Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
10
,
No.
3
,
June
2020
,
pp.
2701
~
2709
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v10
i
3
.
pp2701
-
27
09
2701
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om/i
nd
ex
.ph
p/IJ
ECE
A techni
ca
l
re
view and
comp
arative an
alysis of ma
chine
learnin
g techniq
ues fo
r int
rusion d
etection
systems i
n MANE
T
Sa
f
aa L
aq
tib
1
,
Kh
alid El
Ya
ssini
2
, Moul
ay Lah
ce
n H
as
n
aoui
3
1,2
Inform
at
ic
s
an
d
Applicati
ons
L
abor
at
or
y
(IA),
Depa
rtment
o
f M
at
hematics a
n
d
Com
pute
r
Sci
e
nce
,
Facul
t
y
of
Sci
en
ce
s,
Moul
a
y
Ism
ai
l
Univer
sit
y
,
Morocc
o
3
ISIC E
STM,
L2
MI L
abor
a
tor
y
,
ENSA
M,
Moulay
Ism
ai
l
Univ
ersity
,
Morocc
o
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
hist
or
y:
Re
cei
ved
J
un
28
, 2
019
Re
vised
N
ov 20
, 2
019
Accepte
d
Dec
4,
2019
Mac
hine
le
a
rnin
g
te
chn
ique
s
ar
e
bei
ng
wide
l
y
u
sed
to
develop
a
n
int
rusion
det
e
ct
ion
s
y
st
e
m
(IDS
)
for
det
e
ct
ing
and
class
if
y
ing
c
y
b
er
at
t
ac
ks
at
the
net
work
-
l
ev
el
and
the
host
-
le
ve
l
in
a
ti
m
ely
and
aut
om
a
ti
c
m
anne
r.
How
eve
r,
Tra
d
it
ional
In
trusio
n
Detect
ion
Sy
stems
(IDS
),
base
d
on
tra
ditiona
l
m
ac
h
ine
l
ea
rn
ing
m
ethods
,
lacks
relia
bil
ity
and
accurac
y
.
Inst
ea
d
of
the
tra
d
it
ion
a
l
m
ac
hin
e
l
ea
rn
ing
used
in
pr
e
vious
rese
ar
che
s
,
we
thi
nk
dee
p
l
ea
rn
ing
h
as
the
po
te
nt
ia
l
to
per
form
be
tter
in
ext
r
ac
t
ing
fea
tu
res
of
m
assive
dat
a
co
nsideri
ng
the
m
assive
c
y
b
er
traffic
in
real
li
fe
.
Gene
rall
y
Mobile
Ad
Hoc
Networks
have
give
n
the
low
ph
y
sic
al
sec
uri
t
y
fo
r
m
obile
devi
c
es,
b
ec
ause
of
th
e
prop
ert
i
e
s
suc
h
as
node
m
obil
ity
,
lack
of
ce
n
tra
l
ized
m
ana
gement
an
d
li
m
it
ed
band
width.
To
t
ac
k
le
the
se
se
cur
i
t
y
issues,
tra
ditiona
l
cr
y
pt
ogra
ph
y
sch
emes
ca
n
-
not
complet
e
l
y
saf
egua
rd
MA
NETs
in
te
rm
s
of
novel
t
hre
at
s
and
vuln
era
bi
li
t
ie
s,
thus
b
y
apply
ing
De
ep
learni
ng
me
thods
te
chniques
in
IDS
are
ca
p
abl
e
of
ada
pti
ng
the
d
y
n
amic
envi
ronm
ent
s
of
MA
NETs
and
ena
bl
es
the
s
y
st
em
to
m
ake
de
ci
sions
on
int
rusion
while
cont
inui
ng
to
l
e
arn
about
their
m
obil
e
envi
ron
m
ent
.
An
IDS
in
MA
NET
is
a
sensoring
m
ec
hani
sm
tha
t
m
onit
ors
nodes
and
net
work
ac
t
ivi
ties
in
orde
r
to
detec
t
m
al
ic
ious
actio
ns
and
m
al
ic
io
us
at
te
m
pt
per
form
ed
b
y
I
ntrude
rs.
Recent
l
y
,
m
ult
iple
deep
le
arn
ing
appr
oac
hes
hav
e
bee
n
proposed
t
o
enha
n
ce
th
e
per
form
anc
e
of
int
rusion
de
tection
s
y
stem.
In
thi
s
pape
r
,
we
m
ade
a
s
y
s
t
ematic
compari
son
of
thre
e
m
odel
s,
Inc
epr
t
ion
arc
hi
te
c
ture
co
n
volut
ional
neur
a
l
net
work
(Inc
e
pti
on
-
CNN
),
Bi
dire
c
ti
ona
l
long
short
-
te
rm
m
emory
(BLSTM)
and
de
ep
bel
i
ef
ne
twork
(DBN
)
on
the
de
ep
learni
n
g
-
base
d
int
rusio
n
det
e
ct
ion
s
y
st
ems
,
using
the
NS
L
-
KD
D
dat
ase
t
cont
a
in
ing
informati
o
n
about
int
rusion
and
reg
ul
ar
net
work
conne
c
ti
ons,
the
goal
is
to
prov
ide
b
asic
gu
idance
on
th
e
cho
i
ce
of
de
ep
le
arn
ing
m
odel
s
in
MA
NET.
Ke
yw
or
ds:
Atta
ck
BLSTM
DBN
Deep l
ear
ning
In
tr
us
i
on
detect
ion
syst
em
ID
S
In
ce
ptio
n
-
C
N
N
MANET
Copyright
©
202
0
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
:
Laqti
b Safaa
,
Inform
at
ic
s an
d Appli
cat
ion
s
Lab
or
at
ory
(I
A
),
Dep
a
rtm
ent o
f M
at
hem
a
ti
cs and Com
pu
te
r S
ci
ence,
Fa
culty
of S
ci
e
nces,
Mou
la
y
I
sm
ail U
ni
ver
sit
y,
Me
kn
es
, M
oro
cco.
Em
a
il
: l
aq.
safaa@g
m
ai
l.co
m
1.
INTROD
U
CTION
A
M
ob
il
e
ad
ho
c
N
et
w
ork
(MA
NET)
is
gen
e
rall
y
def
ined
as
a
netw
ork
that
has
m
any
fr
ee
or
auto
no
m
ou
s
node
s
[1
]
,
of
te
n
com
po
sed
of
m
ob
il
e
dev
ic
es
or
ot
her
m
ob
il
e
nodes
that
can
arr
a
nge
them
sel
ve
s
in
va
rio
us
way
s
an
d
op
e
rate
without
stric
t
t
op
-
down
ne
tw
ork
a
dm
inist
rat
ion
.
T
her
e
a
re
m
any
diff
e
re
nt
ty
pes
of
set
ups that
c
ou
l
d be call
ed M
AN
ET
and t
he po
te
ntial
f
or
this s
or
t
of n
et
work is sti
ll
b
e
ing
st
udie
d
.
I
nhere
nt
vu
l
ner
a
bili
ty
of
M
ob
il
e
a
d
ho
c
N
et
wor
ks
intr
oduc
es
new
secu
rity
pro
blem
s,
wh
ic
h
m
os
tl
y
address
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
3
,
J
une
2020
:
27
01
-
2709
2702
the
netw
ork
a
nd
data
li
nk
l
ay
er
of
t
he
prot
oco
l
sta
ck
.
B
ecause
eac
h
pac
ket
m
us
t
be
pas
sed
t
hro
ugh
interm
ediat
e
no
de
s
quic
kly,
t
hat
pac
ket
has
to
tra
vel
f
ro
m
the
s
ource
to
the
destin
at
ion.
Ma
li
c
iou
s
routing
at
ta
cks
m
a
y
tar
get
the
r
outi
ng
detect
io
n
or
m
ai
ntenan
ce
process
by
fail
ing
to
fo
ll
ow
t
h
e
sp
eci
ficat
ion
s
of
the
r
ou
ti
ng
pro
tocol
[
2,
3]
.
T
hi
s
increases
t
he
possibil
it
y
of
att
acks
su
c
h
as
eaves
dro
pp
i
ng,
spo
of
in
g,
de
ni
al
of
serv
ic
e,
an
d
i
m
per
so
natio
n.
Com
par
ed
to
fixe
d
net
wor
ks,
M
obil
e
ad
ho
c
N
et
work
secur
it
y
is
ta
ke
n
int
o
account
from
var
i
ou
s
po
i
nts
su
ch
as
avai
la
bili
ty
,
pr
ivacy
,
reli
abili
ty
,
encr
y
ption,
au
thentic
at
ion,
acce
ss
con
t
ro
l,
and
usa
ge
c
on
t
ro
l.
D
ue
to
the
prom
i
n
ent c
ha
racteri
sti
cs o
f
M
obil
e ad h
oc
N
et
wor
ks
, se
cu
rity
m
e
thods
us
e
d
to
sec
ur
e
fixe
d
net
wor
ks
are
no
t
f
easi
ble
f
or
M
ANE
T
[4]
.
Ne
w
th
r
eat
s
su
ch
as
at
ta
cks
f
ro
m
internal
m
al
ic
iou
s
node
s,
By
zanti
ne
,
an
d
w
or
m
ho
l
e
at
ta
cks
a
re
di
ff
ic
ult
to
de
fe
nd.
An
I
ntr
us
i
on
Detect
io
n
Syst
e
m
(IDS) is a
n
e
ffec
ti
ve
way t
o
i
den
ti
fy
w
hen a
n
at
ta
ck occ
urs
in
a M
ANET.
