I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
,
p
p
.
2618
~
2
6
3
2
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
3
.
p
p
2
6
1
8
-
2
6
3
2
2618
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Driv
eShield
:
a
tt
e
ntion
-
ba
sed hy
bri
d neural n
etwork
for
intrusio
n de
tec
tion in au
tom
o
tive
c
o
ntroller a
rea ne
t
wo
rks
Vis
m
a
y
a
K
o
o
t
a
y
i K
un
na
cher
i
1
,
Arul
L
ee
na
Ro
s
e
P
et
er
J
o
s
eph
2
1
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
F
a
c
u
l
t
y
o
f
S
c
i
e
n
c
e
a
n
d
H
u
ma
n
i
t
i
e
s
,
S
R
M
I
n
st
i
t
u
t
e
o
f
S
c
i
e
n
c
e
a
n
d
Te
c
h
n
o
l
o
g
y
,
C
h
e
n
n
a
i
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
A
p
p
l
i
c
a
t
i
o
n
s,
Fa
c
u
l
t
y
o
f
S
c
i
e
n
c
e
a
n
d
H
u
ma
n
i
t
i
e
s
,
S
R
M
I
n
st
i
t
u
t
e
o
f
S
c
i
e
n
c
e
a
n
d
Te
c
h
n
o
l
o
g
y
,
C
h
e
n
n
a
i
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Dec
1
7
,
2
0
2
4
R
ev
is
ed
Feb
1
2
,
2
0
2
6
Acc
ep
ted
Ap
r
2
0
,
2
0
2
6
Ve
h
icle
n
e
two
rk
se
c
u
rit
y
is
imp
o
rtan
t
a
s
i
n
c
re
a
sin
g
a
m
o
u
n
ts
o
f
c
o
n
n
e
c
ted
tec
h
n
o
l
o
g
y
a
re
b
e
in
g
a
d
d
e
d
to
v
e
h
icle
s
n
o
wa
d
a
y
s,
p
u
tt
i
n
g
t
h
e
m
a
t
risk
o
f
c
y
b
e
ra
tt
a
c
k
s.
T
h
is
p
a
p
e
r
p
re
se
n
t
s
Driv
e
S
h
ield
,
a
n
o
v
e
l
re
a
l
-
ti
m
e
in
tr
u
sio
n
d
e
tec
ti
o
n
s
y
ste
m
(IDS)
t
h
a
t
is
t
h
e
first
t
o
c
o
m
b
in
e
g
a
ted
re
c
u
r
re
n
t
u
n
it
s
(G
RU),
c
o
n
v
o
lu
t
io
n
a
l
n
e
u
ra
l
n
e
two
rk
s
(CNN
)
,
a
n
d
lo
n
g
sh
o
rt
-
ter
m
m
e
m
o
ry
(LS
TM
)
with
a
n
a
tt
e
n
ti
o
n
m
e
c
h
a
n
ism
.
Th
e
sy
ste
m
a
ti
c
p
re
-
p
ro
c
e
ss
in
g
p
ip
e
li
n
e
,
wh
ich
i
n
c
lu
d
e
s
fe
a
tu
re
e
n
g
in
e
e
rin
g
,
t
h
e
sy
n
t
h
e
ti
c
m
in
o
rit
y
o
v
e
rsa
m
p
li
n
g
tec
h
n
i
q
u
e
(S
M
OT
E)
fo
r
c
las
s
b
a
lan
c
in
g
,
a
n
d
n
o
r
m
a
li
z
a
ti
o
n
.
Th
e
m
o
d
e
l
wa
s
v
a
li
d
a
ted
o
n
th
e
o
p
e
n
train
in
g
i
n
tru
si
o
n
d
e
tec
ti
o
n
s
y
ste
m
(OTIDS)
d
a
tas
e
t
a
n
d
th
e
Ha
c
k
in
g
a
n
d
C
o
u
n
term
e
a
su
re
Re
s
e
a
rc
h
Lab
(HCRL)
c
a
r
h
a
c
k
in
g
d
a
tas
e
t.
I
n
th
e
HCRL
d
a
tas
e
t,
th
e
m
o
d
e
l
h
a
d
a
n
a
c
c
u
ra
c
y
o
f
9
6
.
3
0
%
with
F
1
-
sc
o
r
e
s
a
s
h
ig
h
a
s
9
6
%
f
o
r
a
ll
k
in
d
s
o
f
a
tt
a
c
k
s.
On
th
e
OTIDS
d
a
tas
e
t,
it
p
e
rfo
r
m
e
d
v
e
ry
we
ll
i
n
term
s
o
f
g
e
n
e
ra
li
z
a
ti
o
n
,
with
a
h
ig
h
e
st
a
c
c
u
ra
c
y
o
f
9
9
.
7
8
%
a
n
d
a
we
ig
h
te
d
F
1
-
sc
o
re
o
f
9
9
.
7
8
%
.
T
h
e
a
d
d
it
i
o
n
o
f
a
n
a
tt
e
n
ti
o
n
m
e
c
h
a
n
is
m
e
n
a
b
led
th
e
m
o
d
e
l
t
o
c
o
n
c
e
n
tr
a
te
o
n
t
h
e
m
o
st
sig
n
ifi
c
a
n
t
fe
a
tu
re
s,
p
ro
v
id
i
n
g
b
e
tt
e
r
a
d
a
p
ta
b
il
it
y
t
o
c
h
a
n
g
i
n
g
th
re
a
ts.
Th
e
se
fin
d
in
g
s
d
e
m
o
n
stra
te
t
h
e
e
ffica
c
y
,
sc
a
lab
il
it
y
,
a
n
d
re
li
a
b
i
l
it
y
o
f
th
e
sy
ste
m
fo
r
in
-
v
e
h
icle
n
e
two
r
k
s
e
c
u
rit
y
.
T
h
e
fu
t
u
re
re
se
a
rc
h
will
fo
c
u
s
o
n
p
e
rfo
rm
a
n
c
e
o
n
lo
we
r
-
fre
q
u
e
n
c
y
a
tt
a
c
k
s
th
r
o
u
g
h
th
e
st
u
d
y
o
f
u
n
s
u
p
e
rv
ise
d
lea
rn
in
g
m
e
th
o
d
s a
n
d
re
a
l
-
wo
rld
d
e
p
lo
y
m
e
n
t
tri
a
ls.
K
ey
w
o
r
d
s
:
C
o
n
tr
o
ller
ar
ea
n
etwo
r
k
Dee
p
lear
n
in
g
E
lectr
o
n
ic
co
n
tr
o
l u
n
it
I
n
tr
u
s
io
n
d
etec
tio
n
s
y
s
tem
In
-
v
eh
icle
n
etwo
r
k
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Ar
u
l L
ee
n
a
R
o
s
e
Peter
J
o
s
ep
h
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
A
p
p
licatio
n
s
,
Facu
lty
o
f
Scien
ce
a
n
d
Hu
m
an
ities
SR
M
I
n
s
titu
te
o
f
Scien
ce
an
d
T
ec
h
n
o
lo
g
y
Kattan
k
u
lath
u
r
,
C
h
e
n
n
ai,
I
n
d
i
a
E
m
ail: le
en
a.
r
o
s
e5
2
7
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
s
ig
n
if
ican
ce
o
f
c
y
b
er
s
ec
u
r
ity
in
th
ese
in
n
o
v
ativ
e
tec
h
n
o
lo
g
ies
h
as
g
r
o
wn
as
th
e
au
to
m
o
tiv
e
in
d
u
s
tr
y
p
r
o
g
r
ess
es
with
co
n
n
ec
ted
an
d
d
r
i
v
er
less
v
eh
icl
es.
T
h
e
in
cr
ea
s
in
g
in
teg
r
atio
n
o
f
co
n
tem
p
o
r
ar
y
au
to
m
o
tiv
e
tech
n
o
lo
g
ies,
s
u
ch
as
elec
tr
o
n
ic
co
n
tr
o
l
u
n
it
s
(
E
C
U)
an
d
co
n
tr
o
ller
ar
ea
n
etw
o
r
k
(
C
AN
)
b
u
s
,
h
as
m
ad
e
au
t
o
m
o
b
iles
v
u
ln
er
a
b
le
to
a
n
u
m
b
er
o
f
cy
b
e
r
s
ec
u
r
ity
is
s
u
es
[
1
]
.
I
n
-
v
eh
i
cle
n
etwo
r
k
s
a
r
e
co
m
p
u
tatio
n
ally
s
u
p
p
o
r
ted
b
y
th
e
E
C
U,
wh
ich
r
eg
u
lates
a
n
u
m
b
er
o
f
s
u
b
s
y
s
tem
s
,
in
clu
d
in
g
en
ter
tain
m
e
n
t,
en
g
in
e
m
an
a
g
em
en
t,
an
d
b
r
a
k
in
g
;
b
ec
au
s
e
o
f
its
elec
tr
o
n
ic
co
m
p
o
n
en
ts
an
d
n
etwo
r
k
ed
s
y
s
tem
s
,
E
C
U
s
h
av
e
b
ec
o
m
e
a
p
r
o
m
in
en
t
tar
g
et
f
o
r
attac
k
er
s
[
2
]
.
B
ein
g
d
e
v
o
i
d
o
f
en
cr
y
p
tio
n
a
n
d
au
th
e
n
ticatio
n
s
u
p
p
o
r
t,
th
e
co
n
v
en
tio
n
al
C
AN
b
u
s
p
r
o
to
co
ls
ar
e
v
u
ln
er
ab
le
to
m
alicio
u
s
in
ter
f
e
r
en
ce
a
n
d
u
n
a
u
th
o
r
ized
ac
ce
s
s
;
with
g
r
ea
ter
c
o
n
n
ec
tiv
ity
,
th
er
e
is
g
r
ea
ter
ex
p
o
s
u
r
e
to
c
y
b
er
attac
k
s
,
wh
ich
in
tu
r
n
r
esu
lt
i
n
r
is
k
o
f
in
ju
r
y
to
d
r
iv
er
s
an
d
o
cc
u
p
a
n
ts
o
r
co
m
p
r
o
m
is
e
p
r
iv
ac
y
o
r
in
teg
r
ity
[
3
]
.
T
h
e
C
AN
b
u
s
i
s
a
m
es
s
ag
e
b
r
o
ad
ca
s
tin
g
b
u
s
th
at
ca
n
tr
an
s
m
it
u
p
to
6
4
b
its
o
f
d
ata
i
n
f
r
a
m
es
,
as
s
h
o
wn
in
Fig
u
r
e
1
,
an
d
C
AN
id
en
tifie
s
th
e
p
h
y
s
ical
lay
er
an
d
d
ata
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Dri
ve
S
h
ield
:
a
tten
tio
n
-
b
a
s
ed
h
yb
r
id
n
eu
r
a
l n
etw
o
r
k
fo
r
in
tr
u
s
io
n
…
(
V
is
ma
ya
K
o
o
ta
yi
K
u
n
n
a
ch
eri)
2619
lin
k
o
f
o
p
en
s
y
s
tem
in
ter
co
n
n
ec
tio
n
(
OSI
)
an
d
p
r
o
v
id
es
a
s
im
p
le
n
etwo
r
k
s
o
lu
tio
n
f
o
r
q
u
ick
in
-
v
e
h
icle
co
m
m
u
n
icatio
n
[
4
]
.
