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.
3315
~
3322
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v10
i
3
.
pp3315
-
33
22
3315
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om/i
nd
ex
.ph
p/IJ
ECE
LSTM d
ee
p l
earn
ing meth
od for n
etwork in
trusion
detecti
on
sy
stem
Alaeddi
ne B
oukh
alfa
,
Ab
de
rrahim
Abdel
laoui
, Na
bil
H
mi
na
,
H
ab
ib
a Ch
aoui
S
y
stem E
ng
ine
er
ing
L
abor
at
or
y
,
AD
SI T
ea
m
,
Nat
iona
l
School
of
Applie
d
Sc
ie
nc
e
s,
Ibn Tofail Uni
ver
sit
y
,
Morocc
o
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
J
un
15
, 2
019
Re
vised
Dec
4
,
20
19
Accepte
d
Dec
14
, 20
19
The
sec
ur
ity
of
t
he
net
work
has
bec
om
e
a
primar
y
con
ce
rn
for
or
gani
z
at
ions.
Atta
ck
ers
use
d
i
ffe
ren
t
m
ea
ns
to
disrupt
serv
ices,
th
ese
v
ari
ous
a
tt
a
cks
push
to
thi
nk
of
a
ne
w
wa
y
to
b
loc
k
the
m
a
ll
in
one
m
anne
r.
In
add
it
ion,
th
es
e
int
rusions
c
an
c
hange
and
p
e
netrat
e
the
dev
ices
of
sec
u
rity
.
To
solve
the
se
issues,
we
sugge
st,
in
thi
s
pape
r
,
a
new
idea
for
Network
Intrusion
Detect
i
o
n
S
y
stem
(NID
S)
base
d
on
Long
Short
-
Te
rm
Mem
ory
(LSTM)
t
o
rec
ogni
z
e
m
ena
ce
s
and
to
obta
in
a
long
-
t
er
m
m
e
m
ory
on
th
em,
in
orde
r
to
s
t
op
the
new
at
t
ac
ks
th
at
are
l
ike
the
exi
sting
ones,
and
a
t
th
e
sam
e
ti
m
e,
to
h
a
ve
a
singl
e
m
ea
n
to
blo
ck
int
rusions.
Acc
o
rding
to
th
e
res
ult
s
of
the
experim
ent
s
of
det
e
ct
ions
th
at
we
have
r
ea
l
ize
d,
the
A
cc
ura
c
y
rea
ch
es
up
to
99.
98
%
and
99.
93
%
for
respe
ct
ive
l
y
th
e
class
ifi
cation
of
two
cl
asses
and
seve
ral
c
la
ss
es
,
a
lso
the
False
P
ositi
ve
Ra
te
(FP
R)
rea
ch
es
up
to
onl
y
0
,
068
%
a
nd
0,
023
%
for
respe
ctively
the
class
ifi
c
at
io
n
of
two
cl
asses
and
seve
ra
l
cla
ss
es,
which
prove
s
tha
t
the
p
roposed
m
odel
i
s
eff
ec
t
ive
,
it
h
as
a
gre
a
t
ab
il
i
t
y
t
o
m
emorize
and
diffe
ren
ti
a
te
bet
wee
n
norm
al
tra
ffic
and
a
tt
a
c
ks
,
and
it
s
ide
ntification
is
m
ore
accura
t
e th
an
oth
er
Ma
chi
n
e
L
ea
rn
ing
c
la
ss
i
fie
rs
.
Ke
yw
or
d
s
:
Deep Lear
ning
LSTM
Ma
chine
Le
ar
ni
ng
NIDS
RNN
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
:
Alaed
din
e B
ou
kh
al
fa,
Syst
e
m
En
gin
e
erin
g
La
borato
ry
,
A
DSI T
ea
m
,
Nati
on
al
Scho
ol of
A
pp
li
ed
S
ci
ences
,
Ibn
T
ofai
l U
niv
ersit
y
,
Ken
it
r
a,
Mo
ro
cc
o.
Em
a
il
:
al
aedd
ine.
boukhal
fa@
gm
ail.co
m
1.
INTROD
U
CTION
Nowa
days,
th
e
world
is
e
xperie
ncin
g
a
great
revoluti
on
in
the
fiel
d
of
inf
orm
ation
te
chnolo
gy,
ever
y
body
is
exch
a
ngin
g
co
nt
inu
ously
inf
orm
at
ion
acro
ss
t
he
net
work.
Th
is
i
m
plies
the
e
sta
blishm
ent
of
ne
w
too
ls
a
nd
m
echan
ism
s
of
pre
ven
ti
on
an
d
de
te
ct
ion
,
a
nd
th
e
stre
ng
t
hen
i
ng
of
th
os
e
t
hat
exist,
li
ke
Ne
twork
In
tr
us
i
on
Dete
ct
ion
Syst
em
(N
I
DS),
in
ord
er
to
e
nh
a
nce
secur
it
y
an
d
protect
the
net
work
f
ro
m
intru
si
on
s
.
The
f
unct
ion
of
a
NIDS
is
to
ob
s
er
ve,
e
valu
at
e
and
cl
assif
y
traff
ic
transit
ing
th
rou
gh
t
he
netw
ork,
it
is
based,
in
ad
va
nce,
on
est
ablishe
d
m
et
hods
a
nd
te
c
hn
i
qu
e
s
in
ord
er
to
dif
fe
ren
ti
at
e
betwee
n
norm
al
and
s
uspic
ious
traff
ic
.
F
ur
t
herm
or
e,
at
ta
cker
s
are
at
tract
ed
by
inf
or
m
at
ion
and
kn
ow
le
dge
passing
thr
ough
the
net
wor
k,
a
nd
to
ex
plo
it
an
d
prof
it
f
ro
m
them
,
they
are
forced
to
overc
om
e
ob
sta
cl
es
and
barriers
of
secur
it
y
by
cre
at
ing
new
at
ta
cks,
a
nd
ev
olv
i
ng
t
he
existi
ng
ones
.
Wh
il
e
the
c
urren
t
NIDS
a
re
no
t
e
voluti
onar
y,
their
ide
ntifi
cat
ion
al
gorithm
s
do
no
t
progress
to
identify
auto
m
at
ic
ally
new
m
enaces,
w
hic
h
pus
hes
us
to
think
a
bout
a
dvance
d
and
intel
li
gen
t
m
et
ho
ds
of
detect
ion
that
can
ide
ntify
ne
w
at
ta
ck
s
an
d
acc
om
pan
y
the
pr
ogr
ession
of the e
xisti
ng
on
e
s.
More
ov
e
r,
at
ta
cks
ca
n
be
of
diff
e
re
nt
ty
pes,
li
ke
D
oS
(
De
nial
-
of
-
Ser
vice
)
an
d
U
2R
(
Us
er
to
R
oot)
et
c…,
this
pro
blem
of
div
er
s
it
y
le
ads
us
to
fin
d
a
re
so
l
ution
to
detect
a
nd
sto
p
t
hem
al
l
in
a
uniq
ue
way.
Currentl
y,
Dee
p
Lear
ning
is
exp
e
rienci
ng
huge
s
uccess
i
n
sever
al
do
m
ain
s,
it
is
a
set
of
te
ch
niques
use
d
to
recog
nize
obj
e
ct
s,
extract
inf
or
m
at
ion
hidd
en
in
the
data,
and
m
ake
pr
edict
ive
analy
tics
[1
]
,
one
of
these
m
et
ho
ds
cha
ra
ct
erized
by
it
s
long
-
te
rm
m
e
m
or
y
is
the
L
ong
S
hort
-
Te
rm
Me
m
or
y
(LST
M
)
[
2]
.
