I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
E
lec
t
r
ical
an
d
Com
p
u
t
e
r
E
n
gin
e
e
r
in
g
(
I
JE
CE
)
Vol.
15
,
No.
1,
F
e
br
ua
r
y
20
25
,
pp.
689
~
699
I
S
S
N:
2088
-
8708,
DO
I
:
10
.
11591/i
jec
e
.
v
15
i1
.
pp
6
89
-
699
689
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
e
c
e
.
iaes
c
or
e
.
c
om
Qu
ad
r
a
t
ic
m
u
ltivar
ia
t
e
l
in
e
a
r
r
e
gr
e
ssi
ve
d
is
t
r
ib
u
t
e
d
p
r
oxi
m
ity
f
e
at
u
r
e
e
n
gi
n
e
e
r
i
n
g f
o
r
c
y
b
e
r
c
r
im
e
d
e
t
e
c
t
io
n
i
n
d
ig
ital f
u
n
d
t
r
an
sa
c
t
io
n
s wi
t
h
b
ig
d
at
a
Ar
u
l
Je
yan
t
h
i
P
au
lra
j
1
,
B
alaj
i
T
h
alai
m
a
lai
1
,
2
1
D
e
pa
r
tm
e
nt
of
C
omput
e
r
A
ppl
ic
a
ti
ons
a
nd S
c
hool
of
M
a
n
a
ge
me
nt
, M
a
dur
a
i
K
a
ma
r
a
j
U
ni
ve
r
s
it
y, M
a
dur
a
i,
I
ndi
a
2
P
os
t
G
r
a
dua
te
D
e
pa
r
tm
e
nt
of
C
omput
e
r
S
c
ie
nc
e
, G
ove
r
nme
nt
A
r
ts
C
ol
le
ge
, M
e
lu
r
,
I
ndi
a
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
M
a
y
7,
202
4
R
e
vis
e
d
Aug
9,
2024
Ac
c
e
pted
Aug
20,
2024
D
i
g
i
t
al
fu
n
d
t
ra
n
s
ac
t
i
o
n
s
i
n
v
o
l
v
e
t
h
e
el
ect
r
o
n
i
c
t
ran
s
f
er
o
f
fu
n
d
s
b
et
w
een
p
art
i
es
t
h
r
o
u
g
h
d
i
g
i
t
a
l
ch
a
n
n
e
l
s
s
u
c
h
as
o
n
l
i
n
e
b
a
n
k
i
n
g
p
l
a
t
fo
rm
s
,
mo
b
i
l
e
ap
p
l
i
ca
t
i
o
n
s
,
an
d
e
l
ec
t
ro
n
i
c
p
ay
me
n
t
s
y
s
t
em
s
.
H
o
w
e
v
er,
t
h
e
ra
p
i
d
ad
v
a
n
cemen
t
o
f
d
i
g
i
t
a
l
t
ran
s
act
i
o
n
s
h
a
s
al
s
o
d
i
rect
e
d
cy
b
ercri
m
i
n
a
l
s
t
o
ex
p
l
o
i
t
v
u
l
n
erab
i
l
i
t
i
es
,
en
g
a
g
i
n
g
i
n
mo
n
ey
l
a
u
n
d
eri
n
g
an
d
o
t
h
er
i
l
l
e
g
al
act
i
v
i
t
i
e
s
,
res
u
l
t
i
n
g
i
n
s
u
b
s
t
a
n
t
i
al
f
i
n
a
n
ci
a
l
l
o
s
s
e
s
.
T
h
e
i
mp
r
o
v
e
accu
rac
y
o
f
cy
b
ercr
i
mi
n
al
d
e
t
ect
i
o
n
b
y
l
es
s
er
t
i
me
co
n
s
u
m
p
t
i
o
n
,
a
n
o
v
e
l
t
ech
n
i
q
u
e
ca
l
l
e
d
q
u
a
d
rat
i
c
mu
l
t
i
v
ari
a
t
e
l
i
n
ear
reg
res
s
i
v
e
d
i
s
t
r
i
b
u
t
e
d
p
ro
x
i
m
i
t
y
feat
u
re
en
g
i
n
eer
i
n
g
(Q
ML
RD
PFE
)
i
s
d
ev
e
l
o
p
ed
.
T
h
e
p
ro
p
o
s
ed
Q
ML
RD
PF
E
t
ech
n
i
q
u
e
co
m
p
ri
s
es
t
w
o
p
ri
mar
y
s
t
e
p
s
n
amel
y
d
at
a
p
rep
ro
ce
s
s
i
n
g
an
d
feat
u
re
e
n
g
i
n
eer
i
n
g
.
A
n
al
y
zed
re
s
u
l
t
s
p
r
o
v
e
t
h
a
t
t
h
e
Q
ML
RD
PFE
t
ec
h
n
i
q
u
e
o
u
t
p
erf
o
rms
e
x
i
s
t
i
n
g
me
t
h
o
d
s
i
n
a
t
t
a
i
n
i
n
g
s
u
p
er
i
o
r
acc
u
racy
a
n
d
p
rec
i
s
i
o
n
.
Fu
rt
h
ermo
re,
Q
ML
RD
PFE
me
t
h
o
d
s
h
o
w
s
effec
t
i
v
e
i
n
re
d
u
c
i
n
g
t
i
me
u
t
i
l
i
zat
i
o
n
an
d
s
p
ace
co
m
p
l
e
x
i
t
y
fo
r
fra
u
d
u
l
e
n
t
t
r
an
s
ac
t
i
o
n
d
e
t
ect
i
o
n
co
mp
are
d
t
o
ex
i
s
t
i
n
g
ap
p
ro
ac
h
es
.
R
es
u
l
t
s
t
o
p
r
o
v
i
d
e
e
ffect
i
v
e
i
n
red
u
ci
n
g
t
i
me
u
t
i
l
i
z
at
i
o
n
an
d
s
p
ace
c
o
mp
l
ex
i
t
y
fo
r
fra
u
d
u
l
e
n
t
t
ran
s
ac
t
i
o
n
d
et
ec
t
i
o
n
t
h
a
n
t
h
e
c
o
n
v
en
t
i
o
n
a
l
met
h
o
d
s
.
K
e
y
w
o
r
d
s
:
Ada
pti
ve
Z
iggur
a
t
s
ynthetic
s
a
mpl
ing
B
ig
da
ta
C
ybe
r
c
r
im
e
de
tec
ti
on
Digit
a
l
f
und
tr
a
ns
a
c
ti
ons
Qua
dr
a
ti
c
mul
ti
va
r
iate
l
inea
r
r
e
gr
e
s
s
ion
a
ppr
oa
c
h
S
oka
l
–
M
iche
ne
r
’
s
dis
tr
ibut
e
d
pr
oxim
it
y
f
e
a
tur
e
e
nginee
r
ing
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e.
C
or
r
e
s
pon
din
g
A
u
th
or
:
Ar
ul
J
e
ya
nthi
P
a
ulr
a
j
De
pa
r
t
ment
o
f
C
om
pute
r
App
li
c
a
t
ions
a
nd
S
c
hoo
l
of
I
n
f
o
r
ma
ti
o
n
T
e
c
h
nolo
gy
,
M
a
dur
a
i
Ka
ma
r
a
j
Un
i
ve
r
s
it
y
M
a
dur
a
i,
T
a
mi
l
Na
du
,
I
ndia
E
mail:
jeya
nthi
jaya
ba
l@gm
a
il
.
c
om
1.
I
NT
RODU
C
T
I
ON
Digit
a
l
pa
yment
s
c
he
mes
a
r
e
f
ur
ther
popular
due
t
o
the
incr
e
a
s
ing
us
a
ge
o
f
s
mar
tphones
,
magne
ti
z
e
a
tt
e
nti
on
of
f
r
a
uds
ter
s
.
A
f
r
a
ud
de
tec
ti
on
f
r
a
m
e
wor
k
ba
s
e
d
on
XG
B
oos
t
with
r
a
ndom
unde
r
-
s
a
mpl
ing
(
R
US+X
GB
oos
t)
wa
s
de
v
e
loped
by
Ha
jek
e
t
al.
[
1]
with
the
a
im
of
im
pr
oving
f
r
a
ud
de
tec
ti
on
s
ys
tem
s
dur
ing
mobi
le
pa
yment
tr
a
ns
a
c
ti
ons
.
T
he
hybr
id
i
z
a
ti
on
of
c
ompetit
ive
s
wa
r
m
opti
mi
z
a
ti
on
a
s
we
ll
a
s
de
e
p
c
onvolut
ional
ne
ur
a
l
ne
twor
k
(
C
S
O
-
DC
NN
)
wa
s
de
ve
loped
by
Ka
r
thi
ke
ya
n
e
t
al
.
[
2
]
to
e
nha
nc
e
a
c
c
ur
a
c
y
of
f
r
a
udulent
tr
a
ns
a
c
ti
on
de
tec
ti
on.
A
B
a
ye
s
ian
opti
mi
z
a
ti
on
method
wa
s
de
ve
loped
by
Ha
s
he
mi
e
t
a
l.
[
3]
f
o
r
c
r
e
dit
c
a
r
d
f
r
a
ud
r
e
c
ognit
ion
with
we
ight
-
tuni
ng
h
ype
r
pa
r
a
mete
r
s
to
mention
the
pr
ob
lem
of
unba
lan
c
e
d
da
ta
while
c
ons
umi
ng
les
s
e
r
memor
y
a
nd
ti
me.
E
xplo
r
a
tor
y
a
na
lys
is
a
nd
mac
hine
lea
r
ning
(
ML
)
methods
we
r
e
de
s
igned
by
M
or
e
ir
a
e
t
al
.
[
4]
f
or
pr
e
dicting
f
r
a
ud
withi
n
the
ba
nking
s
ys
tem.
R
a
ndom
f
or
e
s
t
(
R
F
)
model
wa
s
p
r
e
s
e
nted
in
[
5]
to
c
las
s
if
y
onli
ne
c
r
e
dit
c
a
r
d
tr
a
ns
a
c
ti
ons
a
s
f
r
a
udulent.
A
ge
ne
ti
c
a
lgo
r
it
hm
(
GA
)
-
ba
s
e
d
f
e
a
tur
e
s
e
lec
ti
on
tec
hnique
incor
po
r
a
ted
by
M
L
methods
wa
s
de
s
igned
[
6]
with
the
a
im
o
f
c
r
e
dit
c
a
r
d
f
r
a
ud
de
tec
ti
on.
A
logi
s
ti
c
r
e
gr
e
s
s
ion
method
wa
s
de
s
igned
[
7]
to
f
or
e
c
a
s
t
tr
a
ns
a
c
ti
on
f
lagge
d
a
s
not
dur
ing
mo
bil
e
c
a
s
h
tr
a
ns
mi
ts
.
A
pe
r
s
ona
li
z
e
d
a
lar
m
met
hod
wa
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
689
-
699
690
int
r
oduc
e
d
in
[
8]
to
dis
ti
nguis
h
f
r
a
uds
withi
n
onl
in
e
f
und
tr
a
ns
f
e
r
s
by
uti
li
z
ing
s
e
que
nc
e
pa
tt
e
r
n
mi
ni
ng
ba
s
e
d
on
us
e
r
s
'
nor
mal
tr
a
ns
a
c
ti
on
log
f
il
e
s
.