I
ts
g
o
al
was
to
r
ed
u
ce
wir
in
g
co
s
ts
b
y
allo
win
g
in
d
ep
en
d
e
n
t
elec
tr
o
n
ic
co
n
tr
o
l
m
o
d
u
les
(
E
C
Ms
)
to
co
m
m
u
n
icate
u
s
in
g
s
in
g
le
wir
e
p
air
s
,
im
p
r
o
v
in
g
d
r
iv
in
g
c
o
m
f
o
r
t
an
d
s
af
ety
v
i
a
s
m
o
o
th
in
p
u
t
an
d
o
u
tp
u
t
co
o
r
d
in
atio
n
[
5
]
;
h
o
we
v
er
,
s
ec
u
r
ity
-
wis
e
it
lack
s
in
te
g
r
ated
au
th
en
ticatio
n
an
d
en
cr
y
p
tio
n
,
em
p
l
o
y
s
a
b
r
o
ad
ca
s
t
tech
n
iq
u
e
th
at
m
ak
es
m
ess
ag
es
ac
ce
s
s
ib
le
to
all
n
o
d
es,
an
d
d
u
e
to
its
lack
o
f
an
ac
ce
s
s
co
n
tr
o
l
m
ec
h
an
is
m
C
AN
is
s
u
s
ce
p
tib
l
e
to
s
p
o
o
f
in
g
an
d
d
en
ial
-
of
-
s
er
v
ice
(
Do
S)
attac
k
s
[
6
]
.
A
ty
p
i
ca
l CAN f
r
am
e
h
as
an
id
en
tific
atio
n
o
f
1
1
b
its
[
7
]
,
[
8
]
(
an
d
an
ex
p
an
d
e
d
C
AN
f
r
am
e
h
as
a
2
9
-
b
it
id
en
tifie
r
)
,
an
d
b
ec
a
u
s
e
an
y
n
o
d
es
lin
k
ed
to
th
e
C
AN
b
u
s
will
r
ec
eiv
e
m
ess
ag
es
i
s
s
u
ed
b
y
an
y
o
th
er
n
o
d
e,
it
is
s
im
p
le
f
o
r
an
attac
k
er
to
jo
in
n
etwo
r
k
tr
af
f
ic
an
d
r
ea
d
d
ata
f
r
am
es;
f
u
r
th
er
m
o
r
e,
m
ess
ag
es
tr
an
s
f
er
r
ed
b
etwe
e
n
E
C
Us
in
a
C
AN
n
etwo
r
k
lack
au
th
e
n
ticato
r
in
f
o
r
m
atio
n
,
s
o
an
attac
k
er
ca
n
s
to
p
a
f
r
am
e
f
r
o
m
g
ettin
g
to
th
e
p
er
s
o
n
it'
s
m
ea
n
t
f
o
r
[
9
]
,
[
1
0
]
.
Fig
u
r
e
2
illu
s
tr
ates
a
ty
p
ical
in
-
v
eh
icle
C
AN
n
etwo
r
k
an
d
th
e
c
o
m
m
o
n
attac
k
s
u
r
f
ac
es
co
n
s
id
er
ed
in
th
is
s
tu
d
y
.
Fig
u
r
e
1
.
Stan
d
a
r
d
an
d
ex
te
n
d
ed
C
AN
f
r
am
e
f
o
r
m
ats
Fig
u
r
e
2
.
T
y
p
ical
in
-
v
eh
icle
C
AN
ar
ch
itectu
r
e
with
k
ey
E
C
Us an
d
m
ajo
r
attac
k
v
ec
to
r
s
,
i
n
clu
d
in
g
OB
D
-
II
co
n
tact
attac
k
s
an
d
r
e
m
o
te
n
o
n
-
co
n
tact
attac
k
s
R
ec
en
t
s
tu
d
ies
h
av
e
ev
alu
ated
m
ac
h
in
e
lear
n
in
g
an
d
d
ee
p
le
ar
n
in
g
(
DL
)
m
eth
o
d
s
f
o
r
C
AN
in
tr
u
s
io
n
d
etec
tio
n
,
s
h
o
win
g
p
r
o
m
is
in
g
co
m
p
ar
ativ
e
r
esu
lts
f
o
r
d
ata
-
d
r
iv
en
ap
p
r
o
ac
h
es
[
1
1
]
,
as
well
as
d
em
o
n
s
tr
atin
g
th
at
h
y
b
r
id
d
ee
p
ar
c
h
itectu
r
es
ca
n
s
ig
n
if
ican
tly
im
p
r
o
v
e
d
etec
tio
n
ac
c
u
r
ac
y
i
n
r
ea
l
-
wo
r
ld
in
-
v
eh
icle
en
v
ir
o
n
m
en
ts
[
1
2
]
.
W
ith
p
r
o
m
is
in
g
d
etec
tio
n
b
y
DL
m
o
d
e
ls
,
th
e
ab
ilit
y
to
id
en
tify
em
er
g
in
g
th
r
ea
ts
,
an
d
th
e
em
er
g
en
ce
o
f
DL
,
th
e
in
-
v
eh
i
cle
in
tr
u
s
io
n
d
etec
tio
n
s
y
s
tem
(
I
DS
)
p
ar
ad
ig
m
tak
es
o
n
a
n
ew
d
im
en
s
io
n
[
1
3
]
.
Fed
er
ated
an
d
m
eta
-
lear
n
in
g
ap
p
r
o
ac
h
es
h
av
e
also
b
ee
n
e
x
p
lo
r
ed
to
en
h
an
ce
p
er
s
o
n
ali
za
tio
n
an
d
ze
r
o
-
d
ay
ad
ap
tab
ilit
y
in
I
o
T
an
d
au
to
m
o
tiv
e
I
DSs
[
1
4
]
,
r
ein
f
o
r
cin
g
th
e
n
ee
d
f
o
r
I
DS
d
esig
n
s
t
h
at
r
em
ain
ef
f
ec
tiv
e
u
n
d
er
e
v
o
lv
in
g
attac
k
s
tr
ateg
i
es a
n
d
o
p
er
atio
n
al
co
n
s
tr
ain
ts
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
:
2
6
1
8
-
2
6
3
2
2620
I
n
s
p
ite
o
f
g
r
o
win
g
in
ter
est
in
DL
-
b
ased
I
DS
s
o
lu
tio
n
s
f
o
r
C
AN
b
u
s
n
etwo
r
k
s
,
m
o
s
t
wo
r
k
is
b
ased
o
n
s
in
g
le
-
m
o
d
el
ar
c
h
itectu
r
e
s
,
wh
ich
p
r
ev
en
ts
th
em
f
r
o
m
s
im
u
ltan
eo
u
s
ly
r
ec
o
g
n
izin
g
b
o
th
s
p
atial
an
d
tem
p
o
r
al
p
atter
n
s
.
Fo
r
ex
a
m
p
le,
So
n
g
et
a
l
.
[
1
5
]
d
e
v
el
o
p
ed
an
e
n
tire
ly
c
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN
)
-
b
ased
I
DS
o
p
tim
ized
f
o
r
s
p
atial
f
ea
tu
r
e
ex
tr
ac
tio
n
an
d
Ho
s
s
ain
et
a
l
.
[
1
6
]
in
tr
o
d
u
ce
d
a
n
l
ong
s
h
o
r
t
-
ter
m
m
e
m
o
r
y
(
L
STM
)
-
o
n
ly
m
o
d
el
th
at
was
s
o
lely
c
o
n
ce
r
n
e
d
with
s
eq
u
e
n
tial
d
ep
en
d
en
cies
in
C
AN
m
ess
ag
es.
L
ik
ewise,
a
lig
h
tw
eig
h
t
g
ate
d
r
ec
u
r
r
e
n
t
u
n
it
(
GR
U)
-
b
ased
s
y
s
tem
was
p
r
o
p
o
s
e
d
to
r
ed
u
ce
laten
c
y
in
r
ea
l
-
tim
e
d
etec
tio
n
[
1
7
]
,
y
e
t
lack
s
s
p
atial
co
n
tex
t
in
teg
r
atio
n
.
Su
ch
s
in
g
le
-
ar
ch
itectu
r
e
m
eth
o
d
s
ten
d
to
f
ail
to
g
en
e
r
alize
o
v
e
r
d
if
f
er
en
t
ty
p
es
o
f
attac
k
s
an
d
d
atasets
,
an
d
ca
n
f
ail
to
ca
p
tu
r
e
c
r
u
cial
cr
o
s
s
-
f
ea
tu
r
e
d
ep
en
d
e
n
cies
in
C
AN
tr
af
f
ic.
T
ab
le
1
p
r
esen
ts
a
co
m
p
ar
ativ
e
s
u
m
m
ar
y
o
f
th
e
ex
is
tin
g
lite
r
atu
r
e,
h
ig
h
lig
h
tin
g
k
ey
asp
ec
ts
s
u
ch
as
m
o
d
el
ar
ch
itectu
r
e,
d
atasets
,
d
etec
ti
o
n
p
e
r
f
o
r
m
an
ce
,
f
alse
p
o
s
itiv
e
r
ates,
an
d
class
im
b
alan
ce
h
an
d
lin
g
.
A
v
is
u
a
l
tax
o
n
o
m
y
s
u
m
m
ar
izin
g
th
e
ev
o
lu
tio
n
o
f
I
DS
m
o
d
els
f
o
r
in
-
v
eh
icle
C
AN
n
etwo
r
k
s
is
p
r
o
v
id
e
d
in
Fig
u
r
e
3
.
R
ep
r
esen
tativ
e
ex
am
p
les
in
clu
d
e
a
VGG
-
16
-
b
ased
C
NN
m
o
d
el
b
y
Ah
m
ed
et
a
l.
[
1
8
]
(
h
i
g
h
ac
c
u
r
ac
y
f
o
r
Do
S
an
d
f
u
zz
y
with
0
.
6
%
f
alse
p
o
s
itiv
e
r
ate
(
FPR
)
,
b
u
t
m
is
s
in
g
tem
p
o
r
al
m
o
d
elin
g
an
d
r
ea
l
-
wo
r
ld
ap
p
licatio
n
)
,
a
C
NN
–
L
STM
m
o
d
el
b
y
Alf
er
aid
i
et
a
l.
[
1
9
]
(
u
p
to
9
9
.
7
%
ac
cu
r
ac
y
b
u
t
with
o
u
t
class
i
m
b
alan
ce
h
an
d
lin
g
o
r
tem
p
o
r
al
atten
tio
n
)
,
an
d
an
L
ST
M
-
GR
U
m
o
d
el
b
y
Alk
h
atib
et
a
l.
