Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
13
,
No.
3
,
Ma
rc
h
201
9
, p
p.
919
~
926
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v1
3
.i
3
.pp
919
-
926
919
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Anomal
y
-
bas
ed i
ntrusion
detecto
r system
u
sing restric
ted
growin
g self o
rgan
izing m
ap
Tomi Y
ahy
a Christ
yaw
an,
Ah
m
ad
A
fif
S
upianto
, Wa
yan
Fir
daus
M
ah
mu
dy
Facul
t
y
of
Com
pute
r
Sc
ie
nc
e, Br
awij
a
y
a
Univ
ersi
t
y
,
Ma
la
ng
,
Indo
nesia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
hist
or
y:
Re
cei
ved
J
ul
06, 2
018
Re
vised
N
ov
1
1,
2018
Accepte
d
Dec
8,
2018
The
rap
id
dev
elopm
ent
of
int
ern
et
and
net
wor
k
te
chnol
og
y
fo
ll
owed
b
y
m
al
ic
ious
thr
ea
t
s
and
at
t
ac
ks
on
net
works
and
co
m
pute
rs.
Intrusi
on
det
e
ct
io
n
s
y
stem
(IDS
)
was
deve
lope
d
to
s
olv
e
th
at
prob
lem
s.
The
dev
el
op
m
ent
of
IDS
using
m
ac
hine
l
ea
rning
is
nee
d
e
d
for
cl
assif
y
ing
the
a
tt
a
cks.
On
e
m
et
hod
of
the
c
la
ss
ifi
c
at
i
o
n
is
Self
-
Organ
iz
ing
Map
(SO
M).
SO
M
abl
e
to
per
form
cl
assifi
ca
t
ion
an
d
visual
i
zation
i
n
le
a
rning
pro
ces
s
to
gai
n
n
ew
knowle
dge
.
How
eve
r,
the
SO
M
has
le
ss
eff
ic
ie
n
t
in
le
arn
ing
proc
ess
when
appl
ie
d
in
Big
Data
.
Thi
s
stud
y
proposes
Restri
ct
ed
Grow
ing
SO
M
m
et
hod
with
cl
usteri
n
g
ref
ere
n
ce
ve
ct
or
(RGS
O
M
-
CRV
)
and
Para
ll
e
l
RGS
OM
-
C
RV
t
o
improve
SO
M
eff
ic
ie
n
c
y
in
c
la
ss
ifica
t
io
n
with
a
c
cur
acy
conside
r
at
ion
t
o
solve
Big
Data
proble
m
.
G
rowing
proc
ess
in
RGS
O
M
is
res
tri
cted
b
y
m
axim
um
nodes
and
growing
th
reshold,
the
r
eu
pdat
e
w
ei
ght
pr
oce
ss
will
upd
a
te
unused
ref
ere
n
ce
ve
ct
or
when
m
ap
size
al
rea
d
y
m
axi
m
um
,
the
se
two
proc
esses
solve
the
cons
um
ing
ti
m
e
of
reg
ular
GS
O
M.
From
the
result
s
of
t
his
rese
arc
h
aga
inst
KD
D
Cup
1999
dat
ase
t
,
proposed
m
ethod
Para
llel
R
GS
OM
-
CR
V
abl
e
to
give
91.
86%
ac
cur
acy
,
20.
58%
fal
s
e
alarm
rat
e
,
95.
32
%
rec
a
ll
o
r
det
e
ct
ion
rate,
a
nd
pre
c
ision
is
94.
35%
and
ti
m
e
consum
in
g
is
outpe
rform
tha
n
r
egul
ar
Gr
owing
SO
M.
T
his
proposed
m
et
hod
is
v
er
y
p
rom
ising
to
handl
e
big
data problems
compare
d
with
oth
er
m
et
hods.
Ke
yw
or
ds:
Bi
g
data
Cl
us
te
rin
g
re
fe
ren
ce
v
ect
or
Grow
i
ng s
om
In
tr
us
i
on D
et
e
ct
or
Syst
e
m
Self
-
orga
nizin
g
m
ap
Copyrig
ht
©
201
9
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed.
Corres
pond
in
g
Aut
h
or
:
Tom
i Yah
ya
Chr
ist
ya
wa
n
,
Faculty
of Com
pu
te
r
Scie
nc
e,
Brawijaya
U
niv
ersit
y,
Vetera
n
street
,
Ma
la
ng
,
Indo
ne
sia
.
Em
a
il
:
to
m
i@rek
avisit
am
a.n
et
1.
INTROD
U
CTION
The
sec
ur
it
y
threat
to
i
nter
ne
t
us
age
a
nd
c
om
pu
te
r
net
w
orks
is
inc
reas
ing
.
Seve
ral
ty
pes
of
new
at
ta
cks
on
t
he
netw
ork
ap
pea
r
pe
rio
dical
ly
,
this
m
ake
a
chalang
e
to
deve
lop
a
fle
xib
le
a
nd
a
da
ptive
ne
twork
secur
it
y.
De
vel
op
i
ng
te
ch
ni
ques
to
detect
ano
m
aly
-
b
ased
ne
twork
intr
us
i
on
to
protect
a
com
pu
te
r
syst
em
and
netw
ork
from
m
al
ic
iou
s
act
ivit
y
at
ta
cks
cal
l
In
tr
us
i
on
Dete
ct
ion
Syst
em
(I
DS),
as
the
de
te
ct
ion
of
s
usp
ic
iou
s
netw
ork
tra
ff
ic
and c
om
pu
te
r usa
ge
ca
n no
t
be done
by c
onven
ti
onal
fire
w
al
ls.
So
m
e
dev
el
op
m
ent
of
I
DS
ba
sed
on
m
achine
le
ar
ning
te
c
hn
i
qu
e
.
T
he
m
et
hods
us
e
d
for
a
no
m
al
y
-
base
d
intr
us
i
on
detect
ion
are
com
m
on
ly
dif
fer
e
ntiat
ed
int
o
cl
assifi
cat
io
n
a
nd
cl
us
te
rin
g.
H
ow
e
ve
r,
t
her
e
is
al
so
a
hy
br
i
dizat
ion
betwee
n
cl
us
te
rin
g
a
nd
cl
assifi
cat
ion
f
or
i
ntr
us
i
on
de
te
ct
ion
syst
em
.
