’
(-
1
)
)
…
C
o
m
put
e
(
β
’
(-
n
)
)
h2
= H
2
(
β
’
(-
n
)
)
4)
If
(
β
’)
h2
= r
h2
or
(
β
’
(-
1
)
)
h2
= r
(-
1
)
h
2
or …
or
(
β
’
(-
n
)
)
h2
= r
(
-
n)
h2
t
h
en
out
put
“
Y
es”
w
h
i
c
h
indicates that t
h
e
user is a
u
thentic, or
“
N
o”
whi
c
h m
eans t
h
at
t
h
e
use
r
i
s
not
.
4.4. Correctne
ss
In
schem
e
1 a
n
d
3, i
f
M =
β
th
en th
e
fo
llo
wi
n
g
equ
a
tion
is
co
rrect.
(
β
’)
h2
= H
2
(
β
’)
=
H
2
(
β
h1
⊕
M’)
=
H
2
(H
1
(
β
)
⊕
M’)
=
H
2
(H
1
(
β
)
⊕
H
1
(M
)
⊕
r)
=
H
2
(r)
=
r
h2
.
An
d i
n
sc
hem
e
3,
i
f
M
(-
i
)
=
β
th
en th
e
fo
llo
wi
n
g
equ
a
tion
is
co
rrect.
(
β
’
(-
i
)
)
h2
= H
2
(
β
’
(-
i
)
)
=
H
2
(
β
h1
⊕
M’
(-
i
)
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Fi
nger
p
ri
nt
Au
t
h
ent
i
c
at
i
o
n Sc
heme
sf
or Mo
bi
l
e
Devi
ces
(
J
i
n
ho
H
a
n)
58
4
=
H
2
(H
1
(
β
)
⊕
M’
(-
i
)
)
=
H
2
(H
1
(
β
)
⊕
H
1
(M
(-
i
)
)
⊕
r
(-
i
)
)
=
H
2
(r
(-
i
)
)
=
r
(-
i
)
h
2
.
Fin
a
lly, in
sch
e
me 2
,
if
M =
β
and
π
=
t
h
en
th
e fo
llowing
eq
u
a
tion
is correct.
(
β
’’)
h2
= H
2
(
β
’’
)
=
H
2
(
β
’
⊕
h1
)
=
H
2
(
β
h1
⊕
M’’
⊕
h1
)
=
H
2
(
β
h1
⊕
M’
⊕
π
h1
⊕
h1
)
=
H
2
(
β
h1
⊕
M’
⊕
H
1
(
π
)
⊕
H
1
(
))
=
H
2
(
β
h1
⊕
M’)
=
H
2
(
β
h1
⊕
M
h1
⊕
r)
=
H
2
(H
1
(
β
)
⊕
H
1
(M)
⊕
r)
=
H
2
(r)
=
r
h2
.
5.
SECU
RIT
Y
A
NAL
YSI
S
In
t
h
i
s
sect
i
o
n
we
per
f
o
r
m
securi
t
y
anal
y
s
i
s
of
o
u
r
fi
nge
rp
r
i
nt
t
e
m
p
l
a
t
e
an
d
ou
r
pr
o
p
o
s
ed
schem
e
s.
Security for
Fingerprint Template
:
I
n
t
h
e pre
v
i
o
us
sect
i
on w
e
ha
ve sh
o
w
n t
h
e
pr
ocess
of
selectin
g
stro
ng
features
for th
e tem
p
late fro
m
th
e o
r
ig
ina
l
finge
rp
rint im
ages as
an e
x
am
ple. The propose
d
cancelable fingerpri
n
t te
m
p
late,
C
anFi
n
g
T
emp
co
nsists o
f
two
co
m
p
on
en
ts, g
e
o
m
etr
i
c in
fo
rm
atio
n
G
and
random
i
zed minutiae
M’
i
n
schem
e
1, 3 (or
M’’
in sche
m
e
2). First, we discuss security for
M’
(o
r
M’’
).