[
2
0
]
(
h
i
g
h
AU
C
o
n
s
y
n
th
etic
tr
af
f
ic
b
u
t
with
o
u
t
r
ea
l
-
tim
e
v
er
if
icatio
n
an
d
r
ea
l
C
AN
tr
af
f
ic
g
en
er
aliza
b
ilit
y
)
.
R
elate
d
ef
f
o
r
ts
in
clu
d
e
I
DS
-
I
VN
b
y
Alq
ah
tan
i
an
d
Ku
m
ar
[
2
1
]
(
au
to
en
co
d
er
-
b
ased
C
NN
–
L
STM
with
9
9
%
ac
c
u
r
ac
y
a
n
d
0
.
3
2
%
FP
R
,
b
u
t
with
o
u
t
d
ep
lo
y
m
en
t
f
ea
s
ib
ilit
y
an
d
s
ca
lab
ilit
y
in
s
ig
h
ts
)
,
an
ML
P
-
b
ased
I
DS
b
y
An
ze
r
an
d
E
lh
a
d
ef
[
2
2
]
(
g
o
o
d
class
if
icatio
n
b
u
t
s
h
allo
w
tem
p
o
r
al
p
r
o
ce
s
s
in
g
)
,
a
n
d
L
STM
a
u
to
en
co
d
er
s
b
y
Ash
r
af
et
a
l.
[
2
3
]
(
m
o
r
e
th
a
n
9
8
%
ac
cu
r
ac
y
,
b
u
t
less
r
ea
l
-
tim
e
r
elev
an
ce
)
.
I
n
ad
d
itio
n
,
L
i
et
a
l
.
[
2
4
]
p
r
o
p
o
s
ed
a
n
a
d
ap
tiv
e
d
ee
p
lear
n
in
g
(
ADL
)
m
o
d
el
f
o
r
s
m
ar
t
-
g
r
id
tr
a
f
f
ic
(
n
etwo
r
k
s
ec
u
r
ity
lab
o
r
ato
r
y
-
k
n
o
wled
g
e
d
is
co
v
er
y
in
d
ata
b
ases
(
NSL
-
KDD
)
)
with
o
u
t
d
ir
ec
t
C
AN
u
s
e
o
r
tem
p
o
r
al
f
ea
tu
r
e
m
o
d
elin
g
,
El
-
Gay
ar
et
a
l
.
[
2
5
]
p
r
o
p
o
s
ed
D
FS
E
Net
(
9
8
–
9
9
%
ac
cu
r
ac
y
v
i
a
p
ar
allel
en
s
em
b
le
lear
n
er
s
b
u
t
with
o
u
t
ex
p
licit
tem
p
o
r
al
s
eq
u
en
ce
lear
n
in
g
)
,
an
d
Kan
g
an
d
Kan
g
[
2
6
]
i
n
tr
o
d
u
ce
d
an
ea
r
ly
DNN
-
b
ased
I
DS
with
d
ee
p
b
elief
n
etwo
r
k
(
DB
N
)
in
itializatio
n
f
o
r
C
AN
p
ac
k
et
class
if
icatio
n
,
b
u
t
with
o
u
t
cu
r
r
en
t
h
y
b
r
id
f
ea
tu
r
e
ex
tr
ac
ti
o
n
m
eth
o
d
s
an
d
th
o
r
o
u
g
h
ev
a
lu
atio
n
m
etr
ics.
A
C
AN
f
r
am
e
co
n
tain
s
m
u
ltip
le
f
ield
s
,
in
clu
d
in
g
t
h
e
id
en
tifie
r
(
1
1
-
b
it
s
tan
d
ar
d
o
r
2
9
-
b
it
ex
ten
d
ed
)
,
c
o
n
tr
o
l
f
ield
(
6
b
its
,
in
clu
d
in
g
th
e
4
-
b
it
d
ata
len
g
th
co
d
e
)
,
d
ata
f
iel
d
(0
–
8
b
y
tes),
cy
clic
r
ed
u
n
d
a
n
c
y
ch
ec
k
(
CRC
)
f
ield
(
1
5
b
its
)
,
ac
k
n
o
wled
g
m
en
t
(
AC
K
)
s
lo
t,
an
d
en
d
-
of
-
f
r
a
m
e
b
its
.
T
h
e
b
r
o
a
d
ca
s
t
n
atu
r
e
o
f
C
AN
m
ea
n
s
th
at
all
n
o
d
es
r
ec
eiv
e
ev
e
r
y
tr
an
s
m
itted
f
r
am
e,
a
n
d
th
e
p
r
o
to
co
l
d
o
es
n
o
t
p
r
o
v
id
e
b
u
ilt
-
in
au
th
en
ticatio
n
,
wh
ich
m
ak
es
in
jectio
n
an
d
s
p
o
o
f
in
g
attac
k
s
f
ea
s
ib
le.
Fig
u
r
e
3
.
T
a
x
o
n
o
m
y
o
f
I
DS m
o
d
els f
o
r
a
u
to
m
o
tiv
e
C
AN
n
et
wo
r
k
s
I
n
o
r
d
er
t
o
o
v
er
c
o
m
e
t
h
e
s
e
d
r
a
w
b
a
c
k
s
,
t
h
i
s
p
ap
e
r
in
t
r
o
d
u
c
es
D
r
i
v
e
S
h
i
e
ld
,
a
DL
-
b
a
s
ed
h
y
b
r
i
d
I
D
S
t
h
a
t
co
m
b
i
n
e
s
C
N
N
,
L
S
T
M
,
a
n
d
G
R
U
l
a
y
er
s
w
i
t
h
an
ad
d
e
d
a
t
t
en
t
i
o
n
m
e
ch
a
n
i
s
m
t
o
d
y
n
am
i
c
a
l
l
y
w
e
i
g
h
t
i
m
p
o
r
t
a
n
t
f
ea
t
u
r
e
s
i
n
s
p
a
t
ia
l
–
t
e
m
p
o
r
a
l
C
A
N
d
a
t
a
.
T
h
i
s
e
n
a
b
le
s
t
h
e
s
y
s
t
e
m
t
o
p
r
i
o
r
i
t
i
z
e
t
h
e
m
o
s
t
d
i
s
c
r
i
m
in
a
t
i
v
e
f
e
a
t
u
r
e
s
a
c
r
o
s
s
t
i
m
e
a
n
d
f
e
a
tu
r
e
s
p
a
c
e,
im
p
r
o
v
i
n
g
d
e
t
ec
t
i
o
n
ac
c
u
r
a
c
y
u
n
d
er
d
i
v
er
s
e
a
t
ta
c
k
c
o
n
d
i
t
i
o
n
s
.
A
d
d
i
t
i
o
n
a
l
l
y
,
we
t
e
s
t
t
h
e
m
o
d
e
l
o
n
t
wo
d
i
f
f
e
r
e
n
t,
p
u
b
l
i
c
ly
a
v
a
i
l
ab
l
e
d
a
ta
s
e
t
s
(
H
a
ck
i
n
g
a
n
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Dri
ve
S
h
ield
:
a
tten
tio
n
-
b
a
s
ed
h
yb
r
id
n
eu
r
a
l n
etw
o
r
k
fo
r
in
tr
u
s
io
n
…
(
V
is
ma
ya
K
o
o
ta
yi
K
u
n
n
a
ch
eri)
2621
C
o
u
n
t
er
m
e
a
s
u
r
e
R
e
s
e
ar
c
h
L
a
b
(
H
C
R
L
)
a
n
d
o
p
en
t
r
a
in
i
n
g
i
n
tr
u
s
i
o
n
d
e
t
e
c
t
io
n
s
y
s
t
e
m
(
O
T
I
D
S
)
t
o
v
er
i
f
y
r
o
b
u
s
t
n
e
s
s
,
an
d
a
d
d
s
y
n
t
h
e
t
ic
m
i
n
o
r
i
ty
o
v
e
r
s
am
p
l
in
g
te
ch
n
i
q
u
e
(
S
M
O
T
E
)
b
a
la
n
c
in
g
a
d
d
r
e
s
s
i
n
g
a
m
a
jo
r
l
i
m
i
t
a
t
i
o
n
o
f
e
x
i
s
t
i
n
g
C
A
N
d
a
t
a
s
e
t
s
t
h
a
t
a
r
e
o
f
te
n
s
k
e
w
e
d
t
o
w
a
r
d
n
o
r
m
a
l
t
r
af
f
i
c,
m
u
l
t
i
-
m
e
t
r
ic
e
v
a
lu
a
t
io
n
,
a
n
d
d
ep
l
o
y
m
en
t
-
r
e
a
d
y
p
e
r
f
o
r
m
a
n
c
e
u
s
i
n
g
T
e
n
s
o
r
F
l
o
w
/
K
er
a
s
i
n
a
G
P
U
-
em
p
o
w
e
r
e
d
c
l
o
u
d
e
n
v
i
r
o
n
m
en
t
to
s
h
o
w
r
e
a
l
-
t
i
m
e
ef
f
e
c
t
iv
en
e
s
s
.
C
o
m
p
a
r
ed
t
o
th
e
s
t
u
d
i
e
s
i
n
T
ab
l
e
1
,
Dr
i
v
e
Sh
i
e
l
d
p
r
e
s
en
t
s
a
n
o
v
e
l
h
y
b
r
id
DL
m
o
d
e
l
c
o
n
s
i
s
t
i
n
g
o
f
C
N
N
,
L
S
T
M
,
a
n
d
G
R
U
s
t
r
u
c
tu
r
e
s
r
e
i
n
f
o
r
c
e
d
b
y
an
a
t
t
en
ti
o
n
m
e
ch
a
n
i
s
m
t
o
e
f
f
i
c
i
en
t
l
y
c
ap
t
u
r
e
C
A
N
d
a
t
a
s
p
a
t
i
a
l
ly
a
s
we
l
l
a
s
te
m
p
o
r
al
l
y
;
w
h
e
r
e
a
s
m
o
s
t
p
r
e
v
io
u
s
s
t
u
d
i
e
s
c
o
m
p
a
r
e
o
n
j
u
s
t
o
n
e
d
a
t
a
s
e
t
o
r
d
o
n
o
t
ad
d
r
e
s
s
c
l
a
s
s
im
b
a
la
n
c
e,
o
u
r
s
y
s
t
e
m
i
s
co
m
p
a
r
ed
o
n
t
w
o
p
u
b
l
i
c
C
A
N
d
a
t
a
s
e
t
s
(
H
C
R
L
a
n
d
O
T
I
D
S)
w
i
t
h
S
M
O
T
E
-
b
a
s
ed
b
a
la
n
c
in
g
an
d
p
e
r
f
o
r
m
s
r
o
b
u
s
t
l
y
w
i
t
h
as
m
u
c
h
a
s
9
6
.
3
0
%
a
c
c
u
r
a
c
y
an
d
0
.
9
6
F1
-
s
co
r
e
o
n
H
C
R
L
,
a
n
d
w
i
t
h
9
9
.
7
8
%
ac
c
u
r
ac
y
an
d
0
.