Fo
r
the
a
bo
ve
reas
on
s
,
it
is
ver
y
im
po
rtant
to
dep
l
oy
in
MANET
as
a
seco
nd
li
ne
of
defense
an
intru
si
on
dete
ct
ion
syst
em
[5]
.
I
ntr
us
io
n
detect
ion
syst
e
m
s
(I
D
S)
are
a
m
echan
ism
for
m
on
it
or
i
ng
a
nd
inv
est
igati
ng
e
ven
ts
occurri
ng
in
a
com
pute
r
syst
e
m
.
An
I
DS
inc
orp
or
at
es
m
et
ho
ds
for
m
od
el
ing
an
d
disco
ver
i
ng abnorm
al
b
ehav
i
or
s
and c
om
plex
te
ch
niques
. T
hey try t
o deter
m
ine w
het
her
or
not t
he netw
ork
i
s
go
i
ng
th
r
ough
any
m
al
ic
iou
s
act
ivit
y.
This
i
s
ty
pical
ly
accom
plished
by
ga
the
rin
g
data
autom
at
ic
a
ll
y
fr
om
a v
a
riet
y of
syst
e
m
s an
d netw
ork
s
ources
and the
n
a
naly
zi
ng the
in
form
at
i
o
n f
or potenti
a
l security
issue
s
.
Current
int
ru
si
on
detect
io
n
a
nd
preve
ntio
n
m
et
ho
ds,
su
c
h
as
fire
wa
ll
s,
ac
cess
co
ntr
ol
pr
oto
c
ols
an
d
authe
ntica
ti
on
,
hav
e
seve
ral
draw
bac
ks
in
de
fendin
g
netw
or
ks
an
d
de
vices
fr
om
ever
m
or
e
adv
a
nced
at
ta
cks,
su
c
h
as
a
de
nial
of
ser
vice
[
6]
.
H
ow
e
ve
r,
m
os
t
sys
tem
s
base
d
on
su
c
h
te
chn
iq
ues
a
r
e
su
f
fer
i
ng
fro
m
hig
h
false
posit
i
ve
and
false
neg
a
ti
ve
detect
ion
rates
an
d
la
ck
of
c
onti
nuous
ada
ptati
on
to
ev
olv
in
g
m
al
i
ci
ous
beh
a
viors.
Dee
p
l
ear
ning,
t
he
refor
e
,
al
lo
ws
to
quic
kly
perf
or
m
data
analy
sis
an
d
vis
ualiz
at
ion
,
t
he
ai
m
is
to
enab
le
the
det
ect
ion
of
dev
i
ce
vu
lne
ra
bili
t
ie
s
and
f
la
ws
by
secur
it
y
prof
essi
onal
s.
Se
ver
al
D
ee
p
L
e
arn
i
ng
(D
L
)
a
ppr
oac
hes
ha
ve
been
ap
plied
to
th
e
issue
of
int
ru
si
on
detect
i
on
to
i
ncr
ease
detect
io
n
rat
es
an
d
adap
ta
bili
ty
. S
uch tec
hniq
ues
are ofte
n use
d t
o
est
ablish
the
att
acks e
xisti
ng a
nd d
et
ai
le
d
knowle
dg
e
bas
e.
The
rem
ai
nd
er
of
this
pa
per
is
orga
nized
as
f
ollow
s
.
Se
ct
ion
2
is
de
di
cat
ed
to
disc
us
s
Secu
rity
at
ta
cks
in
M
A
NET.
The
n,
S
ect
ion
3
descri
bes
the
i
ntr
usi
on
detect
ion
syst
e
m
arch
it
ect
ur
es
i
n
MA
NET.
Sect
ion
4
pro
vid
es
a
dee
p
l
earn
in
g
m
o
dels
for
intr
usi
on
detect
io
n
syst
e
m
in
MANET
.
I
n
sect
ion
5
the
ex
pe
rim
ent
s and
res
ultas
.
I
n sec
ti
on
6
we
conclu
de by c
on
cl
us
io
n.
2.
SECURIT
Y A
TT
AC
K
S I
N MA
NET
Com
par
ed
to
wire
d
inf
rastr
uctu
re
netw
or
ks
,
Mo
bile
a
d
h
oc
Netw
ork
s
(MA
NET)
[
7]
are
m
or
e
vu
l
ner
a
ble
to
at
ta
cks.
MA
N
ET
face
m
or
e
secur
it
y
threat
s
than
ce
ntrali
zed
net
wor
ks
du
e
t
o
their
dy
nam
ic
topolo
gy
an
d
the
la
ck
of
c
entrali
zed
net
work
adm
inist
rati
on.
I
n
t
he
Mob
il
e
a
d
hoc
netw
orks,
sever
al
char
act
e
risti
cs
cou
l
d
be
us
e
d
to
cl
assif
y
at
ta
cks.
Exam
ples
w
ou
l
d
in
c
lud
e
lo
okin
g
at
the
beh
a
vi
or
of
the att
acks
(
pa
ssive
vs
. a
ct
ive
),
t
he
s
ource
of the att
acks
(
e
xt
ern
al
vs
. i
nter
nal)
.
In
act
i
ve
at
ta
ck,
t
he
at
ta
cke
r
is
act
ively
inv
ol
ved
in
t
he
netw
ork
op
e
ra
ti
on
s
a
nd
trie
s
to
c
hange
the
m
essages
bein
g
tra
ns
m
it
t
ed.
By
disru
pting
the
entire
ne
twork
proc
ess
,
the
at
ta
cker
m
ay
m
od
ify
,
insert,
forg
e
,
dro
p
dat
a.
The
freq
ue
nc
y
of
this
at
ta
c
k
is
hig
h
beca
us
e
the
w
ho
le
netw
ork
can
be
br
ou
gh
t
do
w
n
[8]
.
They
are
easy
to
detect
as
the
netw
ork
pe
rfor
m
ance
de
gr
ades
sig
nifica
ntly
.
I
n
pa
ssiv
e
at
ta
ck,
the
a
tt
acker
do
e
s
not
corr
upt
the
sh
a
red
data
but
li
ste
ns
to
it
.
They
are
at
tem
pting
to
gai
n
co
nf
i
de
ntial
inform
ati
on
a
nd
analy
ze
the
traff
ic
patte
r
ns
tr
ansm
itted.
The
y
are
diff
ic
ult
to
detect
becau
se
they
do
not
inte
rr
upt
or
m
od
ify
the
inf
or
m
at
ion
sent
or
recei
ved
.
It
is
al
so
po
s
sible
to
cl
assify
the
at
ta
c
ks
into
tw
o
cat
egories
de
pend
ing
on
the
dom
ai
n
of
the
at
ta
cks,
na
m
el
y
extern
al
at
ta
cks,
an
d
in
te
rn
al
at
ta
cks.
In
te
r
nal
at
ta
ck
s
are
car
ried
out
by
nodes
that
a
re
no
t
pa
rt
of
t
he
do
m
ai
n
of
the
netw
ork.
Exte
r
nal
at
ta
cks
a
re
trigg
e
red
by
node
s
that
a
re
al
read
y
par
t
of the
net
work. E
xter
nal
att
acks are m
or
e se
ver
e
tha
n
i
nter
nal att
acks
[9
]
.
2.1.
Deni
al of
ser
vi
ce atta
c
ks
(DoS)
A
de
nial
of
se
rv
ic
e
at
ta
ck
is
an
at
te
m
pt
to
m
ake
a
m
achi
ne
or
netw
ork
resou
rce
una
va
il
able
to
it
s
intende
d
us
e
rs
[10,
11]
,
su
c
h
as
to
te
m
porar
il
y
or
i
nd
e
f
init
el
y
interrupt
or
suspe
nd
serv
ic
es
of
a
host
connecte
d
t
o
the
I
nter
net.
Th
is
at
ta
ck
can
be
la
un
c
hed
at
diff
e
re
nt
la
ye
rs.
the
physi
cal
la
ye
r,
netw
ork
la
ye
r
,
trans
port lay
er
.
2.2.
Rem
ot
e
to
us
er a
ttacks
(
R2L)
Occurs
w
hen
an
at
ta
cke
r
does
not
hav
e
a
n
acc
ount
on
t
he
victim
m
ac
hin
e
a
nd
at
te
m
pts
to
gain
acce
ss
by
se
ndin
g
pac
kets
t
o
a
m
achine
over
a
netw
ork
in
orde
r
t
o
ge
ner
at
e
so
m
e
vu
l
ner
a
bili
ty
on
t
hat
m
achine that al
lows
him
/ her
to gain l
ocal ac
cess as a
u
se
r o
f
that m
achine.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A tec
hnic
al rev
ie
w and
c
omp
arati
ve a
naly
sis
of mac
hin
e le
arnin
g
te
c
hn
i
ques... (
Sa
f
aa L
aqti
b
)
2703
2.3.
User
to
r
oot a
ttacks
(
U2R)
This
at
ta
ck
oc
cur
s
w
hen
nor
m
al
syst
e
m
us
er
il
le
gally
gains
acce
ss
to
ei
ther
r
oot’s
or
su
pe
r
us
er
’s
pr
i
vileges s
uc
h as Pe
rl,
xterm
[
12]
.
2.4.
Pro
bing
Pr
ob
i
ng
occ
ur
s
w
he
n
a
n
at
t
acker
sca
ns
a
netw
ork
i
n
order
to
gath
er
i
nfor
m
at
ion
or
fin
d
known
vu
l
ner
a
bili
ti
es
that
al
low
him
/her
to
hack
t
he
entire
net
w
ork.