And,
t
o
s
olv
e
the
issues
ci
te
d
ab
ov
e
,
we
pro
pose
in
this
pap
e
r
a
ne
w
a
ppr
oach
f
or
N
ID
S
base
d
on
the
Dee
p
Lear
ning
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
:
33
1
5
-
33
2
2
3316
m
et
ho
d
LSTM
,
wh
ic
h
will
reco
gniz
e
at
ta
cks
an
d
kee
p
a
lon
g
-
te
rm
m
e
m
or
y
of
the
m
,
in
or
de
r
to
bl
ock
the
oth
e
r
new
at
ta
cks,
and
a
t
the
sa
m
e
time,
will
deal,
with
a
sing
le
way,
with
al
l
t
yp
e
of
the
se
at
ta
cks
.
To
ver
ify
t
he
e
ff
ect
ive
ness
of
our
pro
posed
m
et
ho
d
L
STM
for
NIDS
we
app
ly
it
on
NSL
K
DD
[
3]
dat
aset
,
and
we
giv
e
a
com
par
ison
of
it
s
capaci
ty
to
m
e
m
or
iz
e
and
detect
in
trusion
with
the
fam
ou
s
M
achine
Learn
i
ng
cl
assi
fiers
li
ke
S
upport
Vect
or
Ma
chine
(
SV
M
),
K
-
Near
e
st
Neighb
or
s
(KN
N)
,
and
Decisi
on
Trees
.
The
pa
pe
r
is
s
tructu
red
a
s
f
ol
lows
.
In
sect
i
on
2,
we
gi
ve
a
su
m
m
ary
of
the
relat
ed
work.
W
e
de
fine
our
exp
e
rim
ental
env
i
ronm
ent
i
n
sect
ion
3.
S
ect
ion
4
is
re
serv
e
d
f
or
the
expos
ur
e
of
resu
lt
s
an
d
a
na
ly
sis.
At the e
nd, se
c
ti
on
5
a
nnounc
es the c
oncl
us
i
on and t
he fut
ure
work
.
2.
R
EL
ATED
W
ORK
The
ex
plo
it
at
ion
of
Dee
p
L
earn
i
ng
in
a
NIDS
was
de
bated
by
the
auth
or
s
in
[
4
]
,
on
e
of
Dee
p
Learn
i
ng
m
et
ho
ds
Self
-
Tau
ght
Lear
ning
(
STL)
was
e
va
luate
d
on
NSL
KDD
[3
]
,
wh
ic
h
is
a
ba
sis
that
con
ta
in
s
traf
fic
records
of
t
he
netw
ork.
T
he
stud
y
com
par
e
d
the
le
vel
of
r
ecognit
ion
with
an
a
ncient
m
et
hod
of
cl
assi
ficat
ion
S
of
t
-
Ma
x
R
egr
es
sio
n
(
SMR
),
the
res
ults
of
the
e
val
ua
ti
on
c
onfirm
ed
that
STL
i
de
ntifie
s
at
ta
cks
bette
r.
This
pro
posal
pr
ese
nts
ef
fecti
vely
a
re
m
ark
able
ste
p
to
separ
at
e
norm
al
traff
ic
f
ro
m
doub
t
ful
traff
ic
,
ex
ce
pt t
hat the
so
l
utio
n was
no
t
r
eal
ly
r
eal
iz
ed by a
real N
IDS.
Appro
xim
at
e
ly,
the
sa
m
e
dem
arch
e
wa
s
a
dopted
in
[
5
]
,
the
ver
si
on
previo
us
of
NS
L
KDD
cal
le
d
KDD
C
up
99
[
6
]
,
was
em
plo
ye
d
to
co
m
par
e
the
preci
sion
of
tra
ff
ic
i
den
ti
ficat
ion
.
T
he
res
ul
ts
of
exp
e
rim
entat
io
ns
pro
ve
d
that
Su
pp
or
t
Vect
or
Ma
chi
ne
m
et
hod
base
d
on
Re
stric
te
d
Bolt
z
m
ann
Ma
chin
e
(S
VM
-
R
BM
s)
[
7
]
dete
rm
ines
certai
nly
the
na
ture
of
the
t
ra
ff
ic
bette
r
t
han
old
cl
assifi
cat
ion
al
go
rithm
s,
al
so
it
con
su
m
es
m
ini
m
u
m
tim
e
f
or
treat
in
g
a
la
rg
e
m
ass
of
data.
The
a
uthors
did
no
t
ju
sti
fy
the
reasons
f
or
the choic
e
of the
us
e
d Deep
Learn
i
ng m
et
ho
d.
Anothe
r
dif
fere
nt
con
ce
pt
wa
s
plann
e
d
to
ca
tc
h
cod
e
in
j
ect
ion
at
ta
cks
ass
ociat
ed
with
th
e
JavaScr
i
pt
cod
e
in
the
pa
per
[
8
]
,
a
ne
w
com
bin
at
ion
of
Dee
p
Lear
ning
na
m
ed
Hybr
i
d
De
ep
Lear
nin
g
Netw
ork
(
H
DL
N
)
was
create
d.
Perfo
rm
ances
of
this
la
tt
er
wer
e
j
ud
ge
d
ac
cordin
g
to
t
wo
le
vels,
at
the
first,
relat
ively
to
the
num
ber
of
hidden
la
ye
rs
,
the
num
ber
of
filt
ers
an
d
nu
m
ber
of
ne
uro
ns
,
the
res
ults
pro
ved
th
at
accuracy
increases
as
th
e
nu
m
ber
of
f
il
te
rs
increases
,
seco
nd
ly
,
it
was
co
nfr
on
te
d
with
oth
er
t
rad
it
io
nal
cl
assifi
ers,
the
m
ark
ed
accuracy
wa
s
cl
early
the
gr
ea
te
st.
In
the
e
nd,
they
af
firm
ed
that
the
an
te
ceden
t
w
ork
wa
s
i
m
pr
oved
in
t
erm
s
of
accur
acy
by
this
ne
w
m
od
el
.
Th
e
effor
t
pro
vi
ded
is
c
on
si
de
rab
le
,
e
xce
pt
that,
the
so
l
utio
n
w
as
lim
it
ed
to
inj
ect
io
n
at
t
ac
ks
at
ta
ched
to
J
avaSc
ript
c
od
e
,
an
d
it
is
no
t
ded
ic
at
e
d
to
a
no
t
he
r
ty
pe
of att
ack.
The
a
uthors
e
xpose
,
in
t
he
pa
per
[
9
]
,
a
ne
w
idea
of
Dee
p
L
earn
i
ng,
w
hich
associat
es
A
ut
o
-
E
nc
od
e
r
and
Dee
p
Be
li
ef
Netw
ork
(D
B
N)
.
T
he
Au
t
o
-
E
nc
oder
was
us
e
d
for
the
pu
r
po
s
e
of
dec
r
easi
ng
the
dim
ensional
it
y
of
data
a
nd
ide
ntifyi
ng
the
pr
i
ncipal
featur
e
s
of
it
,
wh
il
e
the
DB
N
had
the
m
issi
on
of
detect
ing
t
he
dubious
c
ode.
The
te
st
of
the
new
sug
ge
sti
on
of
m
od
el
was
done
w
it
h
the
dataset
K
D
D
Cup
99
[
6
]
,
th
e
assessm
ent
of
re
su
lt
s
was
c
om
par
ed
with
on
ly
a
sin
gle
DBN.