In
t
he
cl
assifi
cat
ion
m
et
ho
d,
so
m
e
stud
ie
s
us
e
si
ngle
cl
assifi
er
s
uch
as
KNN
[
1]
,
S
upport
Ve
ct
or
Ma
c
hin
e
(
SV
M)
[
2]
,
arti
fici
al
neural
net
wor
k
[3
-
6]
to
s
olv
e
ID
S
pro
blem
.
Othe
r
resea
rc
he
rs
us
e
hybr
i
d
m
et
ho
ds
of
he
ur
ist
ic
al
gorith
m
with
cl
ass
ifie
r
m
e
tho
d
[7
-
9]
,
Mult
i
-
le
vel
SV
M
a
nd
Ext
rem
e
Lea
rn
i
ng
Ma
c
hin
e
with
K
-
Me
an
[10]
,
Decisi
on
Tree
and
S
VM
[
11
]
,
Tree
A
ugm
e
nted
Naï
ve
Ba
ye
s
(TAN)
an
d
Re
duced
E
rror
P
runin
g
(R
EP)
[
12]
.
IDS
-
relat
e
d
stud
ie
s
us
i
ng c
lusterin
g
in
cl
ude
K
-
Me
a
n,
K
-
Me
do
i
ds
, A
-
S
PO
T
[
13
]
, C
A
NN
[
14
]
.
On
e
m
et
ho
d
of
cl
assifi
cat
ion
and
re
duct
io
n
d
at
a
that
can
visu
al
iz
e
the
le
arn
i
ng
proces
s
is
SO
M.
I
n
so
m
e
stud
ie
s
SO
M
able
to
so
lve
the
pro
blem
of
cl
assif
ic
at
ion
with
bette
r
res
ult
[
15
]
.
Howe
ver,
SO
M
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
13
, N
o.
3
,
Ma
rc
h 201
9
:
919
–
926
920
exp
e
rience
d
c
on
st
raints
in
e
ff
ic
ie
ncy
du
rin
g
the
process
of
le
ar
ning
with
la
rg
e
data
(
Bi
g
Data),
t
his
have
high
ti
m
e
con
s
um
ing
.
T
he
prob
le
m
is
du
e
to
the
c
har
act
er
ist
ic
s
of
SO
M
that
cal
culat
e
the
distance
bet
we
e
n
input
vecto
r
and
re
fe
ren
ce
vecto
r
to
deci
de
the
winnin
g
ne
uro
n
on
the
hidden
la
ye
r
fo
r
eac
h
epo
c
h.
Bi
g
data
pr
ob
l
e
m
s
see
m
to
be
face
d
by
L
i
at
al
.
[1]
,
th
ey
j
us
t
sel
ect
5,552
in
sta
nce
s
from
KD
D
Cup
99
sam
ple
data
as
the
trai
ning
da
ta
,
an
d
5,552
instances
a
s
te
sti
ng
data
f
or
their
e
xperim
ent,
not
f
ro
m
the
entire
KDD
C
up 99 d
at
a.
Ba
sed
on
rese
arch
c
onduct
e
d
by
Alaha
koon
[
16]
,
G
r
ow
i
ng
Self
Orga
nizing
Ma
p
(
GSOM)
ca
n
be
us
e
d
to
buil
d r
efere
nce
vecto
r
g
ra
dual
ly
b
ase
d on trai
ning
da
ta
. How
e
ve
r,
t
his ab
le
t
o
s
olve
the
prob
l
em
i
n
the
first
ep
oc
h
on
ly
,
in
nex
t
e
poch
,
to
po
l
og
ic
al
m
ap
will
be
la
rg
er
,
an
d
t
he
pr
ob
le
m
of
tim
e
con
su
m
e
will
reappear
on
ne
xt
ep
oc
h.
Th
is
stud
y
pro
poses
Re
st
rict
ed
Gro
wing
Se
lf
O
rg
a
nizin
g
Ma
p
with
cl
ust
eri
ng
ref
e
ren
ce
vector
(R
GSOM
-
C
RV)
an
d
Pa
rall
el
RGS
OM
-
C
RV
to
so
lve
t
he
issue
on
regular
G
SO
M
to
handle
big
data
pro
blem
s.
This
resea
rch
will
m
easur
e
RG
SO
M
-
CR
V
an
d
Parall
el
RGSO
M
-
C
RV
ef
fici
ency
of
ti
m
e
consum
ing
a
nd accu
racy, f
al
se ala
rm
r
at
e, p
r
eci
sion
,
and
re
cal
l.
2.
RESEA
R
CH MET
HO
D
This
resea
rch
proce
dure
sho
wn
in
Fi
gure
1.
Feat
ures
f
r
om
dataset
wi
ll
be
sel
ect
ed
accor
ding
to
sel
ect
ed
featu
r
es,
an
d
eac
h
fe
at
ur
e
value
will
be
no
rm
alizi
ng
befor
e
proce
ss
at
RGS
OM
-
CR
V.
D
at
a
trai
n
a
nd
data te
sti
ng w
il
l be treat s
am
e
way b
e
f
or
e
pro
cess in R
GSO
M
-
CR
V.
Figure
1
.
Rese
arch p
ro
ce
dure
2.1.
K
D
D Cu
p 99 D
ataset
Dataset
KDD
Cup
99
f
ro
m
UCI
sepa
rate
with
10%
data
trai
n
(4
94,
02
1
instances
)
and
data
te
st
(4,89
8,431
i
ns
ta
nces),
it
’s
ha
ve
41
feat
ur
es
.
This
dataset
c
at
egory
in
4
at
ta
cks
cl
ass
dos
,
pro
be,
r
2l,
u2
r
an
d
norm
al
cat
ego
r
y.
This
re
searc
h
will
us
e
al
l
da
ta
prov
i
ded
by
this
dataset
to
m
easur
e
the eff
ic
ie
nt
a
nd
e
f
fecti
ve
of prop
os
e
d
m
et
hod
to
h
a
ndl
e b
ig
d
at
a
pro
bl
e
m
s.
2.2.
Sele
c
ted Fe
ature
Data
trai
n
w
il
l
be
pr
oces
ses
with
sel
ect
ed
featur
e
s
.
Accord
i
ng
to
KDD
Cu
p
99
ta
sk
(h
tt
p://
kdd.i
cs.
uci.edu/data
ba
ses/kddc
up99/t
ask.htm
l)
wh
ic
h
a
dap
te
d
fro
m
Stolfo
et
al
pa
per
[
17]
,
th
ere
a
re
three
ty
pes
fe
at
ur
e
sel
ect
io
n,
basic
feat
ures
of
i
nd
i
vidual
TCP
co
nnect
ion,
c
on
te
nt
featu
res
within
a
connecti
on
s
uggested
by
dom
ai
n
knowle
dge
,
an
d
tra
ff
ic
fe
at
ur
e
c
om
pu
te
d
usi
ng
a
tw
o
-
seco
nd
ti
m
e
window
.