U
su
al
ly
,
this
m
et
ho
d
is
us
e
d
in
inf
or
m
at
ion
m
ining
s
uc
h
as
saint,
port s
we
ep,
Msca
n, Nm
ap,
etc
[
13
]
.
3.
INTRUSI
ON
DETE
CTIO
N
SYSTE
M A
R
CHI
TE
CT
U
R
ES IN
MA
NET
Ma
ny
intr
us
io
n
de
te
ct
ion
sys
tem
s
(I
DS)
have
bee
n
de
velo
pe
d
f
or
MA
NE
T
to
detect
var
i
ou
s
ty
pes
of
at
ta
cks,
I
DS
pl
ay
s
an
im
po
rta
nt
r
ole
in
M
A
NET
t
o
detect
any
ty
pe
of
at
t
acks
[14,
15]
.
An
I
DS
is
a
software
sys
tem
us
ed
t
o
a
naly
zes
m
i
sb
e
hav
i
or
an
d
vio
la
ti
on
of
po
li
cy
,
a
nd
th
en
ge
ner
at
e
a
re
port
based
on
it
,
Ba
sic
al
ly
,
intru
sio
n
detect
io
n
syst
em
is
cl
assifi
ed
i
nto
f
ollow
i
ng
th
ree
basic
cat
eg
or
i
es
acco
r
ding
t
o
t
heir
op
e
rati
onal
str
uctu
re.
3.1.
The s
t
andal
one arc
hitect
ur
e
In
this
syst
e
m
,
to
determ
i
ne
intr
us
io
n,
the
intr
us
io
n
detect
i
on
syst
e
m
ru
ns
i
nd
e
pende
ntly
on
the
in
div
id
ual
node.
All
decisi
on
s
m
ade
ab
out
a
pa
rtic
ular
act
ivit
y
dep
en
d
s
olely
on
inf
or
m
at
ion
colle
ct
ed
at
it
s
ow
n
node,
as
there
is
no
colla
borati
on
be
twee
n
no
des
in
the
net
work
[16]
.
The
refo
r
e,
there
is
no
trans
fe
r
of
in
form
at
ion
.
Even,
as
no
al
ert
inform
ati
on
is
trans
ferr
ed,
a
node
in
t
he
sam
e
network
does
no
t
ha
ve
any
inf
or
m
at
ion
ab
ou
t
t
he
oth
e
r
node
s
in
the
net
work.
Be
ca
us
e
of
it
s
li
m
i
ta
ti
o
ns
,
this
m
od
el
is
not
ef
fici
ent,
it
ca
n
be
us
ed
ef
fecti
vely
in
a
netw
ork
wh
e
re
al
l
nodes
al
read
y
hav
e
an
I
DS
i
ns
ta
ll
ed.
C
om
par
ed
t
o
m
ulti
-
l
ay
ered
netw
ork
inf
ras
tructu
re,
this
s
yst
e
m
is
al
so
s
uitable
for
sin
gle
la
ye
r
netw
ork.
Si
nce
the
inf
or
m
at
ion
av
ai
la
ble
on an
y
sin
gle
node
is
no
t s
uffi
ci
ent to det
ect
intr
us
io
ns
,
this
syst
e
m
h
as not
been ch
os
e
n
as
MANET
IDS
.
3.2.
The d
is
tribute
d and c
ollab
orative
a
rc
hit
ecture
In
this
a
rch
it
e
ct
ur
e,
the
i
ntr
us
io
n
detect
io
n
e
ng
i
ne
is
i
nst
al
le
d
on
eac
h
node
i
n
t
his
arc
hitec
ture,
wh
ic
h
m
on
it
or
s
local
aud
it
data
and
detect
s
intru
sio
n
W
e
al
so
par
ti
ci
pate
in
the
coo
pe
ra
ti
ve
detect
ion
and
/
or
respo
ns
e
proc
ess
by
sh
ari
ng
aud
it
inf
or
m
a
ti
on
an
d/or
de
te
ct
ion
resu
lt
s
with
neig
hbori
ng
nodes
to
so
lv
e
the
pr
ob
le
m
.
Wh
e
n
the
i
ntr
usi
on
is
ca
ptu
r
e
d,
ei
t
her
a
loc
a
l
respo
ns
e
(e.
g.
al
erti
ng
t
he
l
ocal
us
er
)
or
a
global
respo
ns
e
m
ay
be
iss
ued
by
a
n
IDS
a
gen
t.
Each
node
is
i
nvolv
e
d
in
t
he
m
et
ho
d
a
nd
r
esp
on
se
of
i
ntru
si
on
detect
ion
as
ha
ving
an
IDS
ag
ent
run
ning
on
it
[
17
]
.
A
n
I
D
S
age
nt
is
respon
si
ble
f
or
detect
ing
an
d
c
ollec
ti
ng
local
inf
orm
ati
on
an
d
data
i
n
order
to
ide
ntify
any
at
ta
ck
i
f
an
at
ta
ck
occ
ur
s
in
t
he
netw
ork,
a
s
well
as
ta
king
an
i
nd
e
pe
nd
e
nt
res
pons
e
.
H
ow
e
ve
r,
w
hen
the
eviden
ce
is
no
n
-
c
oncl
us
ive
,
neig
hbori
ng
ID
S
a
gent
s
al
so
coope
rate
in
gl
ob
al
intr
us
i
on
det
ect
ion.
T
hi
s
syst
e
m
,
li
ke
sta
nd
al
on
e
I
DS
,
is
al
s
o
m
or
e
s
uitable
f
or
flat
netw
ork
syst
e
m
s,
no
t m
ulti
-
l
ay
er s
yst
em
s
.
3.3.
The h
ie
r
archi
cal I
DS
in
trus
ion d
e
tecti
on
architec
tu
re
Hierarc
hical
I
DS
syst
em
Exp
an
d
the
distri
bu
te
d
a
nd
c
ooper
at
ive
I
DS
s
yst
e
m
fu
nctio
ns
an
d
ha
ve
bee
n
i
m
ple
m
ented
fo
r
m
ulti
-
la
ye
r
networ
k
inf
r
ast
ru
ct
ures
where
the
net
work
is
di
vid
e
d
into
var
i
ous
s
m
al
l
netw
orks
kn
own
as
cl
us
te
rs.
Usu
al
ly
,
each
cl
us
te
r
hea
d
ha
s
m
or
e
functi
on
al
it
y
than
ot
her
cl
us
te
r
m
e
m
ber
s,
su
c
h
as
t
ran
s
m
itti
ng
data
pa
ckets
to
ot
her
cl
us
te
rs
.
We
can
the
re
fore
say
that
these
cl
us
te
r
hea
ds
work
i
n
so
m
e
way
as
central
points
si
m
il
ar
to
wired
netw
ork
co
ntr
ol
de
vices
su
c
h
as
r
oute
rs,
s
witc
hes
or
gateways
.
The
m
ulti
-
la
yer
in
g
co
nce
pt
app
li
es
to
i
ntr
usi
on
detect
ion
syst
e
m
s
wh
ere
there
is
a
pro
po
s
al
for
hiera
rch
ic
al
ID
S
.
Each
IDS
agen
t
r
uns
on
a
sp
eci
fic
m
e
m
ber
node
a
nd
is
respo
ns
ibl
e
fo
r
it
s
node
,
i.e.
m
on
it
or
in
g
an
d
decidin
g
on
int
ru
si
on
s d
et
ect
ed
local
ly
.
A
cl
us
te
r
hea
d
is
re
sp
onsi
ble
for
their
no
de
local
ly
as
well
as
glo
bally
for
th
ei
r
cl
ust
er,
s
uch
as
m
on
it
ori
ng
ne
tw
ork
traf
fic
an
d
an
nounci
ng
a
global
res
pons
e
w
he
n
dete
ct
ing
netw
ork
i
ntrusion
[
18]
.
4.
DEEP L
EA
R
NING
MO
DE
LS
FO
R
I
NT
RUSIO
N DE
TE
CTION S
Y
STE
M
I
N MA
NET
Deep
le
a
rn
i
ng
is
a
cl
ass
of
M
achine
l
ear
ning
al
gorithm
s
that
us
es
m
ulti
ple
la
ye
rs
to
pro
gr
essi
vely
extract
higher
le
vel
featur
es
from
the
raw
input.
The
ai
m
is
to
m
ake
m
achines
li
ke
com
pu
te
rs
thin
k
an
d
unde
rstan
d
how
hu
m
ans
thi
nk
by
im
it
ati
ng
the
gri
d
of
the
hum
an
br
ai
n
c
onnecti
on,
Deep
le
ar
ni
ng
arch
it
ect
ures
su
c
h
as
dee
p
neural
net
w
orks
,
dee
p
be
li
ef
netw
ork
s
,
rec
urren
t
neural
net
wor
ks
a
nd
conv
olu
ti
onal
neural
netw
orks
ha
ve
been
app
li
ed
to
fiel
ds
in
cl
udin
g
c
om
pu
te
r
visio
n
,
s
peec
h
recogn
it
io
n
,
natu
ral
la
ngua
ge
processi
ng
,
aud
i
o
recog
niti
on,
s
ocial
net
w
ork
filt
ering,
m
achine
tra
ns
l
at
ion
,
bio
i
nform
at
ic
s
,
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
3
,
J
une
2020
:
27
01
-
2709
2704
dru
g
desi
gn
,
m
edical
i
m
age
analy
sis,
m
at
erial
insp
ect
ion
a
nd
boar
d
gam
e
pr
ogra
m
s,
wh
ere
th
ey
have
pro
du
ce
d
res
ults
co
m
par
able
to
and
in
som
e
cases
su
per
ior
to
hum
an
exp
erts
[
19]
.