T
he
at
ta
inm
ent
has
an
nounce
d
that
the
new
m
et
ho
d
is
com
plete
ly
m
or
e
accurate
with
le
ss
consum
ption
of
tim
e.
Howev
e
r,
the
a
ut
hors
di
d
no
t
sp
eci
fy
w
hy
they p
ref
e
rr
e
d
to
co
m
bin
e
DBN
a
nd
Au
t
o
-
Enc
oder
to fo
r
m
this hyb
rid.
In
t
he
w
orl
d
of
the
I
nter
net
of
Things
(IoT
),
ano
t
her
dis
per
s
ed
m
od
el
to
re
veal
intr
us
io
ns
was
op
te
d
in
[
1
0
]
.
O
n
eac
h
no
de
of
fog
-
to
-
t
hings
ne
tw
orks
,
the
Dee
p
Learn
i
ng
has
be
en
set
up
i
n
orde
r
to
get
aut
onom
y
to
disclose
th
r
eat
s
and
tra
ns
f
er
the
pa
ram
eter
s
to
nei
ghbo
ur
i
ng
nodes
,
so
that
they
can
upda
te
them
sel
ves
against
m
enaces.
T
he
obj
ect
ive
was
to
sp
ee
d
up
t
he
id
enti
ficat
ion
a
nd
ge
t
a
quic
k
a
nd
local
up
date
of
node
s
par
am
et
ers.
Th
e
auth
or
s
valid
at
ed
the
su
cce
ss
of
the
m
odern
str
uctu
re
vi
a
NS
L
K
DD,
and
t
h
ey
convi
nced
that
this
con
c
eption
is
m
or
e
eff
ic
ie
nt
tha
n
the
ot
her
ce
ntrali
zed.
T
he
desig
n
pro
vi
des
a
fasci
nating
a
nd
adv
a
nce
d
visio
n
in
the
fiel
d
of
at
ta
cks
detect
ion
,
but
th
e
propose
rs
did
no
t
sp
eci
fy
e
xact
ly
the
ap
plied
Deep
Learn
i
ng m
et
ho
d.
A
subje
ct
of
se
cur
it
y
m
on
it
or
i
ng
in
a
Bi
g
Da
ta
con
te
xt
was
deb
at
e
d
by
the
m
anu
script
[
1
1
]
.
Firstl
y,
the
auth
ors
e
xp
la
ine
d
the
necessit
ie
s
of
m
on
it
or
in
g
of
sec
ur
it
y,
wh
ic
h
are
a
nt
ic
ipati
on
pro
blem
s,
the
a
dap
ta
ti
on
of
de
vices
a
nd
secu
rity
too
ls
to
a
la
r
ge
am
ount
of
d
at
a,
t
he
rap
i
d
detect
ion
of
t
he
a
bnor
m
al
al
arm
and
the
est
ablishm
ent
of
a
ppr
opriat
e
diag
nosis
of
al
ert,
the
li
m
it
a
tio
n
of
sec
ur
it
y
correla
ti
on
al
gorithm
s
to
m
erely
a
fe
w
com
po
ne
nts
and
not
al
l
network
com
pone
nts.
Sec
ondly,
they
exh
ibit
a
syst
e
m
to
m
on
it
or
secur
it
y,
w
hich
le
a
ns
on
the
ag
gr
e
gatio
n
of
Bi
g
Data,
in
te
gr
at
io
n
a
nd
extracti
ng
knowle
dge
for
de
ci
sio
n
m
aking
.
La
stl
y,
they
descr
i
be
so
m
e
cor
r
el
at
ion
al
gorithm
s
to
insp
ect
the
data.
The
pa
pe
r
an
nounce
s
the
desi
gn
of
a
secur
it
y
m
on
it
or
in
g
syst
em
for
a
la
r
ge
m
ass
of
data,
bu
t,
it
did
not
ta
ke
into
c
onsid
erati
on
the pr
ogress o
f
the c
orrelat
ion m
e
tho
ds t
o f
ollow
t
he
e
vo
l
ution o
f
t
he
at
ta
cks.
On
e
m
or
e
visi
on
to
m
on
it
or
the
sec
uri
ty
of
the
In
te
rn
et
of
T
hings
(Io
T)
was
discu
s
sed
i
n
[
1
2
]
,
a
la
rg
e
qua
ntit
y
and
a
div
e
rsi
ty
of
log
s
of
se
cur
it
y
was
coll
ect
ed
from
el
e
ct
ronic
app
a
rat
us
of
us
e
rs,
herea
fter,
they
were
gat
her
e
d
i
n
a
distrib
uted
way
usi
ng
t
he
fr
am
e
work
Ha
doop.
The
data
was
of
di
ff
e
ren
t
f
or
m
at
s,
the
operati
on
of
sta
ndar
dizat
ion
w
as
car
ried
ou
t
to u
nify
th
e
m
,
the
proces
s
of
a
ggreg
at
i
on
an
d
co
rr
el
at
ion
w
as
cond
ucted
by
Com
plex
Eve
nt
P
ro
ces
sin
g
(CEP)
m
et
ho
d,
the
outc
om
e
was
prese
nted
a
doptin
g
de
velo
ped
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
LS
TM
d
e
ep
lea
rn
in
g meth
od f
or
network…
(
Alaed
din
e
B
oukha
lf
a
)
3317
too
ls
f
or
dis
pl
ay
ing
.
T
he
paper
is
m
ai
nly
or
ie
nted
to
wards
visu
al
iz
in
g
lo
gs
without
any
interve
ntio
n
a
gainst
m
enaces.
O
ur
work
pr
e
sents
a
new
a
pproac
h
com
par
ed
t
o
the
oth
e
r
so
l
utions,
it
is
ori
ented
to
m
e
mo
rize
a long
-
te
rm
attacks
i
n order
to disco
ver the
new o
t
her
s
, a
nd to
d
eal
with
al
l i
ntru
sio
ns i
n
a
uniq
ue
m
ann
er
.
3.
E
X
PERI
MEN
TAL EN
VI
R
O
N
MENT
In
this
sect
io
n,
we
de
scrib
e
the
data
fo
r
exp
erim
entat
i
on,
the
evalua
ti
on
ind
ic
at
ors
,
the
wor
k
env
i
ronm
ent, the
a
dopted
m
eth
od a
nd the
com
par
ed
m
et
ho
ds
.
3.1.
D
ataset a
nd
p
re
-
tre
at
me
nt
s
3.1.1
.
D
atase
t
To
eval
uate
our
m
od
el
,
we
us
e
d
the
NS
L
KDD
[
3
]
dataset
.
As
m
entio
ned
a
bove
in
the
pr
e
vi
ous
sect
ion
,
it
is
a
der
i
ved
ver
si
on
from
KD
D
Cup
99
[
6
]
,
that
gro
up
s
netw
ork
traf
fic
colle
ct
ed
by
1998
D
ARP
A
ID
S
[
4
]
.