S
tatis
ti
c
a
l
a
nd
mac
hine
lea
r
ning
models
we
r
e
int
r
oduc
e
d
in
[
9]
f
o
r
pa
yment
c
a
r
d
f
r
a
ud
de
tec
ti
on
.
Hybr
id
method
c
ombi
ning
ba
gging
a
nd
boos
ti
ng
e
ns
e
mbl
e
c
las
s
if
ier
s
wa
s
de
ve
loped
in
[
10]
f
o
r
c
r
e
dit
c
a
r
d
f
r
a
ud
r
e
c
ognit
ion,
r
e
s
ult
ing
in
higher
a
c
c
ur
a
c
y.
F
or
de
tec
ti
ng
int
e
r
ne
t
f
inanc
ial
de
c
e
pti
on,
I
ntelli
g
e
nt
a
nd
dis
pe
r
s
e
d
big
da
ta
method
wa
s
de
ve
loped
in
[
11]
.
An
uns
upe
r
vis
e
d
M
L
method
wa
s
de
s
ign
e
d
by
Ha
na
e
[
12]
f
or
de
tec
ti
ng
t
r
a
ns
a
c
ti
ona
l
f
r
a
ud
thr
ough
be
ha
vior
a
l
a
na
lys
is
.
B
a
c
k
p
r
opa
ga
ti
on
ne
ur
a
l
ne
t
wor
k
(
B
P
NN
)
model
wa
s
de
s
igned
Xiong
e
t
al.
[
13]
f
or
int
e
r
ne
t
f
inanc
ial
f
r
a
ud
identi
f
ica
ti
on.
XG
B
oos
t
a
nd
li
ght
gr
a
dient
boos
ti
ng
mac
hine
(
L
GB
M
)
meth
ods
we
r
e
de
ve
loped
by
Hs
in
e
t
al
.
[
14]
to
a
c
hieve
im
p
r
ov
e
d
f
r
a
ud
r
e
c
ognit
ion
out
c
omes
thr
ough
e
li
mi
na
ti
ng
nois
y
f
e
a
tur
e
s
a
nd
a
ddr
e
s
s
ing
da
ta
im
ba
lanc
e
pr
ob
lem.
An
int
e
ll
igent
s
a
mpl
ing
a
nd
s
e
lf
-
s
upe
r
vis
e
d
lea
r
ning
method
wa
s
de
ve
loped
by
C
he
n
e
t
al.
[
15
]
to
a
c
c
ur
a
tely
identif
y
c
r
e
di
t
c
a
r
d
t
r
a
ns
a
c
ti
ons
by
e
xtr
a
c
ti
ng
s
pa
ti
a
l
a
nd
tempor
a
l
f
e
a
tu
r
e
s
.
F
or
e
nha
nc
ing
a
c
c
ur
a
c
y
o
f
f
r
a
ud
de
tec
ti
on
by
b
a
lanc
ing
the
major
it
y
a
nd
mi
nor
it
y
c
las
s
e
s
,
dua
l
a
utoenc
ode
r
s
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ial
ne
twor
k
wa
s
de
ve
loped
in
s
tudy
[
16]
.
Hyb
r
idi
z
a
ti
on
o
f
bio
-
ins
pir
e
d
opti
mi
z
a
ti
on
method
a
s
we
ll
a
s
s
uppor
t
ve
c
tor
mac
hine
(
S
VM
)
wa
s
de
ve
loped
in
[
17
]
to
e
nha
nc
e
a
c
c
ur
a
c
y
of
c
r
e
dit
c
a
r
d
tr
a
ns
a
c
ti
on
de
tec
ti
on.
A
ne
ur
a
l
ne
twor
k
-
ba
s
e
d
f
e
a
tur
e
e
xtr
a
c
ti
on
method
wa
s
de
s
igned
in
[
18]
that
lea
r
ns
f
e
a
tur
e
f
or
f
r
a
ud
c
las
s
if
ica
ti
on
tas
k.
S
pa
ti
o
-
tempor
a
l
a
tt
e
nti
on
gr
a
ph
ne
ur
a
l
ne
twor
k
(
S
T
A
GN
)
wa
s
int
r
oduc
e
d
in
[
19
]
.
C
r
e
dit
c
a
r
d
de
c
e
pti
on
r
e
c
ognit
ion
method
wa
s
int
r
oduc
e
d
in
[
20]
.
A
de
e
p
c
onv
olut
ional
ne
ur
a
l
ne
twor
k
(
C
NN
)
model
wa
s
de
s
igned
in
[
21]
to
pe
r
c
e
ive
a
nomalies
a
s
o
f
us
ua
l
pa
tt
e
r
ns
c
r
e
a
ted
thr
ough
c
ompetit
ive
s
wa
r
m
opti
mi
z
a
ti
on.
L
e
ve
r
a
ging
M
L
a
s
we
ll
a
s
big
da
ta
a
na
lyt
ics
wa
s
pe
r
f
or
med
in
[
22]
.
In
s
tudy
[
23]
,
big
da
ta
-
dr
iven
ba
nking
ope
r
a
ti
ons
we
r
e
int
r
oduc
e
d
int
o
a
c
c
e
s
s
ibi
li
ty
of
a
ddit
ional
da
ta
i
mpr
ove
d
dif
f
iculty
of
s
e
r
vice
a
dm
ini
s
tr
a
ti
on
a
s
we
ll
a
s
pr
o
duc
ing
f
ier
c
e
c
ompetit
ion
,
a
nd
s
o
on
.
tele
c
omm
u
nica
ti
on
ne
twor
k
f
r
a
ud
de
pe
nd
on
big
da
ta
f
o
r
kil
l
ing
pigs
a
nd
plate
s
wa
s
e
xa
mi
ne
d
in
[
24
]
.
In
s
tudy
[
25]
,
s
pe
c
if
ics
a
nd
pa
tt
e
r
ns
of
c
ybe
r
c
r
im
e
we
r
e
de
s
igned
to
e
xa
mi
ne
the
int
e
r
na
ti
ona
l
c
omm
unit
y
a
nd
a
numbe
r
of
s
tate
s
in
c
ombating
c
ybe
r
c
r
im
e
in
the
f
ield
of
pa
yment
p
r
oc
e
s
s
ing.
T
he
main
c
ont
r
ibut
ion
o
f
thi
s
p
r
opos
e
d
QM
L
R
DP
F
E
method
is
f
oll
ow
a
s
:
−
T
o
im
p
r
ove
a
c
c
ur
a
c
y
of
c
ybe
r
c
r
im
e
de
tec
ti
on
i
n
digi
tal
f
und
tr
a
ns
a
c
ti
ons
with
big
da
ta,
qua
dr
a
ti
c
mul
ti
va
r
iate
li
ne
a
r
r
e
gr
e
s
s
ive
dis
tr
ibut
e
d
p
r
oxi
mi
ty
f
e
a
tu
r
e
e
nginee
r
ing
(
QM
L
R
DPF
E
)
meth
od
is
de
ve
loped
de
pe
nd
on
pr
e
p
r
oc
e
s
s
ing
a
s
we
ll
a
s
f
e
a
t
ur
e
e
nginee
r
ing.
−
T
o
mi
nim
ize
ti
me
f
o
r
f
r
a
udulent
a
c
ti
vit
ies
de
tec
ti
on,
QM
L
R
DPF
E
method
pe
r
f
or
ms
da
ta
p
r
e
pr
oc
e
s
s
ing.
T
he
qua
dr
a
ti
c
mul
ti
va
r
iate
li
ne
a
r
r
e
gr
e
s
s
ion
is
a
ppli
e
d
f
or
de
ter
mi
ning
the
mi
s
s
ing
da
ta.
T
he
Z
i
ggur
a
t
s
ynthetic
s
a
mpl
ing
method
to
s
olve
the
da
ta
im
ba
l
a
nc
e
.
−
T
he
QM
L
R
DPF
E
tec
hnique
uti
li
z
e
s
S
oka
l
–
M
ic
he
ne
r
’
s
dis
tr
ibut
e
d
pr
oxim
it
y
f
e
a
tur
e
e
nginee
r
in
g
f
or
mi
nim
izing
dim
e
ns
ionalit
y
o
f
da
taba
s
e
by
s
e
lec
ti
ng
s
igni
f
ica
nt
f
e
a
tur
e
s
.
−
F
inally,
e
xpe
r
im
e
ntal
a
s
s
e
s
s
ment
is
c
onduc
ted
t
o
c
a
lcula
te
pe
r
f
or
manc
e
of
QM
L
R
DPF
E
metho
d
in
c
ompar
is
on
to
c
onve
nti
ona
l
methods
.
T
he
pr
oblem
s
tate
ment
of
our
wor
k
is
pr
ovi
de
d
a
s
:
W
it
h
a
dva
nc
e
ments
in
mac
hine
lea
r
ning,
dif
f
e
r
e
nt
a
lgor
it
h
ms
ha
ve
be
e
n
e
nha
nc
e
d
to
c
onc
lu
de
whe
ther
tr
a
ns
a
c
ti
ons
in
digi
tal
s
ys
tem
s
a
r
e
f
r
a
udulent
or
not.
C
onve
nienc
e
a
ls
o
br
ings
a
n
incr
e
a
s
e
d
r
is
k
of
c
ybe
r
c
r
im
e
,
a
s
f
r
a
uds
ter
s
e
xploi
t
vulner
a
bil
it
ies
i
n
digi
tal
s
ys
tems
.
T
he
model
lac
ks
in
p
r
ovidi
ng
im
p
r
ove
d
a
c
c
ur
a
c
y
in
big
da
ta
a
ppli
c
a
ti
ons
f
or
f
r
a
ud
d
e
tec
ti
on
s
ys
tems
.
T
he
method
of
C
S
O
-
DC
NN
is
f
a
il
e
d
to
uti
li
z
e
pr
e
pr
oc
e
s
s
ing
tec
hniques
to
mention
pr
oblem
of
une
ve
n
or
unba
lanc
e
d
inf
or
mation
.
T
o
ove
r
c
om
e
thes
e
is
s
ue
s
,
our
p
r
opos
e
d
QM
L
R
DPF
E
tec
h
nique
is
im
pr
oving
the
a
c
c
ur
a
c
y
of
c
ybe
r
c
r
im
inal
de
tec
ti
on
with
les
s
e
r
ti
me
c
ons
umpt
ion
in
digi
tal
f
und
tr
a
ns
a
c
ti
ons
with
big
da
ta.
M
a
nus
c
r
ipt
is
s
tr
uc
tur
e
d
to
f
ive
pa
r
ts
a
s
pur
s
ue
:
S
e
c
ti
on
2
a
ppr
a
is
a
l
li
ter
a
tur
e
r
e
view
.
QM
L
R
DPF
E
method
is
e
xplaine
d
in
s
e
c
ti
on
3.