Acco
r
d
i
n
g t
o
t
h
e er
ro
r-c
or
rec
t
i
on q
u
a
n
t
i
zat
ion
of
m
i
,
we assu
m
e
th
at o
r
ig
in
al on
e m
i
n
u
tia in
form
atio
n
m
a
y
have
t
h
e
si
ze
o
f
15
bi
t
s
s
u
c
h
as
m
i
= (
x
i
,y
i
,
θ
i
,t
i
)
, where at
least
x
i
and
y
i
m
a
y
hav
e
16
rel
a
t
i
v
e di
f
f
ere
n
t
val
u
es
(4
bits, 0
-
15
),
θ
i
3
2
relativ
e v
a
lu
es
(5
b
its, 0-31
),
and
t
i
4
val
u
es (
2
bi
t
s
, 0
-
3)
, res
p
ec
tively.
Whe
n
t
h
e number
of
st
ro
ng feat
ure
s
n
= 5,
M
is a 7
5
b
it lo
ng
v
a
lue an
d
in
th
e case o
f
n
= 10,
M
i
s
a 150 bi
t
l
o
n
g
val
u
e. Th
e
l
e
ngt
h
of
M
i
s
l
o
ng e
n
ou
g
h
, w
h
i
c
h m
a
kes i
t
har
d
er
f
o
r at
t
acke
r
s t
o
t
a
ke a gue
ss. R
a
nd
om
i
zed
m
i
nut
i
ae
M’
(o
r
M’’
) is
com
puted as follows.
M
is tran
sfo
r
m
e
d
with
th
e on
e-way h
a
sh
fun
c
tio
n
(
H
1
(·
)
)
an
d
r
a
ndomized
u
s
ing
r
a
n
d
o
m
num
ber
r
wi
t
h
X
O
R
ope
rat
i
o
n. I
n
sc
hem
e
2,
M
i
s
ra
n
dom
ized
o
n
ce m
o
re
usi
n
g
ha
sh
va
l
u
e
of
π
with
XOR
ope
rat
i
o
n. R
a
n
dom
num
ber
r
also
is tran
sfo
r
m
e
d
with the h
a
sh
fu
n
c
tion
(
H
2
(·
)
) an
d t
h
i
s
hash
val
u
e
r
h2
is
store
d
in t
h
e device. After t
h
at,
r
is d
e
leted fro
m
th
e d
e
v
i
ce, so
no
body can com
pute t
h
e origi
n
al m
i
nutiae
in
fo
rm
atio
n
M
.
Next
, not
e
t
h
at
G
is th
e in
formatio
n
o
f
relat
i
v
e
d
i
stan
ces (
d
i
) an
d a
ngl
es (
a
i
) bet
w
ee
n strong feat
ures
.
Whe
n
G
is c
o
m
p
romised by an attack, the “
o
ld
G
” ca
n be canceled a
nd a
new fi
nge
rpri
nt te
m
p
late with “ne
w
G
” can
be enro
lled
,
wh
ich
m
ean
s th
at
o
t
h
e
r strong
feat
u
r
es will b
e
select
ed
for th
e
n
e
w te
m
p
late. If
we find
20 st
r
o
ng feat
ures f
r
o
m
t
h
e user
’s fi
n
g
e
r
p
r
i
n
t
and sel
ect
10 fe
at
ure
s
fr
o
m
am
ong t
h
em
, we can
m
a
ke ne
w
t
e
m
p
l
a
t
e
s
m
o
re t
h
an 1
8
0
,
0
0
0
t
i
m
e
s (com
binat
i
on
of
20 el
em
ent
s
cho
o
se
10
). I
n
sh
o
r
t
,
ou
r
C
a
nFi
n
gT
emp
is
cancelable a
nd it keeps the
us
e
r’s
fin
g
e
r
p
r
int
data sa
fe.