N
SL
KDD
co
ntains
norm
al
reco
rd
s
,
and
r
eco
rds
of
at
ta
cks
nam
e
ly
:
Do
S
(D
e
ni
al
-
of
-
Se
rv
ic
e
)
wh
ic
h
des
tr
oy
the
se
r
vice
avail
abili
ty
[
13
]
,
Pro
be
wh
ic
h
ext
racts
detai
le
d
inf
or
m
at
ion
from
t
he
ser
ve
rs
[
14
]
,
U2
R
(U
se
r
to
Ro
ot)
wh
ic
h
try
to
exp
l
oit
vu
l
ner
a
bi
li
t
ie
s
in
the
syst
e
m
in
or
de
r
to
obta
in
s
up
e
r
us
er
pr
i
vileges
[
15
]
,
and
R2
L
(Rem
ote
to
Local)
wh
ic
h
se
nd
p
ackets
to
a
m
a
chine
over
a
ne
twork
w
ho
ha
ve
no
acco
unt
on
i
n
order
t
o
le
ad
to
vulne
rab
il
it
y
issues
an
d
ac
cess
secu
re
in
form
ation
[
16
]
.
The
distrib
ution
is
il
lustrate
d
i
n
Table
1
an
d
Ta
ble
2,
Ta
ble
1
sh
ows
t
he
distr
ibu
ti
on
in
t
wo
cl
asses,
w
he
re
as
Table
2
s
hows
the
distri
buti
on
in
five
cl
asses
.
Table
1
. Dist
ri
bu
ti
on
of
dataset
in
tw
o
cl
asse
s
Tr
af
f
ic
Nu
m
b
e
r
o
f
sa
m
p
le
s
No
r
m
al
6
7
3
4
3
Attack
5
8
6
3
0
Total
1
2
5
9
7
3
Table
2
. Dist
ri
bu
ti
on
of
dataset
in
fi
ve
cl
ass
es
Tr
af
f
ic
Nu
m
b
e
r
o
f
sa
m
p
le
s
No
r
m
al
6
7
3
4
3
Attack
Do
S
4
5
9
2
7
Prob
e
1
1
6
5
6
R2
L
995
U2
R
52
Total
1
2
5
9
7
3
The
dataset
does
not
in
cl
ude
redu
nd
a
nt
records,
it
c
on
ta
in
s
43
c
olu
m
ns
,
42
c
olu
m
ns
def
i
ne
t
he
c
har
act
e
risti
cs
of
t
he
re
co
rd
i
ng
as
D
ur
at
io
n,
Protoc
ol_
Ty
pe
,
Ser
vice,
Flag,
et
c.
..,
an
d
one
c
olu
m
n
defi
nes
if
it
is
a
norm
al
record
or a
n
at
t
ack,
t
his c
olu
m
n rep
rese
nts th
e label
of t
he r
ecord
.
3.1.2
.
Pre
-
treatmen
t
Fo
r
data
pr
e
pa
rati
on
re
qu
ire
m
ents
befor
e
t
r
eatm
ent,
on
t
he
on
e
ha
nd,
we
hav
e
trie
d
t
o
norm
al
iz
e
ou
r
dataset
by
c
onver
ti
ng
c
har
act
er
col
um
ns
to
nu
m
eric
colu
m
ns
with
the
help
of
the
fa
m
ou
s
1
-
to
-
n
e
ncodin
g
te
chn
iq
ue,
Fig
ure
1(
a
)
an
d
Fi
g
ure
1(b
)
ex
pl
ai
n
the
dif
fer
e
nces
be
f
or
e
a
nd
after
pr
e
-
trea
t
m
ent
.
And
on
ano
t
her
hand, we
h
a
ve se
par
at
e
d
th
e l
abel col
um
n
from
the o
the
r
c
olu
m
ns
v
ia
a
n Extr
act
-
Tra
ns
f
or
m
-
Loa
d
(
ET
L)
.
0
tcp
f
tp
_
d
ata
SF
491
…
n
o
r
m
al
20
0
udp
o
th
er
SF
146
…
n
o
r
m
al
15
0
tcp
p
rivate
S0
0
…
n
ep
tu
n
e
19
0
tcp
h
ttp
SF
232
…
n
o
r
m
al
21
0
tcp
h
ttp
SF
199
…
n
o
r
m
al
21
0
1
1
1
491
…
1
20
0
2
2
1
146
…
1
15
0
1
3
2
0
…
2
19
0
1
4
1
232
…
1
21
0
1
4
1
199
…
1
21
Figure
1
.
(a
) D
at
aset
r
ows
befor
e
pre
-
treat
m
ent
,
(b)
Dataset
rows
a
fter
pre
-
treatm
ent
a
b
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
:
33
1
5
-
33
2
2
3318
3.2.
Wor
k en
viro
nment
, e
va
lu
ati
on met
hod
and per
forma
n
ce indi
ca
to
rs
3.2.1
.
Wor
k e
nv
ir
onme
n
t
The
c
onfig
urat
ion
of
our
m
achine
is:
t
he
op
e
rati
ng
syst
em
i
s
W
i
ndows
7,
with
In
te
l
(
R
)
Core
(TM
)
i3
2370M
CP
U
@
2.4
0
G
H
Z
(
4
CPU
s
)
,
and
4096MB
of
RAM.
The
E
TL
is
Tal
en
d
Op
e
n
St
ud
i
o
(
TOS)
f
or
Data I
nteg
rati
on, w
hich
is
an
op
e
n so
ur
ce
s
oft
war
e
[1
7
],
ef
f
ic
ie
nt
,
flexi
ble
and easy
to
h
a
nd
le
[
18
]
.
3.2.
2
.
E
valua
tion
m
eth
od
The
k
-
f
old
c
r
os
s
validat
io
n
m
e
tho
d
is
e
m
plo
ye
d
to
m
easur
e
the
su
c
cess
of
a
cl
as
sifie
r
,
it
s
plit
s
the
dataset
int
o
two
subsets,
t
he
fi
rst
f
or
trai
ning
a
nd
t
he
s
econd
for
te
sti
ng
[1
9
]
.
The
operati
on
is
repea
te
d
k
tim
es
separ
at
el
y,
and
the
a
ve
rag
e
of
the
k
perform
ances
is
cal
culat
ed
and
retu
rn
e
d.
T
he
ad
va
ntage
of
this
m
et
ho
d
is
that
the
entire
dataset
is
us
ed
f
or
bo
t
h
trai
ni
ng
an
d
te
sti
ng,
wh
ic
h
m
akes
the
eval
uatio
n
m
or
e
accurate.
We
a
do
pted
5
-
f
old
cro
ss
validat
io
n
to
e
valuate
our
m
od
el
,
if
w
e
increase
t
he
k,
s
om
e
at
ta
ck
s
li
ke
U2
R
a
nd
R2
L
will
decr
eas
e
for
eac
h
s
ubs
et
,
an
d
they
c
an
be
ne
glect
e
d
du
rin
g
the
t
r
eatm
ent.
To
se
par
ate
the traini
ng s
ubset
a
nd test
in
g
s
ubset
, we al
so
em
plo
ye
d
t
he
ETL
(T
OS
)
.
3.2.
3
.
Perf
orm
an
ce in
dica
to
r
s
The
m
od
el
ass
essm
ent ind
ic
at
or
s a
re:
Accuracy
=TP+
TN/(T
P+TN+
F
P+FN)
(1)
Sens
it
ivit
y=
TP/(TP+F
N)
(2)
False
P
os
it
ive
Ra
te
=FP/(FP
+
TN)
(3)
Pr
eci
sio
n=TP/(
TP+FP
)
(4)
Re
cal
l =TP/
(T
P+FN)
(5)
F
-
Me
asu
re=
2*(P
recisi
on.Reca
ll
)/(Pr
eci
sio
n+
Re
cal
l)
(6)
Wh
e
re:
Accuracy
is th
e fr
act
io
n o
f
tr
ue dete
ct
ion o
ve
rall
d
at
a in
sta
nces.
