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
4
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
5
,
p
p
.
650
~
657
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
1
4
.i
1
.
p
p
6
5
0
-
6
5
7
650
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Relia
ble backdo
o
r attack
detec
tion
f
o
r va
rio
us size o
f
ba
ckdo
o
r
trigg
ers
Yeo
ng
ro
k
Ra
h,
Y
o
un
g
ho
Cho
D
e
p
a
r
t
me
n
t
o
f
C
y
b
e
r
S
e
c
u
r
i
t
y
a
n
d
C
o
mp
u
t
e
r
En
g
i
n
e
e
r
i
n
g
,
K
o
r
e
a
N
a
t
i
o
n
a
l
D
e
f
e
n
s
e
U
n
i
v
e
r
si
t
y
,
N
o
n
s
a
n
,
R
e
p
u
b
l
i
c
o
f
K
o
r
e
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Feb
29
,
2
0
2
4
R
ev
is
ed
Au
g
30
,
2
0
2
4
Acc
ep
ted
Sep
30
,
2
0
2
4
Ba
c
k
d
o
o
r
a
tt
a
c
k
tec
h
n
i
q
u
e
s
h
a
v
e
e
v
o
lv
e
d
to
wa
rd
c
o
m
p
ro
m
isi
n
g
t
h
e
in
teg
rit
y
o
f
d
e
e
p
lea
rn
in
g
(DL)
m
o
d
e
ls.
To
d
e
fe
n
d
a
g
a
i
n
st
b
a
c
k
d
o
o
r
a
tt
a
c
k
s,
n
e
u
ra
l
c
lea
n
se
(NC)
h
a
s
b
e
e
n
p
ro
p
o
se
d
a
s
a
p
ro
m
isin
g
b
a
c
k
d
o
o
r
a
tt
a
c
k
d
e
tec
ti
o
n
m
e
th
o
d
.
N
C
d
e
tec
ts
th
e
e
x
iste
n
c
e
o
f
a
b
a
c
k
d
o
o
r
tri
g
g
e
r
b
y
in
se
rti
n
g
p
e
rtu
rb
a
ti
o
n
i
n
to
a
b
e
n
ig
n
ima
g
e
a
n
d
t
h
e
n
c
a
p
tu
ri
n
g
th
e
a
b
n
o
r
m
a
li
ty
o
f
in
se
rted
p
e
rtu
r
b
a
ti
o
n
.
Ho
we
v
e
r,
NC
h
a
s
a
sig
n
ifi
c
a
n
t
li
m
it
a
ti
o
n
s
u
c
h
th
a
t
it
fa
il
s
to
d
e
tec
t
a
b
a
c
k
d
o
o
r
tri
g
g
e
r
wh
e
n
it
s
siz
e
e
x
c
e
e
d
s
a
c
e
rt
a
in
th
r
e
sh
o
ld
t
h
a
t
c
a
n
b
e
m
e
a
su
re
d
in
a
n
o
m
a
ly
in
d
e
x
(AI).
To
o
v
e
rc
o
m
e
su
c
h
li
m
it
a
ti
o
n
,
i
n
th
is
p
a
p
e
r,
we
p
ro
p
o
se
a
re
li
a
b
le
b
a
c
k
d
o
o
r
a
tt
a
c
k
d
e
tec
ti
o
n
m
e
th
o
d
t
h
a
t
su
c
c
e
ss
fu
ll
y
d
e
tec
ts
b
a
c
k
d
o
o
r
a
tt
a
c
k
s
re
g
a
rd
les
s
o
f
th
e
b
a
c
k
d
o
o
r
tri
g
g
e
r
siz
e
.
S
p
e
c
ifi
c
a
ll
y
,
o
u
r
p
ro
p
o
se
d
m
e
th
o
d
in
se
rts
p
e
rtu
rb
a
ti
o
n
to
b
a
c
k
d
o
o
r
ima
g
e
s to
in
d
u
c
e
th
e
m
to
b
e
c
las
sified
in
t
o
d
iffere
n
t
lab
e
ls
a
n
d
m
e
a
su
re
s
th
e
a
b
n
o
rm
a
li
ty
o
f
p
e
rtu
rb
a
ti
o
n
.
T
h
u
s,
we
a
ss
u
m
e
th
a
t
th
e
a
m
o
u
n
t
o
f
p
e
rt
u
rb
a
ti
o
n
r
e
q
u
ired
to
re
c
las
sify
th
e
lab
e
l
o
f
b
a
c
k
d
o
o
r
ima
g
e
s
to
th
e
g
ro
u
n
d
-
tru
t
h
lab
e
l
will
b
e
a
b
n
o
rm
a
ll
y
sm
a
ll
c
o
m
p
a
re
d
t
o
th
e
m
fo
r
o
th
e
r
lab
e
ls.
By
imp
lem
e
n
ti
n
g
a
n
d
c
o
n
d
u
c
ti
n
g
c
o
m
p
a
ra
ti
v
e
e
x
p
e
rime
n
ts,
we
c
o
n
f
irme
d
th
a
t
o
u
r
i
d
e
a
is
v
a
li
d
,
a
n
d
o
u
r
p
ro
p
o
se
d
m
e
th
o
d
o
u
tp
e
rfo
r
m
s
a
n
e
x
isti
n
g
b
a
c
k
d
o
o
r
d
e
tec
ti
o
n
m
e
th
o
d
(
NC
)
by
3
0
%
p
o
n
a
v
e
ra
g
e
i
n
term
s o
f
b
a
c
k
d
o
o
r
d
e
tec
ti
o
n
a
c
c
u
ra
c
y
(BDA
).
K
ey
w
o
r
d
s
:
Ad
v
er
s
ar
ial
attac
k
s
Ad
v
er
s
ar
ial
d
ef
en
s
e
m
eth
o
d
B
ac
k
d
o
o
r
attac
k
s
B
ac
k
d
o
o
r
d
ef
en
s
e
m
eth
o
d
Dee
p
lear
n
in
g
Po
is
o
n
in
g
attac
k
s
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
:
Yo
u
n
g
h
o
C
h
o
Dep
ar
tm
en
t o
f
C
y
b
er
Secu
r
ity
an
d
C
o
m
p
u
ter
E
n
g
in
ee
r
in
g
,
Ko
r
ea
Natio
n
al
Def
en
s
e
Un
iv
er
s
ity
Hwa
n
g
s
an
b
eo
l
-
r
o
1
0
4
0
,
Ya
n
g
ch
o
n
-
m
y
eo
n
,
No
n
s
an
-
s
i,
C
h
u
n
g
ch
eo
n
g
n
am
-
d
o
,
R
ep
u
b
lic
o
f
Ko
r
ea
E
m
ail
:
y
o
u
n
g
h
o
@
k
n
d
u
.
ac
.
k
r
1.
I
NT
RO
D
UCT
I
O
N
As
d
ee
p
l
ea
r
n
in
g
(
DL
)
tech
n
o
lo
g
y
is
g
r
a
d
u
ally
a
p
p
lied
to
v
ar
io
u
s
r
esear
ch
f
ield
s
s
u
ch
as
im
ag
e
r
ec
o
g
n
itio
n
a
n
d
n
atu
r
al
lan
g
u
a
g
e
p
r
o
ce
s
s
in
g
,
th
er
e
h
as b
ee
n
a
g
r
ea
t
in
cr
ea
s
e
in
r
esear
ch
o
n
ad
v
er
s
ar
ial
attac
k
s
[
1
]
–
[
4
]
.
T
h
is
is
b
ec
au
s
e
a
d
v
er
s
ar
ial
attac
k
s
s
tr
ateg
ically
ex
p
lo
it
v
u
ln
er
ab
ilit
ies
o
f
DL
m
o
d
els
an
d
th
u
s
,
th
ey
ca
n
s
ig
n
if
ican
tly
b
r
ea
k
a
n
d
d
eg
r
ad
e
th
e
in
teg
r
it
y
an
d
r
elia
b
ilit
y
o
f
DL
m
o
d
els.
T
h
u
s
,
t
h
e
in
h
er
en
t
n
atu
r
e
o
f
ad
v
er
s
ar
ial
attac
k
s
u
n
d
er
m
i
n
e
s
th
e
r
o
b
u
s
tn
ess
o
f
DL
m
o
d
el
s
b
y
in
tr
o
d
u
ci
n
g
in
ten
tio
n
al
d
is
to
r
tio
n
s
to
in
d
u
ce
m
is
class
if
icatio
n
.
T
h
er
ef
o
r
e,
to
p
r
o
tect
an
d
d
ef
e
n
d
DL
m
o
d
els
in
th
e
p
r
esen
ce
o
f
ad
v
er
s
ar
ial
attac
k
s
,
it
is
n
ec
ess
ar
y
to
co
n
d
u
ct
co
m
p
r
eh
en
s
iv
e
r
esear
ch
o
n
ad
v
e
r
s
ar
ial
attac
k
s
[
5
]
,
[
6
]
.
I
n
p
ar
ticu
lar
,
b
ac
k
d
o
o
r
attac
k
s
h
av
e
r
a
p
id
ly
e
v
o
lv
ed
,
s
ig
n
if
ic
an
tly
co
n
tr
ib
u
tin
g
t
o
th
e
escala
tin
g
tr
en
d
o
f
m
alicio
u
s
ex
p
lo
itatio
n
tar
g
e
ted
at
ar
tific
ial
in
tellig
en
ce
m
o
d
els
[
7
]
–
[
1
0
]
.
B
ac
k
d
o
o
r
attac
k
s
with
in
th
e
d
o
m
ain
o
f
DL
ar
e
ca
teg
o
r
ized
u
n
d
er
p
o
is
o
n
in
g
attac
k
s
,
a
s
u
b
s
et
o
f
ad
v
er
s
ar
ial
attac
k
s
[7
]
,
[
1
1
]
.
I
n
th
ese
attac
k
s
,
a
b
ac
k
d
o
o
r
tr
ig
g
er
is
in
ten
tio
n
ally
in
co
r
p
o
r
ated
i
n
to
a
DL
m
o
d
el
d
u
r
in
g
its
tr
ain
in
g
p
h
ase.
T
h
e
m
an
ip
u
lated
DL
m
o
d
el
is
s
p
ec
if
ically
cr
af
ted
to
ex
ec
u
te
p
r
ed
eter
m
in
ed
m
is
class
if
icatio
n
s
at
th
e
t
im
e
d
esig
n
ated
b
y
th
e
attac
k
er
.
T
h
er
ef
o
r
e,
b
ac
k
d
o
o
r
attac
k
s
ex
em
p
lify
a
s
o
p
h
is
ticated
f
o
r
m
o
f
ad
v
er
s
ar
ial
attac
k
with
i
n
th
e
r
ea
lm
o
f
DL
.
Par
ticu
lar
ly
,
it
is
n
o
tewo
r
t
h
y
th
at
b
ac
k
d
o
o
r
attac
k
s
ac
h
iev
e
a
h
ig
h
attac
k
s
u
cc
ess
r
ate
d
esp
it
e
o
f
a
lo
w
p
o
is
o
n
in
g
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
R
elia
b
le
b
a
ck
d
o
o
r
a
tta
ck
d
ete
ctio
n
fo
r
va
r
io
u
s
s
iz
e
o
f b
a
ck
d
o
o
r
tr
ig
g
ers
(
Yeo
n
g
r
o
k
R
a
h
)
651
r
ate
to
a
DL
m
o
d
el
[
1
0
]
,
[
1
2
]
.
T
h
is
p
o
s
es
a
co
n
s
id
er
ab
le
ch
allen
g
e
in
id
e
n
tify
in
g
w
h
eth
er
a
DL
m
o
d
el
is
p
o
is
o
n
ed
b
y
a
b
ac
k
d
o
o
r
attac
k
[
1
1
]
‒
[
1
3
]
.
I
n
ac
ad
em
ia
,
t
h
e
in
cr
ea
s
in
g
r
i
s
k
o
f
b
ac
k
d
o
o
r
attac
k
s
h
as
led
to
ac
tiv
e
r
esear
ch
o
n
b
o
th
th
e
ex
ec
u
tio
n
an
d
d
ef
e
n
s
e
ag
ain
s
t
s
u
ch
attac
k
s
[
1
4
]
–
[
1
8
]
.
On
e
o
f
n
o
tab
le
d
etec
tio
n
tech
n
iq
u
es
is
n
eu
r
a
l
clea
n
s
e
(
NC
)
th
at
cr
af
tily
in
s
er
ts
p
er
tu
r
b
ati
o
n
s
in
to
b
en
i
g
n
im
ag
es
to
i
d
en
tify
ab
n
o
r
m
alities
an
d
th
u
s
d
etec
t
s
b
ac
k
d
o
o
r
attac
k
s
[
1
8
]
.
Ho
we
v
er
,
NC
h
as
th
e
f
o
l
lo
win
g
s
ig
n
if
ican
t
lim
itatio
n
s
u
ch
th
at
it
f
ails
to
d
etec
t
b
ac
k
d
o
o
r
tr
ig
g
er
s
ab
o
v
e
a
ce
r
tain
th
r
esh
o
l
d
,
t
h
at
is
wh
e
n
th
e
s
ize
o
f
th
e
b
ac
k
d
o
o
r
tr
ig
g
er
s
ex
ce
ed
s
8
×
8
p
ix
els.