Su
pe
rv
ise
d
le
arni
ng
al
gorithm
is
app
li
ed
to
a
datas
et
that
h
as f
eat
ur
es
an
d
eac
h
of
t
hose
feat
ures
associat
ed
w
it
h
a
la
bel.
H
ow
e
ve
r
,
deep
le
a
rn
i
ng
al
gorithm
s
com
es
un
de
r
un
s
up
e
r
vised
le
ar
ning
al
gorithm
s
wh
ic
h
are
a
ppli
ed
to
a
dataset
wh
ic
h
has
m
any featu
res
in
or
der
t
o
l
earn use
f
ul propertie
s
from
the stru
ct
ur
e
of
t
he data
set
[
20
]
.
The
Sec
ur
it
y
app
li
cat
io
ns
of
deep
le
ar
ning
m
od
el
s
li
ke
In
trusion
Detect
i
on
Syst
em
(I
D
S)
,
m
al
war
e
detect
ion,
sp
a
m
-
fil
te
ring
ha
ve
bec
om
e
ess
entia
ls
in
design
i
ng
ta
s
ks
f
or
data
protect
ion,
cl
assifi
cat
ion,
an
d
pr
e
dicti
on. Th
e
se d
iffe
ren
t t
yp
e
s o
f
ta
sk
s
depend
i
ng
on
the i
ntell
igence to
bu
il
d
a m
od
el
t
hat u
sua
ll
y cl
a
ssifie
s
and
disc
rim
ina
te
s
betwee
n
s
a
m
ples
of
"
be
nign"
a
nd
"m
al
ign
,"
su
c
h
a
s
at
ta
cks
a
nd
ben
i
gn
pa
cket
s
[
21
]
.
The
c
om
plexit
y
of
at
ta
ck
te
chn
i
qu
e
s
too
ls
i
s
increase
d
wit
h
the
ra
pid
i
nc
rease
with
t
he
us
e
of
D
eep
le
arn
i
ng
m
od
el
s
.
Ther
e
are
lots
of
popula
r
va
riant
s
of
Dee
p
le
arn
i
ng
m
od
el
s
li
ke
CNNs
and
RN
Ns
.
To
so
lv
e
the
ha
rdness
of
trai
ni
ng,
BL
STM
is
pro
po
sed
to
al
le
viate
so
m
e
lim
it
a
t
ion
s
of
the
ba
sic
RNN,
I
nce
ption
conv
olu
ti
onal
neural
ne
t
wor
k
(C
NN)
w
hi
ch
a
va
riant
of
basic
CN
N
and
dee
p
bel
ie
f
netw
ork
(
DBN
).
This
sect
io
n
gi
ves
a
br
ie
f
intr
oductio
n
of
the
m
od
el
s
we
hav
e
us
e
d
in
our
e
xp
e
ri
m
ents:
the
inc
eption
arch
it
ect
ure C
NN, BLS
TM a
nd D
B
N
[22].
4.1.
The
ince
p
tion archi
tectu
re
CNN
Szege
dy
et
al
[
23
]
s
uggest
t
he
I
nce
ptio
n
a
rchi
te
ct
ur
e
CN
N
t
o
s
olv
e
t
he
pro
blem
of
a
la
rge
num
ber
of
par
am
et
ers
and
sp
ee
d
up
th
e
le
arn
in
g
of
CNN.
An
In
ce
ption
netw
ork
is
ty
pical
l
y
a
netw
ork
co
ns
i
sti
ng
of
m
od
ules
of
t
he
ab
ov
e
ty
pe
sta
cked
on each
ot
her
,
with
occa
sion
al
m
ax
-
pooling
la
ye
rs
wi
th
phase
two
to
halv
e
gr
i
d
res
olu
ti
on.
It
seem
ed
use
fu
l
to
sta
rt
usi
ng
In
ce
ptio
n
m
od
ules
only
at
higher
la
ye
r
s
f
or
te
c
hn
ic
al
reason
s
(m
e
m
or
y ca
pacit
y du
ri
ng trai
ning)
whil
e ret
ai
nin
g t
he
l
ow
er lay
ers
i
n
tra
diti
on
al
c
onvol
ution
al
fas
hi
on.
As
we
ca
n
see in F
igure 1,
t
he
1
x1 convoluti
on
s a
re u
se
d
to
co
m
pu
te
r
educt
ion
s b
e
fore the
ex
pensi
ve
3x3
a
nd
5x5
c
onvoluti
ons.
B
esi
des
bein
g
use
d
as
re
duct
ion
s
,
t
hey
al
so
inclu
de
t
he
use
of
recti
fied
li
near
act
ivati
on
wh
i
ch
m
akes
them
du
al
-
pur
pose
[
24
]
.
One
of
th
e
m
ai
n
ben
e
fits
of
t
his
arc
hitec
ture
is
that
it
m
akes
it
po
ssi
ble
to
dr
am
at
ic
ally
i
ncr
ease
the
nu
m
ber
of
un
it
s
at
each
sta
ge
without
a
n
un
con
t
ro
ll
ed
bl
ow
-
up
in
com
pu
ta
ti
on
al
com
plexity
.
Anothe
r
pr
act
i
cal
ly
us
efu
l
a
sp
ect
of
t
his
de
sig
n
is
t
hat
it
is
in
li
ne
with
the
pr
i
nciple
that
visu
al
in
form
ation
sho
uld
be
processe
d
on
dif
fer
e
nt
scal
es
and
th
en
ag
gr
e
gat
ed
so
tha
t
the
ne
xt
sta
ge
can
sim
ultaneo
us
ly
abstract
fe
at
ur
es
from
diff
ere
nt
scal
es.
The
im
pr
oved
us
e
of
c
om
pu
ta
ti
on
al
resou
rces
al
lo
ws
f
or
inc
r
easi
ng
both
t
he
width
of
each
sta
ge
as
well
as
t
he
nu
m
ber
of
sta
ges
without
getti
ng
into
com
pu
ta
ti
on
al
diff
ic
ulti
es.
A
no
t
her
wa
y
to
util
iz
e
the
incepti
on
arc
hi
te
ct
ur
e
is
to
cr
eat
e
sli
gh
tl
y
infer
io
r,
bu
t c
om
pu
ta
ti
onal
ly
ch
ea
per
ve
rsions
of it
.
Figure
1. A
rch
i
te
ct
ur
e
of
t
he
i
ncep
ti
on m
odul
e w
it
h dim
entio
ns re
duct
ions
[25]
4.2.
Bi
dire
ction
al
long sh
ort
-
ter
m mem
ory
(B
LST
M)
In
ste
a
d
of
r
unni
ng
a
n
RN
N
only
in
the
f
orw
ard
m
od
e
sta
rti
ng
from
the
first
sy
m
bo
l,
we
sta
rt
ano
t
he
r
on
e
from
the
la
st
sy
m
bo
l
ru
nn
i
ng
from
back
to
f
ront.
B
idirect
ion
al
re
cu
r
ren
t
neural
networ
ks
intr
oduc
e
a
hidden
la
ye
r
to
m
or
e
rob
ust
pr
oc
esses
by
passing
i
nfo
rm
ation
in
a
r
ever
se
directi
on.
Fig
ure
2
il
lustrate
s
the
arc
hitec
tur
e
of
a
bi
directi
on
al
rec
urren
t
neural
netw
ork
.
I
n
fact,
t
his
i
s
not
to
o
dissi
m
il
ar
to
the
f
orwa
rd
and
bac
kw
a
rd
recursio
n
we
encou
ntered
above.
T
he
m
ai
n
disti
nctio
n
is
that
in
the
pr
e
vious
case
these
equ
at
io
ns
had
a
sp
eci
fic
sta
ti
sti
cal
m
eaning.
Now
they
are
devoid
of
s
uc
h
easi
ly
acce
ssi
ble
interp
retat
on
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A tec
hnic
al rev
ie
w and
c
omp
arati
ve a
naly
sis
of mac
hin
e le
arnin
g
te
c
hn
i
ques... (
Sa
f
aa L
aqti
b
)
2705
we
ca
n
jus
t
tr
eat
them
as
ge
ner
ic
f
un
ct
io
ns
.
This
tra
ns
it
ion
e
pitom
iz
e
s
m
any
of
t
he
pri
nciples
gu
iding
the
desig
n
of
m
od
ern
dee
p
netw
orks:
first
,
us
e
the
ty
pe
of
functi
onal
dep
e
ndencies
of
cl
assic
al
sta
ti
sti
cal
m
od
el
s,
an
d
th
en use t
he
m
odel
s in
a
ge
ner
ic
for
m
[26]
.
Figure
2. A
rch
i
te
ct
ur
e
of
B
LS
TM
[
27
]
4.3.
Deep
beli
ef
n
etwork
(
DBN
)
To
overc
om
e
t
he
ov
e
rf
it
ti
ng
pro
blem
in
M
ulti
-
Lay
er
Pe
r
ceptr
on
(
MLP
)
,
we
ca
n
set
up
a
DBN
,
do
unsuper
vis
ed
pret
raini
ng
to
get
a
decent
set
of
feature
rep
re
sentat
io
ns
f
o
r
the
in
puts,
the
n
finet
un
e
on
the
trai
ning
se
t
to
act
ually
get
predict
io
ns
from
the
network.