In
this
resea
rc
h
basic
featu
re
s
of
i
nd
i
vidual
TCP
c
onnecti
on
cat
eg
or
y
ty
pe
be
c
hose
n.
Feat
ur
es
us
ed
in
this
ty
pe
are
durati
on, pr
oto
c
ol ty
pe,
ser
vice,
flag,
s
rc
byte
s, ds
t byt
es, lan
d, w
ron
g
f
ra
gm
ent, urg
e
nt.
2.3.
No
r
mali
z
e Feat
ure
Af
te
r
dataset
s
el
ect
ed
by
basi
c
featur
e
ty
pe,
and
the
n
this
f
eat
ur
e
will
be
norm
al
izing
wi
th
(1)
befor
e
processe
d
wit
h
RGSO
M
.
Where
n
is
norm
alized
value
,
x
is
real
value
,
ar
gm
in
(F
i)
is
sm
al
le
st
value
fro
m
i
-
th
featur
e
s,
a
r
gma
x
(F
i)
is
la
r
ge
st
value
f
ro
m
i
-
th
featu
res.
A
fter
norm
al
iz
e
data
will
proce
ss
in
RGS
OM
-
CR
V
to b
e
train
and
then
te
st
with
da
ta
test
.
n
=
x
−
arg
m
in
(
Fi
)
arg
m
ax
(
Fi
)
−
arg
m
in
(
Fi
)
(1)
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
250
2
-
4752
An
omaly
-
base
d
intr
us
io
n det
ect
or
syste
m us
ing
restri
ct
e…
(Tom
i Y
ahy
a
C
hr
ist
yaw
an
)
921
2.4.
G
SO
M
Grow
i
ng
S
O
M
is
m
od
ifie
d
regular
S
OM
w
hich
pro
pos
ed
by
Ala
hakoon
[16]
.
GSOM
pr
ocedure
accor
ding
to
A
la
hakoon
are
I
niti
al
iz
at
ion
phase,
Gro
wing
ph
a
se,
an
d
Sm
oo
t
hing
phase.
Figure
2
sho
w
ne
w
node
ge
ner
at
io
n
from
the
bo
unda
ry
of
t
he
netw
ork.
Fig
ure
2a
is
init
ia
l
node
us
i
ng
4
nodes
as
ref
e
r
ence
vecto
r.
Fig
ur
e
2b
s
how
that
hi
gh
er
ror
occ
urs
wh
e
n
winne
r
node
distance
with
ref
e
re
n
ce
vecto
r
is
m
or
e
than
grow
i
ng
thres
ho
l
d.
Fig
ur
e
2c
sh
ow
the
grow
i
ng
s
hcem
a
of
GSOM.
G
rowin
g
phase
will
occu
r
if
gro
wi
ng
thres
ho
l
d
sm
al
l
er th
a
n dist
anc
e of
winner
no
de.
Figure
2. Ne
w node
ge
ner
at
io
n from
the b
ou
nd
a
ry
of
t
he ne
twork
2.5
.
RGSOM
-
C
RV
This
researc
h
pro
po
se
RG
S
OM
-
CR
V
,
this
m
et
ho
d
is
extend
f
r
om
reg
ul
ar
SO
M
[15]
and
re
gula
r
GSOM
[16]
.
T
he
dif
fer
e
nt
wi
th
regular
GSOM
is
the
gro
wing
of
ref
e
re
nce
vect
or
is
r
est
rict
ed
by
m
axim
u
m
siz
e
of
node
s
(
MN),
a
nd
tw
o
gr
i
d
m
ap
di
m
e
ns
io
ns
.
Ma
p
w
il
l
be
gr
owin
g
if
distance
fro
m
winn
in
g
node
with
input
node
is
m
or
e
that
gr
owin
g
th
reshold
(G
T
).
If
th
e
le
ng
t
h
of
nodes
i
n
m
ap
m
or
e
th
at
MN,
m
ap
will
stop
grow
i
ng
a
nd
r
eupdate
by
sel
ect
ed
ra
ndom
l
y
on
e
re
fer
e
nc
e
vecto
r
w
hich
nev
e
r
h
it
(
not
be
wi
nn
e
r
ye
t)
by
input.
Af
te
r
s
el
ect
winn
e
r
node
,
sam
e
with
re
gula
r
S
O
M,
weig
ht
of
neig
hborh
ood
will
be
up
date
.
O
ne
ref
e
ren
ce
vect
or
c
ould
be
m
or
e
t
ha
n
one
ti
m
e
sel
ect
ed
to
be
winner
no
de
cal
le
d
cl
us
t
erin
g
re
fer
e
nce
vecto
r
(CRV)
.
Cl
us
te
rin
g
re
fer
e
nce
vecto
r
will
be
a
gro
up
of
in
put
with
sim
il
ar
it
y
weigh
t
acc
ordin
g
to
m
ini
m
u
m
GT.
CR
V
m
eth
od
will
reduc
e
the
siz
e
of
t
opog
raphical
m
ap,
t
his
m
et
ho
d
ca
n
de
crease
tim
e
con
su
m
i
ng
w
he
n
sel
ect
w
in
ner
node
pr
ocess.
Figure
3
is
t
he
flo
wc
har
t
of
RGS
OM
proc
edure
,
an
d
Fig
ur
e
3a
is
the
whole
process
of
RG
SO
M
proce
dure.
I
n
i
niti
al
iz
at
ion
process
us
e
r
nee
d
set
ti
ng
the
m
axim
u
m
m
ap
siz
e
and
m
axi
m
um
no
de
f
or
th
e
m
ap
this
propose
f
or
rest
rict
ed
the
siz
e
of
the
gr
owin
g
no
de
in
the
m
ap.
User
al
so
set
ti
ng
the
value
of
sta
rt
le
arn
in
g
rate
and
sto
p
le
a
r
ning
rate,
sta
r
t
gro
wing
t
hresh
old
,
sto
p
gro
wing
t
hr
es
hold,
a
nd
m
axEpoch.
In
it
ia
li
zat
ion
pro
ces
s also
g
e
ne
rate i
niti
al
n
in
e ref
e
re
nce
vec
tor n
od
es
for t
he
m
ap,
as sho
w
n on Fig
ure
4a.