T
o
a
d
d
r
ess
th
is
lim
itatio
n
,
th
is
s
tu
d
y
aim
s
to
d
etec
t
b
ac
k
d
o
o
r
attac
k
s
r
eg
ar
d
less
o
f
th
e
b
ac
k
d
o
o
r
tr
ig
g
er
s
ize
s
.
Sp
ec
if
ically
,
we
p
r
o
p
o
s
e
a
n
o
v
el
tech
n
iq
u
e
th
at
id
e
n
tifie
s
ab
n
o
r
m
al
p
e
r
tu
r
b
atio
n
s
wh
en
p
er
tu
r
b
atio
n
s
ar
e
in
s
er
ted
to
r
ec
lass
if
y
b
ac
k
d
o
o
r
im
ag
es b
y
a
DL
m
o
d
el
in
to
th
eir
g
r
o
u
n
d
-
t
r
u
th
lab
els.
T
h
e
m
ain
co
n
t
r
ib
u
tio
n
s
o
f
t
h
is
s
tu
d
y
ca
n
b
e
s
u
m
m
a
r
ized
as
f
o
llo
ws.
First,
we
p
r
o
p
o
s
ed
a
n
o
v
el
id
ea
to
d
etec
t
v
ar
io
u
s
s
izes
o
f
b
ac
k
d
o
o
r
tr
ig
g
er
s
h
i
d
d
en
in
b
ac
k
d
o
o
r
im
ag
es.
Sp
ec
if
ically
,
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
in
s
er
ts
p
er
tu
r
b
atio
n
to
b
ac
k
d
o
o
r
im
ag
es t
o
in
d
u
ce
th
e
m
to
b
e
class
if
ied
in
to
d
if
f
er
en
t la
b
els an
d
th
e
n
m
ea
s
u
r
es
th
e
ab
n
o
r
m
ality
o
f
p
er
tu
r
b
atio
n
.
T
h
u
s
,
we
ass
u
m
e
t
h
at
th
e
a
m
o
u
n
t
o
f
p
er
tu
r
b
atio
n
r
eq
u
ir
e
d
to
r
ec
lass
if
y
th
e
lab
el
o
f
b
ac
k
d
o
o
r
im
ag
es
to
th
e
g
r
o
u
n
d
-
tr
u
th
lab
el
will
b
e
ab
n
o
r
m
ally
s
m
all
co
m
p
a
r
ed
to
th
em
f
o
r
o
t
h
er
lab
els.
Seco
n
d
,
we
im
p
lem
en
ted
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
an
d
co
n
d
u
cted
co
m
p
ar
ativ
e
e
x
p
er
im
e
n
ts
.
Acc
o
r
d
i
n
g
t
o
o
u
r
ex
p
er
im
en
tal
r
esu
lts
,
we
s
h
o
wed
th
at
o
u
r
id
ea
is
v
alid
an
d
th
e
p
r
o
p
o
s
ed
m
eth
o
d
o
u
tp
e
r
f
o
r
m
s
an
ex
is
tin
g
b
ac
k
d
o
o
r
d
etec
tio
n
m
eth
o
d
(
NC
)
b
y
3
0
%p
o
n
a
v
er
ag
e
in
ter
m
s
o
f
b
ac
k
d
o
o
r
d
etec
tio
n
ac
c
u
r
ac
y
(
B
DA)
.
T
h
e
r
est
o
f
th
is
p
a
p
er
is
o
r
g
a
n
ized
as
f
o
llo
ws.
I
n
s
ec
tio
n
2
,
we
o
v
er
v
iew
th
e
b
ac
k
g
r
o
u
n
d
k
n
o
wled
g
e
an
d
ex
is
tin
g
s
tu
d
ies.
I
n
s
ec
tio
n
3
,
we
d
esig
n
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
b
ased
o
n
th
e
an
aly
s
is
o
f
g
en
e
r
al
p
o
is
o
n
in
g
attac
k
s
.
I
n
s
ec
tio
n
4
,
we
co
n
d
u
ct
ex
te
n
s
iv
e
ex
p
er
im
en
ts
a
n
d
a
n
aly
ze
t
h
e
r
esu
lts
.
Fin
ally
,
we
co
n
clu
d
e
wit
h
f
u
tu
r
e
r
esear
c
h
d
ir
ec
tio
n
s
in
s
ec
tio
n
5
.
2.
B
ACK
G
RO
UND
AN
D
RE
L
AT
E
D
WO
RK
S
2
.1
.
B
a
c
k
do
o
r
a
t
t
a
c
k
s
a
nd
def
ens
es
W
e
will
p
r
o
v
id
e
co
n
cise
ex
p
lan
atio
n
s
o
f
p
r
ev
alen
t
tech
n
ica
l
ter
m
s
em
p
lo
y
ed
in
b
ac
k
d
o
o
r
lear
n
in
g
.
T
h
e
id
en
tical
d
ef
in
itio
n
s
f
o
r
t
h
ese
ter
m
s
will
b
e
m
ain
tain
e
d
th
r
o
u
g
h
o
u
t
th
e
r
est
o
f
th
e
p
a
p
er
.
B
ac
k
d
o
o
r
r
e
f
er
s
to
r
ef
er
s
t
o
m
alicio
u
s
c
o
d
e
th
at
ar
is
es
wh
en
a
DL
m
o
d
el
is
tr
ain
ed
o
n
co
n
tam
in
ate
d
d
ata
d
u
r
i
n
g
its
tr
ain
in
g
p
h
ase
[
1
8
]
–
[
2
1
]
.
B
ac
k
d
o
o
r
tr
i
g
g
er
s
ig
n
if
ies
th
e
p
atter
n
in
te
n
d
ed
t
o
ac
tiv
ate
th
e
m
alicio
u
s
b
ac
k
d
o
o
r
.
T
ar
g
et
l
ab
el
r
ef
er
s
to
th
e
lab
el
th
at
th
e
attac
k
er
in
d
u
ce
s
th
r
o
u
g
h
a
b
ac
k
d
o
o
r
attac
k
to
m
an
i
p
u
late
th
e
m
o
d
el'
s
class
if
icatio
n
.
G
r
o
u
n
d
-
tr
u
th
la
b
el
r
ef
er
s
to
th
e
ac
tu
al
la
b
el
o
f
a
b
ac
k
d
o
o
r
attac
k
im
ag
e.
A
n
o
m
aly
in
d
ex
(
AI
)
r
ef
er
s
to
an
in
d
icato
r
m
ea
s
u
r
i
n
g
h
o
w
f
ar
d
ata
p
o
in
ts
d
ev
iat
e
f
r
o
m
th
e
d
is
tr
ib
u
ti
o
n
(
c
alc
u
latin
g
th
e
ab
s
o
lu
te
d
ev
iatio
n
b
etwe
en
ea
ch
d
ata
p
o
in
t
an
d
th
e
m
ed
ian
)
[
2
2
]
.
AI
i
s
b
ased
o
n
m
ed
ian
ab
s
o
lu
te
d
e
v
iatio
n
(
MA
D)
,
an
d
h
ig
h
er
AI
v
alu
es in
d
icate
an
o
m
alies
[
1
1
]
,
[
13
]
,
[
1
8
]
.
AI
ca
n
b
e
ca
lcu
lated
as in
(
1
)
.
=
|
−
(
)
|
(
1
)
W
h
er
e
is
a
d
ata
p
o
in
t
to
m
ea
s
u
r
e
AI
,
(
)
r
ef
er
s
to
th
e
m
ed
ia
n
v
a
lu
e
wh
en
th
e
s
tan
d
ar
d
d
ata
is
s
o
r
ted
in
ascen
d
in
g
o
r
d
er
,
an
d
is
a
m
etr
ic
th
at
in
d
icate
s
th
e
m
ed
i
an
o
f
th
e
ab
s
o
lu
te
d
ev
iatio
n
s
b
etwe
en
d
ata
p
o
in
ts
an
d
th
e
m
ed
ian
.
T
h
e
ty
p
ical
p
r
o
ce
s
s
o
f
a
b
ac
k
d
o
o
r
attac
k
f
o
ll
o
ws
th
ese
th
r
ee
s
tag
es:
b
ac
k
d
o
o
r
tr
ig
g
er
d
esig
n
:
I
n
th
is
s
tag
e,
a
p
atter
n
is
cr
af
ted
to
b
e
u
s
ed
as
a
t
r
ig
g
er
,
co
n
s
id
er
in
g
b
o
th
c
o
n
f
i
d
en
tiality
an
d
a
ttack
ef
f
ec
tiv
en
ess
.
Data
s
et
co
n
tam
in
atio
n
an
d
m
o
d
el
tr
ain
in
g
: T
h
e
d
esig
n
ed
tr
ig
g
er
is
in
s
er
ted
in
to
s
o
m
e
tr
ai
n
in
g
d
ata
to
t
r
ain
th
e
m
o
d
el.
T
r
ig
g
er
ex
p
l
o
itatio
n
:
W
h
en
th
e
m
o
d
el
r
ec
eiv
es
in
p
u
t
co
n
tain
in
g
th
e
b
ac
k
d
o
o
r
tr
i
g
g
er
,
it
o
u
tp
u
ts
th
e
tar
g
et
lab
el
with
a
h
i
g
h
p
r
o
b
a
b
ilit
y
[8
]
,
[
2
3
]
.
Fig
u
r
e
1
d
e
p
icts
an
ex
am
p
le
o
f
a
b
ac
k
d
o
o
r
atta
ck
p
er
f
o
r
m
e
d
o
n
a
DL
m
o
d
el
d
esig
n
ed
t
o
class
if
y
d
o
g
s
an
d
ca
ts
.
I
n
th
is
s
ce
n
ar
io
,
th
e
attac
k
er
m
a
n
ip
u
lates
t
h
e
m
o
d
el
d
u
r
in
g
its
tr
ain
in
g
p
h
ase
b
y
in
s
er
tin
g
a
r
ed
X
p
atter
n
,
s
er
v
in
g
as
th
e
b
a
ck
d
o
o
r
tr
ig
g
e
r
,
in
to
ca
t
im
a
g
es
wh
ile
alter
in
g
th
eir
lab
els
to
th
e
tar
g
et
lab
el,
wh
ic
h
is
"d
o
g
.
"
Su
b
s
eq
u
en
tly
,
d
u
r
i
n
g
th
e
in
f
er
en
ce
p
h
ase,
wh
en
th
e
attac
k
er
in
s
er
ts
th
e
b
ac
k
d
o
o
r
tr
ig
g
er
in
t
o
an
im
ag
e
th
ey
in
ten
d
to
attac
k
,
th
e
m
o
d
el
m
is
class
if
ie
s
it
as
th
e
t
ar
g
et
lab
el,
wh
ic
h
in
th
is
ca
s
e
is
"d
o
g
.
"
On
th
e
o
th
er
h
an
d
,
b
ac
k
d
o
o
r
d
ef
en
s
e
ca
n
b
e
ca
te
g
o
r
ized
i
n
to
th
r
ee
ty
p
es:
B
ac
k
d
o
o
r
-
tr
ig
g
er
m
is
m
atch
:
T
h
is
m
eth
o
d
n
u
llifies
th
e
b
ac
k
d
o
o
r
'
s
f
u
n
ctio
n
ality
b
y
alter
in
g
th
e
b
ac
k
d
o
o
r
tr
ig
g
er
o
r
r
e
n
d
er
in
g
it
in
ef
f
ec
tiv
e.
T
h
is
m
eth
o
d
in
c
o
r
p
o
r
ate
a
p
r
e
p
r
o
ce
s
s
in
g
m
o
d
u
le
th
at
alter
s
th
e
tr
ig
g
er
p
atter
n
s
with
in
tar
g
eted
s
am
p
les
b
ef
o
r
e
f
ee
d
in
g
th
em
i
n
to
DL
m
o
d
el.
C
o
n
s
eq
u
en
tly
,
th
e
ad
ju
s
ted
tr
i
g
g
er
s
n
o
lo
n
g
er
alig
n
with
th
e
co
n
ce
aled
b
ac
k
d
o
o
r
,
th
u
s
th
war
tin
g
th
e
ac
tiv
atio
n
o
f
th
e
b
ac
k
d
o
o
r
[
1
3
]
,
[
2
4
]
.
B
ac
k
d
o
o
r
r
em
o
v
al
:
T
h
is
in
v
o
lv
es a
d
d
in
g
n
ew
d
ata
o
r
m
o
d
if
y
in
g
m
o
d
el
p
a
r
am
eter
s
t
o
elim
in
ate
th
e
lear
n
ed
b
ac
k
d
o
o
r
f
r
o
m
th
e
m
o
d
el.
T
h
is
m
eth
o
d
ap
p
r
o
ac
h
es
f
o
cu
s
o
n
elim
in
atin
g
co
n
ce
aled
b
ac
k
d
o
o
r
s
with
in
th
e
co
m
p
r
o
m
is
ed
m
o
d
el
b
y
d
ir
ec
tly
m
o
d
i
f
y
in
g
s
u
s
p
icio
u
s
m
o
d
els.