S
e
c
ti
on
4
pr
ovi
de
s
e
xpe
r
im
e
ntal
s
e
tup
a
nd
gives
e
xplana
ti
on
of
da
taba
s
e
.
C
ompar
a
ti
ve
a
na
lys
e
s
of
dis
s
im
il
a
r
pa
r
a
mete
r
s
a
r
e
given
in
s
e
c
ti
on
5.
L
a
s
tl
y,
s
e
c
ti
on
6
p
r
ovides
a
c
on
c
lus
ion
.
2.
M
E
T
HO
D
F
r
a
ud
de
tec
ti
on
a
nd
pr
e
ve
nti
on
in
f
und
tr
a
ns
a
c
ti
ons
a
r
e
c
r
uc
ial
a
s
pe
c
ts
of
the
moder
n
f
inanc
ial
s
ys
tem,
a
s
they
he
lp
a
ve
r
t
moneta
r
y
los
s
e
s
a
s
we
ll
a
s
s
us
tain
c
us
tom
e
r
tr
us
t.
T
he
r
e
f
or
e
,
f
inanc
ial
s
c
he
mes
a
r
e
de
pe
nda
ble
f
or
gua
r
a
ntyi
ng
the
s
a
f
e
ty
a
nd
s
e
c
ur
it
y
of
thei
r
c
us
tom
e
r
s
'
f
unds
.
An
e
f
f
icie
nt
s
ys
tem
is
r
e
qui
r
e
d
f
or
pr
e
ve
nti
ng
f
r
a
ud
de
tec
ti
on
dur
ing
dig
it
a
l
f
un
d
tr
a
ns
a
c
ti
ons
.
I
n
thi
s
s
e
c
ti
on,
a
nove
l
tec
hnique
c
a
ll
e
d
QM
L
R
DPF
E
is
int
r
oduc
e
d
f
or
a
c
c
ur
a
te
f
r
a
ud
d
e
tec
ti
on
in
dig
it
a
l
f
und
tr
a
ns
a
c
ti
ons
with
m
ini
mal
ti
me
c
ons
umpt
ion.
F
igu
r
e
1
il
lus
tr
a
tes
s
tr
uc
tu
r
a
l
de
s
ign
diagr
a
m
of
QM
L
R
DPF
E
tec
hnique
f
or
a
c
c
ur
a
te
d
e
tec
ti
on
of
f
r
a
udulent
tr
a
ns
a
c
ti
ons
or
c
ybe
r
c
r
im
e
s
.
E
f
f
e
c
ti
ve
f
r
a
udulent
t
r
a
ns
a
c
ti
on
de
tec
ti
on
tec
hniques
include
da
ta
a
c
quis
it
ion,
pr
e
pr
oc
e
s
s
ing,
a
nd
f
e
a
tur
e
e
nginee
r
ing
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
Quadr
ati
c
multi
v
ar
iat
e
li
ne
ar
r
e
gr
e
s
s
ive
dis
tr
ibut
e
d
pr
ox
imit
y
featur
e
…
(
A
r
ul
J
e
y
anthi
P
aulr
aj
)
691
F
igur
e
1.
Ar
c
hit
e
c
tur
e
of
p
r
opos
e
d
QM
L
R
DPF
E
tec
hnique
2.
1
.
Dat
a
ac
q
u
is
it
ion
I
t
invol
ve
s
ga
the
r
ing
r
e
leva
nt
t
r
a
ns
a
c
ti
on
inf
or
ma
ti
on
f
r
om
d
if
f
e
r
e
nt
r
e
s
our
c
e
s
s
uc
h
a
s
t
r
a
ns
a
c
ti
on
logs
,
us
e
r
pr
of
il
e
s
,
o
r
ne
twor
k
tr
a
f
f
ic.
T
his
p
r
oc
e
s
s
uti
li
z
e
s
the
f
inanc
ial
pa
yment
s
ys
tem
da
tas
e
t,
whic
h
include
s
s
e
ve
r
a
l
log
f
il
e
s
tot
a
li
ng
594,
643
r
e
c
or
d
s
.
F
e
a
tur
e
s
na
mely
a
ge
,
c
us
tom
e
r
inf
or
mation,
tr
a
ns
a
c
ti
on
a
mount
,
s
our
c
e
of
tr
a
ns
a
c
ti
on,
ta
r
ge
t,
types
of
tr
a
ns
a
c
ti
on
a
nd
labe
ls
a
r
e
e
mpl
oye
d
f
or
da
ta
a
na
lys
is
.
De
pe
nding
on
a
na
lys
is
,
f
r
a
udulent
a
c
ti
vit
y
or
nor
m
a
l
a
c
ti
vit
ies
a
r
e
identif
ied
.
2.
2
.
Dat
a
p
r
e
p
r
oc
e
s
s
in
g
I
t
is
vit
a
l
pa
r
t
in
da
ta
a
na
lys
is
whic
h
include
s
c
lea
ning,
tr
a
ns
f
or
mi
ng
,
or
ga
nizing
r
a
w
inf
or
mation
to
a
ppr
opr
iate
f
o
r
mat
f
or
e
ns
uing
s
tudy
a
nd
modeli
ng.
I
n
it
ially,
lar
ge
number
s
of
tr
a
ns
a
c
ti
on
in
f
or
m
a
ti
on
a
r
e
ga
ther
e
d
a
s
of
da
tas
e
ts
.
How
e
ve
r
,
thi
s
r
a
w
in
f
or
mation
f
r
e
que
ntl
y
include
s
mi
s
s
ing
va
lues
,
incon
s
is
tenc
ie
s
,
a
nd
im
ba
lanc
e
s
.
T
o
ha
ndle
thes
e
is
s
ue
s
,
the
p
r
o
pos
e
d
QM
L
R
DPF
E
pe
r
f
or
ms
da
ta
pr
e
pr
oc
e
s
s
ing,
whic
h
include
s
two
main
tas
ks
s
uc
h
a
s
ha
ndli
ng
mi
s
s
ing
da
ta
a
nd
a
ddr
e
s
s
ing
da
ta
im
ba
lanc
e
pr
oblems
.
2.
2.
1
.
Qu
ad
r
a
t
ic
m
u
lt
ivariat
e
li
n
e
ar
r
e
gr
e
s
s
ion
M
is
s
ing
da
ta
r
e
f
e
r
s
to
de
a
r
th
of
va
lues
in
a
pa
r
ti
c
ular
c
olum
n
o
f
the
da
tas
e
t.
T
he
s
e
mi
s
s
ing
da
ta
s
igni
f
ica
ntl
y
im
pa
c
t
the
a
na
lys
e
s
of
a
c
c
ur
a
te
f
r
a
u
dulent
tr
a
ns
a
c
ti
ons
de
f
e
c
ti
on.
T
he
r
e
f
or
e
,
ha
ndli
ng
mi
s
s
ing
da
ta
is
im
por
tant
to
e
ns
ur
e
a
c
c
ur
a
te
a
nd
r
e
li
a
ble
ou
tcome
s
a
s
of
big
da
ta
a
na
lys
is
.
T
he
pr
opos
e
d
QM
L
R
DPF
E
tec
hnique
uti
li
z
e
s
the
qua
dr
a
ti
c
mul
t
ivar
iate
li
ne
a
r
r
e
gr
e
s
s
ion
f
or
ha
ndli
ng
mi
s
s
ing
da
ta
in
a
given
da
t
a
s
e
t.
Qua
dr
a
ti
c
mul
ti
va
r
iate
li
ne
a
r
r
e
gr
e
s
s
ion
is
the
M
L
method
e
mpl
oye
d
to
pr
e
dict
mi
s
s
ing
va
lues
ba
s
e
d
on
mul
ti
ple
a
va
il
a
ble
da
ta
.
M
ult
ivar
iate
da
ta
indi
c
a
tes
mul
ti
ple
a
va
il
a
ble
da
ta
in
the
da
tas
e
t
us
e
d
f
or
f
indi
ng
the
mi
s
s
ing
va
lues
.
L
e
t
us
a
s
s
ume
input
da
tas
e
t
‘
’
a
s
we
ll
a
s
f
or
mul
a
ted
in
matr
ix,
=
[
1
2
…
11
12
…
1
21
22
…
2
⋮
⋮
…
⋮
1
2
…
]
(
1)
w
he
r
e
,
indi
c
a
tes
a
n
input
da
ta
mat
r
ix,
e
a
c
h
c
olum
n
indi
c
a
tes
a
number
o
f
f
e
a
tu
r
e
s
1
,
2
,
3
,
…
,
e
a
c
h
r
ow
indi
c
a
tes
a
number
of
da
ta
s
a
mpl
e
s
or
ins
t
a
nc
e
s
1
,
2
,
3
,
…
r
e
s
pe
c
ti
ve
ly.
Qua
dr
a
ti
c
li
ne
a
r
r
e
gr
e
s
s
ion
is
us
e
d
to
mea
s
ur
e
the
r
e
lations
hip
be
twe
e
n
the
indepe
nde
nt
va
r
iable
s
i
.
e
.
da
ta
s
a
mpl
e
s
is
modele
d
a
s
(
2)
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
689
-
699
692
=
0
+
1
1
+
2
2
2
+
⋯
+
(
2)
whe
r
e
,
de
notes
a
n
output
of
qua
dr
a
ti
c
li
ne
a
r
r
e
g
r
e
s
s
ion,
1
,
2
,
3
,
…
de
notes
a
numbe
r
of
da
ta
s
a
mpl
e
s
or
ins
tanc
e
s
,
0
,
1
,
3
,
…
de
notes
a
c
oe
f
f
icie
nts
of
th
e
qua
dr
a
ti
c
r
e
gr
e
s
s
ion
e
qua
ti
on,
indi
c
a
tes
the
e
r
r
or
ter
m
whic
h
mi
ni
mi
z
e
s
the
s
um
of
s
qua
r
e
d
va
r
iation
a
mong
e
xa
mi
ne
d
a
s
we
ll
a
s
f
or
e
c
a
s
ted
va
lues
.
=
(
−
)
2
(
3)
T
he
qua
dr
a
ti
c
f
unc
ti
on
invol
ve
s
f
indi
ng
the
va
lu
e
s
of
the
c
oe
f
f
icie
nts
that
mi
nim
ize
i
.
e
.
the
lea
s
t
a
bs
olut
e
de
viation
be
twe
e
n
the
obs
e
r
ve
d
a
nd
pr
e
dicte
d
va
lues
.
I
n
thi
s
wa
y,
thes
e
pr
opos
e
d
a
n
im
putation
tec
hnique
e
f
f
e
c
ti
ve
ly
ha
ndles
a
ll
mi
s
s
ing
va
lues
in
the
given
da
tas
e
t.
2.
2.
2
.
Adap
t
ive
Z
iggu
r
at
s
yn
t
h
e
t
ic
s
am
p
li
n
g
f
or
h
an
d
le
im
b
alan
c
e
d
a
t
a
Da
ta
im
ba
lanc
e
is
a
ddr
e
s
s
e
d
whe
r
e
a
ll
oc
a
ti
on
of
c
las
s
e
s
in
da
taba
s
e
not
e
ve
n.