(
)
=
(
)
(2)
=
{
‖
−
‖
}
(3)
(
+
1
)
=
(
)
+
ℎ
(
(
)
−
(
)
)
(4)
ℎ
=
(
)
−
(
‖
−
‖
2
2
(
)
)
(5)
Figure
3b
is
tr
ai
nin
g
process
flow
c
har
t
.
U
pdat
e
le
arn
in
g
r
at
e
and
up
date
grow
i
ng
th
res
ho
l
d
us
i
ng
m
on
oto
nic
de
crem
ent
fu
nction,
fig
ur
e
ou
t
by
(2
)
,
w
here
S(t)
is
the
updated
le
ar
ni
ng
r
at
e
or
gr
ow
i
ng
thres
ho
l
d,
is
sta
rting
value,
i
s
en
ding
val
ue
,
t
is
curre
nt
epo
c
h,
is
m
ax
epo
c
h.
Lear
ning
rat
e
and
gro
wing
thres
hold
updat
e
us
ed
to
set
c
urren
t
le
ar
ning
rate
and
gro
w
ing
th
res
ho
l
d
at
cur
re
nt
ep
oc
h.
T
he
winner
node
c
al
culat
e
us
e
re
gu
la
r
S
OM
pr
ocedu
re
f
ollo
w
by
(
3).
Wh
e
re
,
win
is
wi
nn
e
r
no
de,
x
is
th
e
input
vecto
r
an
d
is
k
-
th
re
fe
r
ence
vecto
r.
Re
fe
re
nce
ve
ct
or
a
re
li
st
of
no
des
ge
ner
at
e
by
first
node
(n
i
ne
s
quar
e
node
w
hic
h
po
sit
ion
is
in
the
center
of
m
ap
Figure
4a
)
an
d
gen
e
rate
by
gr
ow
i
ng
functi
on
.
If
winner
distance
m
or
e
than
the
grow
i
ng
th
res
ho
l
d
in
the
cu
rr
e
nt
epoch
a
nd
ge
ner
at
e
d
node
s
le
ng
t
h
s
m
al
le
r
than
m
axim
u
m
node,
ref
e
ren
c
e
vector
will
gr
ow
with
r
an
dom
weigh
t.
If
gen
e
rated
node
s
le
ng
th
m
or
e
than
m
axi
m
u
m
no
de
,
unus
e
d
ref
e
rence
vecto
r
will
be
re
updated
.
U
nu
se
d
re
fere
nce
vecto
r
a
r
e
node
s
that
ne
ver
hit
or
no
t
ye
t
sel
ect
ed
to
b
e
w
in
ner.
T
he
sc
hem
e
of
t
he
grow
i
ng o
f
R
GSOM
s
how
n
on Figure 4
c.
N
ei
ghbor
hood of w
in
ning
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
13
, N
o.
3
,
Ma
rc
h 201
9
:
919
–
926
922
neur
on
will
be
update
d
us
e
(
4),
w
he
re
ℎ
is
ga
us
ia
n
nei
ghbo
rho
od
f
un
ct
io
n
decide
by
(
5)
.
Lear
ning
rate
no
ta
ti
on
is
(
)
and
σ(
t
)
is
neighbor
hood
ra
dius,
bo
th
a
re
a
m
on
oto
nical
ly
decr
easi
ng
sc
al
ar
functi
on
of
t
fo
ll
ow
the
r
ule
of
(
2),
is
k
-
t
h
neig
hborh
ood of win
ner node
.
Sele
ct
e
d
winner
n
ode
will
be
pus
h
to
the
li
st
of
us
e
d
nodes
.
Figure
3c
is
te
stin
g
procedu
re
f
lowc
har
t.
At
te
sti
ng
pur
pose,
the
winne
r
will
be
cho
se
n
f
rom
m
ap
gen
e
rate
by
tra
ining
ste
p
an
d
us
e
d
node
s
will
be
sel
ect
ed
to
be
ref
e
re
nce
ve
ct
or
f
or
te
sti
ng
in
pu
t.
T
he
w
inn
e
r
node
will
be
use
d
t
o
cal
culat
e
Tr
ue
Ne
gativ
e
(T
N),
T
ru
e
P
os
it
ive
(TP),
F
al
se
Ne
gative
(
FN
)
,
False
P
osi
ti
ve
(F
P
)
val
ues.
T
ru
e
ne
gative
is
correct
predic
ti
on
w
hich
rea
l
cat
ego
ry
is
norm
al
and
la
be
le
d
as
norm
al
.
Tru
e
po
sit
ive
is
c
orrect
pre
dicti
on
w
hich
real
at
t
ack
cat
e
gory
a
nd
la
bele
d
as
at
ta
ck.
False
ne
gative
is
inc
orrect
pr
e
dicti
on
w
hich
the
real
cat
egory
is
at
ta
ck
bu
t
pr
e
dicte
d
as
norm
al
.
False
po
sit
ive
is
inco
rr
ect
pre
dicti
on
wh
ic
h
the
r
eal
cat
egory is n
or
m
al
b
ut labele
d
as
att
ack.
Figure
3. Flo
w
char
t R
GSOM
-
CR
V
Proce
dur
e.
Figure
4. S
qu
a
re
grow
i
ng no
de
sch
em
a
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
250
2
-
4752
An
omaly
-
base
d
intr
us
io
n det
ect
or
syste
m us
ing
restri
ct
e…
(Tom
i Y
ahy
a
C
hr
ist
yaw
an
)
923
2.6. Me
asure
ment
This
e
xperim
e
nt
has
four
m
e
asur
em
ent
s
to
evaluate
t
his
m
et
ho
d,
acc
uracy
(A
CC
),
fal
se
al
arm
rate
(F
AR
),
detect
ion
rate
(DTR
)
or
recall
,
pr
e
ci
sion
.
Accura
cy
is
total
cor
rectl
y
cl
assifi
e
d
exam
ple
to
total
nu
m
ber
of exa
m
ple. A
ccur
ac
y ca
lc
ulate
w
it
h (6):
ACC
=
TP
+
TN
TP
+
TN
+
FP
+
FN
x
100%
(6)
False
al
arm
rate
is
pe
rcen
ta
ge
of
no
rm
al
categ
ory
that
ha
ve
bee
n
la
bel
a
s
at
ta
ck
to
t
ota
l
nu
m
ber
of
norm
a
l
exam
ples, and c
al
culat
e as (7):
FAR
=
FP
TN
+
FP
x
100%
(7)
Detect
ion
rate
or r
ecal
l i
s sta
ndin
g for c
orrec
tl
y l
abel as att
acks t
o
total
num
ber
o
f
at
ta
c
ks
, use (8
)
as
form
ula.