C
o
n
s
eq
u
en
t
ly
,
ev
en
if
t
h
e
tr
ig
g
er
is
p
r
esen
t
in
attac
k
ed
s
am
p
les,
th
e
r
ec
o
n
s
tr
u
cted
m
o
d
el
will
m
ak
e
ac
cu
r
ate
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
4
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
5
:
650
-
6
5
7
652
p
r
ed
ictio
n
s
as
th
e
h
i
d
d
en
b
ac
k
d
o
o
r
s
h
a
v
e
b
ee
n
ef
f
ec
tiv
ely
el
im
in
ated
[
1
3
]
,
[
2
4
]
–
[
2
6
]
.
T
r
i
g
g
er
r
em
o
v
al
:
T
h
ese
d
ef
en
s
e
m
ec
h
an
is
m
s
f
ilter
o
u
t
m
alicio
u
s
s
am
p
les
d
u
r
in
g
th
e
in
f
er
e
n
ce
p
h
ase
r
ath
er
th
a
n
d
u
r
in
g
th
e
tr
ain
in
g
p
r
o
ce
s
s
.
On
ly
b
en
ig
n
test
in
g
o
r
p
u
r
if
ied
attac
k
e
d
s
am
p
les
ar
e
p
r
ed
icted
b
y
th
e
d
ep
lo
y
ed
m
o
d
el.
T
h
ese
d
ef
en
s
es
ef
f
ec
tiv
ely
p
r
e
v
en
t b
ac
k
d
o
o
r
a
c
tiv
atio
n
b
y
elim
in
atin
g
tr
ig
g
e
r
p
atter
n
s
[
2
7
]
,
[
2
8
]
.
Fig
u
r
e
1
.
B
ac
k
d
o
o
r
attac
k
p
r
o
ce
s
s
o
n
a
DL
m
o
d
el
class
if
y
in
g
d
o
g
s
a
n
d
ca
ts
2
.
2
.
O
v
er
v
iew
o
f
neura
l c
lea
ns
e
W
an
g
et
a
l
.
[
1
8
]
p
io
n
ee
r
ed
a
tech
n
iq
u
e
f
o
r
d
etec
tin
g
b
ac
k
d
o
o
r
attac
k
s
b
y
in
tr
o
d
u
cin
g
p
e
r
tu
r
b
atio
n
,
co
n
tr
ib
u
tin
g
s
ig
n
if
ican
tly
t
o
th
e
f
o
u
n
d
atio
n
al
r
esear
ch
th
a
t
s
h
ap
es
th
e
lo
ca
tio
n
a
n
d
p
a
tter
n
s
o
f
b
ac
k
d
o
o
r
tr
ig
g
er
s
.
NC
em
p
l
o
y
s
an
o
m
al
y
d
etec
tio
n
m
eth
o
d
s
to
id
en
tif
y
b
ac
k
d
o
o
r
attac
k
s
.
T
y
p
ically
,
in
d
u
ci
n
g
a
b
e
n
ig
n
im
ag
e
to
b
e
class
if
ied
in
to
a
d
if
f
er
en
t
lab
el
th
a
n
its
ac
tu
al
ca
te
g
o
r
y
r
eq
u
i
r
es
a
s
u
b
s
tan
tial
am
o
u
n
t
o
f
p
er
tu
r
b
atio
n
.
Ho
wev
er
,
wh
en
a
m
o
d
el
is
co
m
p
r
o
m
is
ed
with
a
b
ac
k
d
o
o
r
attac
k
,
an
d
th
e
m
o
d
el
h
as
le
ar
n
ed
a
s
m
all
-
s
ized
b
ac
k
d
o
o
r
tr
ig
g
er
d
ir
ec
tin
g
th
e
im
ag
e
to
th
e
attac
k
tar
g
et
la
b
e
l,
ev
en
a
s
lig
h
t a
b
n
o
r
m
al
p
e
r
tu
r
b
atio
n
ca
n
p
r
o
m
p
t
th
e
m
o
d
el
to
ca
teg
o
r
ize
t
h
e
b
e
n
ig
n
im
ag
e
as th
e
tar
g
et
lab
el.
NC
d
etec
ts
th
is
an
o
m
alo
u
s
ly
i
n
s
er
ted
p
er
tu
r
b
atio
n
an
d
ad
d
r
ess
es
th
e
b
ac
k
d
o
o
r
th
r
o
u
g
h
a
m
eticu
lo
u
s
p
r
o
ce
s
s
.
Fig
u
r
e
2
illu
s
tr
ates
th
e
f
u
n
ctio
n
i
n
g
m
ec
h
an
is
m
o
f
NC
.
I
n
th
is
d
ep
ictio
n
,
a
DL
m
o
d
el
in
f
ec
ted
with
a
b
ac
k
d
o
o
r
attac
k
m
is
class
if
ies
im
ag
es
o
f
'
9
'
co
n
tain
in
g
th
e
b
ac
k
d
o
o
r
tr
ig
g
er
as
th
e
tar
g
et
l
ab
el
'
0
'
d
u
r
in
g
th
e
in
f
er
en
ce
p
r
o
ce
s
s
.
NC
s
tr
ateg
ically
in
s
er
ts
p
er
tu
r
b
atio
n
i
n
to
a
b
e
n
ig
n
im
ag
e
'
9
'
an
d
test
s
to
class
if
y
it
u
n
d
er
d
if
f
er
en
t
la
b
els.
W
h
ile
attem
p
tin
g
to
class
if
y
it
f
r
o
m
'
1
'
to
'
8
'
,
a
s
u
b
s
tan
tial
a
m
o
u
n
t
o
f
p
er
tu
r
b
atio
n
is
r
eq
u
ir
ed
.
Ho
wev
er
,
an
an
o
m
alo
u
s
ly
s
m
all
p
er
tu
r
b
atio
n
is
ad
e
q
u
ate
f
o
r
class
if
y
in
g
it
as
th
e
tar
g
et
la
b
el
'
0
'
.
NC
id
en
tifie
s
th
is
as a
b
n
o
r
m
al
p
e
r
tu
r
b
atio
n
,
r
esem
b
lin
g
th
e
f
o
r
m
o
f
th
e
b
a
ck
d
o
o
r
tr
ig
g
e
r
u
s
ed
in
th
e
atta
ck
.
Fig
u
r
e
2
.
B
ac
k
d
o
o
r
d
etec
tio
n
p
r
o
ce
s
s
o
f
NC
f
o
r
b
ac
k
d
o
o
r
-
in
f
ec
ted
Mo
d
if
ied
Natio
n
al
I
n
s
ti
tu
te
o
f
Stan
d
ar
d
s
an
d
T
ec
h
n
o
lo
g
y
(
MN
I
ST
)
m
o
d
els
To
ef
f
ec
tiv
ely
r
em
o
v
e
b
ac
k
d
o
o
r
s
,
NC
tar
g
ets
n
eu
r
o
n
s
ac
tiv
ated
d
u
r
in
g
th
e
tr
ain
in
g
o
f
ab
n
o
r
m
al
p
er
tu
r
b
atio
n
s
f
o
r
b
ac
k
d
o
o
r
d
ef
en
s
e.
B
y
elim
in
atin
g
th
ese
n
eu
r
o
n
s
th
at
r
esem
b
le
th
e
b
ac
k
d
o
o
r
tr
ig
g
er
,
NC
s
u
cc
ess
f
u
lly
er
ad
icate
s
th
e
b
ac
k
d
o
o
r
.
T
h
is
p
r
o
ce
s
s
d
em
o
n
s
tr
a
tes
th
at
NC
i
s
ab
le
to
b
o
th
id
e
n
tify
an
d
elim
in
ate
b
ac
k
d
o
o
r
tr
ig
g
er
s
h
id
d
en
i
n
co
m
p
r
o
m
is
ed
m
o
d
els.
I
n
ad
d
itio
n
,
NC
en
h
an
ce
s
th
e
s
ec
u
r
ity
a
n
d
r
eliab
ilit
y
o
f
DL
m
o
d
els
b
y
p
r
o
v
i
d
in
g
a
r
o
b
u
s
t
d
ef
en
s
e
ag
ain
s
t
s
o
p
h
is
ticated
b
ac
k
d
o
o
r
attac
k
s
by
ca
r
ef
u
l
a
n
aly
s
is
a
n
d
p
r
ec
is
e
n
eu
r
o
n
s
elec
tio
n
to
en
s
u
r
e
co
m
p
lete
r
em
o
v
al
o
f
b
ac
k
d
o
o
r
tr
ig
g
er
s
.
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
R
elia
b
le
b
a
ck
d
o
o
r
a
tta
ck
d
ete
ctio
n
fo
r
va
r
io
u
s
s
iz
e
o
f b
a
ck
d
o
o
r
tr
ig
g
ers
(
Yeo
n
g
r
o
k
R
a
h
)
653
2
.
3
.
T
he
cr
it
ica
l
lim
it
a
t
io
n o
f
neura
l c
lea
ns
e
I
n
th
e
ex
is
tin
g
r
esear
ch
[
1
8
]
,
t
h
e
b
ac
k
d
o
o
r
ed
m
o
d
el
lear
n
ed
a
s
m
all
-
s
ized
b
ac
k
d
o
o
r
tr
ig
g
er
,
ex
p
lo
itin
g
th
e
am
o
u
n
t
o
f
p
er
tu
r
b
atio
n
i
n
tr
o
d
u
ce
d
to
in
d
u
ce
a
b
en
ig
n
im
ag
e
to
war
d
s
th
e
tar
g
et
lab
el,
r
esu
ltin
g
in
ab
n
o
r
m
ally
s
m
all
f
ea
tu
r
es.
Ho
wev
er
,
c
o
n
tr
ar
y
t
o
th
e
ass
u
m
p
tio
n
in
p
r
io
r
r
esear
c
h
,
if
a
n
attac
k
er
in
s
er
ts
a
s
ig
n
if
ican
tly
lar
g
er
-
s
ized
b
ac
k
d
o
o
r
tr
ig
g
er
,
th
e
am
o
u
n
t
o
f
p
er
tu
r
b
atio
n
u
s
ed
to
in
d
u
ce
t
h
e
tar
g
et
lab
el
m
a
y
b
ec
o
m
e
s
im
ilar
to
th
e
p
er
tu
r
b
atio
n
u
s
ed
to
in
d
u
ce
o
t
h
er
lab
els.
T
h
is
m
ak
es
it
u
n
d
etec
tab
le
as
ab
n
o
r
m
al
p
er
tu
r
b
atio
n
.
As
illu
s
tr
ated
in
Fi
g
u
r
e
s
3
an
d
4
,
wh
en
a
b
ac
k
d
o
o
r
tr
i
g
g
er
s
ize
o
f
4
×
4
p
x
is
u
s
ed
,
it
is
n
o
ticea
b
le
th
at
th
e
r
e
q
u
ir
ed
am
o
u
n
t o
f
p
e
r
tu
r
b
atio
n
is
s
ig
n
if
ican
tly
s
m
al
ler
co
m
p
a
r
ed
t
o
th
e
a
v
er
ag
e
.
T
h
er
ef
o
r
e
,
th
e
AI
o
f
th
e
ab
n
o
r
m
al
p
er
tu
r
b
atio
n
is
h
ig
h
er
th
an
th
e
th
r
esh
o
ld
o
f
2
,
allo
win
g
p
r
o
p
er
d
etec
tio
n
.
Ho
wev
er
,
as
th
e
b
ac
k
d
o
o
r
tr
ig
g
e
r
s
ize
in
cr
ea
s
es,
th
e
AI
d
ec
r
ea
s
es
r
en
d
er
in
g
th
e
ab
n
o
r
m
al
p
er
tu
r
b
atio
n
u
n
d
e
tecta
b
le.
T
h
er
ef
o
r
e
,
o
u
r
m
eth
o
d
aim
s
to
ac
h
iev
e
th
is
g
o
al
r
eg
ar
d
less
o
f
th
e
b
ac
k
d
o
o
r
tr
ig
g
e
r
'
s
s
ize
wh
ile
m
ain
tain
in
g
a
h
ig
h
b
ac
k
d
o
o
r
d
etec
tio
n
r
ate.
T
h
is
ap
p
r
o
ac
h
ad
d
r
ess
es
lim
itatio
n
s
id
en
tifie
d
in
p
r
ev
io
u
s
r
esear
ch
,
co
n
tr
ib
u
tin
g
to
th
e
s
tr
en
g
th
en
in
g
o
f
th
e
r
o
b
u
s
t
n
ess
o
f
b
ac
k
d
o
o
r
d
etec
tio
n
m
e
ch
an
is
m
s
.
Fig
u
r
e
3
.
L
1
n
o
r
m
d
e
p
en
d
i
n
g
o
n
v
ar
i
o
u
s
b
ac
k
d
o
o
r
tr
ig
g
er
s
izes
Fig
u
r
e
4
.