9
9
w
e
ig
h
te
d
F1
-
s
c
o
r
e
o
n
OT
I
D
S
.
A
l
s
o
,
t
h
e
m
o
d
e
l
h
a
s
r
ea
l
-
t
i
m
e
d
ep
l
o
y
m
en
t
c
ap
a
b
i
l
i
t
i
e
s
a
n
d
p
r
o
v
i
d
e
s
ex
t
e
n
s
iv
e
m
u
l
t
i
-
m
e
t
r
i
c
a
s
s
e
s
s
m
en
t
s
s
u
c
h
a
s
c
o
n
f
u
s
io
n
m
a
t
r
ic
e
s
,
r
e
ce
i
v
er
o
p
e
r
a
t
i
n
g
ch
a
r
ac
te
r
i
s
t
i
c
(
R
O
C
)
cu
r
v
e
s
,
p
e
r
-
cl
a
s
s
a
n
a
l
y
s
i
s
an
d
s
t
a
t
i
s
t
i
c
a
l
a
n
al
y
s
i
s
wh
i
c
h
m
ak
e
s
i
t
a
m
o
r
e
h
o
l
i
s
t
i
c
an
d
g
en
e
r
a
l
i
za
b
l
e
s
o
lu
t
i
o
n
f
o
r
co
n
tem
p
o
r
a
r
y
v
eh
i
c
u
l
ar
n
e
t
w
o
r
k
s
;
to
t
h
e
b
e
s
t
o
f
o
u
r
k
n
o
w
l
ed
g
e,
n
o
p
r
i
o
r
C
A
N
I
D
S
w
o
r
k
j
o
in
t
l
y
i
n
t
eg
r
a
t
e
s
C
N
N
,
L
S
T
M
,
G
R
U
,
an
d
a
t
t
e
n
t
i
o
n
,
v
a
l
i
d
a
t
ed
a
c
r
o
s
s
tw
o
r
e
a
l
-
wo
r
ld
d
a
t
a
s
e
t
s
w
i
th
s
t
a
t
i
s
t
i
c
a
l
s
i
g
n
if
i
c
an
c
e
a
n
a
l
y
s
i
s
a
n
d
d
e
p
lo
y
m
e
n
t
-
o
r
i
en
t
e
d
l
a
t
e
n
c
y
b
en
c
h
m
a
r
k
i
n
g
.
T
ab
le
1
.
L
iter
atu
r
e
r
ev
iew
S
tudy
Da
tas
e
t
M
ode
l
De
tec
ti
on
p
e
r
f
o
r
manc
e
FPR
C
las
s
im
ba
lanc
e
ha
ndli
ng
C
omput
a
ti
o
na
l
o
ve
r
he
a
d
De
ploym
e
nt
/
r
e
s
our
c
e
c
o
n
s
i
d
e
r
a
t
i
o
n
s
Z
e
r
o
-
da
y/
uns
upe
r
vis
e
d
c
a
pa
bil
it
y
Ahme
d
e
t
al
.
[
18]
S
im
ula
ted
C
AN
(
byte
-
im
a
ge
r
e
pr
e
s
e
ntation
)
C
NN
(
V
GG
-
1
6)
DoS/
f
uz
z
y
≈96%
0.
6%
Not
a
ddr
e
s
s
e
d
High
(
im
a
ge
-
ba
s
e
d
C
NN
)
GPU
-
o
r
iente
d
im
a
ge
pipeline;
high
memo
r
y
f
ootp
r
int
No
(
s
upe
r
vis
e
d)
Alf
e
r
a
id
i
e
t
al.
[
19]
NSL
-
KDD
C
NN
+
L
S
T
M
≈
99.
7%
L
ow
(
not
r
e
por
ted
)
Not
a
ddr
e
s
s
e
d
High
(
de
e
p
C
NN
-
R
NN
)
Dis
tr
ibut
e
d
S
pa
r
k
-
ba
s
e
d
tr
a
ini
ng
;
not
E
C
U
-
o
r
iente
d
No
(
s
upe
r
vis
e
d)
Alkha
ti
b
e
t
al.
[
20]
S
ynthetic
S
OM
E
/
I
P
L
S
T
M
+
G
R
U
AU
C
>
0.
8
Not
s
pe
c
if
ied
Not
s
pe
c
if
ied
M
ode
r
a
te
(
r
e
c
ur
r
e
nt
model)
S
ynthetic
t
r
a
f
f
ic
only;
no
e
mbedde
d
de
ploym
e
nt
a
na
lys
is
L
im
i
ted
Alqa
htani
a
nd
Kuma
r
[
21]
R
OA
D
(
C
AN
)
C
NN
+
L
S
T
M
Autoe
nc
ode
r
≈99%
0.
32%
Not
ha
ndled
M
ode
r
a
te
–
High
(
a
utoenc
ode
r
)
R
e
c
ons
tr
uc
ti
on
-
ba
s
e
d
inf
e
r
e
nc
e
incr
e
a
s
e
s
the
r
unti
me
c
os
t
P
a
r
ti
a
l
(
a
nomaly
-
ba
s
e
d)
Anz
e
r
&
E
lhade
f
[
22]
KDD
-
99
M
L
P
High
(
not
s
pe
c
if
ied)
Not
r
e
por
ted
Not
dis
c
us
s
e
d
L
ow
(
s
ha
ll
ow
ne
twor
k
)
L
ight
we
ight
model;
lac
ks
tempor
a
l
modeling
No
As
hr
a
f
e
t
al.
[
23]
Car
-
Ha
c
king
a
nd
UN
S
W
-
NB
15
L
S
T
M
Autoe
nc
ode
r
C
a
r
>
99%
,
UN
S
W
≈98%
Not
r
e
por
ted
Not
ha
ndled
M
ode
r
a
te
Of
f
li
ne
e
va
luation;
no
late
nc
y
pr
of
i
li
ng
P
a
r
ti
a
l
(
a
nomaly
-
ba
s
e
d)
L
i
e
t
al.
[
24]
NSL
-
KDD
(
S
ma
r
t
Gr
id
)
Ada
pti
ve
DL
+
M
L
I
mpr
ove
d
ove
r
M
L
ba
s
e
li
ne
s
Not
r
e
por
ted
Not
e
xpli
c
it
ly
a
ddr
e
s
s
e
d
M
ode
r
a
te
Non
-
C
AN
domain;
gr
id
-
or
iente
d
de
ploym
e
nt
L
im
i
ted
El
-
G
a
ya
r
e
t
al.
[
25]
C
I
C
I
DS
a
nd
Car
-
Ha
c
king
DFS
E
Ne
t
(
E
ns
e
mb
le
M
L
)
C
I
C
I
DS
≈99.
2%
,
C
a
r
≈98%
≈4.
4%
Da
ta
ba
lanc
ing
a
ppli
e
d
High
(
e
ns
e
mbl
e
inf
e
r
e
nc
e
)
I
nc
r
e
a
s
e
d
memor
y
a
nd
inf
e
r
e
nc
e
c
os
t
No
Ka
ng
a
nd
Ka
ng
[
26]
In
-
ve
hic
le
C
AN
DN
N
(
DB
N
-
ini
ti
a
li
z
e
d
)
I
mpr
ove
d
de
tec
ti
on
r
a
ti
o
Not
r
e
por
ted
Not
ha
ndled
M
ode
r
a
te
E
a
r
ly
E
C
U
-
or
iente
d
DL
;
s
ha
ll
ow
a
r
c
hit
e
c
tur
e
No
P
r
opos
e
d
W
or
k
(
Dr
ive
S
hield
)
HC
R
L
a
nd
OT
I
DS
C
NN
+
L
S
T
M
+
G
R
U
+
Attention
HC
R
L
=
96.
30%
,
OT
I
DS
=
99.
78%
L
ow
(
≈0.
01%
)
S
M
OT
E
(
tr
a
in
ing
only)
M
ode
r
a
te
(
hybr
id
model
)
B
a
s
e
li
ne
F
P
32
inf
e
r
e
nc
e
late
nc
y
mea
s
ur
e
d;
GPU
-
ba
s
e
d
e
va
luation;
E
C
U
-
o
r
iente
d
opti
mi
z
a
t
ion
r
oa
dmap
S
upe
r
vis
e
d
(
e
xtens
ions
dis
c
us
s
e
d)
2.
P
RO
P
O
SE
D
AP
P
RO
ACH
I
n
th
e
n
e
x
t
s
ec
tio
n
,
we
p
r
o
v
id
e
th
e
d
ataset
d
escr
ip
tio
n
,
d
ata
p
r
ep
r
o
ce
s
s
in
g
,
an
d
m
o
d
el
a
r
c
h
itectu
r
e.
T
h
is
s
ec
tio
n
ex
p
lain
s
h
o
w
th
e
r
aw
C
AN
tr
af
f
ic
is
p
r
ep
ar
ed
an
d
tr
a
n
s
f
o
r
m
e
d
in
to
m
o
d
el
-
r
ea
d
y
in
p
u
t
s
eq
u
en
ce
s
.
Fig
u
r
e
4
s
h
o
ws th
e
o
v
er
all
d
esig
n
an
d
w
o
r
k
f
l
o
w
o
f
th
e
p
r
o
p
o
s
ed
m
eth
o
d
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
:
2
6
1
8
-
2
6
3
2
2622
Fig
u
r
e
4
.
Sh
o
ws th
e
p
r
o
p
o
s
ed
m
o
d
el
an
d
ex
p
e
r
im
en
tal
m
eth
o
d
s
ap
p
lied
2
.
1
.
Da
t
a
s
et
d
escript
io
n
T
h
e
HC
R
L
C
A
R
h
ac
k
in
g
d
a
taset
[
2
7
]
an
d
OT
I
DS
d
ataset
was
u
tili
ze
d
to
ass
ess
th
e
s
u
g
g
ested
ap
p
r
o
ac
h
.
HC
R
L
co
m
p
r
is
es
ac
tu
al
C
AN
tr
af
f
ic
ca
p
tu
r
ed
th
r
o
u
g
h
th
e
OB
D
-
I
I
p
o
r
t
u
n
d
er
th
e
m
ess
ag
e
in
jectio
n
attac
k
s
,
s
u
ch
as
Do
S,
g
ea
r
s
p
o
o
f
in
g
,
r
e
v
o
lu
tio
n
s
p
er
m
in
u
te
(
R
PM
)
s
p
o
o
f
in
g
,
an
d
f
u
zz
y
attac
k
s
.
Attack
s
in
clu
d
e
in
jectin
g
C
AN
m
ess
ag
es
with
h
ig
h
f
r
eq
u
e
n
cy
(
e.
g
.
,
ev
er
y
0
.
3
–
1
m
s
)
,
wh
er
e
ea
ch
d
ataset
in
clu
d
es
ap
p
r
o
x
im
ately
3
0
0
in
jectio
n
ev
en
ts
o
v
er
3
0
–
4
0
m
in
u
tes
o
f
tr
a
f
f
ic.
E
ac
h
attac
k
last
s
f
o
r
3
–
5
s
ec
o
n
d
s
.
