I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
V
ol
.
10
, N
o.
1
,
M
a
r
c
h
202
1
, pp.
110
~
120
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
10
.i
1
.pp
110
-
120
110
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
T
ow
ar
d
a d
e
e
p
l
e
ar
n
i
n
g
-
b
ase
d
i
n
t
r
u
s
i
on
d
e
t
e
c
t
i
on
sys
t
e
m
f
or
IoT
agai
n
st
b
ot
n
e
t
at
t
ac
k
s
I
d
r
is
s
I
d
r
is
s
i
1
,
M
oh
am
m
e
d
B
ou
k
ab
ou
s
2
,
M
os
t
af
a
A
z
iz
i
3
, O
m
ar
M
ou
s
s
aou
i
4
, H
ak
im
E
l
F
ad
il
i
5
1
,2,3,4
MATSI Researc
h Lab., ESTO, Mohamme
d First
University, Ouj
da, Morocco
5
LIPI Research Lab., ENSAF, Sidi Mohamed Ben Abdell
ah University, Fez, Morocco
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
J
un
9
, 20
20
R
e
vi
s
e
d
D
e
c
3
0, 20
20
A
c
c
e
pt
e
d
F
e
b
2, 20
21
The
massive
network
traffic
data
between
connected
devices
in
the
i
nternet
of
things
have
taken
a
big
challenge
to
many
traditional
intrusion
de
tection
systems
(IDS)
to
find
probable
security
breaches.
However,
security
attacks
lean
towards
unpredictabilit
y.
There
are
numerous
difficulties
to
build
up
adaptable
and powerfu
l IDS fo
r IoT in
order to
avoid fal
se alerts
and e
nsure a
high
recognition
precision
against
attacks,
especially
with
the
ris
ing
of
Botnet
attacks.
These
attacks
can
even
make
harmless
dev
ices
be
coming
zombies
that
send
malicio
us
traffic
and
disturb
the
network.
In
this
pa
per,
we
propose
a
new
IDS
solution,
baptized
BotIDS,
based
on
deep
l
earning
convolut
ional
neural
networks
(CNN).
The
main
interest
of
this
wor
k
is
to
design,
implement
and
test
our
IDS
against
some
well
-
known
Botnet
attacks
using
a
specific
Bot
-
IoT
dataset.
Compared
to
other
deep
l
earning
techniques,
such
as
simple
RNN,
LSTM
and
GRU,
the
obtained
results
of
our
BotIDS
are
promising
with
99.94%
in
validation
accuracy,
0.
58%
in
validation loss, and the prediction execution time is less than 0.34 ms.
K
e
y
w
o
r
d
s
:
B
ot
-
I
oT
B
ot
ne
t
C
N
N
DL
I
D
S
I
oT
R
N
N
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
I
dr
is
s
I
dr
is
s
i
M
A
T
S
I
L
a
bor
a
to
r
y, E
S
T
O
M
oha
m
m
e
d F
ir
s
t
U
ni
ve
r
s
it
y
B
P
473
C
a
m
pus
uni
ve
r
s
it
a
ir
e
A
l
Q
od
s
, O
uj
da
60000, M
or
oc
c
o
E
m
a
il
:
id
r
is
s
i@ump.a
c
.m
a
1.
I
N
T
R
O
D
U
C
T
I
O
N
N
ow
a
da
ys
a
n
e
nor
m
ous
num
be
r
of
obj
e
c
ts
a
r
e
di
s
pa
tc
he
d
a
r
ound
th
e
w
or
ld
a
nd
a
r
e
c
onne
c
te
d
be
twe
e
n
th
e
m
a
nd
to
th
e
i
nt
e
r
ne
t.
T
he
y
va
r
y
f
r
om
pe
r
s
ona
l
g
a
dge
ts
,
w
e
a
r
a
bl
e
s
,
s
e
n
s
or
s
,
a
c
tu
a
to
r
s
to
hom
e
a
ppl
ia
nc
e
s
a
nd
m
e
di
c
a
l
de
vi
c
e
s
.
A
s
e
s
ti
m
a
te
d
by
C
I
S
C
O
in
20
25,
it
w
oul
d
be
s
om
e
w
he
r
e
75
bi
ll
io
n
de
vi
c
e
s
c
onne
c
te
d
to
th
e
I
nt
e
r
ne
t
[
1]
.
T
he
I
oT
ha
s
r
a
is
e
d
c
onc
e
r
ns
th
a
t
a
r
e
gr
ow
in
g
r
a
pi
dl
y
w
i
th
out
f
i
tt
in
g
th
ought
of
th
e
s
ig
ni
f
ic
a
nt
s
e
c
ur
it
y c
ha
ll
e
nge
s
[
2]
.
N
ow
a
da
ys
,
m
os
t
of
th
e
s
e
c
ur
it
y
c
onc
e
r
ns
a
r
e
li
ke
th
os
e
o
f
r
e
gul
a
r
s
e
r
ve
r
s
,
w
or
k
s
ta
ti
ons
a
nd
s
m
a
r
tp
hone
s
;
how
e
ve
r
,
s
e
c
ur
it
y
m
ove
s
e
xt
r
a
or
di
na
r
i
l
y
to
th
e
I
oT
,
in
c
lu
di
ng
m
e
c
ha
ni
c
a
l
s
e
c
ur
it
y
c
ont
r
ol
s
,
hybr
id
f
r
a
m
e
w
or
ks
,
I
oT
-
e
xpl
ic
it
bus
in
e
s
s
pr
oc
e
dur
e
s
,
a
nd
e
dg
e
de
vi
c
e
s
[
3]
.
T
r
a
di
ti
ona
l
s
e
c
ur
it
y
pr
ot
e
c
ti
on
te
c
hnol
ogy
is
li
m
it
e
d
due
Z
e
r
o
-
D
a
y
a
tt
a
c
ks
a
nd
vul
ne
r
a
bi
li
ti
e
s
a
nd
f
ut
ur
e
ne
w
a
tt
a
c
ks
th
a
t
a
r
e
c
ont
in
uous
ly
c
ha
ngi
ng
na
tu
r
e
;
s
e
tt
in
g
up
a
s
te
a
dy,
r
e
li
a
bl
e
,
a
nd
pr
e
c
is
e
in
tr
us
io
n
de
te
c
ti
on
i
s
be
c
om
in
g
m
a
nda
to
r
y
f
or
im
pr
ovi
ng t
he
I
oT
s
e
c
ur
it
y
[
4]
.
B
ot
ne
ts
or
z
om
bi
e
s
a
r
e
r
obot
s
of
in
f
e
c
t
e
d
i
nt
e
r
ne
t
-
c
onne
c
te
d
de
vi
c
e
s
,
th
e
y
a
r
e
us
e
d
to
a
c
hi
e
ve
di
s
tr
ib
ut
e
d
de
ni
a
l
-
of
-
s
e
r
vi
c
e
a
tt
a
c
k
(
D
D
oS
a
tt
a
c
k)
,
pa
s
s
w
or
d
c
r
a
c
ki
ng,
ke
y
lo
ggi
ng
(
s
te
a
l
da
ta
)
,
c
r
ypt
oc
ur
r
e
nc
y
m
in
in
g,
a
nd
gi
ve
th
e
a
tt
a
c
ke
r
th
e
po
s
s
ib
il
it
y
of
a
c
c
e
s
s
in
g
th
e
de
vi
c
e
us
in
g
c
om
m
a
nd
a
nd
c
ont
r
ol
(
C
&
C
)
s
of
twa
r
e
[
5]
.
I
n
S
e
pt
e
m
be
r
2016,
“
M
ir
a
i”
m
a
lwa
r
e
,
a
n
I
oT
B
ot
ne
t
a
tt
a
c
ke
d
m
a
ny
s
it
e
s
of
f
li
ne
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
T
ow
ar
d a de
e
p l
e
a
r
ni
ng
-
bas
e
d i
nt
r
us
io
n d
e
te
c
ti
on s
y
s
te
m
f
or
I
o
T
agains
t
bot
ne
t
at
ta
c
k
s
(
I
dr
is
s
I
dr
i
s
s
i
)
111
li
ke
th
e
c
lo
ud
s
e
r
vi
c
e
pr
ovi
de
r
“
O
V
H
”
w
it
h
ne
a
r
ly
1.1
T
B
p
s
,
th
e
w
e
bs
it
e
of
c
om
put
e
r
s
e
c
ur
it
y
c
ons
ul
ta
nt
B
r
ia
n K
r
e
bs
w
it
h 620 G
bps
of
t
r
a
f
f
ic
, a
nd ma
ny othe
r
w
e
bs
it
e
s
l
ik
e
dyna
m
ic
D
N
S
pr
ovi
de
r
“
D
yn”
[
6]
.
T
he
de
t
e
c
ti
on
a
nd
pr
e
ve
nt
io
n
f
r
om
di
f
f
e
r
e
nt
a
tt
a
c
ks
a
r
e
a
bi
g c
ha
ll
e
nge
.
I
D
S
us
in
g
m
a
c
hi
ne
-
le
a
r
ni
ng
m
e
th
ods
,
ha
s
ga
in
e
d
a
w
id
e
r
e
put
a
ti
on
[
7
]
.
I
D
S
is
a
n
e
s
s
e
nt
ia
l
c
om
pone
nt
in
th
e
s
e
c
ur
it
y
m
e
c
ha
ni
s
m
,
it
is
us
e
d
f
or
th
e
a
na
ly
s
is
a
nd
de
te
c
ti
on
of
t
he
s
e
c
ur
it
y
br
e
a
c
he
s
on
a
ne
twor
k
[
8]
.
I
D
S
s
ys
te
m
s
c
a
n
b
e
ga
th
e
r
e
d
in
to
two
c
a
te
gor
ie
s
:
th
e
f
ir
s
t
one
“
a
nom
a
ly
de
te
c
ti
on”
a
nd
th
e
s
e
c
ond
i
s
“
M
is
us
e
de
te
c
ti
on”
,
or
ga
th
e
r
e
d
in
to
th
r
e
e
m
a
jo
r
di
s
ti
nc
t
f
a
m
il
ie
s
“
hos
t
-
ba
s
e
d
”
,
“
ne
twor
k
-
ba
s
e
d
”
a
n
d
“
hybr
id
”
.
I
D
S
in
ve
s
ti
ga
te
bot
h
tr
a
f
f
ic
in
th
e
ne
twor
k
a
nd
in
th
e
ope
r
a
t
in
g
s
ys
te
m
s
.
I
D
S
a
r
e
us
e
d
f
or
e
f
f
e
c
ti
ve
ne
twor
k
pr
ot
e
c
ti
on.
N
um
e
r
ous
r
e
s
e
a
r
c
h
w
or
ks
a
r
e
tr
yi
ng
to
a
ppl
y
da
ta
m
in
in
g
a
nd
m
a
c
hi
ne
le
a
r
ni
n
g
a
lg
or
it
hm
s
to
c
ybe
r
s
e
c
ur
it
y.