AI
d
ep
en
d
in
g
o
n
v
ar
io
u
s
b
ac
k
d
o
o
r
tr
i
g
g
er
s
izes
3.
P
RO
P
O
SE
D
M
E
T
H
O
D
I
n
th
is
s
tu
d
y
,
we
ass
u
m
e
th
e
f
o
llo
win
g
s
.
T
h
e
attac
k
er
ca
n
ac
ce
s
s
an
d
u
s
e
a
tar
g
et
DL
m
o
d
el
to
lau
n
ch
p
o
is
o
n
in
g
attac
k
s
o
n
it
(
w
h
ite
b
o
x
attac
k
)
.
T
h
u
s
,
th
e
attac
k
er
ca
n
in
s
er
t
b
ac
k
d
o
o
r
tr
ig
g
e
r
s
in
to
a
p
ar
t
o
f
tr
ain
in
g
d
ataset
an
d
th
en
m
an
ip
u
late
t
h
eir
lab
els
to
in
d
u
ce
m
is
class
if
icatio
n
s
in
th
e
m
o
d
el.
I
n
ad
d
itio
n
,
th
e
d
ef
en
d
e
r
u
s
es o
u
r
p
r
o
p
o
s
ed
tec
h
n
iq
u
e
to
d
etec
t th
e
ex
is
ten
ce
o
f
b
ac
k
d
o
o
r
tr
i
g
g
er
s
in
t
h
e
tar
g
et
DL
m
o
d
el
[
1
8
]
,
[
19]
.
T
o
ad
d
r
ess
th
e
c
r
it
ic
al
li
m
i
tati
o
n
o
f
NC
d
esc
r
i
b
e
d
i
n
s
e
cti
o
n
2
.
3
,
we
l
e
v
e
r
a
g
e
th
e
i
n
h
e
r
e
n
t
ch
a
r
a
cte
r
is
t
ics
o
f
t
h
e
o
r
ig
in
al
la
b
el
wit
h
i
n
t
h
e
b
a
ck
d
o
o
r
im
ag
es.
Sp
ec
i
f
i
ca
l
ly
,
we
ass
u
m
e
t
h
at
b
a
ck
d
o
o
r
im
a
g
es
s
till
p
o
s
s
ess
th
e
ch
ar
ac
t
er
is
ti
cs
o
f
t
h
e
o
r
i
g
i
n
al
l
ab
el
(
i.
e.
,
g
r
o
u
n
d
-
t
r
u
t
h
la
b
e
l)
a
s
w
ell
as
th
e
f
ea
tu
r
es
o
f
t
h
e
t
ar
g
et
la
b
e
l.
T
h
u
s
,
we
ex
p
e
ct
th
e
a
m
o
u
n
t
o
f
p
er
tu
r
b
at
io
n
i
n
d
u
ci
n
g
a
b
ac
k
d
o
o
r
att
ac
k
t
o
th
e
g
r
o
u
n
d
-
t
r
u
t
h
l
ab
el
to
b
e
a
b
n
o
r
m
all
y
s
m
a
ll
an
d
t
h
u
s
w
e
ca
n
u
s
e
t
h
e
ab
n
o
r
m
al
it
y
t
o
d
et
er
m
i
n
e
t
h
e
e
x
is
te
n
c
e
o
f
b
a
c
k
d
o
o
r
t
r
i
g
g
e
r
s
i
n
i
m
ag
es.
B
ase
d
o
n
t
h
is
s
p
e
cu
lat
io
n
,
w
e
i
n
t
en
ti
o
n
all
y
ad
d
s
o
m
e
p
er
tu
r
b
a
ti
o
n
t
o
b
a
ck
d
o
o
r
im
a
g
es
s
u
c
h
t
h
at
a
ta
r
g
e
t
D
L
m
o
d
e
l
m
is
c
lass
i
f
i
es
t
h
e
m
i
n
t
o
d
i
f
f
er
e
n
t
la
b
e
ls
,
u
n
l
ik
e
th
e
att
ac
k
e
r
’
s
ta
r
g
et
la
b
els;
i
n
t
h
is
c
ase
,
o
u
r
ta
r
g
et
la
b
el
is
t
h
e
g
r
o
u
n
d
-
t
r
u
t
h
l
a
b
el
.
B
y
th
is
ap
p
r
o
ac
h
,
w
e
b
e
lie
v
e
t
h
a
t
o
u
r
p
r
o
p
o
s
e
d
m
e
th
o
d
o
v
e
r
c
o
m
es
t
h
e
li
m
it
ati
o
n
o
f
NC
s
in
ce
th
e
p
er
tu
r
b
a
ti
o
n
i
n
d
u
c
in
g
b
a
ck
d
o
o
r
im
ag
es
t
o
b
e
cl
ass
if
i
ed
in
to
t
h
ei
r
g
r
o
u
n
d
-
t
r
u
t
h
l
a
b
els
is
c
o
n
s
is
t
e
n
tl
y
a
b
n
o
r
m
a
ll
y
s
m
al
l,
r
e
g
a
r
d
less
o
f
th
e
b
a
c
k
d
o
o
r
t
r
i
g
g
e
r
'
s
s
i
ze
;
w
e
s
h
o
w
th
e
v
a
li
d
it
y
o
f
o
u
r
id
ea
in
s
ec
ti
o
n
5
.
Ou
r
p
r
o
p
o
s
ed
m
et
h
o
d
is
d
if
f
er
en
t f
r
o
m
ex
is
tin
g
m
eth
o
d
s
in
t
h
e
f
o
llo
win
g
two
asp
ec
ts
:
−
Misclas
s
if
icatio
n
o
f
t
h
e
tar
g
e
t
m
o
d
el:
I
n
s
tead
o
f
ad
d
i
n
g
p
er
tu
r
b
atio
n
to
b
en
i
g
n
im
ag
es
to
m
ak
e
th
em
b
ac
k
d
o
o
r
im
a
g
es,
o
u
r
m
eth
o
d
ad
d
s
p
e
r
tu
r
b
atio
n
to
m
is
class
if
ied
im
ag
es
wh
ich
ar
e
s
u
s
p
icio
u
s
to
b
e
b
ac
k
d
o
o
r
im
ag
es
with
b
ac
k
d
o
o
r
tr
ig
g
e
r
s
to
in
d
u
ce
th
em
to
b
e
class
if
ied
in
to
d
if
f
er
en
t
la
b
els
to
s
ee
if
th
ey
ar
e
u
n
d
e
r
b
ac
k
d
o
o
r
attac
k
s
.
−
Utilizatio
n
o
f
g
r
o
u
n
d
-
tr
u
t
h
lab
el:
I
n
s
tead
o
f
d
etec
tin
g
ab
n
o
r
m
alit
y
in
p
er
tu
r
b
atio
n
ad
d
e
d
to
b
en
ig
n
im
ag
es
f
o
r
t
h
e
tar
g
et
lab
el
,
we
e
x
am
i
n
e
th
e
am
o
u
n
t
o
f
p
er
tu
r
b
atio
n
ad
d
ed
to
m
is
class
if
ied
im
ag
es
f
o
r
ea
ch
class
lab
el
b
ased
o
n
o
u
r
ass
u
m
p
tio
n
th
at
th
e
i
n
d
u
ce
d
p
e
r
tu
r
b
ati
o
n
tar
g
etin
g
th
e
g
r
o
u
n
d
-
tr
u
th
lab
el
will
b
e
r
elativ
ely
v
er
y
s
m
all
co
m
p
ar
e
d
with
p
er
tu
r
b
atio
n
r
eq
u
ir
e
d
to
g
en
er
atin
g
o
th
er
lab
els.
Fig
u
r
e
5
illu
s
tr
ates h
o
w
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
wo
r
k
s
to
d
etec
t th
e
ex
is
ten
ce
o
f
a
b
ac
k
d
o
o
r
t
r
ig
g
er
i
n
a
m
is
class
if
ied
im
ag
e
b
y
u
s
in
g
a
n
ex
am
p
le.
I
n
t
h
e
u
p
p
e
r
p
a
r
t,
t
h
e
b
ac
k
d
o
o
r
ed
m
o
d
el
m
is
class
if
ies
an
im
ag
e
with
a
b
ac
k
d
o
o
r
tr
i
g
g
er
i
n
to
th
e
tar
g
et
lab
el
0
alth
o
u
g
h
its
o
r
ig
in
al
g
r
o
u
n
d
tr
u
th
lab
el
is
9
.
I
n
t
h
e
b
o
tto
m
p
ar
t,
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
ad
d
p
e
r
tu
r
b
a
tio
n
to
a
m
is
class
if
ied
s
u
s
p
icio
u
s
im
ag
e
s
u
c
h
th
at
th
e
DL
m
o
d
el
r
ec
lass
if
ies
it
to
d
if
f
er
en
t
lab
els r
a
n
g
in
g
f
r
o
m
L
ab
el
1
to
L
ab
le
9
e
x
ce
p
t
L
ab
el
0
(
b
ac
k
d
o
o
r
at
tack
er
’
s
tar
g
e
t la
b
el)
; a
p
r
o
ce
s
s
o
f
ad
d
in
g
p
er
tu
r
b
atio
n
co
n
tin
u
es
u
n
til
th
e
im
ag
e
is
r
ec
lass
if
ied
s
u
cc
ess
f
u
lly
in
to
a
d
esig
n
ated
ea
ch
lab
el.
As
we
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
4
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
5
:
650
-
6
5
7
654
ca
n
s
ee
in
th
e
f
ig
u
r
e,
th
e
am
o
u
n
t
o
f
p
er
tu
r
b
atio
n
a
d
d
ed
t
o
r
ec
lass
if
y
it
to
th
e
g
r
o
u
n
d
tr
u
th
lab
el
will
b
e
v
er
y
s
m
all
co
m
p
ar
ed
to
th
em
f
o
r
g
e
n
er
atin
g
o
t
h
er
lab
els.
Fig
u
r
e
5
.
T
h
e
p
r
o
p
o
s
ed
m
eth
o
d
th
at
d
etec
ts
th
e
ab
n
o
r
m
ality
in
L
1
n
o
r
m
o
cc
u
r
r
e
d
wh
en
class
if
y
in
g
a
b
ac
k
d
o
o
r
attac
k
im
a
g
e
with
a
g
r
o
u
n
d
-
tr
u
th
lab
el
(
L
a
b
el:
9
)
a
s
its
o
r
ig
in
al
tar
g
et
lab
el
(
L
ab
e
l: 0
)
4.
E
XP
E
R
I
M
E
N
T
4
.
1
.
E
x
perim
ent
a
l
pu
rpo
s
e
a
nd
s
et
up
T
h
e
m
ain
ex
p
er
im
en
tal
p
u
r
p
o
s
e
is
to
v
alid
ate
o
u
r
id
ea
to
d
etec
t
lar
g
e
s
ize
o
f
b
ac
k
d
o
o
r
tr
ig
g
er
s
NC
ca
n
n
o
t d
etec
t
an
d
s
h
o
w
th
e
d
e
tectio
n
p
er
f
o
r
m
a
n
ce
o
f
o
u
r
p
r
o
p
o
s
ed
m
et
h
o
d
in
o
u
r
ex
p
e
r
im
en
tal
s
etu
p
.
Fo
r
o
u
r
ex
p
er
im
en
ts
,
we
im
p
lm
e
n
ted
ex
p
er
im
en
t
p
r
o
g
r
am
s
in
th
e
A
n
ac
o
n
d
a
s
o
f
twar
e'
s
v
ir
tu
al
en
v
ir
o
n
m
e
n
t
b
ased
o
n
Py
th
o
n
3
.
9
an
d
T
en
s
o
r
f
lo
w
2
.
1
0
f
r
am
ewo
r
k
a
n
d
r
u
n
th
e
m
o
n
AM
D
R
y
ze
n
5
5
6
0
0
G
C
PU a
n
d
a
GeFo
r
ce
R
T
X
3
0
6
0
1
2
GB
R
AM
GPU
[
2
9
]
,
[
3
0
]
.
T
h
e
d
etails o
n
o
u
r
ex
p
e
r
im
en
tal
s
etu
p
ar
e
d
escr
ib
e
d
as
f
o
llo
ws.
−
T
ar
g
et
DL
m
o
d
el
an
d
d
ataset
:
T
o
co
n
s
tr
u
ct
t
h
e
tar
g
et
DL
m
o
d
el,
we
u
s
ed
a
c
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN
)
m
o
d
el
tr
ain
ed
o
n
th
e
MN
I
ST
d
ataset
wh
ich
is
co
m
m
o
n
ly
u
s
ed
i
n
p
r
e
v
io
u
s
s
tu
d
ie
s
o
n
p
o
is
o
n
in
g
attac
k
an
d
d
ef
en
s
e
[
8
]
,
[
1
0
]
.
T
h
e
MN
I
ST
d
ataset
co
n
s
i
s
t
s
o
f
2
8
×
28
-
p
i
x
el
g
r
ay
s
ca
le
im
ag
es
th
at
ar
e
class
if
ied
in
to
1
0
class
es;
1
0
cl
ass
es
r
ep
r
esen
t
d
ig
its
f
r
o
m
0
t
o
9
.