Sens
it
ivit
y de
fines
t
he
a
bili
ty
of the m
od
el
to
d
et
ect
c
orrectl
y.
False
Posi
ti
ve
Ra
te
(F
PR)
is
cal
culat
ed
as
t
he
rati
o
of
ne
ga
ti
ve
eve
nts
w
ron
gly
cl
assifi
ed
as
posit
ive
to
the total
neg
at
i
ve
e
ven
ts
.
Pr
eci
sio
n
is t
he
f
racti
on
of r
el
evan
t i
ns
ta
nce
s
am
on
g
the
all
pro
po
se
d
i
ns
ta
nces.
Re
cal
l i
s the fr
act
ion
of r
el
e
va
nt insta
nces t
hat h
a
ve bee
n f
ound
ov
e
r
t
he
t
otal o
f
rele
va
nt instances
.
F
-
Me
asu
re
give
s the
har
m
on
i
c m
ean o
f pr
e
c
isi
on
a
nd r
ecal
l
.
TP,
T
N,
FP
a
nd
F
N
ar
e
ret
rieve
d
from
the
confusi
on
m
at
rix,
they
m
e
an
res
pecti
vel
y:
Tru
e
Posi
ti
ve,
Tru
e
N
e
gative
,
False P
os
it
ive
and
False
Ne
ga
ti
ve.
3.
3
.
Adop
ted me
t
ho
d
and c
ompared
meth
ods
3.
3
.
1
.
Adop
te
d metho
d
As
m
entioned
above,
L
STM
is
a
Deep
Lea
r
ning
m
et
ho
d,
i
t
is
sp
eci
al
ly
a
Re
current
N
eu
ral
Netw
ork
(RN
N
)
[
20
],
w
hich
is
ch
aract
erized
by
it
s
m
e
m
or
y,
that
w
hy
it
is
ado
pted
in
this
w
ork,
i
n
orde
r
to
m
em
or
iz
e
as
long
as
poss
ible
at
ta
cks
a
nd
pr
e
dict
ne
w
oth
e
rs.
As
s
ho
wn
in
th
e
Fig
ure
2
[
21
]
t
he
L
STM
gat
her
s:
an
I
nput
Gate
wh
ic
h
de
te
rm
ines
if
a
new
input
can
t
ran
sit
o
r
not,
a
Fo
r
get
Gate
wh
ic
h
delet
es
inf
or
m
at
ion
if
is
not
i
m
po
rtant
or
l
et
it
i
m
pact
the
outp
ut,
a
n
Ou
t
pu
t
Gate
wh
ic
h
determ
i
nes
t
he
ou
t
pu
t
,
a
si
ng
le
Ce
ll
w
hic
h
represe
nts
the
Con
sta
nt
Er
r
or
Ca
r
ousel
,
and
the
act
iva
ti
on
functi
ons
wh
ic
h
com
pu
te
the
act
ivati
o
n
of
the th
ree
gates.
3.
3
.
2.
C
ompar
ed
met
hod
s
Ther
e
a
re
se
ve
ral
m
et
ho
ds
i
n
Ma
chine
Lear
ning
dom
ai
n,
so
it
is
diff
ic
ult
to
com
par
e
ou
r
sug
gested
m
et
ho
d
t
o
al
l
t
hese
te
c
hn
i
qu
e
s,
we
will
try
to
c
om
par
e
it
only
with
the
m
os
t
ef
fici
ent
a
nd
popula
r
of
t
hem
,
as
SV
M
,
KNN
[
2
2
]
a
nd
Deci
sion
Tree
.
S
V
M
is
a
cl
assifi
er
est
ablis
hed
on
m
arg
ins,
it
us
es
sm
al
l
sa
m
ple
an
d
achieves
good
gen
e
rali
zat
ion
resu
lt
s
[
23
].
KNN
is
a
sim
ple
an
d
ef
fici
ent
te
chn
i
qu
e
w
hich
us
es
the
cl
os
est
trai
ning
exam
ples
to
cl
assify
obj
ect
in
the
f
ea
ture
spa
ce
[
24
].
A
decisi
on
tree
is
a
m
eth
od
of
cl
assifi
cat
ion
,
the
va
rio
us
pos
sible
decisi
ons
are
locat
e
d
at
the
en
ds
of
the
br
a
nc
hes
(the
le
aves
of
the
tr
ee)
an
d
a
re
rea
ched
accor
ding t
o d
eci
sion
s ta
ke
n at
each
ste
p [
25
]
.
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
LS
TM
d
e
ep
lea
rn
in
g meth
od f
or
network…
(
Alaed
din
e
B
oukha
lf
a
)
3319
Fig
ure
2. LST
M archite
ct
ure
4.
R
ESULT
S
AND A
N
ALYSIS
This
pa
rt
is
de
dicat
ed
to
a
nnou
nce
an
d
di
scuss
t
he
va
riou
s
obta
ine
d
r
esults.
W
e
ha
ve
eval
uate
d
the
m
od
el
for
two
ty
pes
of
cl
assifi
cat
ion
,
bin
ary
an
d
m
ulti
-
cl
assifi
cat
ion.
Fo
r
bin
a
ry
cl
assifi
cat
ion
the
da
ta
set
is
div
ide
d
on
t
wo
cl
ass
es:
cl
ass
of
norm
al
record
s
an
d
cl
ass
of
at
ta
ck
s.
Fo
r
m
ulti
-
cl
assifi
cat
ion
,
a
nd
Give
n
U2
R
are
not
de
ns
e
in
te
rm
of
the
num
ber
of
at
ta
cks,
the
c
la
ssific
at
ion
re
su
lt
s
of
this
ty
pe
of
at
ta
ck
ar
e
no
t
sat
isfact
or
y,
s
o
we
deci
ded
to
gro
up
t
hem
with
R2L
at
ta
ck
s
i
n
one
cl
ass,
so
the
dataset
is
div
ide
d
on
four
cl
asses:
cl
ass
of
norm
al
reco
r
ds
a
nd
th
ree
cl
asses
of
t
hree
cat
egories
of
at
ta
cks
(P
r
ob
e
,
DoS,
U
2R
-
R2L
).
We
ha
ve
eval
ua
te
d
the
m
e
tric
s:
Accu
racy,
S
ensiti
vity
,
False
Po
sit
ive
Ra
te
,
Pr
eci
sio
n
an
d
Re
cal
l,
and
w
e
hav
e
com
par
ed
t
he
m
w
it
h
the o
t
he
rs of t
he othe
r
classi
fiers.
Fig
ure
3
,
Fig
ur
e
4
an
d
Fig
ur
e
5
sho
w
resp
ec
ti
vely
:
Accu
ra
cy
,
ave
rag
e
of
Sens
it
ivit
y
an
d
ave
rag
e
of
False
Po
sit
ive
Ra
te
(F
PR)
f
or
bin
a
ry
and
m
ul
ti
-
cl
assifi
cat
ion
.
Wh
il
e
Table
3
an
d
Ta
ble
4
re
veal
Pre
ci
sion
,
Re
cal
l
and
F
-
Me
asur
e
f
or
bi
nar
y
a
nd
m
ulti
-
cl
assifi
cat
ion.
As
in
dicat
ed
by
Figure
3,
th
e
reache
d
value
s
of
the
Accuracy
are
99,
98
%
for
tw
o
cl
asses
cl
assifi
cat
ion
an
d
99,
93
%
for
f
our
cl
asses
cl
assifi
cat
ion
,
wh
ic
h
m
eans
that
LSMT
ca
n
pr
op
e
rly
m
e
m
or
iz
e
and
id
entify
traf
fics,
and
it
s
detect
ion
ca
pacit
y
is
bette
r
tha
n
th
e
oth
e
r
m
achine
le
ar
nin
g
cl
assifi
ers
.