I
n
da
tas
e
t,
one
or
mor
e
c
las
s
e
s
ha
ve
s
igni
f
ica
ntl
y
f
e
we
r
ins
tanc
e
s
than
other
s
.
T
h
is
im
ba
lanc
e
pos
e
s
c
ha
ll
e
nge
s
f
or
M
L
methods
a
s
be
c
ome
bias
e
d
towa
r
d
mains
tr
e
a
m
c
la
s
s
,
out
c
ome
at
de
pr
ived
r
e
s
ult
s
on
mi
no
r
it
y
c
las
s
.
T
o
s
olve
thi
s
is
s
ue
,
a
da
pti
ve
Z
iggur
a
t
s
ynthetic
s
a
mpl
ing
tec
hnique
is
e
mpl
oye
d
in
the
pr
opos
e
d
QM
L
R
DPF
E
to
ge
ne
r
a
te
s
ynthetic
da
ta
f
or
mi
nor
it
y
c
las
s
,
a
im
in
g
to
ba
lanc
e
c
las
s
a
ll
oc
a
ti
on
in
da
tas
e
t.
T
his
pr
oc
e
s
s
is
pa
r
ti
c
ular
ly
us
e
f
ul
f
or
im
pr
oving
the
a
c
c
ur
a
c
y
of
f
a
ult
de
tec
ti
on
in
dig
it
a
l
f
und
tr
a
ns
a
c
ti
ons
.
I
mbala
nc
e
d
da
ta
is
ha
ndled
by
a
pplyi
ng
a
da
pti
ve
Z
iggur
a
t
s
ynthetic
s
a
mpl
ing
f
o
r
ge
ne
r
a
ti
ng
nu
mber
of
inf
o
r
mation
s
a
mpl
e
s
a
t
mi
nor
it
y
c
las
s
.
I
nit
ially,
de
f
ine
tar
ge
t
a
mount
of
s
ynthetic
da
ta
s
a
mpl
e
s
ne
e
ds
to
be
ge
ne
r
a
ted
f
or
the
mi
nor
i
ty
c
las
s
a
s
(
4)
.
=
−
(
4)
whe
r
e
,
de
notes
tar
ge
t
a
mount
of
s
ynthetic
inf
o
r
m
a
ti
on
s
a
mpl
e
s
ne
e
ds
to
c
r
e
a
te
,
indi
c
a
tes
a
major
it
y
c
ounts
of
da
ta
s
a
mpl
e
s
,
r
e
pr
e
s
e
nts
a
mi
no
r
it
y
c
oun
ts
of
da
ta
s
a
mpl
e
s
in
the
da
tas
e
t.
Af
ter
f
indi
ng
the
c
ounts
to
ge
ne
r
a
te
s
ynthetic
d
a
ta
s
a
mpl
e
s
,
the
s
a
mpl
ing
pr
oc
e
s
s
i
s
e
xe
c
uted.
a
da
pti
ve
Z
iggu
r
a
t
s
ynthetic
s
a
mpl
ing
is
a
meth
od
us
e
d
f
or
ge
ne
r
a
ti
ng
da
ta
s
a
mpl
e
s
f
r
om
a
Ga
us
s
ian
pr
oba
bil
it
y
dis
tr
ibut
ion
of
the
other
da
ta
s
a
mpl
e
s
in
the
da
tas
e
t.
I
t
is
an
e
f
f
icie
nt
method
c
ompar
e
d
to
other
methods
.
F
i
r
s
t,
c
ons
ider
the
r
a
ndom
numbe
r
s
‘
’
f
r
om
0
to
1
i
.
e
.
[
0
,
1]
s
ince
Z
iggur
a
t
s
ynthetic
s
a
mpl
ing
uti
li
z
e
s
the
Ga
us
s
ian
pr
oba
bil
it
y
dis
tr
ibut
ion
.
=
{
1
,
2
,
3
,
…
}
(
5)
S
e
c
ondly,
ini
ti
a
li
z
e
the
number
of
laye
r
s
in
the
Ga
us
s
ian
pr
oba
bil
it
y
dis
tr
ibut
ion
a
s
s
hown
in
F
i
gur
e
2.
F
igur
e
2
il
lus
tr
a
tes
the
laye
r
s
e
gmenta
ti
on
in
Ga
us
s
ian
dis
tr
ibut
ion
whe
r
e
indi
c
a
tes
a
number
of
lay
e
r
s
a
nd
r
e
d
point
indi
c
a
tes
a
bounda
r
y
of
the
laye
r
1
,
2
,
3
r
e
s
pe
c
ti
ve
ly.
F
or
e
a
c
h
r
a
ndom
number
,
then
c
omput
e
the
f
oll
owing
f
unc
ti
on
,
=
∗
(
6)
w
he
r
e
,
de
notes
a
r
a
ndom
number
,
indi
c
a
tes
a
bo
unda
r
y
of
the
laye
r
.
Af
ter
that,
the
p
r
oba
bil
it
y
de
n
s
it
y
f
unc
ti
on
is
c
omput
e
d
with
‘
0
’
mea
n
a
nd
de
viation
‘
1’
.
(
)
=
1
√
2
2
[
(
−
)
2
2
2
]
(
7)
B
y
a
pplyi
ng
‘
0’
mea
n
(
)
a
nd
de
viation
(
)
‘
1’
,
the
a
bove
e
qua
ti
on
be
c
omes
wr
it
ten
a
s
(
8
)
,
(
)
=
[
(
)
2
2
]
(
8)
T
he
pr
oc
e
s
s
then
ve
r
if
ies
that
the
c
omput
e
d
‘
’
f
a
ll
s
withi
n
the
s
pe
c
if
ied
r
a
nge
the
(
)
.
T
he
n
the
va
lue
of
is
s
e
lec
ted
a
s
a
s
ynthetic
da
ta
s
a
mpl
e
.
Othe
r
wis
e
,
it
r
e
jec
ts
the
ge
ne
r
a
ted
s
a
mpl
e
s
a
nd
r
e
pe
a
ts
the
a
bove
pr
oc
e
s
s
unti
l
the
tar
ge
t
a
mount
of
s
ynthetic
inf
o
r
mation
s
a
mpl
e
s
is
r
e
a
c
he
d.
L
ike
thi
s
,
da
ta
i
mbala
nc
e
pr
oblems
a
r
e
ha
ndled
in
the
pr
opos
e
d
tec
hniques
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
Quadr
ati
c
multi
v
ar
iat
e
li
ne
ar
r
e
gr
e
s
s
ive
dis
tr
ibut
e
d
pr
ox
imit
y
featur
e
…
(
A
r
ul
J
e
y
anthi
P
aulr
aj
)
693
F
igur
e
2.
L
a
ye
r
s
of
Ga
us
s
ian
dis
tr
ibut
ion
Algor
it
hm
1
given
c
lea
r
ly
de
s
c
r
ibes
da
ta
pr
e
pr
oc
e
s
s
ing
to
r
e
duc
e
ti
me
uti
li
z
a
ti
on
of
f
inanc
ial
f
r
a
ud
pr
e
diction
dur
ing
the
d
igi
tal
f
und
tr
a
ns
a
c
ti
on.
I
ni
ti
a
ll
y,
a
number
o
f
da
ta
s
a
mpl
e
s
a
r
e
ga
ther
e
d
a
s
of
da
tas
e
t.
Ne
xt,
mi
s
s
ing
va
lues
a
r
e
r
e
c
ognize
d
by
a
pplyi
ng
qua
dr
a
ti
c
li
ne
a
r
r
e
gr
e
s
s
ion.
Onc
e
mi
s
s
ing
va
lues
a
r
e
f
i
ll
e
d,
the
is
s
ue
of
da
ta
im
ba
lanc
e
is
a
ddr
e
s
s
e
d.
F
ir
s
tl
y,
the
tar
ge
t
number
of
s
ynthetic
da
ta
s
a
mpl
e
s
is
de
t
e
r
mi
ne
d.
T
he
n,
r
a
ndom
number
s
a
r
e
ge
ne
r
a
ted
.
S
ubs
e
que
ntl
y,
it
mul
ti
pli
e
d
thr
ough
a
pr
e
de
f
ined
bounda
r
y
.
F
oll
owing
thi
s
;
the
pr
oba
bil
it
y
de
ns
it
y
f
unc
ti
on
is
e
s
ti
mate
d
with
z
e
r
o
mea
n
a
nd
one
s
tanda
r
d
de
viation.
E
a
c
h
e
s
ti
mate
d
da
ta
point
is
va
li
da
ted
a
ga
ins
t
the
pr
oba
bil
it
y
de
ns
it
y
f
unc
ti
on.
I
f
it
s
va
lue
is
les
s
e
r
than
that
of
the
pr
oba
bil
it
y
de
ns
it
y
f
unc
ti
on,
it
is
s
e
lec
ted
a
s
a
s
ynthetic
da
ta
s
a
mpl
e
.
Othe
r
wis
e
,
the
da
ta
s
a
mpl
e
is
r
e
jec
ted.
T
his
pr
oc
e
s
s
c
onti
nue
s
unti
l
the
tar
ge
t
number
of
s
ynthetic
da
ta
s
a
mpl
e
s
is
r
e
a
c
he
d.
Algor
it
hm
1
.
Da
ta
p
r
e
-
pr
oc
e
s
s
ing
Input: Dataset ‘
’, features
1
,
2
,
3
,
…
, data samples or instances
1
,
2
,
3
,
…
Output: Pre
-
processed dataset
Begin
1. For each dataset ‘
’ with features ‘
’
2. Formulate input vector matrix ‘
’
using (1)
3. If missing value in dataset then
4. Apply
quadratic
linear regression using (2)
5. Fill the value to the respective missing column
6. End if
7. Find number of target data samples needs to be generated using (3)
8. Define the random numbers using (5)
9. Define the numbers layers ‘
’ and boundary ‘
’
using
10. Measure the product of the random numbers and boundary using (6)
11. compute the probability density function with zero mean and deviation using (8)
12
.
if (
<
(
)
) then
13. Selected as a synthetic data samples
14. else
15: Reject the data samples
16. end if
17. Go to step 8
18. Obtain the number of synthetic data samples
19. Return (balanced dataset)
20. End for
End
2.
3
.
S
ok
al
–
M
ichener
’
s
d
is
t
r
ib
u
t
e
d
p
r
oxim
i
t
y
f
e
at
u
r
e
e
n
gin
e
e
r
in
g
W
it
h
the
ba
lanc
e
d
da
ta
s
e
t,
the
f
e
a
tur
e
e
nginee
r
ing
pr
oc
e
s
s
is
e
xe
c
uted
f
or
dim
e
ns
ionalit
y
r
e
duc
ti
on.
Dimens
ionalit
y
r
e
duc
ti
on
is
a
tec
hnique
to
mi
n
im
i
z
e
the
number
of
f
e
a
tur
e
s
wi
thi
n
a
big
da
tas
e
t.
B
ig
da
tas
e
ts
include
a
mor
e
number
of
f
e
a
tur
e
s
whic
h
c
a
us
e
s
incr
e
a
s
e
d
c
omput
a
ti
ona
l
c
ompl
e
xit
y
a
nd
c
ha
ll
e
nge
s
in
a
c
hieving
a
c
c
ur
a
te
c
las
s
if
ica
ti
on.