Wh
il
e
weig
hts
of
an
MLP
ar
e
init
ia
li
zed
rand
om
l
y,
a
D
BN u
ses
a
g
ree
dy
la
ye
r
-
by
-
la
ye
r
pret
raini
ng
al
gorithm
to
i
niti
al
iz
e
the
ne
twork
wei
gh
ts
throug
h
pro
bab
il
ist
ic
ge
ner
at
ive
mo
dels
com
po
sed
of
a
visible
la
ye
r
an
d
m
ulti
ple
la
ye
rs
of
stoc
ha
sti
c,
la
te
nt
va
riables,
wh
ic
h
are
cal
le
d
hidden
unit
s
or
feat
ur
e
de
te
ct
or
s.
Re
stric
te
d
Bolt
zm
ann
Ma
chines
(R
BM
)
in
the
D
BN
ar
e
sta
cked,
form
i
ng
a
n
undi
rec
te
d
pro
bab
il
ist
ic
gr
ap
hical
m
od
el
si
m
il
a
r
to
Ma
rk
ov
R
andom
Fiel
ds
(MRF):
the
two
la
ye
rs
are
com
po
sed
of
visible
ne
uro
ns
an
d
the
n
hidden
neur
ons.
The
to
p
two
la
ye
rs
in
a
st
acke
d
RB
M hav
e
undi
rected,
sym
m
e
tric
co
nnect
io
ns betwee
n
t
he
m
an
d
f
orm
an
asso
ci
at
ive m
e
m
or
y, wh
e
reas
lowe
r
la
ye
rs
receiv
e
to
p
-
dow
n,
di
rected
c
onne
ct
ion
s
f
ro
m
the
la
ye
r
a
bove
.
A
hybri
d
m
odel
is
est
ablish
ed
by
sta
ckin
g
up
RB
Ms,
as
il
lustr
at
ed
in
F
ig
ur
e
3
.
The
t
op
tw
o
la
ye
rs
f
or
m
th
e
RB
M
an
d
t
he
lo
we
r
la
ye
r
s
f
or
m
a
directed
belie
f
net
[
28
]
.
T
hi
s
hybr
i
d
m
od
el
is
cal
le
d
a
deep
belie
f
netw
ork
(
DBN
)
.
T
he
deep
-
belie
f
-
networ
k
is
a
si
m
ple,
cl
e
an,
fast
Pyt
hon
i
m
ple
m
entat
ion
of
deep
belie
f
netw
orks
bas
ed
on
bin
a
ry
Re
stric
te
d
Bolt
z
m
ann
Ma
c
hin
es
(RB
M)
.
I
n
our
cas
e,
it
was
based
on
N
um
Py
and
Tens
orFlo
w
li
br
aries
to
ta
ke
adv
a
ntage
of
GP
U
com
pu
ta
ti
on
.
Figure
3. Hy
bri
d
m
od
el
of the
D
BN
afte
r gr
e
edy lay
er
-
wise
le
arn
in
g. Th
e t
op tw
o
la
ye
rs f
or
m
the RBM
an
d
the bott
om
layers form
a
dire
ct
ed
belie
f
n
et
work [
29]
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
3
,
J
une
2020
:
27
01
-
2709
2706
5.
E
X
PERI
MEN
T AND
RES
U
LT
S
5.1.
Data prep
roce
ssing
The
m
et
ho
dolo
gy
disc
us
se
d
i
n
this
pa
per
is
app
li
ed
on
t
he
entire
NS
L
-
K
DD
dataset
.
T
he
NS
L
-
K
D
D
dataset
wa
s
propose
d
t
o
deal
with
in
her
e
nt
pro
blem
s
of
th
e
K
DD
Cu
p
1999
dataset
whic
h
c
on
ta
in
to
o
m
any
redu
nd
a
nt
rec
ords
.
A
n
e
xam
ple
of
d
at
aset
re
cord
is
‘0
tc
p
ftp_data
S
F
49
1
0
0
0
0
0
0
0
0 0
0
0
0
0
0
0
0
0
2
2
0
0
0
0
1
0
0
150
25
0.1
7
0.03
0.1
7
0
0
0
0.05
0
norm
al
’.
As
y
ou
can
s
ee,
data
c
onta
ins
s
om
e
te
xt
values
al
so
.
Pr
e
-
proce
ssing
o
f
ori
gi
nal
NS
L
-
K
DD
dataset
is
ne
cessary
to
m
a
ke
it
a
su
it
a
ble
in
pu
t
f
or
le
arni
ng
m
od
el
s.
W
e
need
t
o
tra
nsfo
rm
the
no
m
inal
feature
s
to
num
eri
c
values
.
O
nl
y
colum
n
nu
m
ber
2(
P
r
oto
c
ol_
ty
pe
),
3(Ser
vices)
,
4(Fla
g) a
nd 42
(A
tt
ack
or
Nor
m
al
)
con
ta
in
s
no
m
inal values
[
30
]
.
5.2.
Ev
aluation
m
etrics
Fo
r
e
valuati
on
pur
po
ses
,
Acc
ur
acy
(
ACC)
,
Pr
eci
sio
n
(P),
Re
cal
l
(R)
m
etr
ic
s
are
us
e
d.
These
m
et
rics
are
cal
culat
ed
by
us
in
g
f
our
diff
e
re
nt
m
eas
ur
es
,
true
posit
ive
(TP),
tr
ue
neg
at
ive
(T
N
),
false
po
sit
iv
e
(F
P
)
and
false
ne
ga
ti
ve
(F
N
).
Acc
ur
acy
is
the
pe
rcen
ta
ge
of
the
rec
ords
num
ber
cl
assifi
e
d
co
rr
ect
ly
ov
er
total
the
rec
ords.
Pre
ci
sion
m
eans
the
pe
rcen
ta
ge
of
yo
ur
res
ults
wh
ic
h
are
re
le
va
nt.
On
t
he
oth
e
r
ha
nd,
r
ecal
l
ref
e
rs
to
the
pe
rcen
ta
ge of
t
ot
al
r
el
eva
nt
res
ul
ts correct
ly
classi
fied by
your
al
gorithm
[31
]
.
Tru
e
P
os
it
iv
e
Rate
(
TPR)
:
al
so
know
n
as
Detect
ion
Ra
te
(D
R)
is
t
he
pe
rcen
ta
ge
of
the
an
om
al
y
record
s
nu
m
ber
c
orrectl
y flag
ged as a
nom
aly o
ve
r
t
he
total
num
ber
of a
nom
al
y recor
ds i
n (4):
A
c
cura
cy
=
TP
+
TN
TP
+
FP
+
FN
+
TN
(1)
P
re
ci
sion
=
TP
TP
+
FP
(2)
Re
call
=
TP
TP
+
Fn
(3)
DR
=
T
PR
=
TP
TP
+
Fn
(4)
False
P
os
it
iv
e
Rate
(
FPR)
:
the
per
c
enta
ge
of
the
nor
m
al
reco
rds
num
ber
wro
ng
l
y
flagg
e
d
as
ano
m
al
y i
s d
iv
ided by the
tot
al
n
um
ber
of nor
m
al
r
ecords i
n (5)
[32].
FPR
=
FP
FP
+
TN
(5)
5.3.
Co
m
pa
r
ati
ve
stu
d
y of t
hree
deep lear
ning
mod
el
s
-
b
as
ed
intrusio
n d
e
te
ction
sy
s
tem
The
researc
h
on
secu
rity
issues
relat
in
g
to
IDS
exists
si
nce
the
b
irt
h
of
c
om
pu
te
r
a
rch
it
ect
ures.
In
rece
nt
days,
ap
plyi
ng
dee
p
le
arn
i
ng
to
I
DS
is
of
pr
im
e
interest
am
ong
sec
ur
it
y
re
se
arch
e
rs
a
nd
s
pe
ci
al
ist
s.
A
com
par
at
iv
e
stud
y
of
th
ree
dee
p
le
ar
ning
m
od
el
s,
In
ce
ptio
n
co
nvol
ution
al
ne
ural
netw
ork
(
CNN
),
Bi
directi
on
al
l
ong
s
hort
-
te
rm
m
e
m
or
y (BLSTM
)
a
nd d
ee
p beli
ef
netw
ork (DB
N)
a
ppli
ed
to
ID
S
.
Tables
1
-
3
il
lu
strat
e
the
accu
racy,
preci
sio
n,
an
d
recall
of
our
th
ree
deep
le
arn
in
g
m
od
e
ls
in
K
DD
+
and
K
D
D
-
21,
as
we
can
see
that
the
ince
pt
ion
C
NN
an
d
BLSTM
e
x
ce
ed
t
he
DBN.
Figures
4
-
7
pr
ov
i
de
a
com
par
ison
of
the
ex
pe
rim
e
ntal
resu
lt
s
of
Tables
1
-
3
.
F
ro
m
the
resu
lt
s,
we
co
uld
fi
nd
that
the
inc
eption
arch
it
ect
ure
CNN
got
the
hi
gh
e
st
overall
accuracy
(
ACC
)
.
Be
sides
,
the
BLSTM
m
od
el
su
r
passe
d
th
e
DB
N
m
od
el
on
bo
t
h
the
ov
e
rall
pr
e
ci
sion
an
d
ove
rall
recall
rate
in
KDD+
an
d
KDD
-
21.
Alth
ough
the
DBN
m
od
el
perform
ed
worse
than
the
oth
e
r
three
m
od
el
s
on
ACC,
pr
eci
sio
n,
recall
rate
in
KDD+
an
d
K
DD
-
21,
it
ob
viously
f
ai
le
d
on
in
for
m
at
ion
on
the
at
ta
ck
.