I
n
M
a
c
hi
ne
le
a
r
ni
ng,
pa
tt
e
r
ns
r
e
c
ogni
ti
on
a
nd
da
ta
m
in
in
g
a
lg
or
it
hm
s
a
r
e
e
xt
e
ns
iv
e
ly
a
ppl
ie
d
to
di
s
ti
ngui
s
h
th
e
nor
m
a
l
tr
a
f
f
ic
f
r
om
th
e
m
a
li
c
io
us
one
.
2.
A
R
T
I
F
I
C
I
A
L
N
E
U
R
A
L
N
E
T
WO
R
K
S
(
A
N
N
)
I
n
la
s
t
r
e
c
e
nt
y
e
a
r
s
,
one
of
th
e
m
o
s
t
r
e
s
ul
ti
ng
a
nd
e
f
f
ic
ie
nt
s
ubs
e
ts
of
a
r
ti
f
ic
ia
l
in
te
ll
ig
e
nt
is
de
e
p
le
a
r
ni
ng
(
D
L
)
w
hi
c
h
it
is
a
ls
o
a
s
ubs
e
t
of
m
a
c
hi
ne
le
a
r
ni
ng
ba
s
e
d
on
a
r
ti
f
ic
ia
l
ne
ur
a
l
ne
twor
ks
(
A
N
N
)
[
9]
;
a
c
om
put
in
g
s
ys
te
m
in
s
pi
r
e
d
f
r
om
bi
ol
ogi
c
a
l
br
a
in
w
he
r
e
th
e
m
a
c
hi
ne
le
a
r
ns
f
r
om
m
a
ny
tr
a
in
in
g
e
xa
m
pl
e
s
,
a
ll
ow
in
g
it
to
c
la
s
s
if
y
ot
he
r
e
xa
m
pl
e
s
[
10]
.
D
L
is
in
c
r
e
a
s
in
gl
y
be
in
g
us
e
d.
I
t
c
a
n
be
a
ppl
ie
d
in
m
a
ny
da
t
a
pr
oc
e
s
s
in
g
la
ye
r
s
in
to
a
hi
e
r
a
r
c
hi
c
a
l
a
r
c
hi
te
c
tu
r
e
to
m
a
ke
a
de
e
p
m
ode
l.
D
L
a
c
c
ount
s
on
it
s
c
a
pa
c
it
y
to
id
e
nt
if
y
id
e
a
l
f
e
a
tu
r
e
s
in
r
a
w
da
ta
th
r
ough
s
uc
c
e
s
s
iv
e
nonl
in
e
a
r
tr
a
ns
f
or
m
a
ti
ons
,
w
it
h
e
ve
r
y
a
lt
e
r
a
ti
on
a
c
hi
e
vi
ng
a
m
or
e
e
le
va
te
d
le
ve
l
of
c
om
pl
e
xi
ty
a
nd
a
bs
tr
a
c
ti
on
[
10]
.
I
t
ha
s
be
e
n
a
ppl
ie
d
e
f
f
ic
ie
nt
ly
to
m
a
ny
di
f
f
e
r
e
nt
r
e
s
e
a
r
c
h
f
ie
ld
s
,
f
r
om
m
e
di
c
a
l
im
a
ge
pr
oc
e
s
s
in
g,
na
tu
r
a
l
la
ngua
ge
pr
oc
e
s
s
in
g,
s
pe
e
c
h
r
e
c
ogni
ti
on,
a
nd
s
ig
na
l
r
e
c
og
ni
ti
on
to
m
a
ny
ot
h
e
r
dom
a
in
s
of
s
c
i
e
nc
e
,
bu
s
in
e
s
s
a
nd
gove
r
nm
e
nt
.
I
n
a
ll
th
e
s
e
f
ie
ld
s
,
D
L
s
how
e
d
tr
e
m
e
ndous
ly
pr
om
is
in
g
r
e
s
ul
t
s
[
11]
.
I
n
th
e
I
oT
s
e
c
ur
it
y
f
ie
ld
,
th
e
m
a
c
hi
ne
tr
a
in
s
on
va
r
io
us
c
ol
l
e
c
te
d a
nd l
a
be
le
d a
tt
a
c
k
s
a
nd a
ls
o nor
m
a
l
tr
a
f
f
ic
t
o l
e
a
r
n t
he
m
, w
e
r
e
f
in
a
ll
y t
hi
s
m
a
c
hi
ne
c
a
n i
de
nt
if
y ne
w
s
im
il
a
r
a
tt
a
c
ks
.
C
onvolut
io
na
l
ne
ur
a
l
ne
twor
ks
(
C
N
N
or
C
onvNe
ts
)
:
it
’
s
a
de
e
p
le
a
r
ni
ng
c
la
s
s
de
v
e
lo
pe
d
in
1998
by
L
e
C
un
in
th
e
L
e
N
e
t
a
r
c
hi
te
c
tu
r
e
[
12]
.
I
n
r
e
c
e
nt
two
de
c
a
de
s
,
C
N
N
ga
in
e
d
bi
g
s
uc
c
e
s
s
.
I
t’
s
c
om
pos
e
d
of
a
n
in
put
la
ye
r
,
m
a
ny
hi
dde
n
la
ye
r
s
in
be
twe
e
n
,
a
nd
a
n
out
put
la
ye
r
a
s
s
how
n
in
F
ig
ur
e
1
li
ke
th
e
m
ul
ti
la
ye
r
pe
r
c
e
pt
r
on
(
M
L
P
)
[
13]
ne
twor
ks
.
B
e
s
t
known
a
nd
us
e
d
la
ye
r
s
a
r
e
:
c
onvolut
io
n,
a
c
ti
va
ti
on
o
r
R
e
L
U
,
a
n
d
pool
in
g
[
14]
.
T
he
c
onvolut
io
na
l
la
ye
r
is
th
e
m
os
t
im
po
r
ta
nt
o
ne
,
it
ta
ke
s
a
c
onvolut
io
n
ke
r
ne
l
a
ls
o
c
a
ll
e
d
a
m
a
s
k or
a
f
i
lt
e
r
t
he
n
pa
s
s
i
t
ove
r
t
he
da
ta
(
us
ua
ll
y i
m
a
ge
s
)
a
nd
t
r
a
ns
f
or
m
i
t
ba
s
e
d on the
va
lu
e
s
f
r
om
t
he
f
il
te
r
a
s
s
how
n
in
F
ig
ur
e
1
.
T
he
n
it
c
a
lc
ul
a
te
s
th
e
f
e
a
tu
r
e
m
a
p
va
lu
e
s
us
in
g
th
e
f
or
m
ul
a
(
1)
,
w
he
r
e
th
e
in
put
da
ta
is
r
e
pr
e
s
e
nt
e
d
by
a
nd
th
e
ke
r
ne
l
by
ℎ
,
a
nd
a
r
e
r
e
s
pe
c
ti
ve
ly
th
e
in
de
xe
s
of
r
ow
s
a
nd
c
ol
um
ns
of
th
e
r
e
s
ul
ta
nt
m
a
tr
ix
[
15]
.
T
he
pool
in
g
la
ye
r
it
is
w
ha
t
a
c
hi
e
ve
s
pr
ogr
e
s
s
iv
e
ly
dow
n
s
a
m
pl
in
g
to
r
e
duc
e
th
e
s
iz
e
of
th
e
s
uc
c
e
e
di
ng
la
ye
r
s
th
r
ough
m
a
x
pool
in
g
or
a
ve
r
a
ge
pool
in
g
to
h
e
lp
ove
r
f
it
ti
ng.
M
a
x
pool
in
g
di
vi
de
s
th
e
in
put
in
to
non
-
ove
r
la
ppi
ng c
lu
s
te
r
s
a
nd s
e
le
c
ts
t
he
m
a
xi
m
um
va
lu
e
f
or
e
a
c
h c
lu
s
te
r
i
n t
he
pr
e
vi
ous
l
a
ye
r
[
16]
.
[
,
]
=
(
∗
ℎ
)
[
,
]
=
∑
∑
ℎ
[
,
]
[
−
,
−
]
(
1)
I
n
p
u
t
O
u
t
p
u
t
I
n
p
u
t
L
a
y
e
r
O
u
t
p
u
t
L
a
y
e
r
H
i
d
e
n
L
a
y
e
r
s
C
o
n
v
o
l
u
t
i
o
n
a
l
l
a
y
e
r
P
o
o
l
i
n
g
l
a
y
e
r
F
u
l
l
y
c
o
n
n
e
c
t
e
d
F
ig
ur
e
1
. C
N
N
la
ye
r
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
10
, N
o.
1
,
M
a
r
c
h
20
2
1
:
110
–
120
112
C
ont
r
a
r
iwi
s
e
to
th
e
tr
a
di
ti
ona
l
f
e
a
tu
r
e
s
e
le
c
ti
on
a
lg
or
it
hm
s
i
t
ha
s
th
e
c
a
pa
bi
li
ty
of
le
a
r
ni
ng
be
tt
e
r
f
e
a
tu
r
e
s
a
ut
om
a
ti
c
a
ll
y
a
nd
c
a
te
gor
iz
e
th
e
tr
a
f
f
ic
.
I
n
a
ddi
ti
on,
it
c
a
n
a
c
hi
e
v
e
be
tt
e
r
c
la
s
s
if
ic
a
ti
on
a
nd
le
a
r
n
a
ddi
ti
ona
l
f
e
a
tu
r
e
s
w
it
h
m
or
e
tr
a
f
f
ic
da
ta
be
c
a
us
e
it
s
ha
r
e
s
th
e
s
a
m
e
c
onvolut
io
n
m
a
tr
ix
(
ke
r
ne
l)
,
th
a
t
w
oul
d
de
c
r
e
a
s
e
th
e
num
be
r
of
pa
r
a
m
e
te
r
s
a
nd
c
a
l
c
ul
a
ti
on
s
um
of
tr
a
in
in
g
s
ig
ni
f
ic
a
nt
ly
.
T
hi
s
gi
ve
s
C
N
N
a
f
a
s
t
r
e
c
ogni
ti
on
of
a
tt
a
c
k
na
tu
r
e
,
c
ont
r
a
r
iwi
s
e
to
ot
h
e
r
de
e
p
-
le
a
r
ni
n
g
a
lg
or
it
hm
s
,
or
m
a
c
hi
ne
le
a
r
ni
ng
a
lg
or
it
hm
s
th
a
t
c
a
n
be
ove
r
-
f
it
te
d
w
it
h
m
a
s
s
iv
e
bi
g
da
ta
.
M
or
e
ove
r
,
th
e
li
te
r
a
tu
r
e
s
how
s
th
a
t
us
in
g
C
N
N
s
in
in
tr
us
io
n
de
te
c
ti
on f
ie
ld
gi
ve
s
be
tt
e
r
r
e
s
ul
ts
t
ha
n ot
he
r
a
lg
or
it
hm
s
[
17
-
18]
.
R
e
c
ur
r
e
nt
ne
ur
a
l
ne
twor
ks
(
R
N
N
)
a
r
e
a
c
la
s
s
of
de
e
p
ne
ur
a
l
ne
twor
ks
th
a
t
c
ont
a
in
s
f
e
e
dba
c
k
c
onne
c
ti
ons
a
s
s
how
n
in
F
ig
ur
e
2
.