T
o
mention
thi
s
is
s
ue
,
the
S
oka
l
–
M
iche
ne
r
’
s
dis
tr
ibut
e
d
pr
oxi
mi
t
y
f
e
a
tur
e
e
nginee
r
ing
method
is
de
ve
loped
in
QM
L
R
DPF
E
f
or
di
mens
ionalit
y
r
e
duc
ti
on
by
s
e
lec
ti
ng
t
he
mos
t
s
igni
f
ica
nt
f
e
a
tur
e
s
.
T
hr
ough
the
identif
ica
ti
on
of
s
igni
f
ica
nt
f
e
a
tu
r
e
s
,
thi
s
a
ppr
oa
c
h
e
nha
nc
e
s
the
a
c
c
ur
a
c
y
of
c
ybe
r
c
r
im
e
de
tec
ti
on,
s
pe
c
if
ica
ll
y
in
c
las
s
if
ying
th
e
f
r
a
udulent
a
c
ti
vit
ies
with
in
digi
tal
f
und
tr
a
ns
a
c
ti
ons
.
F
igur
e
3
f
low
p
r
oc
e
s
s
of
the
S
oka
l
–
M
iche
ne
r
’
s
dis
tr
ibut
e
d
pr
oxim
it
y
f
e
a
tur
e
e
nginee
r
ing
f
or
a
c
c
ur
a
te
f
r
a
udulent
a
c
ti
vit
ies
de
tec
ti
on
.
L
e
t
us
c
on
s
ider
the
number
of
f
e
a
tur
e
s
1
,
2
,
3
,
…
dis
tr
ibut
e
d
in
the
given
da
tas
e
t.
P
r
oxim
it
y
r
e
f
e
r
s
to
the
de
gr
e
e
o
f
c
los
e
ne
s
s
or
s
im
il
a
r
it
y
be
twe
e
n
two
f
e
a
tur
e
s
in
a
d
a
taba
s
e
.
Af
ter
wa
r
d
us
ing
S
oka
l
–
M
iche
ne
r
’
s
f
o
r
de
ter
mi
nin
g
s
im
il
a
r
it
y
be
twe
e
n
f
e
a
tur
e
s
.
=
(
,
)
(
9)
w
he
r
e
,
de
notes
a
f
e
a
tu
r
e
p
r
oxim
it
y
,
(
,
)
ind
ica
tes
a
S
oka
l
–
M
iche
ne
r
’
s
s
im
il
a
r
it
y.
I
t
is
mea
s
ur
e
d
as
(
10)
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
689
-
699
694
(
,
)
=
1
−
|
∆
|
(
10)
w
he
r
e
,
i
n
d
ica
tes
a
S
o
ka
l
–
M
ic
he
n
e
r
’
s
s
im
i
la
r
i
t
y
,
∆
d
e
no
tes
a
di
f
f
e
r
e
nc
e
be
twe
e
n
the
tw
o
f
e
a
t
u
r
e
s
,
d
e
n
ot
e
s
a
to
ta
l
n
um
be
r
o
f
f
e
a
t
u
r
e
s
.
T
he
S
o
ka
l
–
M
iche
ne
r
’
s
s
im
il
a
r
i
ty
p
r
o
vi
de
s
the
ou
tc
om
e
s
r
a
n
ge
s
f
r
om
0
t
o
1
.
=
{
,
(
,
)
>
,
(
,
)
<
(
11)
w
he
r
e
de
notes
a
n
output
f
unc
ti
on
,
indi
c
a
tes
a
th
r
e
s
hold
f
or
s
im
il
a
r
it
y
c
oe
f
f
icie
nt
(
,
)
r
e
s
ult
s
.
I
f
the
c
oe
f
f
icie
nt
(
,
)
e
xc
e
e
ds
the
th
r
e
s
hold,
the
f
e
a
tur
e
i
s
ter
med
a
s
s
igni
f
ica
nt
f
e
a
tur
e
(
)
.
Othe
r
wis
e
,
it
is
ter
med
a
s
ins
igni
f
ica
nt
f
e
a
tur
e
(
).
F
inally,
the
s
igni
f
ica
nt
f
e
a
tur
e
s
a
r
e
s
e
lec
ted
f
or
a
c
c
ur
a
te
f
r
a
u
dulent
tr
a
ns
a
c
ti
on
de
tec
ti
on
a
nd
other
f
e
a
tur
e
s
a
r
e
r
e
mo
ve
d
f
r
om
the
da
tas
e
t.
T
he
a
lgo
r
it
hm
f
or
S
oka
l
–
M
iche
ne
r
’
s
dis
tr
ibut
e
d
pr
oxim
it
y
f
e
a
tur
e
e
nginee
r
ing
is
given
.
F
igur
e
3.
F
low
pr
oc
e
s
s
of
S
oka
l
-
M
iche
ne
r
’
s
dis
tr
ib
uted
pr
oxim
it
y
f
e
a
tur
e
e
nginee
r
ing
Algor
it
hm
2
de
s
c
r
ibes
the
pr
oc
e
s
s
of
s
igni
f
i
c
a
nt
f
e
a
tur
e
s
e
lec
ti
on
us
ing
S
oka
l
–
M
iche
ne
r
’
s
dis
tr
ibut
e
d
pr
oxi
mi
ty
f
e
a
tur
e
e
nginee
r
ing
tec
hni
que
f
or
i
mpr
oving
f
r
a
udulent
t
r
a
ns
a
c
ti
on
de
tec
ti
on
while
r
e
duc
ing
ti
me
uti
li
z
a
ti
on
.
T
he
p
r
e
pr
oc
e
s
s
e
d
da
tas
e
t
c
ompr
is
e
s
s
e
ve
r
a
l
f
e
a
tur
e
s
us
e
d
a
s
th
e
input
.
S
ubs
e
que
ntl
y,
f
e
a
tur
e
pr
oxim
it
y
is
c
omput
e
d
be
twe
e
n
the
f
e
a
tur
e
s
ba
s
e
d
on
S
oka
l
–
M
ich
e
ne
r
’
s
s
i
mi
lar
it
y
mea
s
ur
e
.
T
his
s
im
il
a
r
it
y
mea
s
ur
e
dis
ti
nguis
he
s
the
s
igni
f
ica
nt
a
nd
ins
igni
f
ica
nt
f
e
a
tur
e
s
by
s
e
t
ti
ng
the
thr
e
s
hold
withi
n
the
da
tas
e
t.
F
inally
,
im
po
r
tant
f
e
a
tur
e
s
a
r
e
c
hos
e
n
to
im
pr
ove
a
c
c
ur
a
c
y
of
f
r
a
udulent
de
tec
ti
on
in
digi
tal
f
und
tr
a
ns
a
c
ti
ons
.
Algor
it
hm
2
.
S
oka
l
–
M
iche
ne
r
’
s
dis
tr
ibut
e
d
pr
oxim
it
y
f
e
a
tur
e
e
nginee
r
ing
Input:
Preprocessed datasets
‘
’, features
1
,
2
,
3
,
…
,
data samples or instances
1
,
2
,
3
,
…
Output:
Select
relevant features
Begin
1: Collect the preprocessed dataset as input
2. For each feature ‘
’
3. Measure the proximity using (9)
4. Measure the Sokal
–
Michener’s similarity ‘
(
,
)
’
5.
if
(
(
,
)
>
)
then
6. Features are identified as significant
7. else
8. Features are identified as insignificant
9. End if
10.
Select the significant features and remove other features
11. end for
End
3.
E
XP
E
RI
M
E
NT
AL
S
CE
NA
R
I
O
E
xpe
r
im
e
ntal
a
s
s
e
s
s
ment
of
QM
L
R
DPF
E
tec
hnique
a
nd
e
xis
ti
ng
XG
B
oos
t
-
ba
s
e
d
f
r
a
ud
de
tec
ti
on
f
r
a
mew
or
k
[
1]
a
nd
C
S
O
+
DC
NN
[
2
]
a
r
e
e
xe
c
uted
by
P
ython
c
oding
.
T
o
c
a
r
r
y
out
e
xpe
r
im
e
nt,
f
inanc
ial
pa
yment
s
ys
tem
da
tas
e
t
is
c
oll
e
c
ted
a
s
of
Ka
ggle
da
tas
e
t
[
26]
.
M
a
jor
ob
jec
ti
ve
of
thi
s
da
taba
s
e
is
e
mpl
oye
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
Quadr
ati
c
multi
v
ar
iat
e
li
ne
ar
r
e
gr
e
s
s
ive
dis
tr
ibut
e
d
pr
ox
imit
y
featur
e
…
(
A
r
ul
J
e
y
anthi
P
aulr
aj
)
695
to
identi
f
y
f
r
a
udulent
t
r
a
ns
a
c
ti
ons
a
nd
nor
mal
pa
y
ments
.
T
his
da
tas
e
t
include
s
a
s
e
ve
r
a
l
log
f
i
les
tha
t
include
594
,
643
r
e
c
or
ds
.
Da
tas
e
t
include
s
10
f
e
a
tur
e
s
s
uc
h
a
s
s
tep,
c
us
tom
e
r
,
a
ge
,
ge
nde
r
,
z
ipcode
Or
i
,
a
nd
s
o
on.
I
n
or
de
r
to
c
onduc
t
the
e
xpe
r
im
e
nt,
the
number
o
f
da
ta
s
a
mpl
e
s
is
c
ons
ider
e
d
in
the
r
a
nge
s
f
r
om
1
0
,
000
to
100
,
000.
3.
1.
I
m
p
lem
e
n
t
at
ion
d
e
t
ail
s
I
n
thi
s
s
tudy,
we
de
ve
loped
a
nove
l
tec
hnique
c
a
ll
e
d
qua
dr
a
ti
c
mul
ti
va
r
iate
li
ne
a
r
r
e
gr
e
s
s
ive
dis
tr
ibut
e
d
pr
oxim
i
ty
f
e
a
tu
r
e
e
ngi
ne
e
r
ing
(
Q
M
L
R
DPF
E
)
is
de
ve
loped
to
e
nha
nc
e
the
a
c
c
ur
a
c
y
of
c
ybe
r
c
r
im
inal
de
tec
ti
on
with
mi
nim
um
ti
me
c
ons
u
mpt
ion.
−
T
he
QM
L
R
DPF
E
method
c
ompr
is
e
s
two
pr
i
mar
y
s
teps
na
mely
da
ta
pr
e
pr
oc
e
s
s
ing
a
nd
f
e
a
tur
e
e
nginee
r
ing.
−
W
e
c
ompar
e
d
our
QM
L
R
DPF
E
tec
hnique
c
ompar
e
d
to
e
xis
ti
ng
XG
B
oos
t
-
b
a
s
e
d
f
r
a
ud
de
te
c
ti
on
f
r
a
mew
or
k
[
1]
a
nd
C
S
O
+
DC
NN
[
2]
us
ing
f
inanc
ial
pa
yment
s
ys
tem
da
tas
e
t
to
va
li
da
te
the
r
e
s
ult
s
.