The
pro
per
ex
plan
at
ion
that
BLSTM
trie
d
a
lot
on
the
whole
se
quence
c
om
pr
ehensi
on
a
nd
I
ncep
ti
on
-
CN
N
co
uld
ext
ra
ct
th
e
key
i
nform
at
ion
m
or
e
qu
ic
kly.
In
ce
ptio
n
-
C
N
N
an
d
BLST
M
per
f
or
m
bette
r
than
DB
N.
In
the
or
y,
DB
N
shou
l
d
be
the
best
m
od
el
bu
t
it
is
ver
y
hard t
o
es
tim
a
te
j
oi
nt pr
ob
a
bili
ti
es accur
at
el
y at
the
m
o
m
ent.
We
fou
nd
that
al
l
three
m
od
el
s
had
go
od
perform
ance
on
the
N
or
m
al
dat
a
and
D
oS
dat
a
and
Pro
be.
Table
4
sho
ws
the
DR
of
K
DD
T
est
+;
Tab
le
5
sh
ows
th
e
FPR
of
K
D
DTest+;
Table
6
shows
the
DR
of
KDDTest
-
21;
Table
7
sho
w
s
the
FPR
of
KDDTest
-
21
.
Re
su
lt
s
in
Tables
4
-
7
com
par
e
DR
an
d
FPR
of
KDDTest+
an
d
K
DD
T
est
-
21
we
can
co
ncl
ud
e
that
the
I
nc
eption
-
CN
N
a
nd
BLS
TM
prov
i
de
bette
r
re
su
lt
s
in
DR
a
nd
FPR
c
om
par
ed
t
o
D
BN,
we
ca
n
fi
nd
that
res
ults
of
R2L
an
d
U
2R
are
relat
ive
ly
qu
it
e
sm
al
l
for
al
l
m
od
el
s,
wh
ic
h
m
igh
t
becau
s
e
of
the
insuff
ic
ie
ncy
of
their
record
s
in
th
e
dataset
.
How
ever,
the
m
od
e
ls
sti
l
l
can
detect
so
m
e
of
them
.
Bi
d
irect
ion
al
long
sh
ort
-
te
rm
m
e
m
or
y
(BLSTM)
an
d
In
ce
ptio
n
-
C
NN
,
are
use
d
t
o
detect
ano
m
al
i
es
in
sequ
e
nce.
The
resu
lt
s
show
that
the
I
nc
eption
-
CN
N
a
nd
BLSTM
s
howe
d
supe
rior
i
ty
ov
er
the o
the
r DBN
m
od
el
.
We
ti
e
that to the f
act
that
In
ce
ptio
n
-
CNN
a
nd
BL
S
TM has the
a
bi
li
ty
to d
efi
ne
norm
al
beh
a
vior
f
ro
m
la
rg
e
dataset
s
and
ca
n
be
us
e
d
to
detect
a
ne
w
unsee
n
thr
eat
.
It
can
be
c
on
cl
ud
e
d
that
if
one
on
ly
wan
ts
t
o
cl
assify
the
ne
twork
t
raffic
as
norm
al
or
at
ta
ck,
I
nce
ption
CN
N
or
B
LSTM
they
w
il
l
b
e
a b
et
te
r
c
hoic
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A tec
hnic
al rev
ie
w and
c
omp
arati
ve a
naly
sis
of mac
hin
e le
arnin
g
te
c
hn
i
ques... (
Sa
f
aa L
aqti
b
)
2707
Table
1.
Acc
uracy
f
or eac
h
m
od
el
Incep
tio
n
-
CNN
BLST
M
DBN
KDDTest
+
8
8
,03
%
8
4
,03
%
7
1
.91
%
KDDTest
–
21
7
3
.98
%
7
5
.36
%
6
6
.73
%
Table
2.
Pr
eci
s
ion
f
or
eac
h
m
od
el
Incep
tio
n
-
CNN
BLST
M
DBN
KDDTest
+
8
5
.90
%
9
3
.98
%
7
3
.86
%
KDDTest
–
21
8
3
.66
%
7
7
.89
%
7
4
,.
6
5
%
Table
3.
Reca
ll
for
eac
h
m
od
e
l
Incep
tio
n
-
CNN
BLST
M
DBN
KDDTest
+
8
5
.58
%
8
6
.01
%
8
0
.67
%
KDDTest
–
21
7
2
.11
%
7
2
.98
%
6
9
,23
%
Table
4.
DR
of KDD
Test+
No
r
m
al
Do
s
Prob
e
R2
L
U2
R
Incep
tio
n
-
CNN
8
8
.46
8
%
6
9
.54
8
%
6
1
.35
7
%
1
8
.57
9
%
2
2
.34
8
%
BLST
M
8
7
.97
5
%
7
1
.57
6
%
6
3
.57
4
%
2
9
.11
0
%
2
4
.02
2
%
DBN
7
9
.69
7
%
6
6
.24
1
%
5
7
.95
7
%
1
6
.43
6
%
1
7
.04
3
%
Ta
ble
5.
FPR of KD
D T
e
st
+
No
r
m
a
l
Do
s
Pr
o
b
e
R2L
U2
R
Incep
tio
n
-
CNN
2
8
.57
6
%
2
8
.62
2
%
6
.06
1
%
2
.17
8
%
0
.06
3
%
BLST
M
6
5
.44
8
%
2
4
.65
9
%
1
0
.32
3
%
7
.81
0
%
0
.08
2
%
DBN
4
6
.53
3
%
2
5
.21
0
8
%
1
4
.07
3
%
6
.07
3
%
1
.06
5
%
Ta
ble
6.
DR of KDD
Te
st
-
21
No
r
m
a
l
Do
s
Pr
o
b
e
R2L
U2
R
Incep
tio
n
-
CNN
9
5
.87
0
%
7
7
.18
2
%
6
7
.24
5
%
2
2
.32
3
%
2
0
.81
9
%
BLST
M
8
2
.45
1
%
6
7
.58
4
%
6
1
.86
5
%
2
0
.56
7
%
2
1
.63
4
%
DBN
7
2
.21
1
%
5
4
.98
7
%
6
0
.77
3
%
1
6
.76
5
%
1
9
.98
4
%
Ta
ble
7.
FPR of KD
D T
e
st
-
21
No
r
m
al
Do
s
Prob
e
R2
L
U2
R
Incep
tio
n
-
CNN
3
6
.87
0
%
1
4
.53
%
2
.35
7
%
0
.47
9
%
0
.06
8
%
BLST
M
5
1
.97
5
%
1
9
.57
6
%
7
.57
4
%
1
.11
0
%
0
.02
2
%
DBN
6
3
.43
2
%
2
2
.65
3
%
7
.32
5
%
2
.97
2
%
1
.08
6
%
Figure
4. Acc
uracy
,
pr
eci
sio
n and recal
l
of
KDDTest+
Figure
5. DR
of KD
DTest+
usi
ng
ince
ption
-
CNN,
BLSTM
, DBN
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
3
,
J
une
2020
:
27
01
-
2709
2708
Figure
6. Acc
uracy
,
pr
eci
sio
n and recal
l
of
KDDTest
-
21
Figure
7. DR
of KD
D
Test
-
21
us
in
g
i
ncep
ti
on
-
CN
N,
BLSTM
, DBN
6.
CONCL
US
I
O
N
Re
centl
y,
D
ee
p
le
ar
ning
f
or
intr
us
io
n
detect
ion
syst
e
m
has
receive
d
m
uc
h
deliberati
on.
In
any
I
DS
,
aud
it
data
sa
m
ples
are
ana
ly
zed
to
set
de
te
ct
ion
ru
le
s
in
highly
m
ob
il
e
node
netw
ork
t
o
protect
agai
nst
nu
m
ber
of
nove
l
at
ta
cks.
The
pr
im
ary
adv
a
nt
age
of
usi
ng
Deep
le
a
rn
i
ng
base
d
i
ntr
us
io
n
detect
ion
syst
e
m
s
is
that
it
is
high
ly
accurate
an
d
a
ble
to
detect
or
cat
e
gori
ze
at
ta
cks
wit
hout
a
ny
en
vironm
ental
influ
ence
.
Diff
e
re
nt
Dee
p
le
arn
i
ng
ba
sed
I
DS
a
ppr
oach
e
s
ha
ve
their
ow
n
be
ne
fits
and
disa
dv
a
ntage
s.
T
he
refor
e
,
consi
der
i
ng the
MANET
sce
na
rios, it
is im
p
or
ta
nt to
c
hoose
a precise
m
eth
od
f
or
im
plem
enting IDS.
This
pa
per
is m
ot
ivate
d
by
the
nee
d
to d
ev
el
op
g
oo
d
trai
ni
ng
al
gorit
hm
s
for
dee
p
arc
hitec
tures
-
bas
e
d
ID
S
,
since
the
se
can
be
m
uch
m
or
e
re
pr
es
entat
ion
al
ly
ef
fici
ent
than
sha
ll
ow
ones
su
c
h
as
S
VMs
a
nd
one
-
hiddenlay
er
ne
ur
al
nets.
wh
ic
h
m
ay
be
pro
ve
d
im
po
rtant
f
or
sel
ect
ing
t
he
ap
propriat
e
m
et
ho
ds
on
ba
ses
of
the
sit
uation
in
MAN
ET.