T
he
f
ul
ly
c
onne
c
te
d
la
ye
r
w
or
ks
on
a
f
la
tt
e
ne
d
in
put
w
he
r
e
e
a
c
h
of
th
e
s
e
in
put
s
is
c
onne
c
te
d
to
a
ll
ne
ur
ons
.
T
he
a
c
ti
va
ti
on
f
unc
ti
on
of
a
node
de
s
c
r
ib
e
s
it
s
out
put
gi
ve
n a
one
or
s
e
t
of
in
put
s
.
R
e
c
ti
f
ie
d
li
ne
a
r
uni
t
(
R
e
L
U
)
a
c
ti
va
ti
on
a
s
s
how
n
in
F
ig
ur
e
3
(
a
)
,
it
is
a
R
e
L
U
us
e
d
on
a
ll
e
le
m
e
nt
s
of
th
e
vol
um
e
.
I
t
a
im
s
a
t
in
tr
oduc
in
g
non
-
li
ne
a
r
it
ie
s
to
th
e
ne
twor
k.
L
S
T
M
is
c
om
pos
e
d
of
m
e
m
or
y
bl
oc
ks
th
a
t
a
r
e
a
s
e
t
of
r
e
c
ur
r
e
nt
c
onne
c
te
d s
ubne
twor
ks
. T
he
s
e
bl
oc
ks
a
r
e
c
om
pos
e
d w
it
h a
s
e
lf
-
c
onne
c
te
d m
e
m
or
y c
e
ll
s
(
one
or
m
a
ny)
a
s
s
how
n
in
F
ig
ur
e
4
w
hi
c
h
of
f
e
r
a
m
e
m
or
y
to
r
e
m
e
m
be
r
th
e
pr
e
vi
ous
da
ta
,
a
nd
th
r
e
e
uni
t
s
c
a
ll
e
d
ga
te
s
:
in
put
ga
te
(
3.a
)
,
a
f
or
ge
t
ga
te
(
3.b)
a
nd
a
n
out
pu
t
ga
te
(
3.c
)
w
hi
c
h
th
e
y
p
r
ovi
de
a
c
ont
in
uous
e
qui
va
le
nt
of
w
r
it
e
,
r
e
a
d
a
nd
r
e
s
e
t
ope
r
a
ti
ons
[
21]
.
T
he
s
e
ga
te
s
a
r
e
s
ig
m
oi
d
a
s
s
how
n
in
F
ig
ur
e
3(
b)
a
nd
ta
nh
a
s
s
how
n i
n F
ig
ur
e
3(
c
)
a
c
ti
va
ti
on
f
unc
ti
ons
m
e
a
ni
ng t
ha
t
th
e
ir
out
put
i
s
a
va
lu
e
be
twe
e
n 0 a
nd 1 f
or
s
ig
m
oi
d,
a
nd
be
twe
e
n
-
1
a
nd
1
f
or
ta
nh.
D
e
r
iv
e
d
f
r
om
f
e
e
df
or
w
a
r
d
ne
ur
a
l
ne
twor
ks
(
F
N
N
)
but
unl
ik
e
F
N
N
,
th
e
r
e
a
r
e
lo
ops
(
bi
di
r
e
c
ti
ona
l
da
ta
f
lo
w
)
a
nd
m
e
m
or
ie
s
to
r
e
m
e
m
be
r
pr
e
vi
ous
c
om
put
a
ti
ons
a
s
s
how
n
in
F
ig
ur
e
4
[
19
]
.
A
nd
a
ll
ow
in
g
pr
e
c
e
di
ng
out
put
s
to
be
us
e
d a
s
in
put
s
w
hi
le
ha
v
in
g
hi
dde
n
s
ta
te
s
[
20]
w
he
r
e
f
or
e
a
c
h
ti
m
e
s
te
p
, t
he
a
c
ti
va
ti
on
<
>
a
nd t
he
out
put
<
>
a
r
e
e
xpr
e
s
s
e
d:
(
2)
W
he
r
e
,
,
,
,
a
r
e
c
oe
f
f
ic
ie
nt
s
th
a
t
a
r
e
s
ha
r
e
d
te
m
por
a
ll
y
a
nd
1
,
2
a
c
ti
va
ti
on
f
unc
ti
ons
In
p
u
t
L
a
y
er
O
u
t
p
u
t
L
a
y
er
H
i
d
e
n
L
a
yer
s
In
p
u
t
O
u
t
p
u
t
F
ig
ur
e
2
. R
e
c
ur
r
e
nt
ne
ur
a
l
ne
twor
ks
l
a
ye
r
s
R
N
N
c
a
n
f
a
c
e
th
e
lo
ng
-
te
r
m
de
pe
nd
e
nc
y
pr
obl
e
m
a
nd
th
e
va
ni
s
hi
ng
gr
a
di
e
nt
&
e
xpl
odi
ng
gr
a
di
e
nt
,
s
o
w
e
c
it
e
he
r
e
th
e
be
s
t
known R
N
N
ne
twor
k
s
,
th
e
lo
ng
s
hor
t
-
te
r
m
m
e
m
or
y
(
L
S
T
M
)
a
nd
ga
te
d
r
e
c
ur
r
e
nt
uni
t
(
G
R
U
)
ne
twor
ks
to
s
ol
ve
th
e
s
e
pr
obl
e
m
s
.
T
he
m
a
in
di
f
f
e
r
e
n
c
e
to
s
im
pl
e
R
N
N
is
th
a
t
th
e
nonl
in
e
a
r
uni
ts
in
th
e
hi
dde
n
la
ye
r
s
a
r
e
r
e
pl
a
c
e
d
by
m
e
m
or
y
bl
oc
ks
.
T
he
f
ol
lo
w
i
ng
f
or
m
ul
a
(
3
)
r
e
pr
e
s
e
nt
s
th
e
ga
te
s
in
L
S
T
M
[
22]
.
(3
)
W
he
r
e
:
a
r
e
t
he
i
nput
ga
te
s
(
“
”
f
or
t
he
i
nput
ga
te
, “
”
f
or
t
he
f
or
g
e
t,
a
nd “
”
f
or
t
he
out
put
ga
te
)
;
“
”
it
i
s
t
he
s
ig
m
oi
d f
unc
ti
on;
“
”
i
t
is
t
he
bi
a
s
e
s
f
or
t
he
ga
te
(
x)
;
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
T
ow
ar
d a de
e
p l
e
a
r
ni
ng
-
bas
e
d i
nt
r
us
io
n d
e
te
c
ti
on s
y
s
te
m
f
or
I
o
T
agains
t
bot
ne
t
at
ta
c
k
s
(
I
dr
is
s
I
dr
i
s
s
i
)
113
“
ℎ
−
1
”
i
t
is
t
he
out
put
of
t
he
pr
e
c
e
de
nt
L
S
T
M
bl
oc
k;
“
”
i
t
is
t
he
c
ur
r
e
nt
i
nput
.
G
a
te
d
r
e
c
ur
r
e
nt
uni
t
(
G
R
U
)
is
a
s
im
pl
if
ie
d
ve
r
s
io
n
of
L
S
T
M
,
w
he
r
e
th
e
G
R
U
m
odul
a
te
s
th
e
f
lo
w
of
in
f
or
m
a
ti
on
in
s
id
e
th
e
uni
t
us
in
g
ga
ti
ng
uni
ts
a
s
s
how
n
in
F
ig
ur
e
4
,
w
it
hout
s
e
pa
r
a
ti
ng
th
e
m
e
m
or
y
c
e
ll
s
[
23
-
24]
.
I
t
m
e
r
ge
s
th
e
f
or
ge
t
a
nd
th
e
in
put
ga
te
s
in
to
a
n
“
up
da
te
ga
te
”
,
a
ls
o
m
e
r
ge
s
c
e
ll
a
nd
hi
dde
n
s
ta
te
,
G
R
U
ha
s
f
e
w
e
r
pa
r
a
m
e
te
r
s
t
ha
n t
he
L
S
T
M
. I
t
is
de
f
in
e
d by the
f
ol
lo
w
in
g f
or
m
ul
a
s
.
(
4)
R
e
L
U
(
F
i
g
u
r
e
2
.
a
)
T
a
n
h
(
F
i
g
u
r
e
2
.
c
)
S
i
g
m
o
i
d
(
F
i
g
u
r
e
2
.
b
)
-
1
1
1
0
1
0
0
(
a
)
(
b)
(
c
)
F
ig
ur
e
3
. A
c
ti
va
ti
on f
unc
ti
ons
:
(
a
)
R
e
L
U
, (
b)
S
ig
m
oi
d, (
c
)
T
a
nh
x
+
x
x
x
-
1
+
x
x
x
+
T
a
n
H
S
i
g
m
o
i
d
P
o
i
n
t
w
i
s
e
m
u
l
t
i
p
l
i
c
a
t
i
o
n
P
o
i
n
t
w
i
s
e
a
d
d
i
t
i
o
n
V
e
c
t
o
r
c
o
n
c
a
t
e
n
a
t
i
o
n
R
N
N
L
S
T
M
G
R
U
i
n
p
ut
g
a
t
e
o
ut
p
ut
g
a
t
e
f
o
r
g
e
t
g
a
t
e
ce
l
l
st
a
t
e
up
d
a
t
e
g
a
t
e
r
e
se
t
g
a
t
e
F
ig
ur
e
4
. R
N
N
, L
S
T
M
&
G
R
U
bl
oc
ks
[
25]
3.
R
E
L
A
T
E
D
WO
R
K
S
K
or
oni
ot
is
e
t
al
.
[
26]
a
ppl
ie
d
s
uppor
t
ve
c
to
r
m
a
c
hi
ne
(
S
V
M
)
,
L
S
T
M
a
nd
R
N
N
t
o
e
v
a
lu
a
te
th
e
I
o
T
-
D
a
ta
s
e
t.
T
he
a
ut
hor
s
f
oc
u
s
e
d
on
bi
na
r
y
c
la
s
s
if
ic
a
ti
on
on
th
e
d
a
ta
s
e
t,
a
nd
th
e
ir
pr
e
di
c
ti
on
r
e
s
ul
t
w
a
s
e
it
he
r
a
“
nor
m
a
l
tr
a
f
f
ic
”
or
“
s
om
e
ty
pe
of
a
tt
a
c
ks
”
(
f
or
e
ve
r
y
ty
pe
of
a
tt
a
c
k)
,
w
hi
c
h
is
not
h
e
lp
f
ul
f
or
im
pl
e
m
e
nt
in
g
m
a
ny
m
ode
ls
(
f
or
e
ve
r
y
a
tt
a
c
k
ty
pe
)
to
a
w
or
ki
ng
I
D
S
c
ont
r
a
r
y
to
a
m
ul
ti
-
la
be
l
out
put
(
num
e
r
ous
c
a
te
gor
ie
s
of
a
tt
a
c
ks
)
t
ha
t
gi
ve
s
one
a
nd only on
e
m
ode
l.
I
bi
to
ye
e
t
al
.
[
27]
in
th
e
ir
r
e
s
e
a
r
c
h
c
om
pa
r
e
d
th
e
pe
r
f
o
r
m
a
nc
e
be
twe
e
n
two
de
e
p
le
a
r
ni
ng
m
ode
ls
s
e
lf
nor
m
a
li
z
in
g
ne
twor
ks
(
S
N
N
)
a
nd
f
e
e
d
f
or
w
a
r
d
ne
ur
a
l
n
e
twor
ks
(
F
N
N
)
in
s
id
e
th
e
m
il
ie
u
of
a
n
I
oT
ne
twor
k.