−
T
he
da
taba
s
e
c
ontains
pa
yments
f
r
om
dif
f
e
r
e
nt
c
us
tom
e
r
s
made
a
t
dis
s
im
il
a
r
t
im
e
pe
r
iods
a
s
we
ll
a
s
thr
ough
dive
r
s
e
a
mount
s
.
M
a
in
a
im
of
thi
s
da
ta
ba
s
e
is
us
e
d
to
de
tec
t
the
f
r
a
udulent
tr
a
ns
a
c
ti
ons
a
nd
nor
mal
pa
yments
−
I
nit
ially
the
pr
e
pr
oc
e
s
s
ing
is
c
a
r
r
ied
ou
t,
invol
vin
g
two
ke
y
p
r
oc
e
s
s
e
s
na
mely
ha
ndli
ng
mi
s
s
ing
da
ta
a
nd
ba
lanc
ing
the
da
tas
e
t.
T
he
mi
s
s
ing
inf
or
mation
d
e
pe
nds
on
mul
ti
ple
a
va
il
a
ble
da
ta
a
s
we
ll
a
s
im
p
uted
inf
or
mation
is
to
r
e
duc
e
lea
s
t
a
bs
olut
e
de
viation.
−
Af
ter
that
the
f
e
a
tur
e
e
nginee
r
ing
tec
hnique,
s
e
lec
t
ing
the
mos
t
r
e
leva
nt
f
e
a
tur
e
s
.
Du
r
ing
the
identi
f
ica
ti
on
of
s
igni
f
ica
nt
f
e
a
tur
e
s
,
s
pe
c
if
ica
ll
y
in
c
las
s
if
ying
t
he
f
r
a
udulent
a
c
ti
vit
ies
withi
n
digi
tal
f
und
tr
a
ns
a
c
ti
ons
in
thi
s
da
tas
e
t.
4.
P
E
RF
ORM
AN
CE
COM
P
AR
I
S
I
ON
AN
AL
YSI
S
I
n
thi
s
s
e
c
ti
on,
pe
r
f
or
manc
e
of
the
p
r
opos
e
d
QM
L
R
DPF
E
tec
hnique
a
nd
e
xis
ti
ng
R
US+X
GB
oos
t
[
1]
a
nd
C
S
O
+
DC
NN
[
2]
a
r
e
a
s
s
e
s
s
e
d
with
va
r
ious
metr
ics
,
including
a
c
c
ur
a
c
y,
pr
e
c
is
ion,
e
xe
c
uti
on
t
im
e
a
nd
s
pa
c
e
c
ompl
e
xit
y.
−
De
tec
ti
on
a
c
c
ur
a
c
y:
I
t
is
de
f
ined
to
r
a
ti
o
o
f
a
c
c
ur
a
tely
de
tec
ti
ng
f
r
a
udulent
t
r
a
ns
a
c
ti
ons
a
nd
nor
mal
tr
a
ns
a
c
ti
ons
a
s
of
tot
a
l
number
of
da
ta.
I
t
is
c
ompu
ted
a
s
(
12)
,
=
(
+
+
+
+
)
∗
100
(
12)
whe
r
e
,
indi
c
a
tes
a
de
tec
ti
on
a
c
c
ur
a
c
y,
indi
c
a
tes
tr
ue
pos
it
ive,
s
ymbol
ize
tr
ue
ne
ga
ti
ve
,
de
notes
f
a
ls
e
pos
it
ive,
a
nd
indi
c
a
tes
f
a
ls
e
ne
ga
ti
ve
.
I
t
is
c
a
lcula
ted
in
pe
r
c
e
ntage
(
%
)
.
−
P
r
e
c
is
ion:
I
t
is
de
f
ined
a
s
r
a
ti
o
of
de
tec
ti
ng
f
r
a
udulent
tr
a
ns
a
c
ti
ons
a
nd
no
r
mal
tr
a
ns
a
c
ti
ons
.
I
t
is
c
omput
e
d
a
s
(
13)
,
=
(
+
)
(
13)
whe
r
e
,
de
notes
a
pr
e
c
is
ion,
de
notes
the
tr
ue
pos
it
ive,
a
nd
r
e
pr
e
s
e
nts
the
f
a
ls
e
pos
it
ive.
−
De
tec
ti
on
ti
me:
I
t
is
mea
s
ur
e
d
a
s
the
a
mount
o
f
ti
me
c
ons
umed
by
a
lgor
it
hm
f
o
r
de
tec
ti
ng
the
f
r
a
ud
ulent
tr
a
ns
a
c
ti
ons
a
nd
nor
mal
tr
a
ns
a
c
ti
ons
.
T
he
ti
me
is
c
omput
e
d
a
s
(
14)
,
=
∑
=
1
∗
(
)
(
14)
whe
r
e
,
de
notes
a
de
tec
ti
on
ti
me
de
pe
nd
on
da
ta
s
a
mpl
e
s
a
s
we
ll
a
s
a
c
tual
ti
me
uti
li
z
e
d
in
de
tec
ti
ng
the
f
r
a
udulent
t
r
a
ns
a
c
ti
ons
a
nd
nor
mal
tr
a
ns
a
c
ti
ons
de
noted
by
(
)
.
I
t
is
c
a
lcula
ted
in
mi
ll
is
e
c
onds
(
ms
)
.
−
S
pa
c
e
c
ompl
e
xit
y:
I
t
is
c
a
lcula
ted
a
s
a
mount
o
f
m
e
mor
y
s
pa
c
e
uti
li
z
e
d
thr
ough
method
f
or
de
tec
ti
ng
the
f
r
a
udulent
tr
a
ns
a
c
ti
ons
a
nd
nor
mal
tr
a
ns
a
c
ti
ons
.
T
he
S
pa
c
e
c
ompl
e
xit
y
is
c
omput
e
d
a
s
(
15)
,
=
∑
=
1
∗
(
)
(
15)
whe
r
e
,
de
notes
a
s
pa
c
e
c
ompl
e
xit
y
de
pe
nd
on
da
ta
s
a
mpl
e
s
a
nd
memor
y
s
pa
c
e
uti
li
z
e
d
a
t
de
tec
ti
ng
the
f
r
a
udulent
tr
a
ns
a
c
ti
ons
a
nd
nor
mal
tr
a
ns
a
c
ti
ons
de
noted
by
(
)
.
I
t
is
c
a
lcula
ted
in
kil
obytes
(
k
B
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
689
-
699
696
T
a
ble
1
given
a
bove
il
lus
tr
a
tes
pe
r
f
or
manc
e
c
ompar
is
on
of
de
tec
ti
on
a
c
c
ur
a
c
y
of
f
r
a
udulent
tr
a
ns
a
c
ti
ons
a
nd
nor
mal
pa
yments
us
ing
thr
e
e
methods
na
mely
QM
L
R
DPF
E
tec
hnique
a
nd
e
xis
ti
ng
R
US+X
GB
oos
t
[
1]
a
nd
C
S
O
+
DC
NN
[
2]
.
Among
the
thr
e
e
tec
hniques
,
pe
r
f
or
manc
e
of
QM
L
R
DPF
E
method
is
im
pr
ove
d
than
the
c
onve
nti
ona
l
tec
hniques
.
F
or
e
xa
mpl
e
,
mea
s
ur
ing
10
,
000
da
ta
s
a
mpl
e
s
c
omput
ing
de
tec
ti
on
a
c
c
ur
a
c
y,
QM
L
R
DPF
E
method
a
tt
a
ined
a
c
c
ur
a
c
y
of
90%
.
As
we
ll
,
of
c
onve
nti
ona
l
[
1
]
,
[
2]
wa
s
86%
a
nd
982%
,
r
e
s
pe
c
ti
ve
ly.
F
or
e
a
c
h
met
hod,
ten
dif
f
e
r
e
nt
ou
tcome
s
a
r
e
e
xa
mi
ne
d.
T
he
obs
e
r
ve
d
outcome
s
a
r
e
c
ompar
e
d.
Ove
r
a
ll
c
ompa
r
a
ti
ve
s
tudy
de
notes
whic
h
of
QM
L
R
DPF
E
method
e
nha
n
c
e
d
by
5%
a
nd
3%
than
the
s
tudy
[
1
]
,
[
2
]
.
T
his
is
due
t
o
uti
li
z
ing
S
oka
l
–
M
iche
ne
r
’
s
dis
tr
ibut
e
d
p
r
oxim
it
y
f
e
a
tur
e
e
nginee
r
ing
method
is
de
ve
loped
f
or
dim
e
ns
ionalit
y
r
e
duc
ti
on
by
s
e
lec
ti
ng
the
s
igni
f
ica
nt
f
e
a
tur
e
s
.
De
pe
nd
on
a
c
c
ur
a
tely
pe
r
f
or
ms
c
ybe
r
c
r
im
e
de
tec
ti
on
,
by
d
is
ti
nguis
hing
the
f
r
a
udulent
a
c
ti
vit
ies
or
nor
mal
d
ur
ing
the
digi
tal
f
und
tr
a
ns
a
c
ti
ons
.
F
igur
e
4
de
picts
a
c
ompar
is
on
of
pr
e
c
is
ion.
T
hr
e
e
methods
,
na
mely
QM
L
R
DPF
E
tec
hnique,
e
xis
ti
ng
R
US+X
GB
oos
t
[
1]
,
a
nd
C
S
O+
DC
NN
[
2]
,
a
r
e
uti
li
z
e
d
f
or
c
a
lcula
ti
ng
p
r
e
c
is
ion.
O
utcome
s
de
mons
tr
a
te
whic
h
QM
L
R
DPF
E
tec
hnique
a
c
hie
ve
s
s
upe
r
ior
than
c
onve
nti
ona
l
tec
hniques
.
Obs
e
r
ve
d
r
e
s
ult
s
of
QM
L
R
DPF
E
method
a
r
e
c
ompar
e
d
to
e
xis
ti
ng
methods
.
Ove
r
a
ll
c
ompar
is
on
r
e
ve
a
ls
that
the
pr
e
c
is
ion
pe
r
f
or
manc
e
in
a
c
c
ur
a
tely
de
tec
ti
ng
f
r
a
udulent
a
c
ti
vit
ies
dur
ing
digi
tal
f
und
t
r
a
ns
a
c
ti
ons
is
e
nha
nc
e
d
by
6%
a
nd
3%
than
the
[
1]
,
[
2]
whe
n
a
pplyi
ng
the
QM
L
R
DPF
E
tec
hnique.
T
o
a
c
hi
e
ve
thi
s
im
pr
ove
d
pe
r
f
o
r
manc
e
,
the
QM
L
R
DPF
E
tec
hnique
uti
li
z
e
s
S
oka
l
–
M
iche
ne
r
’
s
dis
tr
ibut
e
d
pr
ox
im
it
y
f
e
a
tur
e
e
nginee
r
ing
tec
hnique
f
or
s
e
lec
ti
ng
tar
ge
t
f
e
a
tur
e
s
,
ther
e
by
e
nha
nc
ing
de
tec
ti
on
thr
ough
im
pr
ove
d
a
nd
mi
nim
izing
outcome
s
dur
ing
f
r
a
udulent
t
r
a
ns
a
c
ti
on
de
tec
ti
on.