D
TR
or
Recal
l
=
TP
TP
+
FN
x
100%
(8)
Pr
eci
sio
n
is
be
ing
the
per
ce
nt
age
of
co
rr
ect
l
y
la
bel
at
ta
cks
as
at
ta
cks
to
total
nu
m
ber
of
instance
la
bel
ed
as
at
ta
cks,
a
nd calc
ulate
w
it
h (
9)
:
Preci
sion
=
TP
TP
+
FP
x
100%
(9)
3.
RESU
LT
S
AND DI
SCU
S
S
ION
RGSO
M
-
CR
V
and
Pa
rall
el
RGS
OM
-
CR
V
i
n
this
ex
per
im
ent
will
init
ia
lize
with
sam
e
s
ta
rt
le
arn
in
g
rate
(LRsta
rt
),
stop
le
arn
i
ng
rate
(LRst
op),
sta
rt
gro
wing
thre
shold
(G
Tsta
rt),
sto
p
grow
i
ng
thr
esh
ol
d
(G
Tst
op),
m
ax
ep
och,
m
ap
siz
e.
T
his
ex
pe
rim
ent
us
es
LR
sta
rt=0.9,
LRst
op=0
.1,
GTstart
=0.05,
GTstop
=0.01,
m
axEpoch
=
4,
m
apSize=100
x100. Ma
xim
u
m
n
od
e f
or RG
SO
M
-
CR
V
is 500
0
an
d
m
axi
m
u
m
n
od
e
for
Parall
el
RGSO
M
-
CR
V
dif
fer
e
nt
bet
ween
prot
oc
ol,
U
dp
ha
ve
m
axim
u
m
no
des
3000,
an
d
tc
p
a
nd
ic
m
p
is
5000
,
diff
e
re
nt
val
ue
of
this
s
et
ti
ng
is
beca
us
e
ud
p
at
trai
ning
in
sta
nce
m
uch
f
ewer
that
tc
p
a
nd
ic
m
p
as
show
n
at
Table
1.
This
exp
e
rim
ent
us
ing
com
pu
te
r
w
it
h
sp
eci
ficat
io
n:
P
ro
ces
sor
Int
el
Core
i
7
-
6500U
CP
U
@
2.5
G
H
z
2.60 G
Hz, M
e
m
or
y 8 G
B,
w
i
th syst
em
6
4
-
bi
t Op
e
rati
ng Sy
stem
.
Tab
le
2
s
how
the
regular
GSOM
co
nsum
e
m
or
e
than
2
da
ys
and
e
nd
e
d
in
seco
nd
e
po
ch
beca
us
e
hav
e
m
e
m
or
y
lim
it
issue
du
e
the
m
ap
gr
ow
ing
bigger
.
Fro
m
this
exp
eri
m
ent,
GSOM
no
t
ca
pab
le
t
o
handle
la
rg
e
data
in
our
e
xp
e
rim
ent
du
e
the
li
m
i
ta
tio
n
of
m
e
m
or
y
and
ti
m
e
con
sum
ing
and
t
his
is
no
t
w
horted
t
o
be
con
ti
nue
d.
Ta
ble
1.
Distri
bu
ti
on
of P
ro
t
oc
ol Type
Proto
co
l T
y
p
e
Co
u
n
t
ic
m
p
2
8
3
.602
tcp
1
9
0
.065
udp
2
0
.35
4
Table
2.
GSO
M E
xperim
ent
ACC
FAR
DTR/Recall
Precisio
n
Tr
ain
in
g
T
i
m
e
ep
o
ch
1
-
-
-
-
0
2
:3
2
:
1
1
ep
o
ch
2
-
-
-
-
More than
2 d
a
y
s a
n
d
then
sto
p
p
ed
The
resu
lt
fro
m
five
exp
e
rim
ents
of
Paral
le
l
RGSO
M
-
C
RV
s
how
in
T
able
4.
T
he
a
ve
rag
e
Parall
el
RGSO
M
-
CR
V
accu
racy
is
91.
86%
an
d
fal
se
al
arm
rate
is
20.
58%,
reca
ll
or
detec
ti
on
rate
is
95.32
%,
an
d
pr
eci
sio
n
is
94.35%.
T
he
a
ve
rag
e
ti
m
e
con
s
um
e
wh
il
e
trai
ning
usi
ng
Pa
r
al
le
l
RGSO
M
-
CR
V
is
6
hour
s
33
m
inu
te
and
18
seco
nd (four
e
po
c
h).
H
ow
e
ve
r,
ti
m
e con
su
m
ing
f
or
test
in
g
i
s 46
m
inu
te
s a
nd 39 se
co
nd.
Fr
om
Table
3
and
Ta
ble
4
Pa
ral
le
l
RGSO
M
-
CR
V
is
ou
t
pe
rfor
m
than
R
G
SO
M
-
CR
V
wit
h
91.86%
i
n
accuracy,
false
al
arm
rate
is
20.58%
,
95.32
%
f
or
recall
,
a
nd
94.
35%
i
n
pr
eci
sio
n.
H
oweve
r,
preci
sion
f
ro
m
bo
t
h
ex
per
im
ent
ha
ve
good
r
esult.
From
Ta
ble
3
at
thir
d
exp
e
rim
ent,
accuracy
of
RG
S
OM
-
CR
V
ha
ve
good
resu
lt
tha
n
othe
r
e
xp
e
rim
ent.
This
ca
n
be
ha
pp
e
ne
d
becaus
e
RGS
OM
-
CR
V
is
gen
e
rate
r
andom
ly
and
in
eac
h
exp
e
rim
ent,
so
differe
nt
res
ul
t
m
ay
be
ob
ta
ined.
RGS
OM
-
CR
V
m
ay
be
cou
l
d
hav
e
dif
fer
e
nt
res
ult
to
o
f
or
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
13
, N
o.
3
,
Ma
rc
h 201
9
:
919
–
926
924
diff
e
re
nt
G
TSt
art
an
d
GTst
op
set
ti
ng
.
T
he
r
es
ult
f
ro
m
the
fifth
e
xperim
e
nt
usi
ng
Pa
rall
el
RGS
OM
-
C
RV
f
or
each p
r
oto
c
ol
s
how
n
at
Ta
ble
5.