T
hi
s
c
om
pa
r
is
on
s
ho
w
s
th
a
t
F
N
N
out
pe
r
f
or
m
s
S
N
N
,
e
ve
n
if
S
N
N
r
e
m
a
in
s
be
tt
e
r
in
r
e
ga
r
ds
to
a
dve
r
s
a
r
ia
l
s
a
m
pl
e
s
.
A
ls
o,
th
e
a
ut
hor
s
e
xa
m
in
e
d
th
e
im
pa
c
t
of
f
e
a
tu
r
e
n
or
m
a
li
z
a
ti
on
on
th
e
a
dve
r
s
a
r
ia
l
s
tr
e
ngt
h a
nd de
m
ons
tr
a
te
d i
ts
ba
d i
nf
lu
e
nc
e
t
o a
dve
r
s
a
r
ia
l
a
tt
a
c
ks
r
e
s
is
ti
ng.
F
e
r
r
a
g
e
t
al
.
[
28]
in
hi
s
pa
pe
r
c
onduc
te
d
a
c
om
pa
r
a
ti
ve
s
tu
dy
w
it
h
two
da
ta
s
e
ts
,
B
ot
-
I
oT
a
nd
C
S
E
-
C
I
C
-
I
D
S
2018
da
ta
s
e
ts
us
in
g
s
om
e
de
e
p
le
a
r
ni
ng
a
ppr
oa
c
he
s
,
s
uc
h
R
N
N
,
C
N
N
,
B
ol
tz
m
a
nn
m
a
c
hi
ne
,
de
e
p
be
li
e
f
ne
twor
ks
(
D
B
N
)
,
de
e
p
B
ol
tz
m
a
nn
m
a
c
hi
n
e
s
(
D
B
M
)
,
de
e
p
a
ut
oe
nc
ode
r
s
a
nd
de
e
p
di
s
c
r
im
in
a
ti
ve
m
ode
ls
, w
it
h 100 hidden la
ye
r
s
t
o ge
t
a
n a
c
c
ur
a
c
y of
98.394%
.
M
e
ngm
e
ng
[
29]
pr
opos
e
d
us
in
g
F
N
N
a
n
in
te
ll
ig
e
nt
bi
na
r
y
a
nd
m
ul
ti
c
la
s
s
c
la
s
s
if
ic
a
ti
on
,
but
w
it
h
ju
s
t
f
e
w
c
la
s
s
e
s
to
ge
t
99%
in
a
ll
e
va
lu
a
ti
on
m
e
a
s
ur
e
s
(
a
c
c
ur
a
c
y,
pr
e
c
is
io
n,
r
e
c
a
ll
a
nd
F
1
s
c
or
e
)
f
or
D
D
oS
/DoS
a
tt
a
c
ks
w
hi
le
t
he
nor
m
a
l
tr
a
f
f
ic
c
la
s
s
if
ic
a
ti
on got a
n
a
c
c
ur
a
c
y of
98
%
.
A
lKa
di
[
30]
pr
opos
e
d
a
s
ys
te
m
n
a
m
e
d
m
ix
tu
r
e
lo
c
a
li
z
a
ti
on
-
b
a
s
e
d
out
li
e
r
s
(
M
L
O
)
on
th
e
B
o
T
-
I
oT
D
a
ta
s
e
t
th
a
t
us
e
s
ut
il
iz
e
s
ga
us
s
ia
n
-
m
ix
tu
r
e
m
ode
ls
f
or
f
it
ti
ng
ne
twor
k
da
ta
a
nd
a
lo
c
a
l
out
li
e
r
f
a
c
to
r
f
unc
ti
on
f
or
di
s
c
ove
r
in
g a
bnor
m
a
l
pa
tt
e
r
ns
i
n ne
twor
k t
r
a
f
f
ic
da
ta
, a
nd gott
e
n a
n a
c
c
ur
a
c
y of
97.98
%.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
10
, N
o.
1
,
M
a
r
c
h
20
2
1
:
110
–
120
114
I
n
f
a
c
t,
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
ks
(
C
N
N
or
C
onvNe
ts
)
a
r
e
a
c
la
s
s
of
de
e
p
ne
ur
a
l
ne
twor
ks
th
a
t
a
r
e
us
e
d
in
m
a
ny
f
ie
ld
s
but
m
os
tl
y
in
pa
tt
e
r
n
r
e
c
ogni
ti
on.
C
N
N
is
a
c
la
s
s
of
ne
ur
a
l
ne
twor
ks
th
a
t
us
e
s
th
e
c
onvolut
io
n
a
nd
th
e
pool
in
g
la
ye
r
s
in
s
te
a
d
of
th
e
f
ul
ly
c
onne
c
te
d
hi
dde
n
la
ye
r
s
[
31]
.
C
ont
r
a
r
iwi
s
e
to
th
e
tr
a
di
ti
ona
l
f
e
a
tu
r
e
s
e
le
c
ti
on
a
lg
or
it
hm
s
it
ha
s
th
e
c
a
pa
bi
li
ty
of
le
a
r
ni
ng
be
tt
e
r
f
e
a
tu
r
e
s
a
ut
om
a
ti
c
a
ll
y
a
nd
c
a
te
gor
iz
e
th
e
tr
a
f
f
ic
.
I
n
a
ddi
ti
on,
it
c
a
n
a
c
hi
e
ve
be
tt
e
r
c
la
s
s
if
i
c
a
ti
on
a
nd
le
a
r
n
a
ddi
ti
ona
l
f
e
a
tu
r
e
s
w
it
h
m
or
e
tr
a
f
f
ic
da
ta
be
c
a
u
s
e
it
s
ha
r
e
s
th
e
s
a
m
e
c
onvolut
io
n
m
a
tr
ix
(
m
a
s
k)
,
th
a
t
w
oul
d
de
c
r
e
a
s
e
th
e
num
be
r
of
pa
r
a
m
e
te
r
s
a
nd
c
a
lc
ul
a
ti
on
s
um
of
tr
a
in
in
g
s
ig
ni
f
ic
a
nt
ly
.
T
hi
s
gi
ve
s
C
N
N
a
f
a
s
t
r
e
c
ogni
ti
on
of
a
tt
a
c
k
na
tu
r
e
,
c
ont
r
a
r
iwi
s
e
to
ot
he
r
de
e
p
-
le
a
r
ni
ng
a
lg
or
it
hm
s
,
or
m
a
c
hi
ne
le
a
r
ni
ng
a
lg
or
it
hm
s
th
a
t
c
a
n
be
ove
r
-
f
it
te
d
w
it
h
m
a
s
s
iv
e
bi
g
da
ta
.
M
or
e
ove
r
,
th
e
li
te
r
a
tu
r
e
s
how
s
th
a
t
us
in
g
C
N
N
s
in
in
tr
us
io
n
de
te
c
ti
on
f
ie
ld
gi
ve
s
b
e
tt
e
r
r
e
s
ul
ts
t
ha
n ot
he
r
a
lg
or
it
hm
s
[
17
-
18]
.
T
he
r
e
la
te
d
w
or
k
li
s
te
d
a
bove
c
a
n
pr
ovi
de
a
good
pr
e
di
c
ti
on
o
f
bot
ne
t
a
tt
a
c
ks
th
a
t
c
a
n
a
f
f
e
c
t
a
n
I
oT
s
ys
te
m
.
H
ow
e
v
e
r
,
th
e
s
e
w
or
ks
c
oul
d
not
r
e
c
ogni
z
e
ty
pe
of
a
tt
a
c
k
due
to
th
e
bi
na
r
y
c
la
s
s
if
ic
a
ti
on.
H
e
nc
e
our
s
tu
dy
c
ons
ti
tu
te
s
a
n
im
por
ta
nt
e
xpe
r
im
e
nt
a
l
e
xt
e
ns
io
n
of
th
e
a
bove
-
m
e
nt
io
ne
d
w
or
ks
by
be
nc
hm
a
r
ki
ng
th
e
B
ot
-
I
oT
da
ta
s
e
t
us
in
g
C
N
N
c
om
pa
r
e
d
to
di
f
f
e
r
e
nt
de
e
p
le
a
r
ni
ng
m
ode
ls
,
us
in
g
th
e
m
ul
ti
la
be
l
c
la
s
s
if
ic
a
ti
on
c
or
r
e
s
ponding t
o va
r
io
us
c
a
te
gor
ie
s
o
f
a
tt
a
c
k
s
i
n t
he
I
oT
.
4.
P
R
O
P
O
S
E
D
M
E
T
H
O
D
B
ot
I
D
S
is
our
pr
opos
e
d ne
twor
k
I
D
S
obt
a
in
e
d by
l
e
a
r
ni
ng
f
r
o
m
t
he
B
ot
-
I
oT
da
ta
s
e
t.
T
hi
s
s
ol
ut
io
n i
s
pl
a
nne
d
to
be
pl
a
c
e
d
in
a
f
og
node
w
he
n
it
w
il
l
be
im
pl
e
m
e
nt
e
d
in
a
r
e
a
l
I
oT
e
nvi
r
onm
e
nt
.
A
s
u
c
h
de
pl
oym
e
nt
gi
ve
s
i
t
th
e
pow
e
r
of
a
na
ly
z
in
g i
n r
e
a
l
ti
m
e
t
he
i
nboun
d a
nd outbound t
r
a
f
f
ic
t
hr
ough
th
e
ne
twor
k
by
s
ni
f
f
in
g
it
a
s
s
how
n
in
F
ig
ur
e
5
. T
hi
s
lo
c
a
ti
on
w
il
l
m
a
ke
our
B
ot
I
D
S
a
bl
e
to
m
oni
to
r
a
ll
tr
a
f
f
ic
to
/f
r
om
th
e
de
vi
c
e
s
bot
h
in
s
id
e
a
nd
out
s
id
e
th
e
ne
twor
k,
a
nd
e
v
e
n
pa
r
ti
c
ul
a
r
ly
th
e
tr
a
f
f
ic
be
twe
e
n
th
is
in
s
id
e
r
d
e
vi
c
e
s
in
c
a
s
e
of
a
l
ur
ki
ng z
om
bi
e
de
vi
c
e
i
n
s
id
e
t
he
ne
twor
k.