I
n
this
wo
r
k,
th
e
pr
act
ic
al
pro
blem
s
of
existi
ng
I
DS
hav
e
been
a
ddresse
d
an
d
diff
e
re
nt
Dee
p
Lear
ning
m
od
el
s
(
I
ncep
ti
on
-
CN
N,
BL
ST
M
an
d
Deep
Be
li
ef
m
od
el
)
are
c
om
par
ed
to
s
olv
e
these
pro
blem
s.
The
m
od
el
s
hav
e
bee
n
im
plem
ented
and
te
ste
d
on
N
S
L
-
K
D
D
datase
t.
The
reas
on
beh
i
nd
the
superi
or
it
y
of
I
ncep
ti
on
-
CNN
i
n
ge
ne
r
al
that,
it
’s
th
e
abili
ty
to
de
fine
norm
al
beh
avi
or
from
a
la
rg
e
dataset
an
d
ca
n
be
use
d
t
o
de
te
ct
a
new
un
seen
th
reat
.
T
his
w
ork
ca
n
be
e
xten
ded
in
two
directi
on
s:
first,
i
m
ple
m
ent
oth
er
dee
p
m
od
el
s
and
c
reate
hybr
i
d
m
od
el
s
and
voti
ng
syst
em
s
acro
ss
dif
f
eren
t
m
od
el
s
to
detect
and
rec
ognize
low
fal
se
al
ar
m
threats.
Se
c
ond,
pr
ov
i
de
e
xisti
ng
syst
em
s
with
real
-
w
orl
d
data
f
or
m
ulti
ple
netw
orks
i
n
s
uch
a
way
t
ha
t
the
m
od
el
can
inc
rease
i
ts
accuracy
by
adap
ti
ng
the
def
i
niti
on
of
norm
a
l
act
ivit
ie
s thr
ou
gh
di
ff
e
ren
t
un
-
cal
ibrate
d datase
ts.
REFERE
NCE
S
[1]
E.
Am
iri
,
H.
Ke
shava
rz,
H
.
Heid
ari
,
E.
Moham
ad
i,
and
H.
Mor
ad
za
deh
,
“
Intrusio
n
Dete
c
ti
on
S
y
st
ems
in
MA
NET:
A Re
vie
w,
”
Proce
dia
-
So
ci
al
an
d
Be
ha
vi
oral
Scienc
es
,
vol
.
129
,
pp.
453
–
459
,
20
14.
[2]
S.
La
qt
ib,
K.
E.
Yass
ini
,
M.
Houm
er,
M.
D.
E
.
Ouadghir
i,
and
M.
L.
Hasnaou
i,
“
Im
pac
t
of
m
obil
ity
m
odel
s
on
Optimize
d
Li
nk
Stat
e
Routi
ng
Protocol
in
MA
NET,
”
2016
In
t
ernati
onal
Con
f
ere
nce
on
Wir
eless
Net
works
an
d
Mobil
e
Comm
unic
ati
ons
(
WINCOM)
,
2016.
[3]
S.
La
q
ti
b,
K.
E.
Yass
ini,
and
M.
L
.
Hasnao
ui,
“
Perform
ance
Ev
al
u
a
ti
on
of
Multi
c
ast
Rout
ing
Protocol
s
i
n
MA
NET
,
”
in
In
te
rnational
Conf
ere
nce
on
Ad
va
nce
d
Intelli
g
ent
Syste
ms
for
Sustainabl
e
Dev
el
op
ment
,
Springer
,
Cham,
2018
.
[4]
S.
La
qti
b
,
K.
E.
Yass
ini
,
and
M.
L.
Hasnaoui
,
“
Li
nk
-
state
QoS
routi
ng
protoc
ol
under
var
ious
m
obili
t
y
m
odel
s,
”
Indon
e
sian
Journal
of
El
e
ct
rica
l
Engi
n
ee
ring
and
Computer
Scienc
e
S
ci
en
ce
(
IJE
ECS)
,
vol.
16
,
no.
2
,
pp
.
906
-
916
,
No
v
.
2019
.
[5]
I.
Bu
tun,
S.
Mo
rge
ra
and
R
.
Sa
nkar
,
“
A
Surve
y
of
Intrusion
Dete
c
ti
on
S
y
st
ems
in
W
ire
le
ss
Se
nsor
Networks
,”
IEE
E
Comm
unications
Surve
ys
&
T
utorial
s
,
vol
.
16,
no.
1,
pp.
26
6
-
282,
2014
.
[6]
S.
Iqba
l,
M
.
L
.
M.
Kiah,
B
.
Dha
ghighi
,
M.
Hus
sain,
S.
Kh
an,
M.
K.
Khan,
and
K
.
-
K.
R.
Choo
,
“
On
cl
oud
sec
ur
i
t
y
at
t
ac
ks:
A
ta
xo
nom
y
an
d
int
ru
sion
det
ection
a
nd
pre
vent
ion
a
s
a
servic
e,”
Jo
urnal
of
Net
work
and
Computer
Appl
ic
a
ti
ons
,
vo
l.
74
,
pp
.
98
–
120
,
2016
.
[7]
A.
Fudholi
and
K.
Sopian,
“
Revi
ew
on
Solar
Co
ll
e
ct
or
for
Agric
ult
ura
l
Produce
,
”
Int
ernati
onal
J
ournal
of
Pow
er
El
e
ct
ronics
and
Dr
iv
e
Syst
ems
(
IJP
EDS)
,
vol
.
9
,
no.
1
,
p
p.
414
-
4
19
,
Jan
.
2018
.
[8]
T.
Laagoubi
,
M.
Bouzi
,
and
M.
B
enc
hagr
a,
“
MP
P
T
and
Pow
er
Fact
or
Control
for
Grid
Connec
t
ed
PV
Sy
stems
with
Fuzz
y
Logic
Co
ntrol
lers
,
”
Inte
rn
ati
onal
Journal
of
Powe
r
El
e
ct
r
onic
s
and
Dr
iv
e
Syste
ms
(
IJP
E
DS
)
,
vol.
9,
no
.
1,
p
p.
105
-
113
,
Jan
.
2018
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A tec
hnic
al rev
ie
w and
c
omp
arati
ve a
naly
sis
of mac
hin
e le
arnin
g
te
c
hn
i
ques... (
Sa
f
aa L
aqti
b
)
2709
[9]
A.
Fudholi,
e
t.
a
l
.
,
“
Prim
ar
y
Stu
d
y
o
f
Tr
ac
k
ing
Photovolt
aic
S
y
s
te
m
for
Mobil
e
Stat
ion
in
Ma
lays
ia
,
”
Int
ernati
on
al
Journal
of
Power
El
e
ct
ronics
an
d
Dr
iv
e
S
yste
ms
(
IJP
EDS)
,
vol.
9
,
no
.
1
,
p
p
.
427
-
432
,
Jan
.
2018
.
[10]
A.
Fudholi
and
K.
Sopian
,
“
Revi
ew
on
Ex
erg
y
a
nd
Ene
rg
y
An
aly
sis
of
Sol
ar
Air
Hea
t
er,”
Int
ern
ati
onal
Journal
of
Powe
r E
le
c
troni
cs
and
Dr
ive
Sys
te
ms
(
IJP
EDS)
,
vol.
9
,
no
.
1
,
p
p
.
420
-
426
,
Jan
.
20
18.
[11]
S.
Sura
y
a
,
P.
S.
Sujat
ha
,
and
B.
K.
P,
“
A
Novel
Control
Str
at
eg
y
for
Com
pensa
ti
on
of
Volt
age
Quali
t
y
P
roble
m
i
n
AC Drive
s,”
Inter
nati
onal Journal
of
Powe
r
Elec
t
ronics
and
Dr
ive
Syste
ms
(
IJP
ED
S)
,
vol.
9
,
no
.
1
,
p
p.
8
-
16
,
2018
.
[12]
S.
Sam
adi
,
et.
a
l.
,
“
Optimum
ran
ge
of
angle
tracki
ng
rad
ars:
a
the
ore
ti
c
al
com
puti
ng,
”
Int
ernat
ional
Journal
of
El
e
ct
rica
l
and
C
omputer
Engi
n
e
ering
(
IJE
CE)
,
v
ol.
9
,
no
.
3
,
p
p
.
1
765,
Jan
.
2019
.
[13]
T.
Zaidi
and
R.
Ramprata
p
,
“
Virtua
l
Mac
hi
ne
Alloc
a
ti
on
Policy
in
Clou
d
Com
puti
ng
Envi
ronm
ent
using
CloudSim
,
”
Int
e
rnational
Journal
of
Elec
tric
al
and
Computer
Engi
ne
ering
(
IJ
ECE
)
,
vol
.
8
,
n
o.
1
,
p
p
.
344
-
45
4
,
Jan.
2018
.
[14]
G.
S.
N.
Rao
,
et
.
al
.
,
“
D
y
namic
Ti
m
e
Sli
ce
Cal
culat
ion
for
Round
Robin
P
roc
ess
Scheduli
ng
Us
ing
NO
C,
”
Inte
rnational
Jo
urnal
of El
e
ct
ri
c
al
and
Comput
er
Engi
n
ee
ring
(
IJE
CE)
,
vol
.
5
,
no
.
6,
p
p.
1480
-
148
5
,
Jan
.
2015
.
[15]
A.
Oukennou,
A
.
Sanda
li
,
and
S.
El
m
oum
en,
“
Coordina
t
ed
Pla
cem
ent
and
Set
ti
ng
of
FA
CTS
in
Elec
tr
ic
a
l
Networ
k
base
d
on
Kala
i
-
s
m
orodinsky
B
ar
gai
ning
Solut
ion
and
Volta
ge
De
via
ti
on
Ind
ex,”
I
nte
rnational
Jou
rnal
of
El
e
ct
rica
l
and
Computer
E
ngine
ering
(
IJECE)
,
vol. 8, no.