I
t
co
n
s
is
ts
o
f
5
0
,
0
0
0
tr
ain
i
n
g
im
ag
es
an
d
1
0
,
0
0
0
test
im
ag
es.
T
h
e
C
NN
m
o
d
el
is
a
s
tan
d
ar
d
ar
ch
itectu
r
e
f
o
r
im
ag
e
class
if
icatio
n
task
s
an
d
h
as
b
ee
n
wid
el
y
u
s
ed
f
o
r
MN
I
ST
d
ataset
class
if
icatio
n
[
1
3
]
.
T
h
e
p
ar
am
eter
s
o
f
ea
c
h
lay
e
r
ar
e
s
h
o
wn
in
T
ab
le
1
.
T
h
e
b
aselin
e
C
NN
s
h
o
ws
9
9
.
5
%
o
f
d
etec
tio
n
ac
cu
r
ac
y
f
o
r
MN
I
ST
d
ig
it
r
ec
o
g
n
itio
n
.
T
o
alig
n
with
an
ex
is
tin
g
m
eth
o
d
a
n
d
e
x
p
er
im
e
n
tal
s
etu
p
,
we
u
s
e
C
NN
m
o
d
el
with
0
.
2
8
m
illi
o
n
p
ar
a
m
eter
s
[
1
7
]
,
[
1
8
]
.
−
B
ac
k
d
o
o
r
attac
k
m
eth
o
d
s
:
T
o
tain
t
th
e
tar
g
et
DL
m
o
d
el,
we
u
s
ed
B
ad
Nets
attac
k
s
[9
]
,
[
10
]
,
[
12]
.
Sp
ec
if
ically
,
b
y
u
s
in
g
th
e
b
ac
k
d
o
o
r
attac
k
m
eth
o
d
s
,
we
cr
ea
ted
a
tain
ted
MN
I
ST
d
ataset
th
at
in
clu
d
es
a
p
o
is
o
n
ed
d
ataset
.
At
th
is
p
o
in
t,
th
e
o
r
ig
in
al
g
r
o
u
n
d
-
tr
u
th
l
ab
el
o
f
is
'
9
'
an
d
th
e
attac
k
er
’
s
tar
g
et
lab
el
is
s
e
t
to
'
0
'
.
T
h
er
ef
o
r
e,
w
e
m
an
ip
u
lated
th
e
lab
els
o
f
an
d
th
en
tr
ain
ed
c
o
n
s
tr
u
cted
a
b
ac
k
d
o
o
r
ed
m
o
d
el
b
y
tr
ain
i
n
g
a
C
NN
m
o
d
el
with
g
en
er
ated
at
p
o
is
o
n
in
g
r
atio
=
3
%.
I
n
ad
d
itio
n
,
we
u
s
ed
a
wh
ite
s
q
u
ar
e
b
a
ck
d
o
o
r
tr
ig
g
e
r
p
atter
n
a
d
d
ed
to
th
e
b
o
tto
m
r
i
g
h
t
co
r
n
er
o
f
im
ag
es,
an
d
th
u
s
v
ar
io
u
s
s
izes
o
f
b
ac
k
d
o
o
r
tr
ig
g
er
(1
×
1
,
2
×
2
,
3
×
3
,
4
×
4
,
8
×
8
,
an
d
1
6
×
1
6
p
x
)
was c
r
ea
ted
[
1
0
]
,
[
1
4
]
.
−
B
a
c
k
d
o
o
r
d
e
t
e
ct
i
o
n
m
e
t
h
o
d
s
:
T
o
c
o
m
p
a
r
e
t
h
e
p
e
r
f
o
r
m
a
n
c
e
o
f
o
u
r
p
r
o
p
o
s
e
d
m
e
t
h
o
d
a
n
d
a
n
e
x
i
s
ti
n
g
m
e
t
h
o
d
N
C
f
o
r
v
a
r
i
o
u
s
b
a
c
k
d
o
o
r
t
r
i
g
g
e
r
s
i
z
es
,
w
e
t
e
s
t
e
d
5
0
ti
m
e
s
f
o
r
ea
c
h
b
a
c
k
d
o
o
r
t
r
i
g
g
e
r
s
i
z
e
a
n
d
av
e
r
a
g
e
d
t
h
e
i
r
d
e
t
e
c
t
i
o
n
r
e
s
u
lt
s
.
NC
m
e
a
s
u
r
es
t
h
e
p
e
r
t
u
r
b
a
t
i
o
n
r
e
q
u
i
r
e
d
w
h
e
n
i
n
d
u
c
i
n
g
a
b
e
n
i
g
n
i
m
a
g
e
t
o
b
e
c
l
a
s
s
i
f
ie
d
i
n
t
o
a
d
i
f
f
e
r
e
n
t
la
b
e
l
a
n
d
ass
e
s
s
es
i
f
i
t
is
a
n
o
m
a
l
o
u
s
l
y
s
m
a
ll
w
h
e
n
in
d
u
c
e
d
t
o
w
a
r
d
s
t
h
e
t
a
r
g
e
t
l
a
b
el.
O
n
t
h
e
o
t
h
e
r
h
a
n
d
,
o
u
r
m
e
t
h
o
d
m
e
as
u
r
e
s
t
h
e
p
e
r
t
u
r
b
a
t
i
o
n
n
e
e
d
e
d
w
h
e
n
i
n
d
u
c
i
n
g
a
b
a
c
k
d
o
o
r
i
m
a
g
e
t
o
b
e
cl
a
s
s
i
f
i
e
d
i
n
t
o
a
d
i
f
f
e
r
e
n
t
l
a
b
e
l
a
n
d
e
v
a
l
u
at
es
i
f
i
t
is
a
n
o
m
a
l
o
u
s
l
y
s
m
a
ll
w
h
e
n
in
d
u
c
e
d
t
o
w
a
r
d
s
t
h
e
g
r
o
u
n
d
-
t
r
u
t
h
l
a
b
e
l
[
1
8
]
.
−
E
v
alu
atio
n
m
etr
ics
:
T
o
m
ea
s
u
r
e
th
e
p
er
f
o
r
m
a
n
ce
o
f
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
,
we
u
s
ed
th
e
f
o
llo
win
g
two
ev
alu
atio
n
m
etr
ics.
T
h
e
f
ir
s
t
m
etr
ic
is
AI
th
at
m
ea
s
u
r
es
th
e
d
eg
r
ee
o
f
ab
n
o
r
m
ality
in
a
b
ac
k
d
o
o
r
e
d
m
o
d
el
[
2
2
]
.
T
o
q
u
an
tif
y
th
e
d
eg
r
ee
o
f
ab
n
o
r
m
ality
,
th
e
am
o
u
n
t
o
f
p
e
r
tu
r
b
atio
n
in
s
er
ted
to
class
if
y
th
e
im
ag
e
in
to
a
d
if
f
er
e
n
t
lab
el
is
ca
lcu
lated
u
s
in
g
L
1
n
o
r
m
.
T
h
e
b
ac
k
d
o
o
r
at
tack
d
etec
tio
n
th
r
esh
o
ld
is
s
et
to
b
e
t
h
e
s
am
e
as
NC
,
with
AI
>
2
.
T
h
e
s
ec
o
n
d
m
etr
ic
is
B
DA
th
at
m
ea
s
u
r
es
th
e
p
r
o
b
ab
ilit
y
o
f
b
ac
k
d
o
o
r
d
etec
tio
n
o
n
b
ac
k
d
o
o
r
ed
m
o
d
el.
B
DA
ca
n
b
e
m
ea
s
u
r
ed
as in
(
2
)
.
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
R
elia
b
le
b
a
ck
d
o
o
r
a
tta
ck
d
ete
ctio
n
fo
r
va
r
io
u
s
s
iz
e
o
f b
a
ck
d
o
o
r
tr
ig
g
ers
(
Yeo
n
g
r
o
k
R
a
h
)
655
(
%
)
=
ℎ
ℎ
×
100
(
2
)
T
ab
le
1
.
Ar
c
h
itectu
r
e
o
f
tar
g
et
DL
m
o
d
el
La
y
e
r
I
n
p
u
t
F
i
l
t
e
r
S
t
r
i
d
e
O
u
t
p
u
t
A
c
t
i
v
a
t
i
o
n
c
o
n
v
1
1
×
28
×
28
16
×
1
×
5
×
5
1
16
×
24
×
24
R
e
LU
p
o
o
l
1
16
×
24
×
24
2
×
2
2
16
×
12
×
12
/
c
o
n
v
2
16
×
12
×
12
32
×
16
×
5
×
5
1
32
×
8
×
8
R
e
LU
p
o
o
l
2
32
×
8
×
8
2
×
2
2
32
×
4
×
4
/
f
c
1
32
×
4
×
4
/
/
5
1
2
R
e
LU
f
c
2
5
1
2
/
/
10
S
o
f
t
ma
x
4
.
2
.
E
x
perim
ent
a
l
re
s
ults a
nd
a
na
ly
s
is
First,
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
s
u
cc
ess
f
u
lly
d
etec
ted
b
ac
k
d
o
o
r
im
ag
es
with
8
×
8
o
r
1
6
×
1
6
b
ac
k
d
o
o
r
tr
ig
g
er
s
wh
o
s
e
A
I
>
2
wh
ile
an
ex
is
tin
g
s
tate
-
of
-
t
h
e
-
ar
t
b
ac
k
d
o
o
r
d
etec
tio
n
m
eth
o
d
(
NC
)
c
o
u
ld
n
o
t
d
etec
t
th
em
.
Fig
u
r
e
6
(
a
)
s
h
o
ws
th
e
p
er
tu
r
b
atio
n
v
al
u
e
in
ter
m
s
o
f
L
1
n
o
r
m
r
eq
u
ir
e
d
wh
e
n
NC
class
if
ies
a
test
im
ag
e
(
L
ab
el
9
)
in
to
a
d
if
f
er
e
n
t
lab
el
(
f
r
o
m
L
a
b
el
0
to
L
ab
el
8
e
x
c
ep
t
L
ab
le
9
)
.
Fo
r
e
x
am
p
le,
in
th
e
ca
s
e
o
f
u
s
in
g
a
4
×
4
p
x
b
ac
k
d
o
o
r
tr
ig
g
e
r
s
ize,
t
h
e
am
o
u
n
t
o
f
p
er
tu
r
b
atio
n
n
ee
d
ed
to
class
if
y
a
test
im
ag
e
in
t
o
th
e
tar
g
et
lab
el
'
0
'
is
s
ig
n
if
ican
tly
lo
wer
co
m
p
ar
ed
to
th
e
am
o
u
n
t
o
f
p
er
tu
r
b
a
tio
n
n
ee
d
ed
f
o
r
o
th
er
lab
els
(
L
ab
el
1
-
L
ab
el
8
)
.
Ho
wev
er
,
as
th
e
b
ac
k
d
o
o
r
tr
ig
g
er
s
ize
g
r
o
ws,
t
h
e
p
er
tu
r
b
atio
n
v
alu
e
b
e
co
m
es
clo
s
er
to
th
o
s
e
f
o
r
o
th
er
lab
els.
On
th
e
o
th
e
r
h
an
d
,
Fig
u
r
e
6
(
b
)
s
h
o
w
th
e
p
er
tu
r
b
atio
n
v
al
u
es
m
ea
s
u
r
ed
wh
e
n
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
was
u
s
ed
.
As
we
ca
n
s
ee
in
th
e
f
ig
u
r
es,
o
u
r
m
et
h
o
d
n
ee
d
s
r
elativ
ely
v
er
y
s
m
all
p
er
t
u
r
b
atio
n
wh
en
c
lass
if
y
in
g
b
ac
k
d
o
o
r
im
ag
es
(
L
ab
el
0
)
in
to
th
e
g
r
o
u
n
d
-
tr
u
th
lab
e
l
(
L
a
b
el
9
)
r
eg
ar
d
less
o
f
b
ac
k
d
o
o
r
tr
i
g
g
e
r
s
izes
.
Similar
ly
,
Fig
u
r
e
7
(
a)
r
ep
r
esen
ts
th
e
ca
lc
u
lated
AI
v
alu
es
f
o
r
p
er
t
u
r
b
ati
o
n
u
s
in
g
NC
tech
n
iq
u
e.
W
h
en
th
e
b
ac
k
d
o
o
r
tr
i
g
g
er
s
izes a
r
e
eq
u
al
to
o
r
g
r
ea
ter
th
an
8
×
8
p
x
,
th
e
AI
v
alu
es we
r
e
lo
wer
th
an
th
e
b
ac
k
d
o
o
r
d
ete
ctio
n
th
r
esh
o
ld
o
f
2
.
I
n
o
th
er
wo
r
d
s
,
it
is
n
o
t
p
o
s
s
ib
le
to
d
etec
t
th
e
b
ac
k
d
o
o
r
attac
k
wh
en
th
e
b
ac
k
d
o
o
r
tr
ig
g
er
s
ize
is
8
×
8
o
r
lar
g
er
[
9
]
.