Security
for
Propose
d
Sc
hemes:
I
n
o
u
r
schem
e
s, ran
d
o
m
num
ber
r
is tran
sfo
r
m
e
d
with
th
e
on
e-
w
a
y h
a
sh
functio
n
(
H
2
(·
)
)
and hash val
u
e
r
h2
(or
r
(-
i
)
h
2
)
i
s
st
ored i
n
a dat
a
base as
user’
s
veri
fi
c
a
t
i
on
i
n
f
o
rm
at
i
on. R
a
nd
om
num
ber
r
(or
r
(-
i
)
) is e
lim
inated from
the device a
n
d
r
(or
r
(-
i
)
) cannot
be
rest
o
r
ed
from
t
h
e has
h
val
u
e
r
h2
(or
r
(-
i
)
h
2
). Each
ti
m
e
u
s
er’s en
ro
llm
en
t is i
m
p
l
e
m
en
t
e
d
,
rand
o
m
r
(or
r
(-
i
)
) is a differe
nt
v
a
lu
e
wh
ich
is
th
en
u
s
ed
to
tran
sform
min
u
tiae in
form
atio
n
M
diffe
re
ntly
. I
n
sh
o
r
t, ra
n
d
o
m
r
(or
r
(-
i
)
) is use
d
as a o
n
e-t
i
m
e pass
wo
r
d
(
O
TP) i
n
o
u
r
sc
hem
e
s whi
c
h
add
s
a hi
gh l
e
vel
o
f
sec
u
ri
t
y
t
o
t
h
e p
r
o
pos
ed
aut
h
e
n
t
i
cat
i
on
schem
e
s.
W
e
c
a
n say
t
h
at
o
u
r
aut
h
e
n
t
i
cat
i
on
schem
e
s based
o
n
one
-t
i
m
e passw
or
ds a
r
e s
ecure
.
6.
CO
NCL
USI
O
N
In
t
h
is p
a
p
e
r,
we pro
p
o
s
e a can
celab
le
fin
g
e
rp
rin
t
tem
p
late wh
ich
u
s
es h
i
gh
q
u
a
lity an
d
easily
di
st
i
n
g
u
i
s
ha
bl
e
feat
u
r
es. The pr
o
pose
d
al
go
r
i
t
h
m
,
(w-k
)
S
elect
o
u
t
p
u
t
s stro
ng
features
of certain prob
ab
ility.
Because we
only use pa
rtial features
of
finge
rprints, e
v
en whe
n
pa
rtia
l inform
ation is lost, the user’s
fingerpri
n
t dat
a
is safe.
Our cancelable te
m
p
la
te is co
mp
o
s
ed
o
f
g
e
ometric in
fo
rm
atio
n
G
wh
ich is th
e
relative val
u
e
of dista
n
ces
and angles
be
tween
st
ro
n
g
feat
ure
s
a
nd
r
a
nd
om
i
zed
m
i
nut
i
ae
M’
of st
ro
ng
featu
r
es.
With
th
is can
celab
le fing
erp
r
i
n
t tem
p
la
te an
d
rand
o
m
n
u
m
b
e
r
r
(an
d
user
’s pas
s
wo
r
d
π
), it is
ev
id
en
t
t
h
at
o
u
r
sc
hem
e
i
s
sec
u
re
.
Ou
r sc
hem
e
s per
f
o
rm
onl
y
bi
t
w
i
s
e X
O
R
o
p
erat
i
ons
i
n
e
n
rol
l
m
ent
a
n
d
ve
ri
fi
c
a
t
i
on
pha
ses,
w
h
i
c
h
m
eans t
h
ey
are l
i
ght
-wei
ght
schem
e
s whi
c
h are
wel
l
s
u
i
t
e
d f
o
r l
o
w
per
f
o
r
m
a
nce m
obi
l
e
devi
ces
.
We b
e
l
i
e
ve t
h
at
o
u
r
schem
e
s are
usef
ul
f
o
r b
o
t
h
t
h
e
p
r
ot
ect
i
o
n
o
f
fi
ng
er
pr
i
n
t
dat
a
a
n
d
h
u
m
a
n
aut
h
e
n
t
i
cat
i
on
on
m
obi
l
e
de
vi
ces.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
5, No
. 3,
J
u
ne 2
0
1
5
:
57
9 – 5
8
5
58
5
REFERE
NC
ES
[1]
Apple Inc. Apple Onlin
e Stor
e, “iPhone6 Tou
c
h I
D
,
at http://www.apple.com
/ipho
ne-6/tou
c
h-id/,” 2015.
[2]
L. Ballard
,
S. Kamara, F. Monrose,
and M. Reiter, “Towards Practical Bi
ometr
i
c K
e
y
Gen
e
ration w
ith Randomized
Bi
ome
t
ri
c T
e
mpl
a
te
s”
,
CCS’08
,
October 27-31
, 2
008, Alex
a
ndria, Virginia, USA,
2008.
[3]
H. Wolfson and I. Rigoutsos, “Geo
metric H
a
shin
g: An Overview
”,
IEEE Computational S
c
ien
ce and
Engineering
,
4(4), pp
. 10-21
,
1997.