Also
,
as
e
xpos
es
Fig
ur
e
4
,
t
he
value
s
of
the
ave
rag
e
of
t
he
Sensi
ti
vity
reached
by
the
m
od
e
l
LSTM
are
99
,
986
%
f
or
bi
nar
y
rec
ogniti
on
a
nd
99
,
738
%
f
or
m
ulti
-
recogn
it
io
n,
this
ex
plains
t
hat
the
sug
ges
te
d
m
od
el
is
ve
ry
able
t
o
dif
fer
e
ntiat
e
corr
ect
ly
between
the
di
ff
e
ren
t
t
ypes
of
traff
ic
,
bette
r
than
the
oth
e
r
m
od
el
s.
In
ad
di
ti
on
,
as
no
te
d
by
Figure
5,
th
e
values
of
ave
rag
e
of
False
P
os
it
ive
Ra
te
(F
PR)
achieve
d
by
LST
M
are
on
ly
0,0
68
%
f
or
tw
o
c
la
sses
cl
assifi
cat
ion
an
d
0,0
23
%
for
m
ulti
-
cl
asses
cl
assifi
cat
ion
,
wh
ic
h
m
eans
that
the
m
arg
in
of
e
rro
r
of
the
detect
ion
of
th
e
m
e
tho
d
is
m
inim
a
l,
and
the
values
achieve
d
a
re mi
ni
m
al
co
m
par
ed
to
o
t
her
cl
as
sifie
rs.
Fig
ure
3
.
Acc
uracy
for bi
nar
y
classi
ficat
ion
and m
ulti
-
cl
ass
ific
at
ion
97
98
99
100
LS
TM
SV
M
TR
E
E
KNN
99.93
98.758
99.617
99.9
99.98
98.189
99.844
99.902
A
c
c
u
r
a
c
y
(
%
)
4 cla
sses
2 cla
sses
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
:
33
1
5
-
33
2
2
3320
Fig
ure
4. A
verage
of Se
ns
it
ivit
y fo
r bina
ry c
la
ssific
at
ion
a
nd m
ulti
-
cl
assifi
cat
ion
Figure
5. A
verage
of
false
po
sit
ive r
at
e
(
FP
R) for bi
nar
y c
la
ssific
at
ion
a
nd m
ulti
-
cl
assifi
cat
ion
Table
3
. Recal
l
, precisi
on a
nd
F
-
Me
asu
re
of t
wo cla
sses
Clas
sif
ier
Clas
s
Precisio
n
(
%)
Recall (%
)
F
-
Measu
r
e (
%
)
LST
M
No
r
m
al
9
9
,99
9
9
9
,97
3
9
9
,98
6
Attack
9
9
,96
9
9
9
,99
8
9
9
,98
3
SVM
No
r
m
al
9
9
,02
9
9
7
,56
9
9
8
,29
4
Attack
9
7
,25
4
9
8
,90
2
9
8
,07
1
Tr
ee
No
r
m
al
9
9
,88
7
9
9
,82
0
9
9
,85
3
Attack
9
9
,79
4
9
9
,87
0
9
9
,83
2
KNN
No
r
m
al
9
9
,90
1
9
9
,91
7
9
9
,90
9
Attack
9
9
,90
4
9
9
,88
6
9
9
,89
5
Table
4
. Recal
l
, precisi
on a
nd
F
-
Me
asu
re
of four cla
sses
Clas
sif
ier
Clas
s
Precisio
n
(
%)
Recall (%
)
F
-
Measu
r
e (
%
)
LST
M
No
r
m
al
100
9
9
,93
8
9
9
,96
9
Do
S
9
9
,92
4
9
9
,90
6
9
9
,91
5
U2
R, R2
L
9
5
,89
6
9
9
,10
6
9
7
,47
5
Prob
e
9
9
,86
3
100
9
9
,93
1
SVM
No
r
m
al
9
9
,17
3
9
8
,59
7
9
8
,88
4
Do
S
9
9
,12
8
9
9
,78
2
9
9
,45
4
U2
R, R2
L
8
9
,50
7
9
5
,32
0
9
2
,32
2
Prob
e
9
5
,80
3
9
5
,96
8
9
5
,88
5
Tr
ee
No
r
m
al
9
9
,82
1
9
9
,63
2
9
9
,72
6
Do
S
9
9
,77
2
9
9
,86
7
9
9
,81
9
U2
R, R2
L
8
6
,82
4
9
5
,03
3
9
0
,74
3
Prob
e
9
9
,08
9
9
8
,96
2
9
9
,02
5
KNN
No
r
m
al
9
9
,91
4
9
9
,92
0
9
9
,91
7
Do
S
9
9
,97
4
9
9
,98
0
9
9
,97
7
U2
R, R2
L
9
7
,48
3
9
6
,18
0
9
6
,82
7
Prob
e
9
9
,74
3
9
9
,80
3
9
9
,77
3
96
97
98
99
100
LST
M
SVM
TREE
KNN
99.738
97.417
98.374
98.971
99.986
98.236
99.845
99.902
Sen
sit
ivity
(%)
4 cla
sses
2 cla
sses
LS
TM
SV
M
TR
E
E
KNN
0.023
0.493
0.138
0.04
0.068
1.765
0.155
0.099
F
alse
P
os
iti
v
e
R
a
t
e
(%)
4 cla
sses
2 cla
sses
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
LS
TM
d
e
ep
lea
rn
in
g meth
od f
or
network…
(
Alaed
din
e
B
oukha
lf
a
)
3321
The
s
uprem
e
values
of
P
recis
ion
,
as
il
lustra
t
ed
by
Table
3
and
Table
4
,
a
re
th
os
e
of
our
pro
posed
Deep
Lea
r
ning
m
od
el
LSTM
.
For
the
ide
ntific
at
ion
of
tw
o
cl
asses,
the
Pr
eci
sio
ns
rea
ch
up
to
99,
999
%
and
99,
969
%
for
resp
e
c
ti
vely
no
rm
al
traff
ic
a
nd
at
ta
ck
traf
fic,
m
or
e
than
the
ot
her
cl
a
ssifie
rs.
Fo
r
the
ide
ntif
ic
at
ion
of
f
our
cl
asses,
the
Pr
eci
sion
re
ache
s
up
to
100
%
fo
r
no
rm
al
reco
r
ds
,
this
is
due
to
the
de
ns
it
y
of
this
cl
ass
of
tr
aff
ic
,
al
so
for
Pr
obe
a
nd
D
O
S
at
ta
cks,
the
m
axi
m
u
m
Pr
eci
sion
s
ac
hiev
ed
are
resp
ect
ively
99,
863
%
an
d
99,
924
%
m
or
e
than
the
oth
e
r
cl
assifi
ers,
on
ly
on
e
m
ini
m
a
l
value
of
P
re
ci
sion
95,89
6
%
no
te
d
by
LS
TM
(le
ss
than
only
that
of
KNN
97,
483
%
)
in
the
case
of
t
he
cl
as
sific
at
ion
of
th
e
cl
ass
U2R
-
R2L
,
ex
pl
ai
ned
by
it
s
m
ini
m
u
m
den
sit
y.
This
j
ust
ifie
s
that
th
e
LS
TM
is
gen
e
rall
y
ver
y
accu
rat
e
m
or
e
than
the
oth
e
rs
.