On
th
e
o
th
er
h
an
d
,
F
ig
u
r
e
7
(
b
)
s
h
o
w
th
e
AI
v
alu
es
m
ea
s
u
r
ed
wh
en
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
was
u
s
ed
.
C
o
n
s
eq
u
en
tly
,
th
e
AI
v
alu
es
also
ex
ce
ed
ed
th
e
b
ac
k
d
o
o
r
d
etec
tio
n
th
r
esh
o
ld
o
f
2
a
n
d
th
u
s
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
s
u
cc
ess
f
u
lly
d
etec
ted
b
ac
k
d
o
o
r
im
ag
es with
v
ar
io
u
s
s
ize
o
f
b
ac
k
d
o
o
r
tr
ig
g
e
r
s
.
(a
)
(b
)
Fig
u
r
e
6
.
C
o
m
p
a
r
is
o
n
o
f
L
1
n
o
r
m
wh
en
test
im
ag
es a
r
e
class
if
ied
in
to
d
if
f
er
e
n
t la
b
els u
s
in
g
(
a)
NC
an
d
(
b
)
o
u
r
m
eth
o
d
(
a)
(
b
)
F
i
g
u
r
e
7
.
C
o
m
p
a
r
i
s
o
n
o
f
A
I
w
h
e
n
t
e
s
t
i
m
a
g
e
s
a
r
e
c
l
a
s
s
i
f
i
e
d
i
n
t
o
d
i
f
f
e
r
e
n
t
l
a
b
e
l
s
u
s
i
n
g
(
a
)
N
C
a
n
d
(
b
)
o
u
r
m
e
t
h
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
4
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
5
:
650
-
6
5
7
656
Seco
n
d
,
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
s
h
o
wed
a
h
ig
h
,
s
tab
le
d
etec
tio
n
p
er
f
o
r
m
an
ce
f
o
r
all
s
izes
o
f
b
ac
k
d
o
o
r
tr
ig
g
er
s
.
Sp
ec
if
ically
,
as
in
d
icate
d
in
T
ab
le
2
,
th
e
av
er
ag
e
B
DA
o
f
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
is
9
6
.
3
%
wh
ile
th
e
N
C
s
h
o
wed
ar
o
u
n
d
6
4
.
3
%
o
f
B
DA
o
n
av
er
a
g
e.
T
h
is
m
ea
n
s
t
h
at
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
o
u
tp
er
f
o
r
m
s
t
h
e
e
x
is
tin
g
s
tate
-
of
-
th
e
-
ar
t
d
etec
tio
n
m
et
h
o
d
NC
b
y
ar
o
u
n
d
3
2
%p
in
ter
m
s
o
f
b
ac
k
d
o
o
r
attac
k
d
etec
tio
n
ac
cu
r
ac
y
.
I
n
p
ar
ticu
lar
,
B
DA
o
f
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
f
o
r
8
×
8
o
r
1
6
×
1
6
b
ac
k
d
o
o
r
tr
ig
g
er
s
izes
ar
e
9
8
%
an
d
9
4
%,
r
esp
e
ctiv
ely
,
wh
ile
th
e
NC
co
u
ld
n
o
t
d
etec
t
th
em
at
all
(
B
D
A
=
0
%).
T
h
is
r
esu
lt
co
n
f
ir
m
s
th
at
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
ca
n
r
eso
lv
e
th
e
cr
itica
l
v
u
ln
er
a
b
ilit
y
o
f
NC
an
d
th
u
s
s
h
o
ws
s
tab
le,
r
eliab
le
d
etec
ti
o
n
p
e
r
f
o
r
m
an
ce
f
o
r
v
ar
io
u
s
b
ac
k
d
o
o
r
tr
ig
g
er
s
izes.
T
ab
le
2
.
C
o
m
p
a
r
is
o
n
o
f
B
DA
f
o
r
v
a
r
io
u
s
b
ac
k
d
o
o
r
tr
ig
g
er
s
izes
Tr
i
g
g
e
r
S
i
z
e
B
D
A
(
%) (#
o
f
d
e
t
e
c
t
e
d
b
a
c
k
d
o
o
r
s
a
mp
l
e
s
/
#
o
f
t
e
st
b
a
c
k
d
o
o
r
sam
p
l
e
s
)
N
e
u
r
a
l
c
l
e
a
n
s
e
O
u
r
p
r
o
p
o
s
e
d
me
t
h
o
d
1
×
1
9
6
(
4
8
/
5
0
)
9
6
(
4
8
/
5
0
)
2
×
2
9
8
(
4
9
/
5
0
)
9
8
(
4
9
/
5
0
)
3
×
3
9
8
(
4
9
/
5
0
)
9
6
(
4
8
/
5
0
)
4
×
4
9
4
(
4
7
/
5
0
)
9
6
(
4
8
/
5
0
)
8
×
8
0
(
0
/
5
0
)
9
8
(
4
9
/
5
0
)
16
×
16
0
(
0
/
5
0
)
9
4
(
4
7
/
5
0
)
A
v
e
r
a
g
e
6
4
.
3
9
6
.
3
5.
CO
NCLU
SI
O
N
I
n
th
is
s
tu
d
y
,
we
in
tr
o
d
u
ce
d
a
n
o
v
el
ap
p
r
o
ac
h
to
d
etec
t
b
ac
k
d
o
o
r
attac
k
s
i
n
DL
m
o
d
els
b
y
u
tili
zin
g
p
er
tu
r
b
atio
n
in
s
er
tio
n
to
id
e
n
tify
ab
n
o
r
m
al
p
er
tu
r
b
atio
n
with
in
th
e
d
ata.
T
h
e
p
r
im
ar
y
o
b
jec
tiv
e
o
f
th
is
m
eth
o
d
was to
en
h
an
ce
th
e
ca
p
ab
ilit
ies o
f
b
ac
k
d
o
o
r
attac
k
d
etec
tio
n
.
B
y
im
p
lem
en
tin
g
th
is
tech
n
i
q
u
e,
o
u
r
g
o
al
was to
o
v
er
co
m
e
t
h
e
l
im
itatio
n
s
o
b
s
er
v
ed
in
p
r
i
o
r
r
esear
ch
an
d
p
r
o
v
id
e
a
r
o
b
u
s
t
m
eth
o
d
ca
p
ab
le
o
f
ef
f
ec
tiv
ely
id
en
tify
in
g
b
ac
k
d
o
o
r
attac
k
s
,
r
eg
ar
d
less
o
f
th
e
s
ize
o
f
th
e
i
n
s
er
ted
b
ac
k
d
o
o
r
tr
ig
g
e
r
.
T
h
e
v
alid
atio
n
o
f
th
is
ap
p
r
o
ac
h
was c
o
n
d
u
cted
th
r
o
u
g
h
a
s
er
ies o
f
p
r
elim
i
n
ar
y
ex
p
er
im
en
ts
,
co
n
f
ir
m
in
g
its
ef
f
ica
cy
an
d
p
o
ten
tial f
o
r
p
r
ac
tical
ap
p
licatio
n
.
Ou
r
f
u
tu
r
e
r
esear
ch
d
ir
ec
tio
n
s
ar
e
as
f
o
llo
ws.
First,
we
aim
to
d
elv
e
d
ee
p
er
in
to
r
e
f
in
in
g
an
d
o
p
tim
izin
g
th
e
p
r
o
p
o
s
ed
tech
n
iq
u
e,
s
tr
iv
in
g
to
im
p
r
o
v
e
its
p
r
ec
is
io
n
,
s
ca
lab
ilit
y
,
a
n
d
a
p
p
licab
ilit
y
ac
r
o
s
s
v
ar
io
u
s
m
o
d
els
an
d
d
atasets
.
Seco
n
d
,
we
p
la
n
to
ex
p
l
o
r
e
ad
d
itio
n
al
d
ef
en
s
iv
e
s
tr
ateg
ie
s
ag
ain
s
t
b
ac
k
d
o
o
r
attac
k
s
,
in
clu
d
in
g
in
v
esti
g
atin
g
m
eth
o
d
s
to
ef
f
ec
tiv
el
y
r
e
m
o
v
e
d
etec
ted
b
ac
k
d
o
o
r
tr
ig
g
er
s
o
r
d
is
r
u
p
t
th
e
co
n
n
ec
tio
n
b
etwe
en
tr
ig
g
er
s
a
n
d
b
ac
k
d
o
o
r
s
.
T
h
ese
s
tr
ateg
ic
m
ea
s
u
r
es
ar
e
d
esig
n
ed
to
f
o
r
tify
th
e
r
esil
ien
ce
o
f
DL
m
o
d
els ag
ain
s
t p
o
ten
tial b
ac
k
d
o
o
r
th
r
ea
ts
in
r
ea
l
-
w
o
r
ld
s
ce
n
ar
io
s
.
RE
F
E
R
E
NC
E
S
[
1
]
D
.
J.
M
i
l
l
e
r
,
Z.
X
i
a
n
g
,
a
n
d
G
.
K
e
s
i
d
i
s
,
“
A
d
v
e
r
s
a
r
i
a
l
l
e
a
r
n
i
n
g
t
a
r
g
e
t
i
n
g
d
e
e
p
n
e
u
r
a
l
n
e
t
w
o
r
k
c
l
a
ssi
f
i
c
a
t
i
o
n
:
a
c
o
mp
r
e
h
e
n
s
i
v
e
r
e
v
i
e
w
o
f
d
e
f
e
n
ses
a
g
a
i
n
s
t
a
t
t
a
c
k
s,
”
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
I
EE
E
,
v
o
l
.
1
0
8
,
n
o
.
3
,
p
p
.
4
0
2
–
4
3
3
,
2
0
2
0
,
d
o
i
:
1
0
.
1
1
0
9
/
JP
R
O
C
.
2
0
2
0
.
2
9
7
0
6
1
5
.
[
2
]
A
.
I
l
y
a
s,
S
.
S
a
n
t
u
r
k
a
r
,
D
.
Ts
i
p
r
a
s,
L
.
En
g
st
r
o
m,
B
.
Tr
a
n
,
a
n
d
A
.
M
a
d
r
y
,
“
A
d
v
e
r
sari
a
l
e
x
a
mp
l
e
s
a
r
e
n
o
t
b
u
g
s,
t
h
e
y
a
r
e
f
e
a
t
u
r
e
s
,
”
Ad
v
a
n
c
e
s i
n
N
e
u
r
a
l
I
n
f
o
rm
a
t
i
o
n
Pro
c
e
ssi
n
g
S
y
st
e
m
s
,
v
o
l
.
3
2
,
2
0
1
9
.
[
3
]
K
.
R
e
n
,
T.
Zh
e
n
g
,
Z.
Q
i
n
,
a
n
d
X
.
L
i
u
,
“
A
d
v
e
r
sar
i
a
l
a
t
t
a
c
k
s
a
n
d
d
e
f
e
n
ses
i
n
d
e
e
p
l
e
a
r
n
i
n
g
,
”
E
n
g
i
n
e
e
ri
n
g
,
v
o
l
.
6
,
n
o
.
3
,
p
p
.
3
4
6
–
3
6
0
,
2
0
2
0
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
e
n
g
.
2
0
1
9
.
1
2
.
0
1
2
.
[
4
]
A
.
C
h
a
k
r
a
b
o
r
t
y
,
M
.
A
l
a
m
,
V
.
D
e
y
,
A
.
C
h
a
t
t
o
p
a
d
h
y
a
y
,
a
n
d
D
.
M
u
k
h
o
p
a
d
h
y
a
y
,
“
A
s
u
r
v
e
y
o
n
a
d
v
e
r
sar
i
a
l
a
t
t
a
c
k
s
a
n
d
d
e
f
e
n
c
e
s,
”
C
AAI Tr
a
n
sa
c
t
i
o
n
s
o
n
I
n
t
e
l
l
i
g
e
n
c
e
T
e
c
h
n
o
l
o
g
y
,
v
o
l
.
6
,
n
o
.
1
,
p
p
.
2
5
–
4
5
,
2
0
2
1
,
d
o
i
:
1
0
.
1
0
4
9
/
c
i
t
2
.
1
2
0
2
8
.
[
5
]
S
.
A
n
e
j
a
,
N
.
A
n
e
j
a
,
P
.
E.
A
b
a
s,
a
n
d
A
.
G
.
N
a
i
m,
“
D
e
f
e
n
s
e
a
g
a
i
n
s
t
a
d
v
e
r
s
a
r
i
a
l
a
t
t
a
c
k
s
o
n
d
e
e
p
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
s
t
h
r
o
u
g
h
n
o
n
l
o
c
a
l
d
e
n
o
i
si
n
g
,
”
I
AE
S
I
n
t
e
r
n
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
f
Ar
t
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
,
v
o
l
.
1
1
,
n
o
.
1
,
p
p
.
1
–
1
4
,
Ju
n
.
2
0
2
2
,
d
o
i
:
1
0
.
1
1
5
9
1
/
i
j
a
i
.
v
1
1
.
i
3
.
p
p
9
6
1
-
9
6
8
.
[
6
]
I
.
J.
G
o
o
d
f
e
l
l
o
w
,
J.