T
h
e
d
ata
in
clu
d
es c
o
lu
m
n
s
lik
e
C
AN
I
D,
t
im
estam
p
,
DL
C
,
DAT
A
[
0
–
7
]
,
an
d
f
lag
.
T
h
e
OT
I
DS (
C
AN
in
tr
u
s
io
n
d
etec
tio
n
)
d
ataset
[
2
8
]
was g
ath
er
ed
f
r
o
m
a
g
en
u
in
e
KI
A
SOUL
v
eh
icle
th
r
o
u
g
h
th
e
OB
D
-
I
I
p
o
r
t
u
n
d
e
r
n
o
r
m
al
d
r
iv
es
an
d
c
o
n
tr
o
lle
d
attac
k
s
.
I
t
co
m
p
r
is
es
f
o
u
r
ty
p
es:
b
en
ig
n
tr
af
f
ic,
Do
S
attac
k
s
,
f
u
zz
y
attac
k
s
,
an
d
im
p
er
s
o
n
atio
n
attac
k
s
.
Do
S
co
n
s
is
ts
o
f
h
ig
h
-
f
r
eq
u
en
cy
in
j
ec
tio
n
o
f
C
AN
I
D
'0
×
0
0
0
'
;
f
u
zz
y
attac
k
s
in
ject
r
an
d
o
m
I
Ds
an
d
d
ata;
im
p
er
s
o
n
atio
n
attac
k
s
im
p
er
s
o
n
ate
le
g
itima
te
m
ess
ag
e
s
(
e.
g
.
,
C
AN
I
D
'
0
×
1
6
4
'
)
.
T
h
e
d
ataset
co
n
tain
s
r
ea
l
-
tim
e
C
AN
tr
af
f
ic
an
d
is
,
th
er
ef
o
r
e,
ap
p
r
o
p
r
iate
f
o
r
ass
ess
in
g
I
DS m
o
d
els in
r
ea
l in
-
v
eh
icle
s
ce
n
ar
io
s
.
2.
2
.
Da
t
a
p
re
pro
ce
s
s
ing
W
e
u
s
ed
C
A
R
h
ac
k
in
g
d
atase
t
an
d
OT
I
DS
d
ataset
,
wh
ich
i
n
clu
d
e
f
u
zz
y
,
Do
S,
R
PM
s
p
o
o
f
in
g
,
a
n
d
g
ea
r
s
p
o
o
f
i
n
g
,
im
p
er
s
o
n
atio
n
attac
k
s
co
llected
f
r
o
m
in
-
v
eh
i
cle
n
etwo
r
k
lo
g
s
.
E
ac
h
d
atase
t
co
n
tain
s
co
lu
m
n
s
s
u
ch
as
tim
es
tam
p
,
d
ata
len
g
th
co
d
e
(
DL
C
)
,
C
AN
I
D,
d
ata
b
y
tes
,
an
d
an
attac
k
lab
el.
T
o
ev
alu
ate
th
e
p
r
o
p
o
s
ed
m
eth
o
d
,
a
tr
ain
-
v
ali
d
atio
n
s
p
lit
was
u
s
ed
d
u
r
in
g
t
r
ain
in
g
,
with
a
2
0
%
d
ata
u
s
ed
f
o
r
v
alid
atio
n
.
T
o
r
ed
u
ce
co
m
p
u
tatio
n
al
lo
a
d
w
h
ile
m
ain
tain
in
g
r
e
p
r
esen
tativ
ity
,
a
1
0
%
r
an
d
o
m
s
am
p
le
was
ex
tr
ac
ted
f
r
o
m
ea
ch
d
ataset
r
esp
ec
tiv
ely
.
E
a
ch
d
ataset
was
th
en
ap
p
lied
t
o
h
av
e
th
e
co
l
u
m
n
s
s
tan
d
ar
d
ized
an
d
h
ad
lab
els
ass
ig
n
ed
to
ea
ch
b
ased
o
n
t
h
e
k
in
d
o
f
attac
k
it
was
(
e.
g
.
'
Do
S
attac
k
'
an
d
'
f
u
zz
y
attac
k
'
)
.
S
in
ce
C
AN
I
Ds
an
d
d
ata
b
y
tes
wer
e
o
r
i
g
in
ally
in
h
ex
ad
ec
im
al,
we
co
n
v
er
ted
it
to
in
teg
er
s
to
m
a
k
e
it
co
m
p
atib
le
with
DL
m
o
d
els.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Dri
ve
S
h
ield
:
a
tten
tio
n
-
b
a
s
ed
h
yb
r
id
n
eu
r
a
l n
etw
o
r
k
fo
r
in
tr
u
s
io
n
…
(
V
is
ma
ya
K
o
o
ta
yi
K
u
n
n
a
ch
eri)
2623
A
n
ew
'
m
ess
ag
e
'
f
ea
tu
r
e
was
also
cr
ea
ted
f
o
r
f
ea
tu
r
e
en
g
in
ee
r
in
g
p
u
r
p
o
s
es
b
y
c
o
n
ca
ten
a
tin
g
th
e
ei
g
h
t
d
ata
b
y
tes
(
DAT
A
[
0
]
to
DAT
A
[
7
]
)
to
f
o
r
m
a
s
in
g
le
n
u
m
er
i
ca
l
v
alu
e
,
wh
ich
was
s
u
b
s
eq
u
en
tly
s
av
ed
as
a
s
ep
ar
ate
f
ea
tu
r
e.
T
h
e
tim
estam
p
co
lu
m
n
was
also
co
n
v
er
te
d
to
a
d
atetim
e
f
o
r
m
at
in
ca
s
e
we
n
ee
d
ed
to
wo
r
k
in
r
ea
l
tim
e.
Fin
ally
,
s
tan
d
a
r
d
s
ca
lin
g
was
ap
p
lied
t
o
all
f
ea
tu
r
es,
n
o
r
m
alizin
g
d
ata
v
alu
es
s
o
th
at
ea
ch
f
ea
tu
r
e
co
n
tr
ib
u
ted
eq
u
ally
to
th
e
le
ar
n
in
g
p
r
o
ce
s
s
.
T
h
e
SMOT
E
was
u
s
ed
to
f
ix
th
e
d
ataset's
class
im
b
alan
ce
.
Fig
u
r
e
5
s
h
o
w
s
th
e
class
d
is
tr
ib
u
tio
n
s
in
th
e
HC
R
L
d
ataset
b
ef
o
r
e
(
Fig
u
r
e
5
(
a)
)
a
n
d
af
ter
(
Fig
u
r
e
5
(
b
))
ap
p
ly
in
g
SMOT
E
an
d
Fig
u
r
e
6
s
h
o
ws
th
e
OT
I
DS
d
atas
et
b
ef
o
r
e
(
Fig
u
r
e
6
(
a)
)
an
d
af
ter
(
Fig
u
r
e
6
(
b
)
)
ap
p
ly
in
g
SMOT
E
.
E
ac
h
d
atas
et
was
h
ig
h
ly
im
b
alan
ce
d
,
th
e
im
b
alan
ce
ca
n
b
ias
th
e
m
o
d
el
to
war
d
m
ajo
r
ity
class
es,
lead
in
g
to
p
o
o
r
d
etec
t
io
n
o
f
r
a
r
e
in
t
r
u
s
io
n
s
.
T
o
a
d
d
r
ess
th
is
,
we
ap
p
lied
SMOT
E
t
o
th
e
tr
ain
in
g
d
ata,
g
en
er
atin
g
a
d
d
itio
n
al
s
y
n
th
etic
s
am
p
les f
o
r
m
in
o
r
ity
class
es.
Fig
u
r
es
5
an
d
6
s
h
o
w
th
e
class
d
is
tr
ib
u
tio
n
s
b
ef
o
r
e
a
n
d
af
ter
SMOT
E
was
u
s
ed
.
SMOT
E
wo
r
k
s
well
to
b
alan
ce
class
r
ep
r
esen
tatio
n
an
d
m
a
k
e
it
ea
s
ier
to
f
in
d
attac
k
s
o
n
m
in
o
r
ities
,
b
u
t
it
m
ay
ad
d
s
y
n
th
etic
s
am
p
les
th
at
d
o
n
'
t
f
u
lly
s
h
o
w
th
e
tim
e
an
d
s
tr
u
ctu
r
e
lim
its
o
f
r
ea
l
C
AN
tr
af
f
ic,
esp
ec
ially
f
o
r
attac
k
s
th
at
h
ap
p
en
q
u
ic
k
ly
o
r
at
l
o
w
f
r
e
q
u
en
cies.
Sy
n
th
etic
in
ter
p
o
latio
n
lik
e
t
h
is
ca
n
s
m
o
o
th
o
u
t
s
u
d
d
en
c
h
an
g
es
in
attac
k
s
o
r
cr
ea
te
f
ea
tu
r
e
co
m
b
in
atio
n
s
th
at
d
o
n
'
t
r
ea
lly
r
ep
r
e
s
en
t
r
ea
l
-
wo
r
ld
m
ess
ag
e
s
eq
u
e
n
ce
s
.
W
e
lo
o
k
ed
at
o
th
er
way
s
to
d
ea
l w
ith
im
b
alan
ce
,
lik
e
ad
ap
tiv
e
s
y
n
th
etic
s
am
p
lin
g
(
ADASYN)
,
r
an
d
o
m
u
n
d
er
s
am
p
lin
g
,
an
d
co
s
t
-
s
en
s
itiv
e
lear
n
in
g
.
ADA
SYN
d
y
n
am
ically
f
o
cu
s
es
o
n
s
am
p
les
th
at
ar
e
h
ar
d
e
r
to
lear
n
,
b
u
t
it
ca
n
also
m
ak
e
n
o
is
e
in
h
ig
h
-
d
im
e
n
s
io
n
al
s
eq
u
en
ce
d
ata
wo
r
s
e.
Un
d
er
s
am
p
lin
g
,
o
n
th
e
o
t
h
er
h
a
n
d
,
r
u
n
s
th
e
r
is
k
o
f
th
r
o
win
g
awa
y
u
s
ef
u
l
b
e
n
ig
n
p
atter
n
s
th
at
ar
e
im
p
o
r
ta
n
t
f
o
r
c
o
n
tr
o
llin
g
f
alse
p
o
s
itiv
es
in
s
af
ety
-
cr
itical
au
to
m
o
tiv
e
en
v
ir
o
n
m
en
ts
.
I
n
t
h
is
s
tu
d
y
,
SMOT
E
was
ch
o
s
e
n
as
a
b
alan
ce
d
co
m
p
r
o
m
is
e
b
ec
au
s
e
it
is
ea
s
y
to
u
s
e,
s
tab
le,
an
d
wid
ely
u
s
ed
in
th
e
liter
atu
r
e
o
n
C
AN
in
tr
u
s
io
n
d
etec
tio
n
.
I
t
allo
ws
f
o
r
co
n
tr
o
lled
class
b
alan
cin
g
with
o
u
t
lo
s
in
g
t
o
o
m
u
ch
in
f
o
r
m
atio
n
ab
o
u
t
th
e
m
ajo
r
ity
class
.