T
he
par
al
le
l R
GS
OM
-
CR
V
ha
ve
total
acc
ur
acy
is 97.27
%,
false
al
arm
rate
is
12.72%
,
recall
or
detect
io
n
r
at
e
is
99.79%
,
an
d
pr
eci
sio
n
is
96.87%
.
Ic
m
p
pr
ot
oco
l
ha
ve
la
r
gest
acc
ur
at
io
n,
false al
arm
r
at
e, d
et
ect
io
n rate
and lar
gest
preci
ssion
.
Five
ex
per
im
e
nts
for
both
m
et
hods
will
be
evaluated
.
Ta
ble
3
sho
w
th
e
resu
lt
of
fiv
e
exp
e
rim
ent
us
in
g
R
GSOM
-
CR
V.
The
av
erag
e
RGS
OM
-
CR
V
acc
ur
ac
y
is
51.
45
%
an
d
false
al
arm
rate
is
11.
80
%
,
recal
l
or
detect
io
n
ra
te
is
41
.78%,
a
nd
pr
eci
sio
n
is
93
.
09%.
T
he
aver
a
ge
of
tim
e
con
s
u
m
e
trai
ning
us
in
g
RG
SO
M
-
CR
V
is
5
ho
urs
and
31
m
inu
te
s
and
1
sec
on
d.
Tim
e
con
sum
e
wh
il
e
te
stin
g
is
1
ho
ur
a
nd
44
m
inu
te
s
and
59
seco
nd.
Table
3.
R
GSOM
-
CR
V
Exp
erim
ents R
esult
ACC
FAR
DTR/Recall
Precisio
n
Tr
ain
in
g
T
i
m
e
Testin
g
T
i
m
e
Exp
eri
m
en
t 1
3
8
.22
%
1
8
.08
%
2
4
.52
%
8
1
.21
%
0
5
:1
7
:
2
4
0
1
:3
9
:
0
5
Exp
eri
m
en
t
2
3
8
.62
%
1
8
.69
%
2
7
.94
%
8
5
.66
%
0
5
:2
5
:
4
2
01
:3
2
:
5
6
Exp
eri
m
en
t
3
9
7
.74
%
9
.71
%
9
9
.61
%
9
7
.60
%
0
5
:3
4
:
3
0
0
1
:4
8
:
1
2
Exp
eri
m
en
t
4
4
1
.12
%
7
.48
%
2
8
.25
%
9
3
.78
%
0
5
:4
3
:
0
3
0
1
:2
7
:
2
6
Exp
eri
m
en
t
5
4
1
.54
%
3
.87
%
2
7
.85
%
9
6
.64
%
0
5
:3
7
:
0
1
0
2
:1
7
:
1
6
Av
erage
5
1
.45
%
1
1
.80
%
4
1
.78
%
9
3
.09
%
0
5
:3
1
:
3
2
0
1
:4
4
:
5
9
Table
4.
Parall
el
RGSO
M
-
C
RV E
xp
e
rim
ents R
esult
ACC
FAR
DTR/Recall
Precisio
n
Tr
ain
in
g
T
i
m
e
Testin
g
T
i
m
e
Exp
eri
m
en
t
1
9
7
.54
%
1
1
.02
%
9
9
.72
%
9
7
.26
%
0
8
:3
2
:
1
1
0
0
:5
2
:
4
8
Exp
eri
m
en
t
2
9
4
.79
%
1
9
.93
%
9
9
.48
%
9
4
.01
%
0
6
:1
5
:
5
9
0
0
:4
2
:
0
2
Exp
eri
m
en
t
3
7
8
.51
%
1
7
.52
%
7
7
.26
%
9
3
.34
%
0
5
:5
1
:
2
2
0
0
:5
7
:
0
7
Exp
eri
m
en
t
4
9
1
.21
%
4
2
.68
%
9
9
.69
%
9
0
.32
%
0
6
:1
8
:
4
6
0
0
:4
7
:
3
8
Exp
eri
m
en
t
5
9
7
.27
%
1
2
.72
%
99.
79%
9
6
.87
%
0
5
:4
8
:
1
6
0
0
:3
3
:
4
2
Av
erage
9
1
.86
%
2
0
.58
%
9
5
.32
%
9
4
.35
%
0
6
:3
3
:
1
8
0
0
:4
6
:
3
9
Table
5.
Parall
el
RGSO
-
CR
V
Re
su
lt
Eac
h
P
ro
t
oco
l at
Ex
pe
rim
ent 5
Proto
co
l
ACC
FAR
DTR/Recall
Precisio
n
ic
m
p
9
9
.74
%
5
6
.80
%
1
0
0
.00
%
9
9
.74
%
tcp
9
3
.39
%
1
4
.97
%
9
9
.44
%
9
0
.17
%
udp
9
8
.48
%
0
.52
%
3
3
.41
%
4
9
.60
%
Total
9
7
.27
%
1
2
.72
%
9
9
.79
%
9
6
.87
%
Fo
r
m
or
e
detai
l
insigh
t
we
can
stu
dy
w
it
h
the
m
ap
gen
e
rated
i
n
each
ep
oc
h,
wh
ic
h
sho
w
n
at
Figure
5
-
7.
Fr
om
the
vis
ua
li
zat
ion
show
n
at
Figure
5
-
7
t
her
e
a
re
ne
w
knowle
dge
of
in
form
ation
a
bout
th
e
tr
ai
ning
proce
s
s
of
Pa
rall
el
RGS
OM.
Fig
ure
5
at
udp
proto
col
show
n
that
nodes
wh
ic
h
s
epar
at
e
ra
ndom
ly
is
nodes
t
hat
ge
ne
rated
by
re
update
unuse
d
re
fer
e
nce
ve
ct
or
process
.
The
r
andom
ly
separ
at
e
of
s
om
e
node
al
so
app
ea
r
at
Fig
ure
7
f
or
TCP
and
ud
p
proto
col.
At
Icm
p
protoc
ol
f
r
om
first
un
ti
l
f
our
th
ep
oc
h
s
how
n
that
decr
easi
ng of
use
d refe
ren
ce
vec
tor n
um
ber
, t
hat m
ean th
ere
are m
or
e sim
i
l
ar
weig
ht in
t
ra
ining data.
Parall
el
RGS
OM
-
CR
V
is
outpe
rfor
m
regular
R
GSOM
in
ef
fici
ency
of
tim
e
con
s
um
e,
it
s
sp
e
nd
aver
a
ge
6
ho
ur
and
33
m
inu
te
s
and
18
sec
ond
w
hile
trai
ning,
an
d
46
m
inu
te
s
and
39
seco
nd
for
te
sti
ng
.