S
h
l
e
n
s
,
a
n
d
C
.
S
z
e
g
e
d
y
,
“
E
x
p
l
a
i
n
i
n
g
a
n
d
h
a
r
n
e
ss
i
n
g
a
d
v
e
r
sar
i
a
l
e
x
a
m
p
l
e
s,
”
3
r
d
I
n
t
e
rn
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
L
e
a
r
n
i
n
g
R
e
p
r
e
se
n
t
a
t
i
o
n
s
,
I
C
L
R
2
0
1
5
-
C
o
n
f
e
r
e
n
c
e
T
r
a
c
k
Pro
c
e
e
d
i
n
g
s
,
2
0
1
5
.
[
7
]
Y
.
L
i
,
S
.
Z
h
a
n
g
,
W
.
W
a
n
g
,
a
n
d
H
.
S
o
n
g
,
“
B
a
c
k
d
o
o
r
a
t
t
a
c
k
s
t
o
d
e
e
p
l
e
a
r
n
i
n
g
mo
d
e
l
s
a
n
d
c
o
u
n
t
e
r
m
e
a
s
u
r
e
s
:
a
su
r
v
e
y
,
”
I
E
EE
O
p
e
n
J
o
u
rn
a
l
o
f
t
h
e
C
o
m
p
u
t
e
r
S
o
c
i
e
t
y
,
v
o
l
.
4
,
p
p
.
1
3
4
–
1
4
6
,
2
0
2
3
,
d
o
i
:
1
0
.
1
1
0
9
/
O
JC
S
.
2
0
2
3
.
3
2
6
7
2
2
1
.
[
8
]
A
.
S
h
a
f
a
h
i
e
t
a
l
.
,
“
P
o
i
so
n
f
r
o
g
s!
T
a
r
g
e
t
e
d
c
l
e
a
n
-
l
a
b
e
l
p
o
i
s
o
n
i
n
g
a
t
t
a
c
k
s
o
n
n
e
u
r
a
l
n
e
t
w
o
r
k
s,”
Ad
v
a
n
c
e
s
i
n
N
e
u
ra
l
I
n
f
o
rm
a
t
i
o
n
Pro
c
e
ssi
n
g
S
y
s
t
e
m
s
,
v
o
l
.
2
0
1
8
,
p
p
.
6
1
0
3
–
6
1
1
3
,
2
0
1
8
.
[
9
]
A
.
Tu
r
n
e
r
,
D
.
T
si
p
r
a
s,
a
n
d
A
.
M
a
d
r
y
,
“
C
l
e
a
n
-
l
a
b
e
l
b
a
c
k
d
o
o
r
a
t
t
a
c
k
s,
”
T
h
e
I
n
t
e
rn
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
L
e
a
r
n
i
n
g
Re
p
r
e
se
n
t
a
t
i
o
n
s
,
p
p
.
1
-
2
1
,
2
0
1
9
.
[
1
0
]
T.
G
u
,
K
.
L
i
u
,
B
.
D
o
l
a
n
-
G
a
v
i
t
t
,
a
n
d
S
.
G
a
r
g
,
“
B
a
d
N
e
t
s
:
e
v
a
l
u
a
t
i
n
g
b
a
c
k
d
o
o
r
i
n
g
a
t
t
a
c
k
s
o
n
d
e
e
p
n
e
u
r
a
l
n
e
t
w
o
r
k
s,”
I
EEE
A
c
c
e
ss
,
v
o
l
.
7
,
p
p
.
4
7
2
3
0
–
4
7
2
4
3
,
2
0
1
9
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
1
9
.
2
9
0
9
0
6
8
.
[
1
1
]
Y
.
L
i
,
Y
.
Ji
a
n
g
,
Z
.
Li
,
a
n
d
S
.
T.
X
i
a
,
“
B
a
c
k
d
o
o
r
l
e
a
r
n
i
n
g
:
a
su
r
v
e
y
,
”
I
E
EE
T
ra
n
s
a
c
t
i
o
n
s
o
n
N
e
u
r
a
l
N
e
t
w
o
rks
a
n
d
L
e
a
r
n
i
n
g
S
y
s
t
e
m
s
,
v
o
l
.
3
5
,
n
o
.
1
,
p
p
.
5
–
2
2
,
2
0
2
4
,
d
o
i
:
1
0
.
1
1
0
9
/
TN
N
LS
.
2
0
2
2
.
3
1
8
2
9
7
9
.
[
1
2
]
E.
W
e
n
g
e
r
,
J.
P
a
ssa
n
a
n
t
i
,
A
.
N
.
B
h
a
g
o
j
i
,
Y
.
Y
a
o
,
H
.
Zh
e
n
g
,
a
n
d
B
.
Y
.
Z
h
a
o
,
“
B
a
c
k
d
o
o
r
a
t
t
a
c
k
s
a
g
a
i
n
s
t
d
e
e
p
l
e
a
r
n
i
n
g
sy
s
t
e
m
s
i
n
t
h
e
p
h
y
s
i
c
a
l
w
o
r
l
d
,
”
T
h
e
I
EEE
C
o
m
p
u
t
e
r S
o
c
i
e
t
y
C
o
n
f
e
r
e
n
c
e
o
n
C
o
m
p
u
t
e
r
V
i
si
o
n
a
n
d
P
a
t
t
e
r
n
Re
c
o
g
n
i
t
i
o
n
,
p
p
.
6
2
0
2
–
6
2
1
1
,
2
0
2
1
,
d
o
i
:
1
0
.
1
1
0
9
/
C
V
P
R
4
6
4
3
7
.
2
0
2
1
.
0
0
6
1
4
.
[
1
3
]
Y
.
Li
u
e
t
a
l
.
,
“
A
su
r
v
e
y
o
n
n
e
u
r
a
l
t
r
o
j
a
n
s,”
2
0
2
0
2
1
st
I
n
t
e
r
n
a
t
i
o
n
a
l
S
y
m
p
o
s
i
u
m
o
n
Q
u
a
l
i
t
y
E
l
e
c
t
r
o
n
i
c
D
e
si
g
n
(
I
S
Q
E
D
)
,
S
a
n
t
a
C
l
a
r
a
,
C
A
,
U
S
A
,
2
0
2
0
,
p
p
.
3
3
–
3
9
,
d
o
i
:
1
0
.
1
1
0
9
/
I
S
Q
ED
4
8
8
2
8
.
2
0
2
0
.
9
1
3
7
0
1
1
.
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
R
elia
b
le
b
a
ck
d
o
o
r
a
tta
ck
d
ete
ctio
n
fo
r
va
r
io
u
s
s
iz
e
o
f b
a
ck
d
o
o
r
tr
ig
g
ers
(
Yeo
n
g
r
o
k
R
a
h
)
657
[
1
4
]
B
.
Z
h
a
o
a
n
d
Y
.
La
o
,
“
T
o
w
a
r
d
s
c
l
a
s
s
-
o
r
i
e
n
t
e
d
p
o
i
so
n
i
n
g
a
t
t
a
c
k
s
a
g
a
i
n
st
n
e
u
r
a
l
n
e
t
w
o
r
k
s
,
”
2
0
2
2
I
EE
E/
C
VF
Wi
n
t
e
r
C
o
n
f
e
re
n
c
e
o
n
Ap
p
l
i
c
a
t
i
o
n
s
o
f
C
o
m
p
u
t
e
r
Vi
s
i
o
n
,
WA
C
V
2
0
2
2
,
p
p
.
2
2
4
4
–
2
2
5
3
,
2
0
2
2
,
d
o
i
:
1
0
.
1
1
0
9
/
W
A
C
V
5
1
4
5
8
.
2
0
2
2
.
0
0
2
3
0
.
[
1
5
]
J.
Li
n
,
L.
X
u
,
Y
.
L
i
u
,
a
n
d
X
.
Zh
a
n
g
,
“
C
o
m
p
o
s
i
t
e
b
a
c
k
d
o
o
r
a
t
t
a
c
k
f
o
r
d
e
e
p
n
e
u
r
a
l
n
e
t
w
o
r
k
b
y
m
i
x
i
n
g
e
x
i
st
i
n
g
b
e
n
i
g
n
f
e
a
t
u
r
e
s,
”
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
AC
M
C
o
n
f
e
re
n
c
e
o
n
C
o
m
p
u
t
e
r
a
n
d
C
o
m
m
u
n
i
c
a
t
i
o
n
s
S
e
c
u
ri
t
y
,
p
p
.
1
1
3
–
1
3
1
,
2
0
2
0
,
d
o
i
:
1
0
.
1
1
4
5
/
3
3
7
2
2
9
7
.
3
4
2
3
3
6
2
.
[
1
6
]
H
.
P
a
r
k
a
n
d
Y
.
C
h
o
,
“
A
d
i
l
u
t
i
o
n
-
b
a
se
d
d
e
f
e
n
s
e
m
e
t
h
o
d
a
g
a
i
n
s
t
p
o
i
s
o
n
i
n
g
a
t
t
a
c
k
s
o
n
d
e
e
p
l
e
a
r
n
i
n
g
s
y
st
e
ms,”
I
n
t
e
r
n
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
f
El
e
c
t
ri
c
a
l
a
n
d
C
o
m
p
u
t
e
r E
n
g
i
n
e
e
ri
n
g
,
v
o
l
.
1
4
,
n
o
.
1
,
p
p
.
6
4
5
–
6
5
2
,
2
0
2
4
,
d
o
i
:
1
0
.
1
1
5
9
1
/
i
j
e
c
e
.
v
1
4
i
1
.
p
p
6
4
5
-
6
5
2
.
[
1
7
]
S
.
H
u
a
n
g
,
W
.
P
e
n
g
,
Z.
J
i
a
,
a
n
d
Z.
Tu
,
“
O
n
e
-
p
i
x
e
l
si
g
n
a
t
u
r
e
:
c
h
a
r
a
c
t
e
r
i
z
i
n
g
C
N
N
mo
d
e
l
s
f
o
r
b
a
c
k
d
o
o
r
d
e
t
e
c
t
i
o
n
,
”
C
o
m
p
u
t
e
r
V
i
si
o
n
–
E
C
C
V
2
0
2
0
,
p
p
.
3
2
6
–
3
4
1
,
2
0
2
0
,
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
3
-
0
3
0
-
5
8
5
8
3
-
9
_
2
0
.
[
1
8
]
B
.
W
a
n
g
e
t
a
l
.
,
“
N
e
u
r
a
l
c
l
e
a
n
se
:
I
d
e
n
t
i
f
y
i
n
g
a
n
d
mi
t
i
g
a
t
i
n
g
b
a
c
k
d
o
o
r
a
t
t
a
c
k
s
i
n
n
e
u
r
a
l
n
e
t
w
o
r
k
s,”
2
0
1
9
I
EEE
S
y
m
p
o
si
u
m
o
n
S
e
c
u
ri
t
y
a
n
d
Pri
v
a
c
y
(
S
P)
,
S
a
n
F
r
a
n
c
i
sc
o
,
C
A
,
U
S
A
,
2
0
1
9
,
p
p
.
7
0
7
-
7
2
3
,
d
o
i
:
1
0
.
1
1
0
9
/
S
P
.
2
0
1
9
.
0
0
0
3
1
.
[
1
9
]
X
.
S
u
n
e
t
a
l
.
,
“
D
e
f
e
n
d
i
n
g
a
g
a
i
n
s
t
b
a
c
k
d
o
o
r
a
t
t
a
c
k
s
i
n
n
a
t
u
r
a
l
l
a
n
g
u
a
g
e
g
e
n
e
r
a
t
i
o
n
,
”
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
3
7
t
h
AA
AI
C
o
n
f
e
re
n
c
e
o
n
Art
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
,
AAAI
2
0
2
3
,
v
o
l
.
3
7
,
p
p
.
5
2
5
7
–
5
2
6
5
,
2
0
2
3
,
d
o
i
:
1
0
.
1
6
0
9
/
a
a
a
i
.
v
3
7
i
4
.
2
5
6
5
6
.
[
2
0
]
Z.
X
i
a
n
g
,
D
.
J.
M
i
l
l
e
r
,
a
n
d
G
.
K
e
s
i
d
i
s,
“
D
e
t
e
c
t
i
o
n
o
f
b
a
c
k
d
o
o
r
s
i
n
t
r
a
i
n
e
d
c
l
a
ss
i
f
i
e
r
s
w
i
t
h
o
u
t
a
c
c
e
ss
t
o
t
h
e
t
r
a
i
n
i
n
g
s
e
t
,
”
I
EE
E
T
ra
n
s
a
c
t
i
o
n
s
o
n
N
e
u
r
a
l
N
e
t
w
o
r
k
s
a
n
d
L
e
a
r
n
i
n
g
S
y
s
t
e
m
s
,
v
o
l
.
3
3
,
n
o
.
3
,
p
p
.
1
1
7
7
–
1
1
9
1
,
2
0
2
2
,
d
o
i
:
1
0
.
1
1
0
9
/
TN
N
LS.2
0
2
0
.
3
0
4
1
2
0
2
.
[
2
1
]
J.
G
u
o
,
A
.