Give
n
a
bove
de
picts
the
pe
r
f
or
manc
e
c
ompar
is
on
of
de
tec
ti
on
ti
me
by
QM
L
R
DPF
E
tec
hnique,
e
xis
ti
ng
R
US+X
GB
oos
t
[
1]
,
a
nd
C
S
O+
DC
NN
[
2
]
.
P
e
r
f
or
manc
e
o
f
f
or
e
ve
r
y
thr
e
e
tec
hniques
obtain
e
nha
nc
e
d
a
s
e
nha
nc
ing
number
of
da
ta
s
a
mpl
e
s
.
E
s
pe
c
ially,
f
or
QM
L
R
DPF
E
tec
hnique
is
mi
nim
ize
d
than
the
[
1]
,
[
2
]
.
L
e
t
us
a
s
s
ume
ini
ti
a
l
it
e
r
a
ti
on
with
10
,
0
00
da
ta
s
a
mpl
e
s
,
whe
r
e
f
or
QM
L
R
DPF
E
metho
d
wa
s
l
ikew
is
e
,
ti
me
uti
li
z
a
ti
on
f
or
[
1]
,
[
2]
r
e
s
pe
c
ti
ve
ly.
T
he
obtaine
d
ove
r
a
ll
r
e
s
ult
s
of
QM
L
R
DPF
E
method
a
r
e
c
ompar
e
d
to
ou
tcome
s
of
c
onve
nti
ona
l
tec
hniques
.
T
he
c
ompar
is
on
outcome
s
de
notes
whic
h
pe
r
f
or
manc
e
of
de
tec
ti
on
ti
me
us
ing
QM
L
R
DPF
E
tec
hnique
is
s
ig
nif
ica
ntl
y
r
e
duc
e
d
by
13%
a
nd
7
%
than
the
s
tudy
[
1]
,
[
2]
.
T
his
is
owing
to
QM
L
R
DPF
E
tec
hnique
pe
r
f
or
m
e
d
the
da
ta
pr
e
pr
oc
e
s
s
ing
a
nd
f
e
a
tur
e
s
e
lec
ti
on
pr
oc
e
s
s
.
I
n
da
ta
pr
e
pr
oc
e
s
s
ing,
mi
s
s
ing
inf
o
r
mation
is
de
ter
mi
ne
d
by
a
pplyi
ng
qua
dr
a
ti
c
mul
ti
va
r
iate
li
ne
a
r
r
e
gr
e
s
s
ion
a
ppr
oa
c
h.
Da
ta
im
ba
lanc
e
pr
oblem
a
ls
o
s
olved
thr
ough
the
a
da
pti
ve
Z
iggur
a
t
s
ynthetic
s
a
mpl
ing
tec
hnique
to
c
r
e
a
te
s
ynthetic
da
ta
s
a
mpl
e
s
.
T
he
tar
ge
t
f
e
a
tur
e
s
e
lec
ti
on
a
ls
o
mi
nim
ize
s
ti
me
c
ons
umpt
ion
o
f
f
r
a
udulent
tr
a
ns
a
c
ti
on
de
tec
ti
on.
T
a
ble
1.
C
ompar
is
on
of
de
tec
ti
on
a
c
c
ur
a
c
y
N
umbe
r
of
da
ta
s
a
mpl
e
s
D
e
te
c
ti
on a
c
c
ur
a
c
y (
%
)
Q
M
L
R
D
P
F
E
R
U
S
+
X
G
B
oos
t
C
S
O
+
D
C
N
N
10000
90
87
88.5
20000
91.22
88.52
89.85
30000
90.33
86.23
87.52
40000
90.88
85.99
87.78
50000
92.12
86.89
89.52
60000
91.05
86.74
88.74
70000
92.05
85.52
87.22
80000
91.5
86.56
88.56
90000
92.2
87.52
89.5
100000
91.88
85.98
87.22
F
igur
e
4.
P
e
r
f
or
manc
e
c
ompar
is
on
o
f
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
Quadr
ati
c
multi
v
ar
iat
e
li
ne
ar
r
e
gr
e
s
s
ive
dis
tr
ibut
e
d
pr
ox
imit
y
featur
e
…
(
A
r
ul
J
e
y
anthi
P
aulr
aj
)
697
T
a
ble
2
a
nd
F
igur
e
5
de
picts
the
pe
r
f
or
manc
e
c
ompar
is
on
of
de
tec
ti
on
ti
me
by
QM
L
R
DPF
E
tec
hnique,
e
xis
ti
ng
R
US+X
GB
oos
t
[
1]
,
a
nd
C
S
O+
DC
NN
[
2]
.
P
e
r
f
o
r
manc
e
of
f
or
e
ve
r
y
thr
e
e
t
e
c
hniques
obtain
e
nha
nc
e
d
a
s
e
nha
nc
ing
numbe
r
o
f
da
ta
s
a
mpl
e
s
.
E
s
pe
c
ially,
f
o
r
QM
L
R
DPF
E
tec
hnique
is
mi
nim
ize
d
than
the
s
tudy
[
1]
,
[
2
]
.
L
e
t
us
a
s
s
ume
ini
ti
a
l
it
e
r
a
t
ion
with
10
,
000
da
ta
s
a
mpl
e
s
,
whe
r
e
f
or
QM
L
R
DPF
E
method
wa
s
li
ke
wis
e
,
ti
me
uti
li
z
a
ti
on
f
or
[
1
]
,
[
2]
r
e
s
pe
c
ti
ve
ly.
T
he
ob
taine
d
ove
r
a
ll
r
e
s
ult
s
of
QM
L
R
DPF
E
method
a
r
e
c
ompar
e
d
to
outcome
s
o
f
c
onve
nti
o
na
l
tec
hniques
.
T
he
c
ompar
is
on
outcome
s
de
not
e
s
whic
h
pe
r
f
or
manc
e
of
de
tec
ti
on
ti
me
us
ing
QM
L
R
DPF
E
tec
hnique
is
s
igni
f
ica
ntl
y
r
e
duc
e
d
by
13%
a
nd
7%
than
the
s
tudy
[
1]
,
[
2]
.
T
his
is
owing
to
QM
L
R
DPF
E
tec
hnique
pe
r
f
or
med
the
da
ta
p
r
e
pr
oc
e
s
s
ing
a
nd
f
e
a
tur
e
s
e
lec
ti
on
pr
oc
e
s
s
.
I
n
da
ta
pr
e
pr
oc
e
s
s
ing,
mi
s
s
ing
inf
or
mation
is
de
ter
mi
ne
d
by
a
pplyi
ng
q
ua
dr
a
ti
c
mul
ti
va
r
iate
l
inea
r
r
e
gr
e
s
s
ion
a
ppr
oa
c
h.
Da
ta
im
b
a
lanc
e
pr
oblem
a
ls
o
s
olved
thr
ough
the
a
da
pti
ve
Z
iggur
a
t
s
ynthetic
s
a
mpl
ing
tec
hnique
to
c
r
e
a
te
s
ynthetic
da
ta
s
a
mpl
e
s
.
T
he
tar
ge
t
f
e
a
tur
e
s
e
lec
ti
on
a
ls
o
m
ini
mi
z
e
s
ti
me
c
ons
umpt
ion
of
f
r
a
udulent
tr
a
ns
a
c
ti
on
de
tec
ti
on.
T
a
ble
3
de
notes
c
ompar
is
on
of
s
pa
c
e
c
o
mpl
e
xit
y
a
mong
the
f
oll
owing
a
lgor
it
hms
li
ke
QM
L
R
DPF
E
,
R
US+X
GB
oos
t
a
nd
C
S
O+
D
C
NN
.
S
a
mpl
e
da
ta
s
e
t
va
lues
r
a
nge
f
r
om
10
,
000
to
100
,
000.
I
t
s
hows
the
s
pa
c
e
c
ompl
e
xit
y
of
the
da
tas
e
t
va
lues
.
T
a
ble
2
.
C
ompar
is
on
of
d
e
tec
ti
on
ti
me
N
umbe
r
of
da
ta
s
a
mpl
e
s
D
e
te
c
ti
on t
im
e
(
ms
)
Q
M
L
R
D
P
F
E
R
U
S
+
X
G
B
oos
t
C
S
O
+
D
C
N
N
10000
43
53
48
20000
50
60
54
30000
63
72
69
40000
72
88
80
50000
80
91
85
60000
88.8
96
91.2
70000
93.1
101.5
96.6
80000
96.8
112
108
90000
106.2
118.8
112.5
100000
116
130
125
F
igur
e
5
.
P
e
r
f
or
manc
e
c
ompar
is
on
o
f
de
tec
ti
on
t
im
e
F
igur
e
6
de
picts
r
e
s
ult
ou
tcome
s
of
s
pa
c
e
c
omp
lexity
ve
r
s
us
number
of
da
ta
s
a
mpl
e
s
e
xtr
a
c
ted.
W
hil
e
the
number
o
f
inf
o
r
mation
s
a
mpl
e
s
incr
e
a
s
e
s
,
the
s
pa
c
e
c
ompl
e
xit
y
of
e
ve
r
y
thr
e
e
methods
g
r
a
dua
ll
y
incr
e
a
s
e
s
.
Nota
bly,
the
s
pa
c
e
c
ompl
e
xit
y
f
or
the
QM
L
R
DPF
E
method
is
s
igni
f
ica
ntl
y
mi
nim
ize
d
than
the
[
1]
,
[
2]
.
L
e
t's
c
ons
ider
the
r
e
s
ult
s
f
r
om
the
f
ir
s
t
it
e
r
a
ti
on
with
10,
000
da
ta
s
a
mpl
e
s
.
T
he
s
pa
c
e
c
ompl
e
xi
ty
f
or
the
QM
L
R
DPF
E
tec
hnique
wa
s
us
e
d
to
c
a
lcula
te
s
pa
c
e
c
ompl
e
xit
y
[
1]
,
[
2]
r
e
s
pe
c
ti
ve
ly.
S
ubs
e
que
ntl
y,
the
ove
r
a
ll
outcome
s
of
QM
L
R
DPF
E
tec
hnique
a
r
e
c
ompar
e
d
to
c
onve
nti
ona
l
tec
hniques
.
T
he
a
ve
r
a
ge
r
e
s
ult
s
de
mons
tr
a
te
that
pe
r
f
or
manc
e
o
f
s
pa
c
e
c
ompl
e
xit
y
is
mi
nim
ize
d
by
20%
a
nd
9%
than
the
e
xis
ti
ng
R
US+X
GB
oos
t
[
1]
a
nd
C
S
O+
DC
NN
[
2
]
,
r
e
s
pe
c
ti
ve
ly.
T
his
r
e
duc
ti
on
in
s
pa
c
e
c
ompl
e
xit
y
is
a
c
hieve
d
due
to
the
QM
L
R
DPF
E
tec
hniques
pe
r
f
or
ms
the
dim
e
n
s
ionalit
y
r
e
duc
ti
on
th
r
ough
S
oka
l
–
M
iche
ne
r
’
s
di
s
tr
ibut
e
d
pr
oxim
it
y
f
e
a
tur
e
e
nginee
r
ing
tec
hnique.