W
S
N
/
M
2
M
F
o
g
n
o
d
e
I
n
t
e
r
n
e
t
AP
B
o
tI
D
S
s
e
r
v
e
r
R
o
u
t
e
r
/
G
a
te
w
a
y
C
l
ou
d
D
a
t
a
b
a
s
e
AP
AP
I
P
c
a
m
e
r
a
(
Z
o
m
b
i
e
)
I
P
c
a
m
e
r
a
(
Z
o
m
b
i
e
)
I
P
c
a
m
e
r
a
(
Z
o
m
b
i
e
)
Sm
a
r
t
T
V
(
Z
om
b
i
e
)
Sm
a
r
t
Pho
ne
(
Z
o
m
b
i
e
)
Sm
a
r
t
Ph
o
n
e
(
Z
o
m
b
i
e
)
C
o
m
p
u
t
e
r
(
Z
o
m
b
i
e
)
T
a
b
l
e
t
(
Z
o
m
b
i
e
)
F
ig
ur
e
5
. A
r
c
hi
te
c
tu
r
e
of
pr
opos
e
d a
ppr
oa
c
h
T
he
B
ot
I
D
S
i
s
a
de
e
p l
e
a
r
ni
ng
-
ba
s
e
d m
e
th
od on a
de
e
p l
e
a
r
ni
ng mode
l
th
a
t
c
ont
a
in
s
t
hr
e
e
pha
s
e
s
:
a
s
s
how
n i
n
F
ig
ur
e
6
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
T
ow
ar
d a de
e
p l
e
a
r
ni
ng
-
bas
e
d i
nt
r
us
io
n d
e
te
c
ti
on s
y
s
te
m
f
or
I
o
T
agains
t
bot
ne
t
at
ta
c
k
s
(
I
dr
is
s
I
dr
i
s
s
i
)
115
U
n
i
f
yi
n
g
d
a
t
a
f
o
rm
a
t
No
rma
l
i
z
i
n
g
t
h
e
d
a
t
a
va
l
u
es
C
o
n
vert
In
g
t
h
e
d
a
t
a
i
n
t
o
i
ma
g
e
s
h
a
p
e
S
p
l
i
t
i
n
g
t
h
e
d
a
t
a
s
e
t
t
o
T
ra
i
n
i
n
g
&
T
e
s
t
i
n
g
s
e
t
B
u
i
l
d
i
n
g
t
h
e
d
eep
l
ea
rn
i
n
g
mo
d
el
T
ra
i
n
i
n
g
t
h
e
mo
d
el
B
a
d
P
e
rf
o
rm
a
n
c
e
?
O
r
T
ra
i
n
i
n
g
s
t
o
p
p
ed
?
No
Y
e
s
T
e
s
t
i
n
g
d
a
t
a
s
e
t
T
h
e
ra
w
D
a
t
a
s
et
Sa
vi
n
g
t
h
e
m
o
d
e
l
T
ra
i
n
i
n
g
d
a
t
a
s
et
A
d
j
u
s
t
i
n
g
t
h
e
p
a
ra
met
ers
T
e
s
t
i
n
g
t
h
e
mo
d
e
l
F
ig
ur
e
6
.
M
ode
l
bui
ld
in
g pr
oc
e
s
s
−
1s
t
P
ha
s
e
:
D
a
ta
s
e
t
pr
e
pr
oc
e
s
s
in
g;
F
ir
s
t
of
a
ll
, w
e
ne
e
d t
o a
lt
e
r
t
he
r
a
w
da
ta
a
nd nor
m
a
li
z
e
i
ts
va
lu
e
s
w
it
h
th
e
goa
l
of
t
he
be
s
t
pe
r
f
or
m
a
nc
e
of
de
e
p l
e
a
r
ni
ng mode
l,
a
nd t
h
e
n c
onve
r
t
it
in
to
i
m
a
ge
s
ha
pe
.
−
2nd
P
ha
s
e
:
B
ui
ld
in
g
th
e
m
ode
l;
th
e
m
ode
l
is
a
t
f
ir
s
t
f
it
on
a
t
r
a
in
in
g
da
ta
s
e
t
(
a
pa
r
t
of
th
e
da
ta
s
e
t)
us
in
g
pa
r
a
m
e
te
r
s
to
a
c
hi
e
ve
th
e
im
pr
ove
m
e
nt
of
th
e
m
ode
l
pe
r
f
or
m
a
nc
e
,
th
e
s
e
p
a
r
a
m
e
te
r
s
a
r
e
c
h
a
nge
d
in
th
e
tr
a
in
in
g
pr
oc
e
s
s
to
r
e
a
c
h
b
e
tt
e
r
pe
r
f
or
m
a
nc
e
.
S
e
c
ondl
y,
th
e
te
s
t
da
ta
s
e
t
(
th
e
r
e
m
a
in
in
g
pa
r
t
of
th
e
da
ta
s
e
t)
i
s
us
e
d t
o va
li
da
te
t
he
a
c
c
ur
a
c
y of
t
he
m
ode
l.
−
3r
d
P
ha
s
e
:
E
va
lu
a
ti
ng
th
e
m
ode
l
by
pr
e
di
c
ti
on;
a
f
te
r
bui
ld
in
g
a
nd
ge
ne
r
a
ti
ng
th
e
m
ode
l,
w
e
e
v
a
lu
a
te
th
is
m
ode
l
w
it
h t
he
t
e
s
t
da
t
a
s
e
t
by pr
e
di
c
ti
ng a
tt
a
c
ks
a
nd c
a
lc
ul
a
ti
ng t
he
t
im
e
ne
e
de
d f
or
t
hi
s
pr
e
di
c
ti
on.
4.1. Dat
as
e
t
p
r
e
p
r
oc
e
s
s
in
g
F
or
c
onduc
ti
ng
pr
opos
e
d
w
or
k,
w
e
ha
ve
us
e
d
th
e
la
te
s
t
B
o
t
-
I
oT
da
ta
s
e
t
[
32]
th
a
t
w
a
s
c
r
e
a
te
d
s
pe
c
if
ic
a
ll
y
f
or
I
oT
s
ys
te
m
s
by
a
n
a
c
tu
a
l
ne
twor
k
m
il
ie
u
a
t
th
e
C
ybe
r
R
a
nge
L
a
b
of
th
e
C
e
nt
e
r
of
U
N
S
W
C
a
nbe
r
r
a
C
ybe
r
.
T
he
e
nvi
r
onm
e
nt
in
c
or
por
a
te
s
a
c
om
bi
na
ti
on of
us
ua
l
nor
m
a
l
a
nd
ba
d
tr
a
f
f
i
c
,
w
it
h
s
ix
ty
pe
s
of
a
tt
a
c
ks
a
nd
10
s
ubc
a
te
gor
ie
s
,
na
m
e
ly
,
r
e
c
onna
i
s
s
a
n
c
e
(
s
e
r
vi
c
e
s
c
a
nni
ng
a
nd
O
S
f
in
ge
r
pr
in
ti
ng)
,
D
D
o
S
(
T
C
P
, U
D
P
a
nd H
T
T
P
)
, D
oS
(
T
C
P
, U
D
P
a
nd H
T
T
P
)
, t
he
f
t
(
ke
y l
oggi
ng a
nd da
ta
e
xf
il
tr
a
ti
on)
.
W
it
h
72
m
il
li
on
r
e
c
or
ds
of
da
ta
tr
a
f
f
ic
s
im
ul
a
te
d
I
oT
e
nvi
r
on
m
e
nt
.
T
he
w
hol
e
da
ta
w
a
s
a
s
c
e
nd
e
d
dow
n
to
5%
in
to
a
“
f
ul
l
-
f
e
a
tu
r
e
”
da
ta
s
e
t
w
it
h
a
r
ound
3.6
m
il
li
on
r
e
c
or
ds
a
nd
a
not
he
r
ve
r
s
io
n
c
a
ll
e
d
“
10
be
s
t
f
e
a
tu
r
e
s
”
is
a
l
s
o
pr
ovi
de
d
w
it
h
s
e
le
c
ti
on
of
be
s
t
f
e
a
tu
r
e
s
f
r
om
th
e
“
F
ul
l
f
e
a
tu
r
e
s
”
ve
r
s
io
n,
bot
h
ve
r
s
io
ns
a
r
e
us
e
d
f
or
our
e
xpe
r
im
e
nt
.
T
he
tr
a
in
in
g
a
nd
te
s
t
da
ta
s
e
t
ha
ve
11 output
c
la
s
s
e
s
w
hi
c
h
r
e
f
le
c
t
th
e
nor
m
a
l
tr
a
f
f
ic
,
a
nd t
he
10 t
ype
s
of
a
tt
a
c
k
s
w
hi
c
h w
e
r
e
c
a
r
r
ie
d out a
ga
in
s
t
th
e
I
oT
ne
twor
k.
T
he
B
ot
-
I
O
T
d
a
ta
s
e
t
c
ont
a
in
s
n
e
twor
k
c
onne
c
ti
on
a
tt
r
ib
u
te
s
;
nom
in
a
l,
num
e
r
ic
a
nd
I
P
a
ddr
e
s
s
e
s
.
W
e
c
onve
r
t
th
e
nom
in
a
l
da
ta
to
num
e
r
ic
da
ta
,
ip
v4
a
nd
ip
v
6
a
ls
o
be
c
onv
e
r
te
d
to
num
e
r
ic
a
l
s
ha
p
e
,
a
nd
m
e
r
gi
ng
c
a
te
gor
y
a
nd
s
ubc
a
t
e
gor
ie
s
f
ie
ld
s
in
to
one
f
ie
ld
th
a
t
c
ont
a
in
s
10
ty
pe
s
of
a
tt
a
c
k
s
a
nd
th
e
11t
h
i
s
a
nor
m
a
l
tr
a
f
f
ic
th
e
n
w
e
c
onve
r
t
th
e
ne
w
c
a
te
gor
y
a
tt
r
ib
ut
e
us
in
g
“
one
-
hot
e
nc
odi
ng”
,
a
nd
dr
oppe
d
th
e
bi
na
r
y
“
A
tt
a
c
k”
f
ie
ld
c
a
us
e
. O
ur
f
oc
us
i
s
on a
m
ul
ti
la
be
l
out
put
a
nd no
t
a
bi
na
r
y one
.
A
f
te
r
e
nc
odi
ng
th
e
da
ta
,
w
e
nor
m
a
li
z
e
d
it
us
in
g
S
c
ik
it
-
l
e
a
r
n;
m
e
a
ni
ng
s
c
a
li
ng
th
e
v
e
c
to
r
s
in
di
vi
dua
ll
y t
o unit
nor
m
, a
nd c
onve
r
ti
ng t
he
nor
m
a
li
z
e
d output
da
ta
i
nt
o i
m
a
ge
da
ta
s
ha
p
e
.
T
he
n
w
e
s
pl
it
th
e
da
ta
a
t
f
ir
s
t
in
to
d
a
ta
X
(
c
ont
a
in
s
a
ll
th
e
f
e
a
tu
r
e
s
e
xc
e
pt
th
e
“
c
a
te
gor
y”
f
e
a
tu
r
e
)
a
nd
la
be
l
Y
(
c
ont
a
in
s
th
e
“
c
a
te
gor
y”
f
e
a
tu
r
e
)
,
a
nd
th
e
n
s
pl
it
it
in
to
r
a
ndom
tr
a
in
in
g
s
ubs
e
t
a
nd
te
s
ti
ng
s
ubs
e
t
w
it
h
75%
f
or
tr
a
in
in
g
s
e
t
a
nd
25%
f
or
th
e
te
s
ti
ng
s
e
t.
F
ig
ur
e
7
s
how
s
th
e
num
be
r
of
da
ta
r
ow
s
f
or
e
a
c
h
s
e
t
a
nd e
a
c
h a
tt
a
c
k t
yp
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
10
, N
o.