6,
p
p
.
4079
-
408
8
,
Jan
.
2018
.
[16]
A
.
Othm
an,
N
.
I
.
S
.
Shaar
i,
A
.
M.
Zobi
la
h
,
N
.
A
.
Shairi,
Z
.
Z
aka
r
ia,
“
Design
of
Co
m
pac
t
Ultra
W
id
eba
nd
Antenna
f
or
Microwa
ve
Med
ic
a
l
Im
agi
ng
Applicati
on
,
”
Indo
nesian
Journal
of
El
e
ct
ri
cal
En
gine
ering
and
C
omputer
Sci
en
c
e
(
IJE
ECS)
,
v
ol
.
1
5,
n
o
.
3
,
pp
.
119
7
-
1202
,
Sep
.
20
19.
[17]
N
.
Bat
a
y
e
v,
“
Axial
Com
pre
ss
or
Fouling
Dete
cti
on
for
Gas
Turbi
ne
Drive
n
Ga
s
Com
pre
s
sion
Unit
,”
Indon
esian
Journal
of
Elec
t
rical
Engi
ne
erin
g
and
Computer
Sci
en
ce
(
IJE
ECS
)
,
v
ol
.
15
,
n
o
.
3
,
pp.
1257
-
1263
,
Sep
.
2019
.
[18]
dje
ll
ou
l
khe
ira
,
“
Perform
anc
e
of
cha
nne
l
sel
e
ct
ion
used
for
Multi
-
class
EEG
signal
class
ifi
cation
of
m
oto
r
imager
y
,”
Indo
nesian
Journal
of
E
lectric
a
l
Engi
ne
ering
an
d
Computer
Sc
ie
nc
e
(
IJE
ECS)
,
v
ol
.
15,
n
o
.
3
,
pp.
1305
-
1312
,
Sep
.
2019
.
[19]
S
.
H
.
Aham
m
ad,
“
Auto
m
at
ic
Se
gm
ent
at
ion
of
Spinal
Cord
Diff
usion
MR
I
m
age
s
for
disea
se
loc
ation
findi
ng
,”
Indone
sian
Journal
of
El
e
ct
ri
ca
l
Engi
ne
ering
and
Computer
Sci
ence
(
IJE
ECS)
,
v
ol
.
15,
n
o
.
3
,
pp.
1313
-
1321
,
Sep
.
2019
.
[20]
M.
La
la
ou
i,
A.E.
Afia,
R.
Chiheb,
“
A
Self
-
Tuned
Sim
ula
te
d
Annea
l
ing
Algorit
h
m
U
sing
Hidden
Markov
Model,
”
Inte
rnational
Jo
urnal
of El
e
ct
ri
c
al
and
Comput
er
Engi
n
ee
ring
(
IJE
CE)
,
vol
.
8
,
no
.
1,
p
p.
291
-
298
,
Jan.
2018
.
[21]
A.
Fudholi
and
K.
So
pia
n,
“
Rev
ie
w
on
Solar
Co
ll
e
ct
or
for
Agri
c
ult
ura
l
Produce
,
”
Inte
rnat
ional
J
ournal
of
Powe
r
El
e
ct
ronics
and
Dr
iv
e
Syst
ems (
IJP
EDS)
,
vol
.
9
,
no.
1
,
p
p
.
414
-
4
19
,
Jan
.
2018
.
[22]
A.F.
Maji
d,
Y
.
Mukhlis
,
“
Ape
rture
Coupli
ng
Rec
t
angul
ar
Slo
tt
ed
Circ
u
la
r
Ri
ng
Microstri
p
Patc
h
Antenna
,
”
Indone
sian
Journal
of
Elec
tri
ca
l
Engi
ne
ering
a
nd
Computer
Sci
ence
(
IJE
ECS)
,
v
ol
.
15,
n
o
.
3
,
pp
.
1419
-
1427
,
Sep
.
2019
.
[
23]
Christi
an
Sz
eged
y
,
W
ei
Li
u
,
Y
angqi
ng
Jia
,
Pi
e
rre
Serm
ane
t
,
S
cot
t
R
ee
d
,
Drag
om
ir
Anguelov,
Dum
it
ru
Erha
n
,
Vince
nt
Vanhou
cke
,
Andrew
R
abi
novi
ch
,
“
MPPT
Going
dee
per
with
convol
u
ti
ons
,
”
Conf
ere
n
ce
on
Computer
Vi
sion a
nd
Patte
rn R
ec
ogn
it
ion
(
CVP
R)
.
IE
E
E
C
onfe
renc
e
on
20
15
.
[24]
Fudholi,
Ahm
ad,
and
Kam
aru
zz
aman
Sopian
,
“
Revi
ew
on
exe
rg
y
and
ene
r
g
y
an
aly
sis
of
solar
ai
r
he
at
er
,”
Inte
rnational
Jo
urnal
of Powe
r
El
e
ct
ronics
and
Dr
iv
e
Syst
ems
(
IJP
EDS)
,
v
ol
.
9
,
n
o
.
1
,
pp
.
420
-
42
6
,
2018
.
[25]
Kuchibha
t
la
,
Sa
m
ant
haka
Mani
,
D.
Padm
ava
thi
,
and
R.
Sriniva
s
a
Rao
,
“
Eff
e
ct
o
f
Carr
ie
r
Frequ
e
nc
y
in
Grid
Inter
Connec
t
ed
W
ind
Sy
st
em
with
SS
FC
Control
le
r
,
”
Inte
rnat
ional
Journal
of
Pow
er
El
ectronics
a
nd
Dr
iv
e
Syste
m
s
(
IJP
EDS)
,
v
ol
.
9
,
n
o
.
3
,
pp
.
1349
-
1355
,
2018
.
[26]
Nee
thu
B
,
“
Cla
ss
ifi
cation
of
int
r
usion
det
ection
dat
ase
t
using
m
ac
hine
l
ea
rning
a
pproa
che
s,
”
Inter
nati
onal
Journal
of
E
le
c
tronic
s a
nd
Computer
Sc
i
enc
e
Engi
n
ee
rin
g,
pp
.
1044
-
105
1,
2012
.
[27]
Ramkum
ar,
Purigil
la
V
enka
t
a,
a
nd
Munaga
la
Su
r
y
a
Kala
v
at
h
i
,
“
Frac
ti
on
al
Order
PID
Control
le
d
Inte
rl
ea
ved
Boo
st
conve
rt
er
Fed
Shunt
Acti
ve
F
il
te
r
S
y
stem
,”
I
nte
rnational
Jo
urnal
of
Powe
r
El
e
ct
ronics
an
d
Dr
iv
e
Syste
m
s
(
IJP
EDS)
,
vol.
9
,
no
.
1
,
pp
.
126
-
138
,
2018
.
[28]
Pane,
S
y
afr
ia
l
Fachr
i,
et
a
l.
,
“
RF
ID
-
base
d
convey
or
belt
for
improve
ware
h
ouse
oper
ations
,
”
TEL
KOMNIKA
(
Tele
communic
ati
on
Computing
El
e
ct
ronics
and
Control
)
,
vol. 17
,
no
.
2
,
pp
.
794
-
800,
2019
.
[29]
Mohs
en,
Mow
afa
k
K.
,
e
t
a
l.
,
“
El
ectroni
ca
l
l
y
cont
ro
ll
ed
rad
ia
ti
on
pa
tt
ern
l
ea
k
y
wav
e
antenna
arr
a
y
for
(C
band)
app
li
c
at
ion
,
”
TEL
KOMNIKA
(
Tele
co
mm
unic
ati
on
Co
mputing
Elec
tro
nic
s
and
Contro
l
)
,
vol.
17
,
no.
2
,
pp
.
573
-
579
,
20
19.
[30]
Raha
rdj
a,
Untun
g,
Eka
Purnam
a
Hara
hap
,
and
Shy
lv
ia
Ra
tna
Dew
i
,
“
The
stra
te
g
y
of
enha
nc
ing
article
citation
an
d
H
-
inde
x
on
SINT
A
to
improve
te
rtiar
y
r
eputat
i
on
,”
TEL
KOMN
IKA
(
Tele
communic
ati
on
Comp
uti
ng
Elec
troni
cs
and
Control
)
,
vo
l.
17
,
no
.
2
,
pp
.
6
83
-
692,
2019
.
[31]
Za
inur
i,
Akhm
a
d,
e
t
a
l.
,
“
VRLA
bat
t
er
y
st
at
e
of
he
alth
es
t
i
m
at
ion
base
d
o
n
cha
rg
ing
ti
m
e
,
”
TEL
KOMNIK
A
(
Tele
communic
ati
on
Computing
El
e
ct
ronics
and
Control
)
,
vol. 17
,
no
.
3
,
pp
.
1577
-
1583,
2019
.
[32]
H.
Sett
i
,
et
.
al
.
,
“
A
new
conf
ig
ura
ti
on
of
a
p
ri
nte
d
dip
le
x
er
de
signed
for
DCS
and
ISM
bands
,
”
TEL
KOMNIK
A
(
Tele
communic
ati
on
Computing
El
e
ct
ronics
and
Control
)
,
vol. 17
,
no
.
3
,
p
p
.
1090
-
1095
,
Jan
.
2019
.
Evaluation Warning : The document was created with Spire.PDF for Python.