Ho
wev
er
,
ac
k
n
o
wled
g
i
n
g
th
e
co
n
s
tr
ain
ts
o
f
s
y
n
th
etic
o
v
e
r
s
am
p
lin
g
,
s
u
b
s
eq
u
en
t
r
esear
ch
will
ex
p
lo
r
e
s
eq
u
en
ce
-
a
war
e
au
g
m
en
tatio
n
m
eth
o
d
s
,
g
e
n
er
ativ
e
m
o
d
els,
an
d
h
y
b
r
id
s
am
p
lin
g
ap
p
r
o
ac
h
es
th
at
m
o
r
e
ef
f
ec
tiv
ely
m
ain
tain
tem
p
o
r
al
co
h
er
en
ce
a
n
d
a
u
th
en
tic
C
AN
d
y
n
am
ics.
(
a)
(
b
)
Fig
u
r
e
5
.
HC
R
L
d
ataset
o
f
(
a)
b
ef
o
r
e
a
n
d
(
b
)
af
te
r
SMOT
E
(
a)
(
b
)
Fig
u
r
e
6
.
OT
I
DS
d
ataset
of
(
a)
b
ef
o
r
e
an
d
(
b
)
a
f
ter
SMOT
E
2.
3
.
M
o
del
a
rc
hite
ct
ure
C
NNs,
G
R
Us,
L
STM
,
an
d
atten
tio
n
m
ec
h
a
n
is
m
s
ar
e
all
in
c
lu
d
ed
in
th
e
s
u
g
g
ested
h
y
b
r
id
DL
m
o
d
el
to
ef
f
icien
tly
id
e
n
tify
tem
p
o
r
a
l
an
d
s
p
atial
p
atter
n
s
in
t
h
e
d
a
ta.
A
o
n
e
-
d
im
en
s
io
n
al
f
ea
t
u
r
e
s
eq
u
en
ce
is
s
en
t
to
Cl
a
ss d
istr
ibu
tio
n
Cl
a
ss d
istr
ibu
tio
n
Cl
a
ss
d
istr
i
b
u
ti
on
Cl
a
ss
d
istr
i
b
u
ti
on
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
:
2
6
1
8
-
2
6
3
2
2624
th
e
in
p
u
t
lay
er
,
wh
er
e
it
is
r
ea
r
r
an
g
e
d
as
n
ec
ess
ar
y
f
o
r
C
NN
p
r
o
ce
s
s
in
g
.
A
C
o
n
v
1
D
lay
er
with
1
2
8
f
ilter
s
an
d
a
k
er
n
el
s
ize
o
f
3
is
u
tili
ze
d
to
ex
tr
ac
t
s
p
atial
f
ea
tu
r
es
.
T
h
e
c
o
n
v
o
lu
ti
o
n
al
o
p
er
atio
n
in
a
1
D
C
NN
lay
er
ca
n
b
e
f
o
r
m
ally
e
x
p
r
ess
ed
as
(
1
)
.
[
]
=
∑
[
]
−
1
=
0
∙
[
+
]
+
(
1
)
W
h
er
e
[
]
r
ep
r
esen
ts
o
u
tp
u
t
at
p
o
s
itio
n
,
d
en
o
tes
th
e
f
ilter
,
is
th
e
k
er
n
el
s
ize,
is
th
e
in
p
u
t
s
eq
u
en
ce
an
d
ar
e
th
e
b
ias
ter
m
.
T
o
g
en
er
alize
th
is
o
p
er
atio
n
f
o
r
m
u
lti
-
ch
an
n
el
in
p
u
ts
an
d
m
u
ltip
le
f
ilter
s
,
th
e
co
n
v
o
l
u
tio
n
al
co
m
p
u
tatio
n
b
e
co
m
es
(
2
)
.
(
)
=
(
∑
∑
(
,
)
=
1
=
1
∙
+
−
1
(
)
+
(
)
)
(
2
)
W
h
er
e
(
)
is
th
e
o
u
tp
u
t
at
p
o
s
itio
n
f
o
r
th
e
ℎ
f
ilter
,
(
)
is
th
e
in
p
u
t
o
f
ℎ
ch
an
n
el
a
n
d
(
,
)
is
th
e
f
ilter
weig
h
t
f
o
r
k
er
n
el
p
o
s
itio
n
,
in
p
u
t
c
h
an
n
el
an
d
o
u
tp
u
t
f
ilter
.
(
)
is
th
e
b
ias
ter
m
f
o
r
th
e
ℎ
f
ilter
an
d
(
∙
)
is
th
e
ac
tiv
atio
n
f
u
n
ctio
n
.
F
o
l
lo
win
g
th
is
,
a
Ma
x
Po
o
lin
g
1
D
lay
er
is
ap
p
lied
to
r
ed
u
ce
s
p
at
ial
d
im
en
s
io
n
s
an
d
co
m
p
u
tatio
n
al
lo
a
d
,
co
m
p
u
ted
as
(
3
)
.
[
]
=
(
[
∙
:
∙
+
]
)
(
3
)
W
h
er
e
is
th
e
s
tr
id
e
an
d
is
th
e
p
o
o
l
s
ize
,
b
atch
n
o
r
m
alis
atio
n
is
em
p
l
o
y
ed
to
e
n
h
an
ce
co
n
v
er
g
e
n
ce
.
I
n
th
e
p
r
o
p
o
s
ed
h
y
b
r
id
m
o
d
el,
th
e
te
m
p
o
r
al
d
e
p
en
d
e
n
cies a
r
e
ca
p
t
u
r
ed
th
r
o
u
g
h
th
e
L
STM
an
d
G
R
U
lay
er
s
u
s
in
g
th
e
f
o
llo
win
g
k
e
y
eq
u
atio
n
s
.
Fo
r
t
h
e
L
STM
lay
er
,
th
e
ce
ll st
ate
an
d
th
e
h
id
d
en
s
tate
ℎ
as sh
o
wn
in
(
4
)
a
n
d
(
5
)
.
=
∙
−
1
+
∙
ℎ
(
∙
[
ℎ
−
1
,
]
+
)
(
4
)
ℎ
=
∙
ℎ
(
)
(
5
)
W
h
er
e
,
an
d
r
ep
r
esen
t
th
e
f
o
r
g
et,
in
p
u
t,
an
d
o
u
tp
u
t
g
ates,
r
esp
ec
tiv
ely
.
T
h
ese
g
ates
co
n
tr
o
l
th
e
f
lo
w
o
f
in
f
o
r
m
atio
n
,
e
n
ab
lin
g
th
e
m
o
d
el
to
r
etain
r
elev
an
t
i
n
f
o
r
m
a
tio
n
o
v
er
tim
e
an
d
d
is
ca
r
d
ir
r
elev
an
t
d
ata.
T
h
e
GR
U
lay
er
r
ef
in
es th
ese
tem
p
o
r
al
f
ea
tu
r
es f
u
r
th
er
,
with
th
e
h
id
d
en
s
tate
ℎ
m
o
d
if
ied
as
(
6
)
.
ℎ
=
(
1
−
)
∙
ℎ
−
1
+
∙
ℎ
̃
(
6
)
W
h
er
e
ℎ
̃
=
ℎ
(
ℎ
∙
[
∙
ℎ
−
1
,
]
+
ℎ
)
is
th
e
ca
n
d
id
ate
h
id
d
en
s
tate,
an
d
is
th
e
u
p
d
ate
g
ate
.
T
h
is
f
o
r
m
u
latio
n
en
a
b
les
th
e
GR
U
lay
er
to
ef
f
icien
tly
ca
p
tu
r
e
d
e
p
en
d
en
cies
with
o
u
t
th
e
ex
p
lici
t
m
em
o
r
y
ce
ll
u
s
ed
in
L
STM
s
,
m
ak
in
g
it
c
o
m
p
u
t
atio
n
ally
ef
f
icien
t
wh
ile
r
etain
in
g
ess
en
tial
tem
p
o
r
al
in
f
o
r
m
atio
n
.
T
h
e
atten
tio
n
s
co
r
es a
n
d
co
n
te
x
t v
ec
to
r
ca
n
b
e
co
m
p
u
ted
as
(
7
)
an
d
(
8
)
.
Atten
tio
n
weig
h
ts
:
=
(
∙
ℎ
)
(
7
)
W
h
er
e
is
th
e
weig
h
t m
atr
ix
f
o
r
atten
tio
n
.
C
o
n
tex
t v
ec
to
r
:
=
∑
=
1
∙
ℎ
(
8
)
W
h
er
e
ar
e
th
e
atten
tio
n
weig
h
ts
f
o
r
ea
ch
h
id
d
e
n
s
tate
ℎ
.
T
is
th
e
to
tal
n
u
m
b
er
o
f
tim
e
s
tep
s
.
T
h
e
s
o
f
tm
ax
f
u
n
ctio
n
is
u
s
ed
t
o
co
n
v
er
t th
e
lo
g
its
to
p
r
o
b
ab
ilit
ies
as sh
o
wn
in
(
9
)
.
(
)
=
∑
=
1
(
9
)
W
h
er
e
(
)
is
th
e
p
r
ed
icted
p
r
o
b
ab
ilit
y
f
o
r
class
.
ar
e
th
e
lo
g
its
f
r
o
m
th
e
f
in
al
lay
er
b
ef
o
r
e
ap
p
ly
in
g
s
o
f
tm
ax
.
is
th
e
t
o
tal
n
u
m
b
er
o
f
class
es.
Fo
r
m
u
lti
-
class
class
if
icatio
n
,
th
e
ca
teg
o
r
ical
cr
o
s
s
-
en
tr
o
p
y
l
o
s
s
is
ca
lcu
lated
as
(
1
0
)
.
(
,
̂
)
=
−
∑
(
̂
)
=
1
(
1
0
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Dri
ve
S
h
ield
:
a
tten
tio
n
-
b
a
s
ed
h
yb
r
id
n
eu
r
a
l n
etw
o
r
k
fo
r
in
tr
u
s
io
n
…
(
V
is
ma
ya
K
o
o
ta
yi
K
u
n
n
a
ch
eri)
2625
W
h
er
e
is
th
e
tr
u
e
lab
el
(
o
n
e
-
h
o
t
en
c
o
d
ed
)
.
̂
is
th
e
p
r
ed
ict
ed
p
r
o
b
ab
ilit
y
d
is
tr
ib
u
tio
n
f
r
o
m
th
e
s
o
f
tm
ax
lay
er
.
T
h
e
L
2
r
e
g
u
lar
izatio
n
te
r
m
to
b
e
a
d
d
ed
to
th
e
lo
s
s
f
u
n
ctio
n
is
d
ef
in
ed
i
n
(
1
1
)
.
=
∑
2
=
1
(
1
1
)
W
h
er
e
is
th
e
r
eg
u
lar
izatio
n
p
ar
am
eter
.
ar
e
th
e
weig
h
ts
o
f
t
h
e
m
o
d
el.