1
,
M
a
r
c
h
20
2
1
:
110
–
120
116
F
ig
ur
e
7
. A
tt
a
c
ks
di
s
tr
ib
ut
io
n i
n t
r
a
in
in
g a
nd t
e
s
ti
ng s
e
ts
4.2. B
u
il
d
in
g ou
r
m
od
e
ls
T
he
C
N
N
m
ode
l
s
w
e
r
e
de
fi
ne
d
to
ha
v
e
a
n
in
put
la
y
e
r
w
it
h
th
e
num
be
r
of
ne
ur
ons
e
qua
l
to
th
e
a
m
ount
of
in
put
f
e
a
tu
r
e
s
,
f
our
hi
dde
n
la
ye
r
s
C
onvolut
io
n2D
la
ye
r
,
M
a
xP
ool
in
g2D
la
ye
r
,
F
la
tt
e
n
la
ye
r
,
D
e
ns
e
la
ye
r
a
nd a
n output
la
ye
r
. F
or
t
he
be
s
t
f
e
a
tu
r
e
s
da
ta
s
e
t,
t
he
m
od
e
l
w
a
s
t
r
a
in
e
d i
n 10 e
poc
hs
(
th
e
w
hol
e
da
ta
s
e
t
is
pa
s
s
e
d
th
r
ough
th
e
ne
ur
a
l
ne
twor
k
10
ti
m
e
s
)
w
it
h
ba
tc
h
s
iz
e
of
32
(
a
m
ount
of
t
r
a
in
in
g
s
im
pl
e
s
in
a
s
in
gl
e
ba
tc
h i
s
32)
a
nd a
ke
r
ne
l
s
iz
e
of
(
1, 10)
. T
he
ne
ur
a
l
ne
twor
k c
o
m
pr
is
e
s
16 i
nput
ne
ur
ons
(
in
t
he
fi
r
s
t
la
ye
r
, t
he
s
a
m
e
num
be
r
a
s
th
e
f
e
a
tu
r
e
s
)
,
w
it
h
4
in
te
r
m
e
di
a
te
(
hi
d
de
n)
la
ye
r
s
w
it
h
32
(
C
onvolut
io
n2D
)
,
32
(
M
a
xP
ool
in
g2D
)
,
480
(
F
la
tt
e
n)
,
22
(
D
e
ns
e
)
ne
ur
ons
,
a
nd
11
out
put
ne
ur
ons
f
or
th
e
m
ul
ti
la
be
l
c
la
s
s
ifi
c
a
ti
on
a
s
s
how
n
in
F
ig
ur
e
8
.
F
or
our
f
ul
l
-
f
e
a
tu
r
e
da
ta
s
e
t,
th
e
m
ode
l
w
a
s
tr
a
in
e
d
in
15
e
poc
h
s
(
ba
tc
h
s
i
z
e
of
32
a
nd
a
ke
r
ne
l
s
iz
e
of
(
1,
10)
)
,
a
nd
ha
d
a
40
-
ne
ur
on
in
put
la
ye
r
,
s
a
m
e
num
be
r
a
nd
c
ons
is
te
nc
y
of
hi
dde
n
la
ye
r
s
a
s
w
it
h t
he
be
s
t
f
e
a
tu
r
e
m
ode
l
a
nd 11 o
ut
put
ne
ur
ons
f
or
t
he
m
ul
ti
la
be
l
c
la
s
s
ifi
c
a
ti
on. I
n both c
a
s
e
s
, f
or
t
he
i
nput
a
nd hidde
n l
a
ye
r
s
, t
he
a
c
ti
va
ti
on f
unc
ti
on t
ha
t
w
a
s
us
e
d w
a
s
‘
R
e
lu
’
, w
hi
le
t
he
out
put
l
a
ye
r
a
c
ti
va
ti
on f
unc
ti
on
w
a
s
‘
S
of
tm
a
x’
a
s
s
how
n i
n
F
ig
ur
e
8
a
nd T
a
bl
e
1
.
T
he
ot
he
r
r
e
c
ur
r
e
nt
ne
ur
a
l
ne
twor
k
(
R
N
N
)
m
ode
ls
w
e
r
e
de
fi
ne
d
to
ha
ve
one
in
put
la
ye
r
w
it
h
th
e
num
be
r
of
ne
ur
ons
e
qua
l
to
th
e
a
m
ount
of
in
put
f
e
a
tu
r
e
s
,
f
our
hi
dde
n
la
ye
r
s
a
nd
a
n
out
put
la
ye
r
.
F
or
th
e
be
s
t
f
e
a
tu
r
e
s
da
ta
s
e
t,
th
e
m
ode
l
w
a
s
tr
a
in
e
d
in
40
e
po
c
hs
w
it
h
32
in
th
e
ba
tc
h
s
iz
e
.
T
h
e
ne
ur
a
l
ne
twor
k
c
om
pr
is
e
d
16
in
put
ne
ur
ons
(
in
th
e
fi
r
s
t
la
ye
r
,
th
e
s
a
m
e
num
be
r
a
s
th
e
f
e
a
tu
r
e
s
)
,
w
it
h
4
(
S
im
pl
e
R
N
N
/L
S
T
M
/G
R
U
)
in
te
r
m
e
di
a
te
(
hi
dde
n)
la
ye
r
s
w
it
h
32,
64,
128,
22
ne
ur
ons
,
a
nd
11
out
put
n
e
ur
ons
f
or
th
e
m
ul
ti
la
be
l
c
la
s
s
ifi
c
a
ti
on a
s
s
ho
w
n i
n F
ig
ur
e
9.
F
or
our
f
ul
l
-
f
e
a
tu
r
e
da
ta
s
e
t,
t
he
m
ode
l
w
a
s
t
r
a
in
e
d i
n
40 e
poc
hs
(
32
in
t
he
ba
tc
h
s
iz
e
)
,
a
nd
ha
d
a
40
-
ne
ur
on
in
put
la
ye
r
,
s
a
m
e
num
be
r
a
nd
c
ons
is
te
n
c
y
of
hi
dde
n
la
y
e
r
s
a
s
w
it
h
th
e
b
e
s
t
f
e
a
tu
r
e
m
ode
l
a
nd
11
out
put
ne
ur
ons
f
or
th
e
m
ul
ti
la
be
l
c
la
s
s
ifi
c
a
ti
on.
I
n
bot
h
c
a
s
e
s
,
f
or
th
e
out
put
la
ye
r
a
c
ti
va
ti
on f
unc
ti
on w
a
s
‘
S
of
tm
a
x’
a
s
s
how
n i
n
F
ig
ur
e
9
, a
nd
T
a
bl
e
1
.
F
ig
ur
e
8
.
C
N
N
m
ode
l
la
ye
r
s
F
ig
ur
e
9
.
R
N
N
m
ode
ls
l
a
ye
r
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
T
ow
ar
d a de
e
p l
e
a
r
ni
ng
-
bas
e
d i
nt
r
us
io
n d
e
te
c
ti
on s
y
s
te
m
f
or
I
o
T
agains
t
bot
ne
t
at
ta
c
k
s
(
I
dr
is
s
I
dr
i
s
s
i
)
117
T
a
bl
e
1
.
M
od
e
ls
pa
r
a
m
e
te
r
s
f
or
t
he
i
m
pl
e
m
e
nt
e
d m
ode
l
s
DL
A
l
gor
i
t
hm
C
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
k
(
C
N
N
)
S
i
m
pl
e
r
e
c
ur
r
e
nt
ne
ur
a
l
ne
t
w
or
k (
R
N
N
)
L
ong s
hor
t
-
t
e
r
m
m
e
m
or
y
(
L
S
T
M
)
ba
s
e
d R
N
N
G
a
t
e
d r
e
c
ur
r
e
nt
uni
t
(
G
R
U
)
ba
s
e
d R
N
N
D
a
t
a
s
e
t
ve
r
s
i
on
F
ul
l
f
e
a
t
ur
e
s
B
e
s
t
f
e
a
t
ur
e
s
F
ul
l
f
e
a
t
ur
e
s
B
e
s
t
f
e
a
t
ur
e
s
F
ul
l
f
e
a
t
ur
e
s
B
e
s
t
f
e
a
t
ur
e
s
F
ul
l
f
e
a
t
ur
e
s
B
e
s
t
f
e
a
t
ur
e
s
E
poc
hs
15
10
40
40
30
25
40
20
L
a
ye
r
s
4
4
4
4
4
4
4
4
N
e
ur
ons
40 I
nput
C
onvol
ut
i
ona
l
l
a
ye
r
2D
M
a
x pool
i
ng s
ha
pe
l
a
y
e
r
D
e
ns
e
L
a
ye
r
(
R
e
L
U
)
11 O
ut
put
40 I
nput
C
onvol
ut
i
ona
l
l
a
ye
r
2D
M
a
x pool
i
ng s
ha
pe
l
a
y
e
r
D
e
ns
e
L
a
ye
r
(
R
e
L
U
)
11 O
ut
put
11 O
ut
put
40 I
nput
4 S
i
m
pl
e
R
N
N
L
a
ye
r
s
11 O
ut
put
16 I
nput
4 S
i
m
pl
e
R
N
N
L
a
ye
r
s
11 O
ut
put
40 I
nput
4 L
S
T
M
L
a
ye
r
s
11 O
ut
put
16 I
nput
4 L
S
T
M
L
a
ye
r
s
11 O
ut
put
40 I
nput
4 G
R
U
L
a
ye
r
s
11 O
ut
put
16 I
nput
4 G
R
U
L
a
ye
r
s
11 O
ut
put
A
c
t
i
va
t
i
on
f
unc
t
i
on
H
i
dde
n l
a
ye
r
s
:
R
e
l
u
O
ut
put
l
a
ye
r
s
:
S
of
t
m
a
x
O
ut
put
l
a
ye
r
s
:
‘
S
of
t
m
a
x’
O
ut
put
l
a
ye
r
s
:
‘
S
of
t
m
a
x’
O
ut
put
l
a
ye
r
s
:
‘
S
of
t
m
a
x’
O
pt
i
m
i
z
e
r
a
da
m
a
da
m
a
da
m
a
da
m
B
a
t
c
h s
i
z
e
32
32
32
32
5.
R
E
S
U
L
T
S
AND
D
I
S
C
U
S
S
I
O
N
5.1. Har
d
w
ar
e
c
h
ar
ac
t
e
r
is
t
ic
s
T
he
r
e
s
ul
ts
w
e
obt
a
in
e
d
a
r
e
pe
r
f
or
m
e
d
on
a
m
a
c
hi
ne
(
la
pt
op)
w
it
h
f
ol
lo
w
in
g
ha
r
dw
a
r
e
c
ha
r
a
c
te
r
is
ti
c
s
.
−
C
P
U
:
I
nt
e
l
i7
8t
h ge
ne
r
a
ti
on (
1 s
oc
ke
t,
6 c
or
e
s
, 12 thr
e
a
ds
)
−
R
A
M
:
8 G
B
−
G
P
U
:
N
V
id
ia
G
e
F
or
c
e
G
T
X
1050 with
c
uda
v10.1
I
n
our
e
xpe
r
im
e
nt
s
w
e
w
or
ke
d
w
it
h
K
e
r
a
s
(
2.2.4)
;
a
n
ope
n
-
s
our
c
e
pyt
hon
de
e
p
le
a
r
ni
ng
li
br
a
r
y
w
hi
c
h
is
r
unni
ng
on
to
p
of
G
oogl
e
’
s
ope
n
-
s
our
c
e
da
ta
f
lo
w
s
of
twa
r
e
;
T
e
n
s
or
F
lo
w
(
1.13.1)
a
s
a
ba
c
ke
nd
e
ngi
ne
.