[4]
A. J
u
els
and
M
.
W
a
ttenb
erg,
“
A
fuzz
y co
m
m
itm
ent s
c
hem
e”, in
Proc. ACM Conf. o
n
Computer a
n
d
Communications Security
, pp
. 28
-36, 1999
.
[5]
A. Juels
and M.
Sudan, “A
fuzzy vault sch
e
me”, in
IEE
E
In
tl.
Sym
p
. On In
formatio
n Theory
, 2002.
[6]
T. Connie, A. Teoh, M. Goh, an
d
D. Ngo, “Palmhashing: a novel appro
ach for
cancelable
b
i
ometr
i
cs”,
In
formatio
n
Processing Letters
, Vol 93, pp. 6
14-634, 2005
.
[7]
Y. Sutcu, T. Sen
car, and
N. Memon,
“A secure biometric
au
then
tication scheme b
a
sed on robust h
a
shing”, In
AC
M
MMSEC Worksh
op
, 2005
.
[8]
N. Ratha, J. Con
n
ell, and R
.
Bolle,
and
S
.
Chikke
rur, “
C
ance
labl
e
biom
etrics
: A
c
a
se stud
y in fing
erprints”
,
In
Intl.
Conf. on
Pa
ttern
Recognition
, pp
. 370-373
, 2006
.
[9]
N. Ratha, S. Chikkerur, J. C
onnell, and R
.
Bolle,
“Generating ca
n
celable f
i
ngerprint templates”,
IEEE Transactions
on Pattern Ana
l
ysis and Ma
chin
e Intelligen
ce
, V
o
l 29. No.4
, pp
.
561-172, 2007
.
[10]
Y.
J.
Lee, H.
G.
Lee, K.
R. Park,
and J.
Kim,
“Inva
riant b
i
om
etri
c k
e
y ext
r
ac
tion b
a
sed on iris
code”,
Proc. o
f
I
EEK
Fall Conf
., Seou
l, Korea,
vol. 28
, no. 2, 2005.
[11]
F. Hao, R. And
e
rson, and J.
D
a
ugm
an, “
C
om
bining cr
ypto wi
t
h
biom
etrics
eff
ect
ivel
y”
,
I
EEE Transactions on
Computers
, vol.
55, no
. 9
,
pp
. 10
81-1088, 2006
.
[12]
A. Gohand D.C
.
L. Ngo,
“
C
om
putation
of
cr
ypt
ographic
ke
ys
fr
om
face b
i
om
etr
i
cs
”,
Int
e
rnation
a
l F
e
deration
fo
r
Information Pro
cessing 2003, Sp
ringer-Verlag, LNCS 2828
, pp
. 1
–13, 2003
.
[13]
A. Venckauskas
and P. Nane
vicius, “Cr
y
ptogr
ap
hic Key
Ge
ner
a
tion from Finger Vein”,
International Journal of
Engineering Sciences and
Resea
r
ch Technolog
y
.
2(4), pp
. 733–73
8, 2013
.
[14]
D. Moon,
et
al
.,
“Fingerprint Template Pr
otectio
n using Fuzzy
V
a
ult”,
LNCS 470
7
, pp
. 1141-115
1, 2007
.
BI
O
G
R
A
P
HY
OF
A
U
T
HO
R
Jinho Han
was born in Seoul,
Korea on April
5th, 1965; He
r
eceived th
e B.A
.
degree in
the
Department of Forestr
y
from
Korea Univers
i
t
y
a
nd M
.
E. de
gree
in the Department of Networ
k
Management at Dongguk
Univer
sity
,
Seou
l, Korea,
in 1990
and
2
006,
resp
ectively
.
He receiv
ed
the Ph.D. d
e
gree at th
e Graduate
School of Infor
m
ation Security
from Korea University
in 2013
.
He is currently
anassistant prof
essor in the depa
rtment of Liber
a
l studies (computer) at Korean
Bible Univers
i
t
y
, S
e
oul, Kore
a.
His
res
earch ar
e
a
s
includ
e broad
cas
t en
cr
yption
,
attribu
t
e-b
a
s
e
d
encr
ypt
i
on,
and
biom
etrics
.
Evaluation Warning : The document was created with Spire.PDF for Python.