T
his
a
p
pr
oa
c
h
s
e
lec
ts
s
igni
f
ica
nt
f
e
a
tur
e
s
while
r
e
movi
ng
other
s
f
r
om
the
da
tas
e
t,
ther
e
by
mi
nim
izing
s
tor
a
ge
s
pa
c
e
in
big
da
ta
a
na
lys
is
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
689
-
699
698
T
a
ble
3.
C
ompar
is
on
of
s
pa
c
e
c
ompl
e
xit
y
N
umbe
r
of
da
ta
s
a
mpl
e
s
S
pa
c
e
c
ompl
e
xi
ty
(
k
B)
Q
M
L
R
D
P
F
E
R
U
S
+
X
G
B
oos
t
C
S
O
+
D
C
N
N
10000
320
420
380
20000
378
462
433
30000
433
510
485
40000
457
546
505
50000
501
612
532
60000
522
675
568
70000
548
724
610
80000
593
763
633
90000
635
812
687
100000
687
824
736
F
igur
e
6.
P
e
r
f
or
manc
e
c
ompar
is
on
o
f
s
pa
c
e
c
ompl
e
xit
y
5.
CONC
L
USI
ON
I
n
th
is
manus
c
r
ipt
,
a
ne
w
tec
hnique
c
a
ll
e
d
QM
L
R
DPF
E
is
de
s
igned
f
or
c
ybe
r
c
r
im
e
de
tec
ti
on
in
digi
tal
f
und
t
r
a
ns
a
c
ti
ons
.
QM
L
R
DPF
E
tec
hnique
include
s
da
ta
pr
e
pr
oc
e
s
s
ing
in
the
f
ir
s
t
s
tage
to
a
r
r
a
nge
the
da
tas
e
t
pr
ope
r
ly
by
f
il
li
ng
in
mi
s
s
ing
da
ta
be
f
o
r
e
uti
li
z
ing
ML
method.
F
oll
owing
thi
s
,
dim
e
n
s
ionalit
y
r
e
duc
ti
on
is
im
pleme
nted
by
S
oka
l
–
M
iche
ne
r
’
s
dis
tr
ibut
e
d
pr
oxim
it
y
f
e
a
tur
e
e
nginee
r
ing
tec
hn
ique
f
or
f
r
a
udulent
tr
a
ns
a
c
ti
on
de
tec
ti
on
with
h
igher
a
c
c
ur
a
c
y
a
nd
.
A
c
ompr
e
he
ns
ive
e
xpe
r
im
e
ntal
a
s
s
e
s
s
m
e
nt
is
pe
r
f
or
med
with
de
tec
ti
on
a
c
c
ur
a
c
y,
pr
e
c
is
ion,
de
tec
ti
on
ti
me,
a
nd
s
pa
c
e
c
ompl
e
xit
y.
T
he
a
na
lyze
d
r
e
s
ult
s
pr
ove
whic
h
QM
L
R
DPF
E
method
is
be
tt
e
r
than
c
onve
nti
ona
l
methods
in
a
c
hieving
higher
a
c
c
ur
a
c
y
a
nd
pr
e
c
is
ion.
I
n
a
ddit
ion
,
QM
L
R
DPF
E
method
pr
o
ve
s
mor
e
e
f
f
e
c
ti
ve
at
r
e
duc
ing
ti
me
ut
il
iza
ti
on
a
nd
s
pa
c
e
c
ompl
e
xit
y
f
or
f
r
a
udulent
tr
a
ns
a
c
ti
on
de
tec
ti
on
tha
n
the
c
onve
nti
ona
l
methods
.
RE
F
E
RE
NC
E
S
[
1]
P
.
H
a
je
k,
M
.
Z
.
A
be
di
n,
a
nd
U
.
S
iv
a
r
a
ja
h,
“
F
r
a
ud
de
te
c
ti
on
in
mobi
le
pa
yme
nt
s
ys
te
ms
us
in
g
a
n
X
G
B
oos
t
-
ba
s
e
d
f
r
a
me
w
o
r
k,”
I
nf
or
m
at
io
n Sy
s
te
m
s
F
r
ont
ie
r
s
, vol
. 25, no. 5, pp. 1985
–
2003,
O
c
t.
2023, doi:
10.1007/s
10796
-
022
-
10346
-
6.
[
2]
T
.
K
a
r
th
ik
e
ya
n,
M
.
G
ovi
nda
r
a
ja
n,
a
nd
V
.
V
ij
a
ya
kum
a
r
,
“
A
n
e
f
f
e
c
ti
ve
f
r
a
ud
de
te
c
ti
on
us
in
g
c
ompe
ti
ti
ve
s
w
a
r
m
opt
im
iz
a
t
io
n
ba
s
e
d de
e
p n
e
ur
a
l
ne
twor
k,”
M
e
as
ur
e
m
e
nt
:
Se
ns
o
r
s
, vol
. 27, J
u
n. 2023, doi:
10.1016/j
.me
a
s
e
n.2023.100793.
[
3]
S
.
K
.
H
a
s
he
mi
,
S
.
L
.
M
ir
ta
he
r
i,
a
nd
S
.
G
r
e
c
o,
“
F
r
a
ud
de
te
c
ti
on
in
ba
nki
ng
da
ta
by
ma
c
hi
ne
le
a
r
ni
ng
te
c
hni
que
s
,”
I
E
E
E
A
c
c
e
s
s
,
vol
. 11, pp. 3034
–
3043, 2023, doi:
10.1109/AC
C
E
S
S
.2022.323
2287.
[
4]
M
.
Â
.
L
.
M
or
e
ir
a
e
t
al
.
,
“
E
xpl
o
r
a
to
r
y
a
na
ly
s
is
a
nd
im
pl
e
me
nt
a
ti
on
of
ma
c
hi
ne
le
a
r
ni
ng
te
c
hni
que
s
f
or
pr
e
di
c
ti
ve
a
s
s
e
s
s
me
n
t
of
f
r
a
ud i
n ba
nki
ng s
ys
te
ms
,”
P
r
oc
e
di
a C
om
put
e
r
Sc
ie
nc
e
, vol
. 21
4, pp. 117
–
124, 2022, doi:
10.1016/j
.pr
oc
s
.2022.11.156.
[
5]
J
.
K
.
A
f
r
iy
ie
e
t
al
.
,
“
A
s
up
e
r
vi
s
e
d
ma
c
hi
ne
le
a
r
ni
ng
a
lg
or
it
h
m
f
or
de
te
c
ti
ng
a
nd
pr
e
di
c
ti
ng
f
r
a
ud
in
c
r
e
di
t
c
a
r
d
tr
a
ns
a
c
ti
o
ns
,”
D
e
c
is
io
n A
nal
y
ti
c
s
J
ou
r
nal
, vol
. 6, M
a
r
. 2023, doi:
10.1016/j
.da
jo
ur
.2023.100163.
[
6]
E
.
I
le
be
r
i,
Y
.
S
un,
a
nd
Z
.
W
a
ng,
“
A
ma
c
hi
ne
le
a
r
ni
ng
ba
s
e
d
c
r
e
di
t
c
a
r
d
f
r
a
ud
d
e
te
c
ti
on
u
s
in
g
th
e
G
A
a
lg
or
it
hm
f
or
f
e
a
tu
r
e
s
e
le
c
ti
on,”
J
ou
r
nal
of
B
ig
D
at
a
, vol
. 9, no. 1, De
c
. 2022, doi:
10.1186/s
40537
-
022
-
00573
-
8.
[
7]
M
.
E
.
L
oka
na
n,
“
P
r
e
di
c
ti
ng
mobi
le
mon
e
y
tr
a
ns
a
c
ti
on
f
r
a
ud
us
in
g
ma
c
hi
ne
le
a
r
ni
ng a
lg
or
it
hms
,”
A
ppl
ie
d
A
I
L
e
tt
e
r
s
,
vol
.
4,
n
o.
2,
A
pr
. 2023, doi:
10.1002/ail2.85.
[
8]
J
.
K
im
,
H
.
J
ung,
a
nd
W
.
K
im
,
“
S
e
que
nt
ia
l
pa
tt
e
r
n
mi
ni
ng
a
p
pr
oa
c
h
f
or
pe
r
s
ona
li
z
e
d
f
r
a
udul
e
nt
tr
a
n
s
a
c
ti
on
d
e
te
c
ti
on
in
on
li
ne
ba
nki
ng,”
Sus
ta
in
abi
li
ty
, vol
. 14, no. 15, Aug. 20
22, doi:
10.3390/s
u14159791.
[
9]
M
.
S
e
e
r
a
, C
.
P
.
L
im
,
A
.
K
um
a
r
,
L
.
D
ha
mot
ha
r
a
n, a
nd
K
.
H
.
T
a
n,
“
A
n
in
te
ll
ig
e
nt
p
a
yme
nt
c
a
r
d
f
r
a
ud
d
e
te
c
ti
on s
ys
te
m,”
A
nnal
s
of
O
pe
r
at
io
ns
R
e
s
e
a
r
c
h
, vol
. 334, no. 1
–
3, pp. 445
–
467, M
a
r
. 2024, doi:
10.1007/s
10479
-
021
-
04149
-
2.
[
10]
V
.
S
.
S
.
K
a
r
th
ik
,
A
.
M
is
hr
a
,
a
nd
U
.
S
.
R
e
ddy,
“
C
r
e
di
t
c
a
r
d
f
r
a
ud
de
t
e
c
ti
on
by
m
ode
l
li
ng
b
e
h
a
vi
our
p
a
tt
e
r
n
u
s
in
g
h
ybr
id
e
ns
e
mb
le
mode
l
,”
A
r
abi
an
J
ou
r
nal
f
o
r
S
c
ie
nc
e
an
d E
ngi
ne
e
r
i
ng
,
vol
.
47,
no. 2
, pp
. 19
87
–
1
997,
F
e
b
. 20
22,
doi
:
10.1
007/
s
13
369
-
0
21
-
061
4
7
-
9.
[
11]
H
.
Z
hou,
G
.
S
un,
S
.
F
u,
L
.
W
a
ng,
J
.
H
u,
a
nd
Y
.
G
a
o,
“
I
nt
e
r
ne
t
f
in
a
nc
ia
l
f
r
a
ud
de
te
c
ti
on
ba
s
e
d
on
a
di
s
tr
ib
ut
e
d
bi
g
da
ta
a
ppr
o
a
c
h
w
it
h N
ode
2ve
c
,”
I
E
E
E
A
c
c
e
s
s
, vol
. 9, pp. 43378
–
43386, 2021, doi:
10.1109/AC
C
E
S
S
.2021.3062467.
0
200
00
00
00
1000
c
e
c
e
i
e
r
s
e
s
M
L
R
D
P
F
E
RU
S
X
G
Boos
t
C
S
D
C
N
N
Evaluation Warning : The document was created with Spire.PDF for Python.