L
i
,
a
n
d
C
.
L
i
u
,
“
A
e
v
a
:
B
l
a
c
k
-
b
o
x
b
a
c
k
d
o
o
r
d
e
t
e
c
t
i
o
n
u
s
i
n
g
a
d
v
e
r
sari
a
l
e
x
t
r
e
m
e
v
a
l
u
e
a
n
a
l
y
si
s
,
”
I
C
L
R
2
0
2
2
-
1
0
t
h
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
L
e
a
r
n
i
n
g
Re
p
res
e
n
t
a
t
i
o
n
s
,
p
p
.
1
-
2
4
,
2
0
2
2
.
[
2
2
]
F
.
R
.
H
a
m
p
e
l
,
“
T
h
e
i
n
f
l
u
e
n
c
e
c
u
r
v
e
a
n
d
i
t
s
r
o
l
e
i
n
r
o
b
u
st
e
st
i
ma
t
i
o
n
,
”
J
o
u
r
n
a
l
o
f
t
h
e
Am
e
r
i
c
a
n
S
t
a
t
i
s
t
i
c
a
l
Ass
o
c
i
a
t
i
o
n
,
v
o
l
.
6
9
,
n
o
.
3
4
6
,
1
9
7
4
,
d
o
i
:
1
0
.
2
3
0
7
/
2
2
8
5
6
6
6
.
[
2
3
]
W
.
G
u
o
,
B
.
T
o
n
d
i
,
a
n
d
M
.
B
a
r
n
i
,
“
A
n
o
v
e
r
v
i
e
w
o
f
b
a
c
k
d
o
o
r
a
t
t
a
c
k
s
a
g
a
i
n
st
d
e
e
p
n
e
u
r
a
l
n
e
t
w
o
r
k
s
a
n
d
p
o
ssi
b
l
e
d
e
f
e
n
c
e
s,
”
I
E
EE
O
p
e
n
J
o
u
rn
a
l
o
f
S
i
g
n
a
l
Pro
c
e
ss
i
n
g
,
v
o
l
.
3
,
p
p
.
2
6
1
–
2
8
7
,
2
0
2
2
,
d
o
i
:
1
0
.
1
1
0
9
/
O
JS
P
.
2
0
2
2
.
3
1
9
0
2
1
3
.
[
2
4
]
B
.
G
.
D
o
a
n
,
E
.
A
b
b
a
s
n
e
j
a
d
,
a
n
d
D
.
C
.
R
a
n
a
si
n
g
h
e
,
“
F
e
b
r
u
u
s
:
i
n
p
u
t
p
u
r
i
f
i
c
a
t
i
o
n
d
e
f
e
n
s
e
a
g
a
i
n
s
t
t
r
o
j
a
n
a
t
t
a
c
k
s
o
n
d
e
e
p
n
e
u
r
a
l
n
e
t
w
o
r
k
s
y
st
e
ms,”
A
C
M
I
n
t
e
rn
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
Pro
c
e
e
d
i
n
g
S
e
r
i
e
s
,
p
p
.
8
9
7
–
9
1
2
,
2
0
2
0
,
d
o
i
:
1
0
.
1
1
4
5
/
3
4
2
7
2
2
8
.
3
4
2
7
2
6
4
.
[
2
5
]
D
.
W
u
a
n
d
Y
.
W
a
n
g
,
“
A
d
v
e
r
sari
a
l
n
e
u
r
o
n
p
r
u
n
i
n
g
p
u
r
i
f
i
e
s
b
a
c
k
d
o
o
r
e
d
d
e
e
p
m
o
d
e
l
s
,
”
Ad
v
a
n
c
e
s
i
n
N
e
u
ra
l
I
n
f
o
r
m
a
t
i
o
n
Pr
o
c
e
ss
i
n
g
S
y
s
t
e
m
s
,
v
o
l
.
2
0
,
p
p
.
1
6
9
1
3
–
1
6
9
2
5
,
2
0
2
1
.
[
2
6
]
Y
.
Ze
n
g
,
S
.
C
h
e
n
,
W
.
P
a
r
k
,
Z
.
M
.
M
a
o
,
M
.
J
i
n
,
a
n
d
R
.
Ji
a
,
“
A
d
v
e
r
sari
a
l
u
n
l
e
a
r
n
i
n
g
o
f
b
a
c
k
d
o
o
r
s
v
i
a
i
mp
l
i
c
i
t
h
y
p
e
r
g
r
a
d
i
e
n
t
,
”
I
C
L
R
2
0
2
2
-
1
0
t
h
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
L
e
a
rn
i
n
g
Re
p
res
e
n
t
a
t
i
o
n
s
,
p
p
.
1
-
2
8
,
2
0
2
2
.
[
2
7
]
M
.
Ja
v
a
h
e
r
i
p
i
,
M
.
S
a
mr
a
g
h
,
G
.
F
i
e
l
d
s,
T
.
J
a
v
i
d
i
,
a
n
d
F
.
K
o
u
s
h
a
n
f
a
r
,
“
C
l
e
a
N
N
:
A
c
c
e
l
e
r
a
t
e
d
t
r
o
j
a
n
sh
i
e
l
d
f
o
r
e
mb
e
d
d
e
d
n
e
u
r
a
l
n
e
t
w
o
r
k
s,”
2
0
2
0
I
EEE
/
A
C
M
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
O
n
C
o
m
p
u
t
e
r
Ai
d
e
d
D
e
si
g
n
(
I
C
C
AD
)
,
S
a
n
D
i
e
g
o
,
C
A
,
U
S
A
,
2
0
2
0
,
p
p
.
1
-
9
.
[
2
8
]
Y
.
G
a
o
,
C
.
X
u
,
D
.
W
a
n
g
,
S
.
C
h
e
n
,
D
.
C
.
R
a
n
a
si
n
g
h
e
,
a
n
d
S
.
N
e
p
a
l
,
“
S
t
r
i
P
:
A
d
e
f
e
n
c
e
a
g
a
i
n
s
t
t
r
o
j
a
n
a
t
t
a
c
k
s
o
n
d
e
e
p
n
e
u
r
a
l
n
e
t
w
o
r
k
s,”
AC
M
I
n
t
e
rn
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
P
ro
c
e
e
d
i
n
g
S
e
r
i
e
s
,
p
p
.
1
1
3
–
1
2
5
,
2
0
1
9
,
d
o
i
:
1
0
.
1
1
4
5
/
3
3
5
9
7
8
9
.
3
3
5
9
7
9
0
.
[
2
9
]
Y
.
S
u
n
,
X
.
W
a
n
g
,
a
n
d
X
.
Ta
n
g
,
“
D
e
e
p
l
e
a
r
n
i
n
g
f
a
c
e
r
e
p
r
e
s
e
n
t
a
t
i
o
n
f
r
o
m
p
r
e
d
i
c
t
i
n
g
1
0
,
0
0
0
c
l
a
ss
e
s,”
2
0
1
4
I
EE
E
C
o
n
f
e
re
n
c
e
o
n
C
o
m
p
u
t
e
r
Vi
s
i
o
n
a
n
d
P
a
t
t
e
r
n
Re
c
o
g
n
i
t
i
o
n
,
C
o
l
u
m
b
u
s,
O
H
,
U
S
A
,
2
0
1
4
,
p
p
.
1
8
9
1
-
1
8
9
8
,
d
o
i
:
1
0
.
1
1
0
9
/
C
V
P
R
.
2
0
1
4
.
2
4
4
.
[
3
0
]
J.
S
t
a
l
l
k
a
m
p
,
M
.
S
c
h
l
i
p
si
n
g
,
J.
S
a
l
me
n
,
a
n
d
C
.
I
g
e
l
,
“
M
a
n
v
s.
c
o
m
p
u
t
e
r
:
B
e
n
c
h
m
a
r
k
i
n
g
ma
c
h
i
n
e
l
e
a
r
n
i
n
g
a
l
g
o
r
i
t
h
ms
f
o
r
t
r
a
f
f
i
c
s
i
g
n
r
e
c
o
g
n
i
t
i
o
n
,
”
N
e
u
r
a
l
N
e
t
w
o
r
k
s
,
v
o
l
.
3
2
,
p
p
.
3
2
3
–
3
3
2
,
2
0
1
2
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
n
e
u
n
e
t
.
2
0
1
2
.
0
2
.
0
1
6
.
B
I
O
G
RAP
H
I
E
S O
F
AUTH
O
RS
Ye
o
n
g
r
o
k
Ra
h
re
c
e
iv
e
d
B.
S
.
d
e
g
re
e
in
in
tern
a
ti
o
n
a
l
re
latio
n
s
f
ro
m
Ko
re
a
Na
v
a
l
Ac
a
d
e
m
y
,
Jin
h
a
e
,
Re
p
u
b
li
c
o
f
K
o
re
a
.
He
is
c
u
rre
n
t
ly
a
L
ieu
ten
a
n
t
in
Re
p
u
b
l
ic
o
f
Ko
re
a
Na
v
y
a
n
d
p
u
rs
u
in
g
th
e
M
.
S
.
d
e
g
re
e
i
n
De
p
a
rtme
n
t
o
f
Cy
b
e
r
S
e
c
u
ri
ty
a
n
d
C
o
m
p
u
ter
En
g
i
n
e
e
rin
g
with
Ko
re
a
Na
ti
o
n
a
l
De
fe
n
se
Un
iv
e
rsity
,
No
n
sa
n
,
Re
p
u
b
li
c
o
f
Ko
re
a
.
H
is
re
se
a
r
c
h
in
tere
sts
in
c
lu
d
e
d
e
e
p
lea
rn
in
g
,
a
d
v
e
rsa
rial
m
a
c
h
i
n
e
lea
rn
in
g
,
a
n
d
c
y
b
e
rwa
rfa
re
.
He
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
sk
d
u
d
fh
r
7
8
9
@g
m
a
il
.
c
o
m
.
Yo
u
n
g
h
o
Cho
re
c
e
iv
e
d
t
h
e
B.
S
.
d
e
g
re
e
in
in
d
u
str
ial
e
n
g
in
e
e
rin
g
fr
o
m
Ko
re
a
Air
F
o
rc
e
Ac
a
d
e
m
y
,
Re
p
u
b
l
ic
o
f
K
o
re
a
,
in
1
9
9
8
a
n
d
th
e
M
.
S
.
d
e
g
r
e
e
in
c
o
m
p
u
ter
sc
ien
c
e
a
n
d
in
d
u
strial
s
y
ste
m
s
e
n
g
i
n
e
e
rin
g
fr
o
m
Yo
n
se
i
Un
iv
e
rsity
,
Re
p
u
b
li
c
o
f
Ko
re
a
,
in
2
0
0
6
a
n
d
t
h
e
P
h
.
D.
d
e
g
re
e
i
n
e
lec
tri
c
a
l
a
n
d
c
o
m
p
u
ter
e
n
g
in
e
e
ri
n
g
fr
o
m
Un
i
v
e
rsity
o
f
M
a
r
y
lan
d
,
Co
l
leg
e
P
a
rk
,
M
D,
USA,
in
2
0
1
3
.
He
is
a
n
As
so
c
iate
P
ro
fe
ss
o
r
wit
h
De
p
a
rt
m
e
n
t
o
f
Cy
b
e
r
S
e
c
u
ri
ty
a
n
d
Co
m
p
u
ter
E
n
g
i
n
e
e
rin
g
,
G
ra
d
u
a
te
S
c
h
o
o
l
o
f
De
fe
n
se
M
a
n
a
g
e
m
e
n
t,
Ko
re
a
Na
ti
o
n
a
l
De
fe
n
se
Un
iv
e
r
sity
,
No
n
sa
n
,
Re
p
u
b
li
c
o
f
Ko
re
a
.
His
re
se
a
rc
h
in
tere
sts
in
c
lu
d
e
wire
les
s
n
e
two
rk
se
c
u
rit
y
,
tru
st
m
e
c
h
a
n
ism
,
b
o
tn
e
t
d
e
tec
ti
o
n
,
ste
g
a
n
o
g
ra
p
h
y
-
b
a
se
d
c
o
v
e
rt
c
o
m
m
u
n
ica
ti
o
n
,
a
d
v
e
rsa
rial
m
a
c
h
in
e
lea
rn
in
g
,
a
n
d
AI
se
c
u
rit
y
.
He
h
a
s
a
u
th
o
re
d
o
r
c
o
a
u
th
o
re
d
m
o
re
th
a
n
7
0
p
e
e
r
-
re
v
iew
e
d
p
a
p
e
rs
p
u
b
li
sh
e
d
in
j
o
u
r
n
a
ls
a
n
d
in
t
h
e
p
r
o
c
e
e
d
in
g
s
o
f
c
o
n
fe
re
n
c
e
s.
He
is
a
n
IEE
E
se
n
io
r
m
e
m
b
e
r.
He
c
a
n
b
e
c
o
n
tac
ted
a
t
t
h
e
e
m
a
il
:
y
o
u
n
g
h
o
c
h
o
@
k
o
re
a
.
k
r.
Evaluation Warning : The document was created with Spire.PDF for Python.