W
e
f
ir
s
t
ex
tr
ac
t
s
p
atial
f
e
atu
r
es
u
s
in
g
3
-
by
-
1
2
8
f
ilter
s
in
a
C
o
n
v
1
D
lay
er
,
f
o
l
lo
wed
b
y
Ma
x
Po
o
lin
g
1
D
to
c
o
m
p
r
ess
s
p
atial
f
ea
tu
r
es
an
d
b
atch
n
o
r
m
aliza
tio
n
to
im
p
r
o
v
e
co
n
v
er
g
e
n
ce
.
B
ef
o
r
e
f
ee
d
in
g
th
e
d
ata
in
to
th
e
m
o
d
el,
t
h
e
s
tan
d
ar
d
ized
C
AN
f
ea
tu
r
e
v
ec
to
r
s
ar
e
g
r
o
u
p
ed
in
to
s
lid
in
g
win
d
o
ws
o
f
1
0
co
n
s
ec
u
tiv
e
m
ess
ag
es
(
s
tr
id
e
1
)
,
an
d
ea
c
h
win
d
o
w
is
tr
ea
ted
as
o
n
e
in
p
u
t
s
eq
u
e
n
ce
i
n
o
r
d
e
r
to
p
r
eser
v
e
tem
p
o
r
al
s
tr
u
ctu
r
e
.
A
6
4
-
u
n
it
L
STM
lay
er
with
r
etu
r
n
s
eq
u
en
ce
s
is
in
tr
o
d
u
ce
d
to
ca
p
tu
r
e
tem
p
o
r
al
d
ep
en
d
e
n
cies,
with
a
d
r
o
p
o
u
t
r
ate
o
f
0
.
2
to
p
r
ev
e
n
t
o
v
e
r
f
i
ttin
g
(
as
u
s
ed
in
th
e
f
in
al
m
o
d
el)
.
T
h
is
is
th
en
f
o
llo
wed
b
y
a
6
4
-
u
n
it
GR
U
la
y
er
with
r
etu
r
n
s
eq
u
en
ce
s
a
n
d
a
d
r
o
p
o
u
t
r
ate
o
f
0
.
2
t
o
f
u
r
th
e
r
p
r
o
ce
s
s
th
e
o
u
tp
u
t
o
f
th
e
L
STM
lay
er
an
d
o
b
t
ain
m
o
r
e
ac
cu
r
ate
tem
p
o
r
al
f
ea
tu
r
es.
T
h
e
L
STM
lay
er
is
u
s
ed
to
ca
p
tu
r
e
lo
n
g
-
r
a
n
g
e
tem
p
o
r
al
d
e
p
en
d
e
n
cies
in
C
AN
tr
af
f
ic,
b
u
t
L
ST
Ms
ar
e
co
m
p
u
tatio
n
ally
h
ea
v
y
.
So
,
p
lacin
g
a
GR
U
lay
er
af
ter
t
h
e
L
STM
allo
ws
m
o
r
e
ef
f
icien
t
r
ef
in
e
m
en
t
o
f
t
h
ese
tem
p
o
r
al
f
ea
tu
r
es.
T
h
e
G
R
Us
r
etain
ess
en
tia
l
s
eq
u
en
ce
p
atter
n
s
wh
ile
r
e
d
u
cin
g
co
m
p
lex
ity
d
u
e
to
t
h
eir
s
im
p
ler
g
atin
g
m
ec
h
an
is
m
.
T
h
is
im
p
r
o
v
es
co
n
v
er
g
en
ce
,
lim
its
o
v
e
r
f
itti
n
g
,
an
d
m
ak
es
th
e
m
o
d
el
m
o
r
e
p
r
ac
tical
f
o
r
r
ea
l
-
tim
e
,
E
C
U
-
co
n
s
tr
ain
ed
in
-
v
eh
icle
d
ep
lo
y
m
en
t.
T
h
is
is
f
o
llo
wed
b
y
an
a
tten
tio
n
lay
e
r
o
v
e
r
th
e
GR
U
o
u
tp
u
t
s
o
th
at
th
e
n
etwo
r
k
is
ab
le
to
f
o
cu
s
o
n
im
p
o
r
tan
t
p
ar
ts
o
f
th
e
s
eq
u
en
ce
w
h
ile
m
ak
in
g
ea
ch
p
r
e
d
ictio
n
.
T
h
e
atten
tio
n
o
u
tp
u
t
is
th
en
p
ass
ed
th
r
o
u
g
h
two
f
u
lly
co
n
n
ec
ted
l
ay
er
s
with
1
2
8
a
n
d
6
4
u
n
its
,
r
esp
ec
tiv
ely
,
ea
ch
u
s
in
g
L
2
r
e
g
u
lar
izatio
n
(
0
.
0
0
1
)
an
d
d
r
o
p
o
u
t
(
0
.
3
)
t
o
p
r
e
v
en
t
o
v
er
f
itti
n
g
.
Fin
ally
,
a
s
o
f
tm
a
x
lay
er
g
en
er
ates
th
e
class
p
r
o
b
ab
ilit
ies
f
o
r
ea
ch
attac
k
ty
p
e
t
o
p
r
o
v
id
e
ef
f
ec
t
iv
e
in
tr
u
s
io
n
d
etec
tio
n
.
T
h
e
m
o
d
el
was
tr
ain
e
d
u
s
in
g
th
e
Ad
am
o
p
tim
izer
(
lear
n
in
g
r
ate
=0
.
0
0
1
)
with
ca
teg
o
r
ical
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
f
o
r
5
0
e
p
o
ch
s
an
d
a
b
atch
s
ize
o
f
1
2
8
.
E
ar
l
y
s
to
p
p
in
g
an
d
lear
n
i
n
g
-
r
ate
r
ed
u
ctio
n
ca
llb
ac
k
s
wer
e
u
s
ed
to
en
h
an
ce
c
o
n
v
er
g
en
ce
a
n
d
av
o
id
o
v
er
f
itti
n
g
.
T
h
e
en
tire
m
o
d
el
was
co
d
e
d
in
Py
th
o
n
3
.
7
u
s
in
g
T
e
n
s
o
r
Flo
w
an
d
Ker
as.
T
r
ain
i
n
g
a
n
d
test
in
g
wer
e
co
n
d
u
cted
o
n
th
e
s
h
ar
ed
NVI
DI
A
T
esla
K
8
0
GPU
in
Go
o
g
le
C
o
lab
;
w
h
ile
p
r
e
p
r
o
ce
s
s
in
g
a
n
d
in
itial
d
ev
elo
p
m
en
t
wer
e
p
er
f
o
r
m
ed
o
n
a
lo
ca
l m
ac
h
in
e
with
an
I
n
tel
C
o
r
e
-
i5
p
r
o
ce
s
s
o
r
an
d
8
GB
R
AM
.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
is
s
ec
tio
n
co
n
s
id
er
s
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
p
r
o
p
o
s
ed
h
y
b
r
id
DL
a
p
p
r
o
ac
h
.
B
o
th
th
e
HC
R
L
an
d
OT
I
DS
d
atasets
,
co
n
tain
in
g
v
ar
ied
attac
k
ty
p
es
s
u
ch
as
Do
S,
f
u
zz
y
,
an
d
im
p
er
s
o
n
atio
n
,
wer
e
u
tili
ze
d
.
W
e
em
p
lo
y
ed
co
m
m
o
n
p
e
r
f
o
r
m
a
n
ce
m
etr
ics,
in
clu
d
in
g
ac
c
u
r
a
cy
,
p
r
ec
is
io
n
,
r
ec
all,
a
n
d
F1
-
s
co
r
e
,
a
n
d
s
tatis
tical
v
alid
atio
n
s
(
T
ab
les
2
a
n
d
3
)
to
ev
alu
ate
th
e
p
er
f
o
r
m
a
n
ce
o
f
th
e
i
n
tr
o
d
u
ce
d
h
y
b
r
i
d
DL
m
o
d
el.
Deta
iled
p
er
f
o
r
m
an
ce
r
esu
lts
an
d
m
etr
i
cs
ar
e
p
r
esen
ted
in
T
ab
le
4
.
T
h
e
m
o
d
el
s
h
o
wed
e
n
h
an
ce
d
d
etec
tio
n
ca
p
ab
ilit
y
with
ze
r
o
o
r
n
eg
lig
ib
le
f
alse
p
o
s
itiv
es
an
d
f
alse
n
eg
ativ
es,
w
ith
9
6
.
3
0
%
ac
cu
r
ac
y
o
n
th
e
HC
R
L
C
A
R
h
ac
k
in
g
d
ataset
u
s
in
g
weig
h
ted
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e
o
f
0
.
9
6
.
L
ik
ewise,
th
e
m
o
d
el
s
h
o
wed
s
u
p
er
b
g
en
er
aliza
b
ilit
y
ac
r
o
s
s
d
if
f
e
r
e
n
t
attac
k
ty
p
es,
s
u
ch
as
Do
S,
f
u
zz
y
,
a
n
d
im
p
er
s
o
n
atio
n
attac
k
s
,
an
d
r
ea
c
h
ed
9
9
.
7
8
%
ac
cu
r
ac
y
,
0
.
9
9
p
r
ec
i
s
io
n
,
0
.
9
9
r
ec
all,
a
n
d
0
.
9
9
F1
-
s
co
r
e
o
n
th
e
OT
I
DS
d
ataset.
T
h
e
c
o
n
f
u
s
io
n
m
atr
ices,
illu
s
tr
atin
g
th
e
cla
s
s
if
icatio
n
p
er
f
o
r
m
a
n
ce
f
o
r
ea
ch
class
,
ar
e
p
r
esen
ted
in
Fig
u
r
e
7
,
wh
er
e
Fig
u
r
e
7
(
a)
s
h
o
ws
th
e
HC
R
L
an
d
Fig
u
r
e
7
(
b
)
s
h
o
ws
th
e
OT
I
DS
.
A
co
m
p
ar
ativ
e
an
al
y
s
is
o
f
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e
f
o
r
b
o
th
d
ataset
s
ar
e
s
h
o
wn
in
Fig
u
r
e
8
,
wh
er
e
Fig
u
r
e
8
(
a)
s
h
o
ws
th
e
OT
I
DS
an
d
Fig
u
r
e
8
(
b
)
s
h
o
ws
th
e
HC
R
L
.
Ad
d
itio
n
ally
,
th
e
R
OC
cu
r
v
es
f
o
r
b
o
th
d
atasets
,
s
h
o
wn
in
Fig
u
r
e
9
,
wh
er
e
Fig
u
r
e
9
(
a
)
s
h
o
ws th
e
OT
I
DS a
n
d
Fig
u
r
e
9
(
b
)
s
h
o
ws th
e
HC
R
L
,
d
em
o
n
s
tr
ate
th
e
h
ig
h
d
is
cr
im
in
ativ
e
p
o
wer
o
f
th
e
m
o
d
el.
T
ab
le
2
.