Trainin
g
tim
e
consum
e
wh
e
n
us
in
g
R
GSOM
-
CR
V
is
be
tt
er
than
Pa
ra
ll
el
RGSO
M
-
CR
V,
this
bec
ause
i
n
par
al
le
l
RGS
O
M
-
CR
V
the
re
are
proce
dure
to
sel
ect
in
g
input
acc
ordin
g
to
prot
oco
l
ty
pe.
H
oweve
r
,
tim
e
consum
ing
f
or
te
sti
ng
usi
ng
Parall
el
RGSOM
-
CR
V
is
bett
er
tha
n
RG
SOM
,
this
beca
use
at
par
al
le
l
RGS
OM
gen
e
rate
le
ss
use
d
nodes
in
t
he
m
ap,
s
o
ti
m
e
fo
r
sca
nn
i
ng
the
wi
nn
e
r
node
m
or
e
ef
fici
ent.
T
he
prob
l
e
m
of
Re
gu
la
r
G
SOM
for
cl
as
sifi
ed
big
data
ha
s
bee
n
fi
xe
d
by
RGSO
M
-
CR
V
an
d
Parall
el
RGSO
M
-
C
RV,
th
e
restrict
ed
of
node
s
le
ngth
ge
ner
at
e
by
gro
wing
th
reshol
d
m
ake
the
li
m
i
ta
ti
on
of
m
ap
to
gr
ow
i
ng
bigger
a
nd
bigger
.
Cl
us
te
rin
g
ref
e
ren
ce
vect
or
m
ake
RGSO
M
-
CR
V
capa
ble
to
ge
ner
al
iz
e
weig
h
t
base
on
grow
i
ng
thres
ho
l
d.
Fr
om
Tab
le
6
, Paral
le
l R
GS
O
M
-
CR
V
has
l
ower acc
ur
acy
t
han
oth
e
r
m
et
h
od
s
, but let
see the n
um
ber
of
te
sti
ng
data,
pro
posed
m
et
hod
ha
ve
4,898,431
i
ns
ta
nce
s
as
te
sti
ng
dat
a,
an
d
ha
ve
le
s
s
featu
re
to
pr
ocess.
W
it
h
a
la
rg
e
r
am
ou
nt
of
te
ste
d
data,
Pa
rall
el
RGSO
M
-
CR
V
is
capab
le
of
pr
od
ucin
g
91.
86%
accuracy,
so
this
m
et
ho
d ve
ry pr
om
isi
ng
to
s
olv
e the
b
i
g data
pro
blem
s in
cl
assifi
cat
ion
.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
250
2
-
4752
An
omaly
-
base
d
intr
us
io
n det
ect
or
syste
m us
ing
restri
ct
e…
(Tom
i Y
ahy
a
C
hr
ist
yaw
an
)
925
Table
6.
C
om
par
iso
n of p
rop
ose
d
m
et
ho
d wit
h othe
r
m
et
ho
ds by
nu
m
ber
of
trainin
g data, t
esi
ng d
at
a,
featur
e
s,
a
nd a
ccur
acy
.
Metho
d
Tr
ain
in
g
Data
Testin
g
Data
Featu
res
ACC
%
KNN
[
1
]
5
,55
2
5
,55
2
15
9
7
.69
%
SVM+
EL
M
b
ase
K
-
m
e
an
[
1
0
]
4
9
4
,021
3
1
1
,029
41
9
5
.75
%
TAN+R
EP
[
1
2
]
3
2
6
,053
1
6
7
,968
No
t pro
v
id
ed
9
8
.99
%
Parallel
RGSO
M
-
CRV
4
9
4
,021
4
,89
8
,431
9
9
1
.86
%
Figure
5. Ma
p gen
e
rated
w
it
h Parall
el
RGS
OM
-
CR
V
for f
irst ep
och at e
xperim
ent 5
Figure
6. Ma
p gen
e
rated
w
it
h Parall
el
RGS
OM
-
CR
V
for s
econd e
poch
at
experim
ent 5
Figure
7. Ma
p gen
e
rated
w
it
h Parall
el
RGS
OM
-
CR
V
for
thir
d
e
po
c
h
at
e
xp
e
rim
ent 5
4.
CONCL
US
I
O
N
Fr
om
this
exp
erim
ent
Parall
el
and
RGSO
M
-
CR
V
ou
t
perform
than
reg
ula
r
GSOM
in
tim
e
consum
ing
,
s
o
this
pro
pose
m
et
ho
d
is
m
ore
eff
ic
ie
nt
tha
n
GSOM
m
et
ho
d,
a
nd
f
ro
m
com
par
at
ion
with
ot
he
r
m
et
ho
d,
the
re
su
lt
of
Parall
e
l
RGSO
M
is
acce
ptable
with
91.
86%
for
accuracy,
false
al
arm
rate
aro
un
d
20.58%
,
re
cal
l
or
detect
ion
r
at
e
is
95.
32%,
an
d
94.35%
in
pr
eci
sio
n.
T
his
stu
dy
al
s
o
con
cl
ud
e
that
f
ind
t
he
best of m
axi
m
um
n
od
e
w
il
l i
ncr
ease
the e
ff
i
ci
ency, and R
GSOM ge
ne
rali
ze capab
il
it
y d
epend
on
GTst
art an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
13
, N
o.
3
,
Ma
rc
h 201
9
:
919
–
926
926
GTS
t
op.
T
he
capab
il
it
y
to
gen
erali
ze
ref
e
re
nce
vecto
r
m
a
ke
accura
cy
and
detect
ion
rat
e
acce
ptable.
F
ind
in
g
the opti
m
u
m
p
aram
et
er s
et
ti
n
g of g
rowin
g
t
hr
es
holds
can
be use
d
as a
r
e
fer
e
nce
for
th
e
fu
t
ur
e
resea
rch.
REFEREN
CE
S
[1]
L.
Li
,
H.
Zha
ng
,
H.
Peng
,
and
Y.
Yang.
Nea
r
e
st
Neighbor
s
ba
sed
density
peaks
appr
oac
h
to
i
ntrusion
detec
ti
o
n.
Chaos,
Soli
tons
&
Frac
ta
ls
.
2018
;
110:
33
–
40.
[2]
E.
Kabi
r,
J.
Hu,
H.
W
ang,
and
G.
Zhuo.
A
novel
sta
ti
sti
ca
l
t
ec
hniqu
e
for
in
t
rusion
det
e
ct
ion
s
y
stem
s.
Futur
e
Gene
ration
Com
pute
r
S
y
stems
.