The
val
ues
of
Re
cal
l
of
LS
TM
fo
r
t
he
id
entifi
cat
ion
of
two
cl
asses,
as
sh
ow
n
in
Tab
le
3
,
are
ver
y
high,
99,
973
%
f
or
norm
al
traff
ic
an
d
99,99
8
%
f
or
at
ta
c
k
tr
aff
ic
.
Als
o,
t
he
value
s
of
Re
cal
l
for
the
identific
at
ion
of
four
cl
a
sses,
as
sho
wn
in
Table
4
,
a
r
e
al
so
ver
y
hi
gh,
99,
938
%
for
the
norm
al
cl
ass,
99,10
6
%
f
or
U2
R
-
R2L
cl
as
s,
10
0
%
f
or
P
robe
cl
ass
,
an
d
99,
906
%
f
or
D
OS
cl
ass
(le
ss
tha
n
on
ly
th
at
of
KNN
99,
980
%).
This
pro
ve
s
that
L
STM
c
an
fin
d
nor
m
al
instance
s
a
nd
at
ta
ck
insta
nce
s
m
or
e
tha
n
t
he
ot
her
m
od
el
s.
The
e
xp
e
rim
ent
resul
t
has
pr
ov
e
d
t
hat
the
ne
w
m
et
hod
LS
TM
is
ve
ry
ef
fici
en
t,
it
can
ef
fect
ively
m
e
m
or
iz
e
and
diff
e
re
ntiat
e
be
tween
t
raffic
s:
norm
al
and
at
t
ack,
i
n
t
he
bo
t
h
cases
of
cl
as
sific
at
i
on
,
bi
na
ry
an
d
m
ul
ti
-
cl
assifi
cat
ion
.
5.
CONCL
US
I
O
N AND F
UT
U
RE W
ORK
In
t
his
pa
per,
we
pr
opos
e
d
a
new
i
dea
f
or
NIDS
est
a
blished
on
the
De
ep
Lear
ning
m
et
hod
LSTM
,
wh
ic
h
will
rec
ognize
at
ta
cks
and
kee
p
a
lo
ng
-
te
rm
m
e
m
or
y
of
them
,
in
orde
r
to
blo
c
k
t
he
ot
her
ne
w
a
tt
acks,
and
at
the
sam
e
tim
e,
will
tre
at
,
with
a
un
i
que
m
ann
er,
al
l
ty
pe
of
t
hese
a
tt
acks.
T
o
vali
date
the
e
ff
ect
i
ven
e
ss
of
our
ne
w
sugg
e
ste
d
ap
proa
ch,
we
em
plo
ye
d
the
fam
ou
s
NS
L
KDD
as
dataset
for
trai
ning
an
d
te
sti
ng
,
a
nd
Accuracy
,
Se
nsi
ti
vity
,
False
P
os
it
ive
Ra
te
,
Pr
eci
sio
n
a
nd
Re
cal
l
as
m
et
r
ic
s
for
the
e
va
luati
on
,
an
d
w
e
ha
ve
com
par
ed
th
e
new
m
et
ho
d
LSTM
to
t
he
oth
er
Ma
chine
Lea
r
ning
cl
assifi
ers.
The
e
xp
e
rim
e
nt
has
dem
on
strat
ed
t
hat
the
m
et
ric
s
of
t
he
detec
ti
on
of
the
L
STM
m
e
tho
d
reach
ver
y
hig
h
value
s
m
or
e
tha
n
the
oth
e
r
cl
assi
fiers.
wh
ic
h
prov
e
s
that
our
ne
w
pro
pose
d
m
et
hod
is
ef
fecti
ve
f
or
N
I
DS.
I
n
the f
ut
ur
e, we
plan
to
im
ple
m
ent
reall
y
a
ne
w
i
ntell
igent
NID
S
in
the
real
world
us
i
ng
our
ne
w
pro
po
sed
Dee
p
Lea
rn
i
ng
m
od
el
LSTM.
REFERE
NCE
S
[1]
Yann
L.,
et al
.
,
“
Dee
p
l
ea
rn
ing,”
Nature,
In
te
rnati
onal
Journal
of
Sci
en
ce,
vol
.
52
1,
pp
.
436
-
444
,
2015.
[2]
Gensler
A.,
et
al.
,
“
Dee
p
L
ea
r
n
ing
for
Solar
Pow
er
Forec
asting
–
An
Approac
h
Us
ing
Autoen
code
r
and
LST
M
Neura
l
Network
s,”
in
:
2016
IE
EE
Inte
rnat
iona
l
Con
fe
renc
e
o
n
Syste
ms
,
Man,
and
Cybe
rnet
i
cs,
SMC
2016
,
Artic
l
e
n
o.
7844
673,
pp
.
2858
-
2
865,
2017
.
[3]
NS
L
KD
D,
[Online
]
.
Avai
la
bl
e:
htt
ps://
g
it
hub.
co
m
/de
fco
m
17/NS
L_KDD
.
[4]
Ni
y
a
z
Q
.
,
Sun
W
.
,
Java
id
A
.
an
d
Alam
M
.
,
“
A
Dee
p
Learni
ng
Approac
h
for
Ne
twork
Intrusion
Dete
c
ti
on
S
y
s
tem
,”
I
n
Proceedi
ngs
of
th
e
9th
EAI
Inte
rnationa
l
Confe
renc
e
on
Bi
o
-
inspired
In
f
orm
ati
on
and
Comm
unic
ati
ons
Technol
ogi
es
(
BICT'15)
,
Unite
d
Stat
es
ACM
,
pp
.
21
-
26,
2015.
[5]
Dong
B.
and
Wa
ng
X.,
“
Com
pa
rison
Dee
p
Le
ar
ning
Method
to
Tra
di
ti
ona
l
Met
hod
s
using
for
Network
Intrusion
Dete
c
ti
on
,”
8
th
I
EE
E
Inte
rnat
ion
al
Conf
ere
nce o
n
Comm
unic
ati
o
n
Soft
ware
and
Net
works,
pp
.
58
1
-
585,
2016
.
[6]
KD
D Cup
99,
[
Online
]
.
Avai
la
b
le
:
htt
p
:/
/kdd
.
ic
s
.
u
ci
.
edu/
da
ta
b
ase
s/kddcup99/
kdd
cup99.
html
.
[7]
Yang
J.,
Deng
J.,
Li
S.
and
Hao
Y.,
“
Im
prove
d
Tra
ffic
Det
ection
with
Support
Vec
tor
Mac
hin
e
ba
sed
on
Restri
ct
e
d
Bolt
zmann
Ma
c
hine
,”
So
ft
Com
puti
ng,
vol
.
21
,
no.
11
,
pp
.
3101
–
3112,
2017
.
[8]
Yan
R.
,
Xiao
X.
,
Hu
G.,
Peng
S.
and
Jiang
Y.,
“
New
Dee
p
Le
ar
ning
Method
to
Dete
c
t
Code
Inj
ec
t
ion
Atta
cks
o
n
H
y
brid
Applications
,
”
Journal
of
Syste
ms
and
Sof
t
ware,
vol
.
137
,
p
p.
67
–
77
,
2018
.
[9]
Li
Y.
,
Ma
R
.
an
d
Jiao
R.
,
“
A
Hy
brid
Malicious
Code
Det
ec
t
ion
Metho
d
Based
o
n
Dee
p
L
ea
rn
in
g
,”
In
te
rnationa
l
Journal
of
Se
curit
y
and
Its
Appli
cat
ions (
IJS
IA)
,
vol.