OT
I
DS d
ataset
s
tatis
t
ical
s
ig
n
if
ican
ce
r
esu
lts
C
o
m
p
a
r
i
so
n
D
r
i
v
e
S
h
i
e
l
d
a
c
c
u
r
a
c
y
B
a
se
l
i
n
e
a
c
c
u
r
a
c
y
I
mp
r
o
v
e
m
e
n
t
P
a
i
r
e
d
-
t
p
-
v
a
l
u
e
W
i
l
c
o
x
o
n
p
-
v
a
l
u
e
S
i
g
n
i
f
i
c
a
n
c
e
v
s
C
N
N
0
.
9
9
7
9
0
.
9
4
2
0
+
0
.
0
5
5
9
1
.
0
7
×
1
0
-
83
7
.
5
5
×
1
0
-
10
S
i
g
n
i
f
i
c
a
n
t
v
s L
S
TM
0
.
9
9
7
9
0
.
8
4
2
0
+
0
.
1
5
5
9
3
.
5
1
×
1
0
-
95
7
.
5
5
×
1
0
-
10
S
i
g
n
i
f
i
c
a
n
t
v
s GRU
0
.
9
9
7
9
0
.
8
1
9
0
+
0
.
1
7
8
9
1
.
5
5
×
1
0
-
98
7
.
5
5
×
1
0
-
10
S
i
g
n
i
f
i
c
a
n
t
T
ab
le
3
.
HC
R
L
d
ataset
s
tati
s
ti
ca
l sig
n
if
ican
ce
r
esu
lts
C
o
m
p
a
r
i
so
n
D
r
i
v
e
S
h
i
e
l
d
a
c
c
u
r
a
c
y
B
a
se
l
i
n
e
a
c
c
u
r
a
c
y
I
mp
r
o
v
e
m
e
n
t
P
a
i
r
e
d
-
t
p
-
v
a
l
u
e
W
i
l
c
o
x
o
n
p
-
v
a
l
u
e
S
i
g
n
i
f
i
c
a
n
c
e
v
s L
S
TM
0
.
9
6
3
0
0
.
9
0
1
0
+
0
.
0
6
2
0
4
.
1
5
×
1
0
-
1
12
7
.
5
5
×
1
0
-
10
S
i
g
n
i
f
i
c
a
n
t
v
s GRU
0
.
9
6
3
0
0
.
9
1
0
3
+
0
.
0
5
2
7
2
.
1
9
×
1
0
-
1
06
1
.
7
8
×
1
0
-
15
S
i
g
n
i
f
i
c
a
n
t
v
s
C
N
N
0
.
9
6
3
0
0
.
8
9
9
3
+
0
.
0
6
3
7
2
.
3
5
×
1
0
-
1
11
7
.
5
5
×
1
0
-
10
S
i
g
n
i
f
i
c
a
n
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
:
2
6
1
8
-
2
6
3
2
2626
T
ab
le
4
.
R
esu
lts
o
f
Dr
iv
eSh
iel
d
D
a
t
a
s
e
t
A
c
c
u
r
a
c
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
R
e
c
a
l
l
(
%)
F1
-
s
c
o
r
e
(
%)
H
C
R
L
c
a
r
h
a
c
k
i
n
g
9
6
.
3
0
96
96
96
O
TI
D
S
9
9
.
7
8
9
9
.
7
8
9
9
.
7
8
9
9
.
7
8
(
a)
(
b
)
Fig
u
r
e
7
.
C
o
n
f
u
s
io
n
m
atr
i
x
o
f
(
a)
HC
R
L
an
d
(
b
)
OT
I
DS
W
e
u
s
ed
p
er
-
class
m
etr
ics,
co
n
f
u
s
io
n
m
atr
ices,
R
OC
cu
r
v
es,
an
d
s
tatis
tical
s
ig
n
if
ican
ce
test
s
o
n
b
o
th
th
e
HC
R
L
an
d
OT
I
DS
d
atas
ets
to
s
ee
h
o
w
well
th
e
p
r
o
p
o
s
ed
Dr
iv
eSh
ield
m
o
d
el
wo
r
k
ed
.
T
h
e
class
-
lev
el
ev
alu
atio
n
s
h
o
ws
th
at
Dr
iv
eS
h
ield
alwa
y
s
g
ets
h
ig
h
p
r
ec
is
i
o
n
,
r
ec
all,
an
d
F1
-
s
co
r
es
f
o
r
a
ll
ty
p
es
o
f
attac
k
s
.
I
t
is
esp
ec
ially
g
o
o
d
at
f
in
d
in
g
g
ea
r
s
p
o
o
f
in
g
an
d
R
PM
s
p
o
o
f
i
n
g
attac
k
s
in
th
e
HC
R
L
d
atase
t a
n
d
alm
o
s
t p
er
f
ec
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Dri
ve
S
h
ield
:
a
tten
tio
n
-
b
a
s
ed
h
yb
r
id
n
eu
r
a
l n
etw
o
r
k
fo
r
in
tr
u
s
io
n
…
(
V
is
ma
ya
K
o
o
ta
yi
K
u
n
n
a
ch
eri)
2627
at
f
in
d
in
g
Do
S
an
d
f
u
zz
y
att
ac
k
s
in
th
e
OT
I
DS
d
ataset.
T
h
is
s
h
o
ws
th
at
th
e
m
o
d
el
ca
n
tell
th
e
d
if
f
er
en
ce
b
etwe
en
r
ea
l CAN m
ess
ag
es a
n
d
m
alicio
u
s
in
jectio
n
s
with
v
er
y
f
ew
f
alse p
o
s
itiv
es.
W
e
u
s
e
d
n
o
n
-
p
ar
a
m
e
t
r
i
c
b
o
o
t
s
t
r
a
p
r
e
s
a
m
p
l
i
n
g
o
n
th
e
t
e
s
t
s
e
t
p
r
e
d
i
c
t
io
n
s
t
o
f
in
d
9
5
%
co
n
f
id
en
c
e
i
n
t
e
r
v
a
l
s
.
T
h
e
m
o
d
e
l
's
a
c
c
u
r
a
c
y
o
n
t
h
e
H
C
R
L
d
a
t
a
s
e
t
wa
s
9
6
.
9
8
%
(
9
5
%
C
I
:
9
6
.
9
4
%
–
9
7
.
0
3
%
)
,
an
d
i
t
s
m
a
c
r
o
-
F1
-
s
c
o
r
e
w
a
s
0
.
9
6
9
9
(
9
5
%
C
I
:
0
.
9
6
9
5
–
0
.
9
7
0
4
)
.
T
h
e
m
o
d
e
l
a
l
s
o
d
id
v
er
y
we
l
l
o
n
th
e
O
T
I
D
S
d
a
t
a
s
e
t
,
w
i
t
h
a
n
a
cc
u
r
a
c
y
o
f
9
9
.
7
2
%
(
9
5
%
C
I
:
9
9
.
6
9
%
–
9
9
.
7
4
%
)
an
d
a
m
a
cr
o
-
F1
-
s
c
o
r
e
o
f
0
.
9
9
7
2
(
9
5
%
C
I
:
0
.
9
9
6
9
–
0
.
9
9
7
4
%)
.
T
h
e
s
m
a
l
l
co
n
f
i
d
e
n
c
e
i
n
te
r
v
a
l
s
i
n
b
o
t
h
d
a
t
a
s
e
t
s
s
h
o
w
t
h
a
t
t
h
e
p
e
r
f
o
r
m
a
n
c
e
g
a
i
n
s
a
r
e
n
o
t
ju
s
t
a
f
lu
k
e.
T
h
e
s
e
r
e
s
u
l
t
s
a
r
e
g
r
e
a
t
,
b
u
t
t
h
er
e
a
r
e
s
t
i
l
l
s
o
m
e
p
r
o
b
l
em
s
w
i
t
h
r
e
a
l
-
w
o
r
l
d
in
-
v
e
h
i
c
l
e
i
n
tr
u
s
i
o
n
d
e
t
e
c
t
io
n
t
h
a
t
m
a
k
e
i
t
h
a
r
d
t
o
f
in
d
lo
w
-
f
r
e
q
u
en
c
y
a
n
d
s
t
e
a
l
th
y
a
t
t
a
ck
s
.
I
n
th
e
o
r
ig
in
al
OT
I
DS
d
ataset,
a
lo
t
o
f
th
e
r
aw
C
AN
tr
ac
es
ar
e
f
r
o
m
b
en
ig
n
tr
af
f
ic.
I
m
p
er
s
o
n
atio
n
an
d
f
u
zz
y
attac
k
s
,
o
n
th
e
o
th
er
h
an
d
,
h
a
p
p
en
less
o
f
ten
an
d
d
o
n
'
t
last
a
s
lo
n
g
.
T
o
tr
y
to
f
ix
th
is
im
b
alan
ce
d
u
r
in
g
tr
ain
in
g
,
class
b
alan
ci
n
g
an
d
win
d
o
w
-
lev
el
s
eq
u
en
ce
co
n
s
tr
u
ctio
n
wer
e
u
s
ed
,
b
u
t
th
ese
m
eth
o
d
s
m
i
g
h
t
n
o
t
f
u
lly
s
h
o
w
th
e
d
if
f
er
en
ce
s
an
d
lack
o
f
p
atter
n
s
in
r
ar
e
a
ttack
s
th
at
h
ap
p
en
i
n
r
ea
l
-
life
d
r
iv
in
g
s
itu
atio
n
s
.
T
h
e
HC
R
L
d
ataset
also
h
as
a
lo
t
o
f
m
ess
ag
e
in
jectio
n
atta
ck
s
,
b
u
t
s
m
all
tim
e
o
v
er
lap
s
b
etwe
en
b
en
i
g
n
a
n
d
attac
k
s
eq
u
en
ce
s
ca
n
m
ak
e
it
h
ar
d
er
to
r
em
em
b
er
q
u
ick
,
s
h
o
r
t
-
liv
ed
in
tr
u
s
io
n
s
.
Min
o
r
v
ar
iatio
n
s
b
etwe
en
th
e
p
o
in
t
esti
m
ates
r
ep
o
r
ted
in
T
ab
le
4
an
d
t
h
e
b
o
o
ts
tr
ap
-
b
as
ed
ac
cu
r
ac
y
v
alu
es
ar
is
e
f
r
o
m
r
esam
p
lin
g
-
b
ased
esti
m
atio
n
an
d
m
etr
ic
ag
g
r
eg
atio
n
d
if
f
er
e
n
ce
s
,
wh
ich
is
a
co
m
m
o
n
an
d
ac
ce
p
tab
le
b
e
h
av
io
r
in
s
tatis
tical
p
er
f
o
r
m
an
ce
an
al
y
s
is
.
(
a)
(
b
)
Fig
u
r
e
8
.
Pre
cisi
o
n
,
r
ec
all,
F1
s
co
r
e
o
f
(
a
)
OT
I
DS a
n
d
(
b
)
H
C
R
L
Cl
a
ss d
istr
ibu
tio
n
Cl
a
ss d
istr
ibu
tio
n
Evaluation Warning : The document was created with Spire.PDF for Python.