5.2. E
val
u
at
in
g t
h
e
m
od
e
l
In
F
ig
ur
e
s
10
-
11
(
ge
ne
r
a
te
d
by
T
e
ns
or
boa
r
d)
a
nd
T
a
bl
e
2
,
w
e
pr
e
s
e
nt
th
e
a
c
c
ur
a
c
y
tr
a
in
in
g,
lo
s
s
tr
a
in
in
g,
a
c
c
ur
a
c
y
va
li
da
ti
on,
lo
s
s
va
li
da
ti
on
a
lo
ng
w
it
h
tr
a
in
in
g
t
im
e
of
a
ll
m
ode
ls
th
a
t
a
r
e
tr
a
in
e
d
ove
r
s
e
ve
r
a
l
e
poc
h
s
.
W
e
c
on
s
id
e
r
t
he
num
be
r
of
e
poc
hs
f
o
r
w
hi
c
h
our
m
ode
l
r
e
a
c
he
s
th
e
be
s
t
r
e
s
ul
ts
f
or
e
a
c
h
ve
r
s
io
n of
t
he
da
ta
s
e
t
(
f
ul
l
f
e
a
tu
r
e
s
a
nd be
s
t
f
e
a
tu
r
e
s
)
.
W
e
i
ni
ti
a
ll
y t
e
s
te
d e
a
c
h m
ode
l
w
it
h a
ba
tc
h
s
iz
e
of
1
28
,
w
hi
c
h a
ll
ow
e
d us
t
o ge
t
good ti
m
in
g f
o
r
a
ll
m
ode
ls
but
w
i
th
a
p
oor
pe
r
f
or
m
a
nc
e
, c
om
pa
r
e
d t
o a
s
m
a
ll
e
r
ba
tc
h
s
iz
e
l
ik
e
32 r
e
c
or
ds
t
ha
t
le
a
d
s
to
a
n i
m
pr
ove
d r
e
s
ul
t
but
w
it
h a
h
ig
he
r
c
om
put
a
ti
on t
im
e
.
A
s
s
how
n
in
F
ig
ur
e
10
a
nd
T
a
bl
e
2
, C
N
N
m
ode
l
s
in
bot
h
d
a
ta
s
e
t
ve
r
s
io
ns
“
F
ul
l
F
e
a
tu
r
e
s
”
a
nd
“
B
e
s
t
F
e
a
tu
r
e
s
”
r
e
a
c
he
s
r
e
s
p
e
c
ti
ve
ly
99,430
a
nd
99,935
a
s
a
c
c
ur
a
c
y
in
ju
s
t
1395s
a
nd
823s
(
15
a
nd
10
e
poc
hs
f
o
r
e
a
c
h)
.
C
om
pa
r
e
d
to
our
C
N
N
,
G
R
U
is
a
li
tt
le
m
or
e
a
c
c
ur
a
t
e
b
ut
it
to
ok
m
or
e
ti
m
e
to
be
tr
a
in
e
d
(
12412s
a
nd
5818s
in
40
a
nd
20
e
poc
hs
)
.
T
he
ti
m
e
of
our
C
N
N
is
a
lm
o
s
t
twi
c
e
c
om
pa
r
e
d
to
s
im
pl
e
R
N
N
f
or
th
e
“
F
ul
l
F
e
a
tu
r
e
s
”
ve
r
s
io
n
(
6089s
in
40
e
poc
h
s
)
,
a
nd
a
lm
os
t
e
qu
a
l
t
o
th
e
“
B
e
s
t
F
e
a
tu
r
e
s
”
ve
r
s
io
n
(
5708s
in
40
e
poc
hs
)
.
O
n
th
e
ot
he
r
ha
nd,
L
S
T
M
m
ode
ls
ha
ve
th
e
ir
be
s
t
a
c
c
ur
a
c
y
in
a
ti
m
in
g
be
twe
e
n
s
im
pl
e
R
N
N
a
nd
G
R
U
(
w
it
h
10627s
a
nd
8302s
in
30
a
nd
15
e
poc
hs
)
.
F
or
th
e
L
os
s
a
s
s
how
n
in
F
ig
ur
e
11
a
nd
T
a
bl
e
2,
C
N
N
m
ode
ls
a
r
e
a
lwa
ys
th
e
be
s
t
r
e
a
c
hi
ng
m
ode
l
s
in
bot
h
da
ta
s
e
t
ve
r
s
io
ns
,
“
F
ul
l
F
e
a
tu
r
e
s
”
a
nd
“
B
e
s
t
F
e
a
tu
r
e
s
”
,
w
it
h
r
e
s
pe
c
ti
ve
ly
0,582%
a
nd
1,663%
c
om
pa
r
e
d
to
ot
he
r
R
N
N
m
ode
ls
(
be
tw
e
e
n
1.7%
a
nd
2.9%
)
.
B
y
e
xa
m
in
in
g
th
e
s
e
r
e
s
ul
ts
,
it
is
c
le
a
r
th
a
t
th
e
C
N
N
m
ode
l
us
in
g
“
F
u
ll
F
e
a
tu
r
e
s
”
is
th
e
be
s
t
m
ode
l
a
c
c
or
di
ng
to
it
s
hi
ghe
r
a
c
c
ur
a
c
y
(
0.9993)
,
a
nd
th
e
be
s
t
ti
m
e
pe
r
f
or
m
a
nc
e
w
he
n
c
ons
id
e
r
in
g
th
e
w
hol
e
ti
m
e
f
or
obt
a
in
in
g
th
e
s
e
r
e
s
ul
ts
.
I
n
a
ddi
ti
on,
w
he
n
c
om
pa
r
in
g
th
e
r
e
s
ul
ti
ng
m
e
tr
ic
s
in
a
ll
m
ode
ls
w
it
h
th
e
two
-
da
ta
s
e
t
ve
r
s
io
ns
,
w
e
c
le
a
r
ly
s
e
e
th
a
t
th
e
“
F
ul
l
F
e
a
tu
r
e
s
”
ve
r
s
io
n
ge
ts
b
e
s
t
pe
r
f
or
m
a
nc
e
r
e
la
te
d
to
“
B
e
s
t
F
e
a
tu
r
e
s
”
ve
r
s
io
n.
T
hi
s
m
e
a
ns
th
a
t
d
e
e
p
le
a
r
ni
ng
a
lg
or
it
hm
s
c
a
n
a
c
hi
e
v
e
be
tt
e
r
c
l
a
s
s
if
ic
a
ti
on
a
nd
le
a
r
n
a
ddi
ti
ona
l
f
e
a
tu
r
e
s
w
it
h
a
ddi
ti
ona
l
tr
a
f
f
ic
da
ta
a
s
a
lr
e
a
dy
e
xpl
a
in
e
d
in
th
e
pr
e
vi
ou
s
s
e
c
ti
on,
but
in
c
om
put
a
ti
on
ti
m
e
a
nd
e
ne
r
gy
c
ons
um
pt
io
n, i
t
c
ons
um
e
s
m
or
e
be
c
a
us
e
m
or
e
da
ta
t
o t
r
a
in
a
r
e
ne
e
de
d.
In
F
ig
ur
e
12
,
w
e
c
om
pa
r
e
d
th
e
ti
m
e
th
a
t
a
pr
e
di
c
ti
on
w
oul
d
ta
k
e
f
or
a
ll
th
e
c
ons
id
e
r
e
d
m
ode
ls
(
how
m
uc
h
ti
m
e
th
e
I
D
S
w
oul
d
ta
ke
to
id
e
nt
if
y
a
n
a
tt
a
c
k)
w
hi
c
h
c
oul
d
be
a
not
he
r
im
por
ta
nt
m
e
tr
ic
f
o
r
th
e
I
D
S
im
pl
e
m
e
nt
a
ti
on
in
a
r
e
a
l
I
oT
e
nvi
r
onm
e
nt
.
W
e
th
e
n
c
onc
lu
de
d
th
a
t
C
N
N
ha
s
not
onl
y
th
e
be
s
t
a
c
c
ur
a
c
y
a
nd
th
e
lo
w
e
s
t
lo
s
s
,
but
a
ls
o
th
e
lo
w
e
s
t
ti
m
e
to
pr
e
di
c
t
a
n
a
tt
a
c
k
w
it
h
ju
s
t
a
f
r
a
gm
e
nt
of
m
il
li
s
e
c
onds
(
w
it
h
a
n
a
ve
r
a
ge
of
0.34m
s
)
.
T
he
ot
he
r
m
ode
l
s
m
a
ke
a
d
e
te
c
ti
on
of
a
n
a
tt
a
c
k
in
m
or
e
th
a
n
twi
c
e
or
tr
ip
le
th
a
t
ti
m
e
of
C
N
N
(
0.78ms
f
or
s
im
pl
e
R
N
N
, 1.03ms
f
or
G
R
U
, a
nd 1.08ms
f
or
L
S
T
M
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
10
, N
o.
1
,
M
a
r
c
h
20
2
1
:
110
–
120
118
T
he
a
m
ount
of
da
ta
in
th
e
da
ta
s
e
t
is
c
o
ns
id
e
r
a
bl
y
unba
la
nc
e
d
r
e
ga
r
di
ng
th
e
di
f
f
e
r
e
nt
ty
pe
s
of
a
tt
a
c
ks
.
F
or
e
xa
m
pl
e
,
D
oS
(
H
T
T
P
)
,
D
D
oS
(
H
T
T
P
)
,
ke
y
lo
ggi
ng,
a
nd
da
ta
e
xf
il
tr
a
ti
on
a
tt
a
c
ks
in
th
e
tr
a
in
in
g
a
nd
te
s
t
s
e
ts
ha
v
e
ve
r
y
s
m
a
ll
a
m
ount
of
da
ta
a
s
s
how
n
in
F
ig
ur
e
7
.
T
he
r
e
f
or
e
,
th
e
m
ode
l
ha
s
a
li
m
it
e
d
c
a
pa
bi
li
ty
to
le
a
r
n
a
c
c
ur
a
te
ly
th
e
s
e
ty
pe
s
of
a
tt
a
c
ks
. T
he
de
te
c
ti
on
a
ve
r
a
ge
of
th
e
s
e
a
tt
a
c
ks
i
s
one
of
th
e
m
a
in
f
a
c
to
r
s
r
e
s
tr
ic
ti
ng t
he
ove
r
a
ll
de
te
c
ti
on a
c
c
ur
a
c
y.
F
ig
ur
e
10
.
A
c
c
ur
a
c
y s
c
a
la
r
F
ig
ur
e
11
. L
os
s
s
c
a
la
r
T
a
bl
e
2
.