9
,
no
.
5
,
pp
.
205
–
216,
2015
.
[10]
Diro
A.
A.
and
Chil
amkurti
N.,
“
Distribut
ed
Att
ac
k
Detect
ion
S
che
m
e
using
Dee
p
Le
arn
ing
Approac
h
for
Inte
r
net
of
Thi
ngs,
”
In
te
r
nati
onal Journal
Fut
ure
Gene
rat
i
on
Computer
Sy
stems (
FGCS
)
,
vol.
82
,
pp
.
761
-
7
68,
2018
.
[11]
La
n
L
.
and
Jun
L.
,
“
Som
e
Special
Iss
ues
of
Network
Secur
ity
Monitori
ng
on
Big
Data
Env
iro
nm
ent
s
,”
In:
20
13
IEE
E
11th
Inte
r
nati
onal
Con
fe
re
nce
on
Depe
nda
ble
Aut
onom
ic a
nd
Sec
ure
Comp
uti
ng,
pp.
10
-
15,
2013.
[12]
Saenko
I.
,
Kote
nko
I.
,
and
Kus
hner
evi
ch
A.
,
“
Para
ll
el
Proce
s
sing
of
Big
He
te
roge
n
eous
Data
for
Secur
i
t
y
Monitori
ng
of
I
oT
Networks
,”
In:
2017
25th
E
urom
ic
ro
Inte
rnational
Conf
ere
nce
on
Parall
e
l,
Distribute
d
an
d
Net
works
-
Based Proce
ss
ing,
pp
.
329
-
336,
2017
.
[13]
El
-
Sofan
y
H.F.
,
El
-
Seoud
S.A,
T
aj
-
Edd
in,
I
.
A.T.
F.,
“
A
Case
Stud
y
of
th
e
Im
pact
of
Denial
of
Service
Attacks
i
n
Cloud
Appli
ca
t
i
ons
,”
Journal
of
Comm
unic
ati
ons
,
vol
.
14
,
no
.
2
,
2
019.
[14]
Alja
warne
h
S.
,
Aldwairi
M.,
Ya
ss
ei
n,
M.B.
,
“
Anom
aly
-
b
ase
d
in
trusion
det
e
ct
ion
sy
st
em
through
fea
ture
se
le
c
ti
o
n
ana
l
y
sis
and
bu
ilding h
y
brid
eff
icient m
odel
,”
Jou
rnal
of
Computa
ti
onal
S
cienc
e,
v
ol.
25
,
pp
.
152
-
1
60,
2018
.
[15]
Sorns
uwit
P.,
Jai
y
en
S.,
“
Intrusion
det
e
ction
m
odel
bas
ed
on
ens
e
m
ble
learni
ng
fo
r
U2R
and
R2L
at
t
ac
ks
,”
in
:
201
5
7th
Inte
rnationa
l
Confe
renc
e
on
Information
Technol
ogy
and
E
le
c
tric
al
Eng
ineering,
En
vi
sioni
ng
the
Tr
end
of
Computer,
Infor
mation
and
Enginee
ring (
ICITEE
)
,
Artic
l
e
n
o.
74
08971,
pp
.
354
-
359,
2015
.
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
:
33
1
5
-
33
2
2
3322
[16]
Dube
y
S.
,
Dube
y
J.
,
“
KBB
:
A
h
y
brid
m
et
hod
for
int
rusion
d
e
te
c
ti
on
,”
in
:
IE
EE
Int
ernati
ona
l
Confe
ren
ce
on
Computer,
Com
municat
ion
and
Control
(
IC4
-
2015)
,
Artic
l
e
n
o.
r 737
5704,
2016
.
[17]
Venka
te
sh Pulla
V.S.,
Varol
C.
,
Al
M.
,
“
Open
so
urc
e
data
qualit
y
tool
s: Re
visit
ed
,”
Adv
anc
es
in
Inte
lligent
S
yste
ms
and
Computing
,
vol.
448
,
pp
.
893
-
902,
2016
.
[18]
Majc
hrz
ak
T
.
A.,
Jansen
T.
and
Kuche
n
H.,
“
Eff
ic
i
ency
Eva
lu
ation
of
Open
So
urc
e
E
TL
Tool
s
,
”
In
:
(SA
C
'
11
)
Proce
ed
ings o
f
t
he
2011
ACM
S
y
mpos
ium
on
App
li
ed
Computing,
pp.
287
-
294
,
20
11
.
[19]
Refa
e
il
z
ade
h
P.,
Ta
ng
L.
and
Li
u
H.,
“
Cross
-
Vali
d
at
ion
,”
Encyc
lo
p
edi
a
o
f
Databas
e
Syst
ems
,
Sprin
ger
,
2009
.
[20]
Zha
o
Z.,
Chen
W
.
,
W
u
X.,
Chen
P.C.
Y.,
Li
u
J.
,
“
LSTM
net
wor
k:
a
dee
p
l
ea
rni
ng
appr
oac
h
for
short
-
te
rm
tra
ff
ic
fore
ca
st
,”
IET
In
te
lligent Tr
anspor
t
Syste
ms
,
vol. 11,
no.
2,
pp.
68
-
75,
2017
.
[21]
Greff
K.,
Srivast
ava
R.
K.
,
Koutn
ik
J.,
Steune
b
rin
k
B.
R.
and
Sch
m
idhube
r
J.,
“
LSTM:
A
Sear
ch
Space
Od
y
ss
e
y
,
”
2017
IEEE
Tr
ansacti
ons on
N
eural
Ne
tworks
and
Learning
Syste
m
s,
vol. 28, pp. 22
22
-
2232,
2017
.
[22]
Xie
L
.
,
Fu
Z
.
H.
,
Feng
W
.
and
Luo
Y.
,
“
Pitc
h
-
Densit
y
-
B
ase
d
Feat
ure
s
and
an
SV
M
Bina
r
y
T
ree
Appro
ac
h
fo
r
Multi
-
Cla
ss
Aud
io
Cl
assific
a
ti
on
in
Broad
ca
st
Ne
ws
,
”
Multimedia
Syste
ms
,
vol. 17
,
no
.
2
,
pp
.
101
-
112,
2011
.
[23]
Kuang
F.,
Xu
W
.
,
and
Zha
ng
S.,
“
A
Novel
H
y
brid
KP
CA
and
SV
M
with
G
A
Mode
l
for
Intrusion
Dete
ction
,”
Appl
ie
d
Sof
t
Co
mputing,
vo
l. 18, pp. 178
-
184,
20
14.
[24]
Am
in
Aburo
m
ma
n
A.,
and
Ibne
Rea
z
M.
B
.
,
“
A
novel
SV
M
-
kN
N
-
PSO
Ense
m
ble
Method
for
In
trusion
Detect
io
n
S
y
stem,
”
Applie
d
Soft Computin
g
,
vol
.
38
,
pp
.
36
0
-
372,
2016
.
[25]
Tsai
C
.
F.,
Hs
u
Y.F.,
Li
n
C
.
Y.,
and
Li
n
W
.
Y
.
,
“
Intrusion
Det
e
ct
ion
b
y
Mac
h
i
ne
L
ea
rning
:
A
Rev
i
ew
,”
Ex
p
e
rt
Syste
ms
wit
h
Ap
pli
cations,
vo
l. 3
6,
no
.
10
,
pp
.
11
994
-
12000,
200
9.
Evaluation Warning : The document was created with Spire.PDF for Python.