2
018;
79:
303
–
31
8.
[3]
M.
Ahm
ed,
A.
Naser
Mahm
ood,
and
J.
Hu
.
A
surve
y
of
ne
twork
anomal
y
de
te
c
ti
on
t
ec
hniqu
es.
Journal
of
Ne
tw
ork
and
Computer
A
ppli
cations
.
201
6;
60:
19
–
31.
[4]
E.
de
la
Hoz,
E
.
de
la
Hoz,
A.
Ortiz
,
J.
Orteg
a,
and
A.
Martínez
-
Álvar
ez.
Feat
ure
sel
ec
t
ion
b
y
m
ult
i
-
obj
ec
t
i
ve
opti
m
isat
ion:
Applicati
on
to
n
e
twork
anomal
y
det
e
ct
ion
b
y
hi
e
rar
chica
l
self
-
or
gani
sing
m
aps.
Knowle
dge
-
Base
d
Syste
ms
.
2014;
7
1:
322
–
338.
[5]
Z.
Chib
a,
N.
A
bghour,
K.
Mous
said,
A.
El
O
m
ri,
and
M.
Ri
da.
A
novel
archit
e
ct
ure
combined
with
opti
m
al
par
amete
rs
for
bac
k
propa
g
atio
n
neur
al
n
et
wor
ks
appl
ie
d
to
an
om
aly
n
et
work
int
rusion
de
te
c
tion.
Computers
&
Sec
urit
y
.
2018;
7
5:
36
–
58.
[6]
F.
Al
Huda,
W
.
Firdaus
Mahmud
y
,
and
H.
To
ll
e
.
Android
M
al
ware
De
te
c
ti
o
n
Us
ing
Bac
kpropa
gation
Neura
l
Network.
Indon
e
sian J
ournal
o
f El
e
ct
rica
l
Eng
in
ee
ring a
nd
Computer
Sc
ie
nc
e
.
2
016;
4(1):
240
-
2
44.
[7]
C.
Xiang
,
Y.
Xi
a
o,
P.
Qu,
and
X.
Qu.
Network
Int
rusion
Detect
ion
Based
on
PS
O
-
SV
M.
T
ELKOMNIKA
Indone
sia
n
Jou
rnal
of
Elec
t
rical
Engi
ne
erin
g
.
2014;
12(2):
1
502
-
1508.
[8]
A.
Kara
m
i
and
M.
Guerre
ro
-
Zapat
a
.
A
fuz
z
y
an
om
aly
de
te
c
ti
on
s
y
stem
base
d
on
h
y
brid
PS
O
-
K
m
ea
ns
al
gorit
hm
in
cont
en
t
-
ce
n
tric
n
et
works
.
Ne
uroc
omputing
.
2015
; 149:
1253
–
1269.
[9]
A.
H.
Ham
amot
o,
L.
F.
Carva
lh
o,
L.
D.
H.
Sampa
io,
T
.
Abrão,
and
M.
L.
Proenç
a.
Network
Anom
aly
Detect
io
n
S
y
stem using
Ge
net
i
c
Algori
thm
and
Fuzz
y
Log
ic.
Exp
ert S
y
st
ems
with
Appl
icati
o
ns.
2018;
92
:
39
0
–
402.
[10]
W
.
L.
Al
-
Yasee
n,
Z.
A.
Othm
a
n,
and
M.
Z.
A
.
Naz
ri
.
Multi
-
l
eve
l
h
y
b
rid
sup
port
vec
tor
m
achine
and
ex
tre
m
e
le
arn
ing
m
ac
hin
e
base
d
on
m
odifi
ed
K
-
m
ea
ns
for
int
rusion
detec
t
ion
s
y
stem.
Ex
pert
Syst
ems
wit
h
Appl
i
cation
s
.
2017;
67:
296
–
3
03.
[11]
G.
Kim
,
S.
Le
e
,
and
S.
Kim
.
A
novel
h
y
brid
in
tr
usion
det
e
ct
ion
m
et
hod
int
egr
at
i
ng
anomal
y
d
ete
ct
ion
with
m
isuse
det
e
ct
ion
.
Ex
per
t
Syst
ems wi
th
A
ppli
cations
.
201
4;
41(4):
1690
–
1
700.
[12]
K.
Raj
ase
k
ara
n
and
K.
Nirm
al
a
.
A
Novel
and
Advanc
ed
Dat
a
Mining
Model
base
d
H
y
brid
In
trusion
Detect
io
n
Fram
ework.
TE
L
KOMNIKA
Indone
sian J
ournal
of
E
le
c
tric
al
En
gine
ering
.
2015;
13(2):
223
-
231.
[13]
J.
Jabe
z
and
B
.
Muthukum
ar.
Intrusion
Detect
i
on
S
y
stem
(IDS
):
Anom
aly
Detect
ion
Us
ing
Outli
er
De
te
c
ti
o
n
Approac
h.
Proc
edi
a
Comput
er
S
ci
en
ce
.
2015;
48
:
338
–
346.
[14]
W.
-
C.
Li
n,
S.
-
W
.
Ke,
and
C.
-
F.
Tsai
.
CAN
N:
A
n
int
rusion
det
ection
s
y
stem
base
d
on
combining
cl
uster
c
ent
ers
a
nd
nea
rest
ne
ighbor
s.
Knowle
dg
e
-
B
ased
Syste
ms
.
20
15;
78:
13
–
21.
[15]
T.
Kohonen
.
Essentials
of
the sel
f
-
orga
nizing
m
a
p.
Neural
Ne
two
rks
.
2013;
37:
52
–
65.
[16
]
D.
Alaha
koon,
S.
K.
Halga
m
uge,
and
B.
Srin
iv
asa
n.
D
y
namic
self
-
orga
ni
zi
ng
m
aps
with
cont
roll
ed
growth
fo
r
knowledge
d
isco
ver
y
.
IEEE
Tr
ansacti
ons on
N
eural
Ne
tworks
.
200
0;
11(3):
601
–
61
4.
[17]
S.
J.
Stolfo,
W
.
Fan,
W
.
Lee,
A.
Prodrom
idi
s,
and
P.
K.
Chan.
C
ost
-
base
d
Model
ing
and
Ev
al
ua
tion
for
Data
Mini
ng
with
Appli
ca
t
ion
to
Fraud
and
Int
rusion
Detect
ion
:
Result
s
from
th
e
JA
M Proje
c
t. 2
000.
Evaluation Warning : The document was created with Spire.PDF for Python.