T
r
a
in
in
g a
nd t
e
s
ti
ng r
e
s
ul
ts
f
or
t
he
i
m
pl
e
m
e
nt
e
d m
ode
l
s
DL
A
l
gor
i
t
hm
C
onvol
ut
i
ona
l
N
e
ur
a
l
N
e
t
w
or
k
(
C
N
N
)
S
i
m
pl
e
R
N
N
(
R
e
c
ur
r
e
nt
N
e
ur
a
l
N
e
t
w
or
k)
L
ong S
hor
t
-
T
e
r
m
M
e
m
or
y
(
L
S
T
M
)
ba
s
e
d R
N
N
G
a
t
e
d R
e
c
ur
r
e
nt
U
ni
t
(
G
R
U
)
ba
s
e
d R
N
N
D
a
t
a
s
e
t
ve
r
s
i
on
F
ul
l
F
e
a
t
ur
e
s
B
e
s
t
F
e
a
t
ur
e
s
F
ul
l
F
e
a
t
ur
e
s
B
e
s
t
F
e
a
t
ur
e
s
F
ul
l
F
e
a
t
ur
e
s
B
e
s
t
F
e
a
t
ur
e
s
F
ul
l
F
e
a
t
ur
e
s
B
e
s
t
F
e
a
t
ur
e
s
E
poc
hs
15
10
40
40
30
25
40
20
L
a
ye
r
s
4
4
4
4
4
4
4
4
A
c
c
ur
a
c
y
t
r
a
i
ni
ng
0.99762
0.99086
0.98167
0.98038
0.98395
0.98285
0.98787
0.98301
L
os
s
t
r
a
i
ni
ng
0.01132
0.03294
0.07101
0.07115
0.06155
0.06526
0.04508
0.06661
A
c
c
ur
a
c
y
va
l
i
da
t
i
on
0.99935
0.99430
0.99406
0.99427
0.99401
0.99747
0.99403
0.99430
L
os
s
va
l
i
da
t
i
on
0.00582
0.01663
0.02021
0.01819
0.01848
0.02278
0.01762
0.02921
T
i
m
i
ng
T
ot
a
l
:
1395s
E
poc
h:
138s
S
t
e
p:
47μ
s
T
ot
a
l
:
823s
E
poc
h:
82s
S
t
e
p:
30μ
s
T
ot
a
l
:
6089s
E
poc
h:
153s
S
t
e
p:
52μ
s
T
ot
a
l
:
5708s
E
poc
h:
142s
S
t
e
p:
52μ
s
T
ot
a
l
:
10627s
E
poc
h:
354s
S
t
e
p:
121μ
s
T
ot
a
l
:
8302s
E
poc
h:
332s
S
t
e
p:
121μ
s
T
ot
a
l
:
12412s
E
poc
h:
310s
S
t
e
p:
106μ
s
T
ot
a
l
:
5818s
E
poc
h:
291s
S
t
e
p:
106μ
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
T
ow
ar
d a de
e
p l
e
a
r
ni
ng
-
bas
e
d i
nt
r
us
io
n d
e
te
c
ti
on s
y
s
te
m
f
or
I
o
T
agains
t
bot
ne
t
at
ta
c
k
s
(
I
dr
is
s
I
dr
i
s
s
i
)
119
F
ig
ur
e
12
.
P
r
e
di
c
ti
on t
im
in
g c
om
pa
r
is
on
6.
C
O
N
C
L
U
S
I
O
N
T
he
num
be
r
of
I
o
T
obj
e
c
t
s
di
s
pa
t
c
he
d
a
r
ound
th
e
w
or
ld
is
gr
ow
in
g
pr
ogr
e
s
s
iv
e
ly
,
w
hi
c
h
m
a
ke
s
it
s
s
e
c
ur
it
y
a
bi
g
c
ha
ll
e
nge
.
O
ur
w
or
k
he
r
e
f
oc
us
e
d
on
in
tr
us
io
n
d
e
te
c
ti
on
s
ys
te
m
s
f
or
I
oT
us
in
g
f
our
va
r
ia
nt
s
of
de
e
p
le
a
r
ni
ng
m
ode
ls
,
a
nd
c
om
pa
r
e
d
th
e
m
e
f
f
ic
ie
nt
ly
to
de
te
c
t
va
r
io
us
ty
pe
s
of
I
oT
n
e
twor
k
a
tt
a
c
ks
,
u
s
ua
ll
y
done
by
bot
ne
ts
.
A
f
te
r
s
e
v
e
r
a
l
e
xpe
r
im
e
nt
s
,
w
e
obt
a
in
e
d
a
r
e
a
s
ona
bl
e
de
te
c
ti
on
r
a
te
on
our
a
ll
f
our
m
ode
ls
.
B
y
a
na
ly
z
in
g
th
e
obt
a
in
e
d
r
e
s
ul
ts
,
w
e
c
on
c
lu
de
d
th
a
t
C
N
N
is
t
he
be
s
t
one
f
or
in
tr
us
io
n
de
te
c
ti
on
s
ys
te
m
s
,
it
w
a
s
a
bl
e
to
id
e
nt
if
y
s
uc
c
e
s
s
f
ul
ly
di
f
f
e
r
e
nt
ty
pe
s
of
a
tt
a
c
k
s
a
nd
s
how
e
d
th
e
hi
ghe
r
a
c
c
ur
a
c
y
(
w
it
h
99.94%
)
in
c
om
pa
r
is
on
w
it
h
ot
he
r
D
L
a
lg
o
r
it
hm
s
,
s
uc
h
a
s
S
im
pl
e
R
N
N
,
L
S
T
M
,
a
nd
G
R
U
,
a
nd
it
a
ll
ow
e
d
a
de
te
c
ti
on
w
it
h
lo
w
e
r
lo
s
s
r
a
te
s
(
0.58%
)
a
nd
a
be
tt
e
r
pe
r
f
o
r
m
a
nc
e
in
te
r
m
s
of
pr
e
di
c
ti
on
t
im
e
.
A
s
f
ut
u
r
e
w
or
ks
,
w
e
w
il
l
a
ppl
y
th
e
C
N
N
m
ode
l
on
a
r
e
a
l
ne
twor
k
tr
a
f
f
ic
da
t
a
,
a
nd
r
e
in
f
or
c
in
g
it
us
in
g
de
e
p
tr
a
ns
f
e
r
le
a
r
ni
ng
by
a
ddi
ng
ot
he
r
c
ha
r
a
c
te
r
is
ti
c
s
f
r
om
ot
he
r
da
ta
s
e
ts
or
f
r
om
di
ve
r
s
e
f
ir
e
w
a
ll
s
,
lo
gs
,
a
nd
I
D
S
s
e
r
ve
r
s
.
I
n
a
ddi
ti
on,
w
e
w
il
l
tr
y
to
us
e
s
e
lf
-
s
upe
r
vi
s
e
d
le
a
r
ni
ng
to
ge
n
e
r
a
te
a
pow
e
r
f
ul
a
nd
upda
te
d
m
ode
l,
a
nd
th
e
n
im
pl
e
m
e
nt
a
n
a
ut
onomous
i
nt
r
us
io
n de
te
c
ti
on s
ys
t
e
m
.
R
E
F
E
R
E
N
C
E
S
[1]
“Internet
of
Things
(IoT)
-
The
future
of
IoT
miniguide:
The
burgeoning
IoT
market
continues
-
Cisco.”
[Online]
.
Available: http
s://www.cisco.com
/c/en/us/sol
utions/int
ernet
-
of
-
things/futur
e
-
of
-
iot.html. [Acc
essed: 06
-
Jun
-
2020].
[2]
J.
Singh,
T.
Pasquier,
J.
Bacon,
H.
Ko,
and
D.
Eyers,
“Twenty
S
ecurity
Considerations
for
Cloud
-
Supporte
d
Internet
of
Things,”
IEEE
Internet
Things
J.
,
vol.
3,
n
o.
3,
pp.
269
–
284,
Jun.
2016.
DOI:
10.1109/JIOT.2015.2460333
.
[3]
J.
Lin,
W.
Yu,
N.
Zhang,
X.
Yang,
H.
Zhang,
and
W.
Zhao.,
“A
Survey
on
Internet
of
Things:
Architecture,
Enabling
Technologies,
Security
and
Privacy,
and
Applications,”
IEE
E
Internet
Things
J.
,
vol.
4,
no.
5,
pp.
1125
–
1142, 2017.
DOI: 10.1109/JIOT.
2017.2683200
.
[4]
Y.
Xiao,
C.
Xing,
T.
Zhang,
and
Z.
Zhao,
“An
Intrusion
Detecti
on
Model
Based
on
Feature
Reduction
and
Convolutiona
l
Neura
l
Network
s,”
IEEE
Access
,
vol.
7,
pp.
42210
–
42219,
2019.
DOI:
10.1109/ACCESS.2019.2904620
.
[5]
E.
Bertino
,
N.
Islam,
“Botnets
and
Internet
of
Things
Security,”
Comp
uter
(Long.
Beach.
Calif).
,
vol.
50,
no.
2,
pp.
76
–
79, 2017.
DOI: 10.1109/MC
.2017.62
.
[6]
C.
Kolias,
G.
Kambourakis,
A.
Stavrou,
and
J.
Voas.,
“DDoS
in
t
he
IoT:
Mirai
and
other
botnets,”
Computer
(Long.
Beach. Calif).
, vol. 50, no. 7, pp. 80
–
84, 2017.
DOI: 10.1109/MC
.2017.201
.
[7]
A.
J.
Malik,
W.
Shahzad,
and
F.
A.
Khan.,
“Network
intrusion
det
ection
using
hybrid
binary
PSO
and
random
forests algo
rithm,”
Secur. Commun. Networks
, vol. 8, no. 16, pp. 2646
–
2660
, 2015.
https://doi.org/10.1002/sec.508
.
[8]
J.
Jabez
,
B.
Muthukumar.,
“Intrusion
detection
system
(ids):
Anomal
y
detection
using
outlier
detection
approach,”
in
Procedia Computer Science
, 2015, vol. 48, no. C, pp. 338
–
346.
DOI
:
10.1016/j.procs.2015.04.191
.
[9]
N.
Ouerdi,
I.
Elfarissi,
A.
Azizi,
and
M.
Azizi.,
“Artificial
neural
ne
twork
-
based
methodology
for
vulnerabilities
detection
in
EMV
cards,”
in
Proceedings
of
the
2015
11th
Internation
al
Conference
on
Information
Assurance
and
S
ecurity,
IAS 2015
, pp. 85
–
90
, 2016
.
DOI: 10.1109/IS
IAS.2015.749275
0.
[10]
S.
Vieira,
W.
H.
L.
Pinaya,
and
A.
Mechelli.,
“Using
deep
learning
t
o
investigate
the
neuroimaging
correlates
of
psychiatric
and
neurological
disorders:
Methods
and
applications,”
N
eurosc
ience
and
Biobehavioral
Reviews
,
vol.
74. Elsevier Ltd, pp. 58
–
75, 2017.
DOI: 10.1016/j.
neubiorev.2017.01.0
02.
[11]
Y.
Lecun,
Y.
Bengio,
and
G.
Hinton.,
“Deep
learning,”
Nature
,
vol.
521,
no.
7553.
Nature
Publishing
Group,
pp.
436
–
444, 2015
.
https://doi.org/10.
1038/nature14539
.
Evaluation Warning : The document was created with Spire.PDF for Python.