Co
m
pu
t
er
Science
a
nd
I
nfo
r
m
a
t
io
n T
ec
hn
o
lo
g
ies
Vo
l.
7
,
No
.
3
,
No
v
em
b
er
20
26
,
p
p
.
271
~
290
I
SS
N:
2722
-
3
2
2
1
,
DOI
:
1
0
.
1
1
5
9
1
/csi
t
.
v
7
i
3
.
p
2
7
1
-
2
9
0
271
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ia
e
s
p
r
ime.
co
m/in
d
ex
.
p
h
p
/csi
t
Q
uishi
ng
Shiel
d:
o
n
-
dev
ice m
ulti
-
m
o
da
l det
ection
of
quick
respo
nse phis
hing
Sh
a
bo
ur
B
a
n
da
1
,
M
a
ro
ng
e
M
us
a
ra
2
,
M
a
info
rd
M
uta
nd
a
v
a
ri
3
1
D
e
p
a
r
t
me
n
t
o
f
D
a
t
a
S
c
i
e
n
c
e
a
n
d
A
n
a
l
y
t
i
c
s,
H
a
r
a
r
e
I
n
st
i
t
u
t
e
o
f
T
e
c
h
n
o
l
o
g
y
,
H
a
r
a
r
e
,
Zi
mb
a
b
w
e
2
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
o
n
S
e
c
u
r
i
t
y
a
n
d
A
ss
u
r
a
n
c
e
,
H
a
r
a
r
e
I
n
st
i
t
u
t
e
o
f
Te
c
h
n
o
l
o
g
y
,
H
a
r
a
r
e
,
Z
i
mb
a
b
w
e
3
D
e
p
a
r
t
me
n
t
o
f
D
a
t
a
S
c
i
e
n
c
e
a
n
d
B
u
s
i
n
e
ss S
y
st
e
ms,
S
R
M
I
n
st
i
t
u
t
e
o
f
S
c
i
e
n
c
e
a
n
d
T
e
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
Ap
r
10
,
2
0
2
6
R
ev
is
ed
J
u
n
12
,
2
0
2
6
Acc
ep
ted
J
u
l
13
,
2
0
2
6
Qu
ick
re
sp
o
n
se
(QR)
c
o
d
e
p
h
is
h
in
g
(
q
u
is
h
in
g
)
a
tt
a
c
k
s
tak
e
a
d
v
a
n
tag
e
o
f
u
se
r
tru
st
in
QR
p
h
y
sic
a
l
a
n
d
d
i
g
it
a
l
m
e
d
ia,
wh
il
e
c
u
rre
n
tl
y
a
v
a
il
a
b
le
p
ro
tec
ti
o
n
m
e
c
h
a
n
ism
s
o
n
l
y
d
e
tec
t
a
sin
g
le
d
ime
n
si
o
n
sig
n
a
l
a
n
d
a
re
u
n
a
b
le
to
d
e
tec
t
c
ro
ss
-
m
o
d
a
l
d
e
c
e
p
ti
o
n
.
T
h
is
p
a
p
e
r
i
n
tr
o
d
u
c
e
s
Qu
ish
in
g
S
h
ield
,
a
n
o
n
-
d
e
v
ice
q
u
ish
in
g
d
e
tec
ti
o
n
s
y
ste
m
b
a
se
d
o
n
m
u
lt
i
-
m
o
d
a
l
d
e
e
p
lea
rn
i
n
g
wh
ich
p
re
se
rv
e
s
p
r
iv
a
c
y
o
n
m
o
b
i
le
p
latf
o
rm
s.
Cr
o
ss
-
m
o
d
a
l
a
tt
e
n
ti
o
n
fu
si
o
n
is
u
se
d
to
c
o
m
b
i
n
e
v
is
u
a
l
p
o
st
e
r
fe
a
tu
re
s,
o
p
ti
c
a
l
c
h
a
ra
c
ter
re
c
o
g
n
i
ti
o
n
(
O
CR
)
re
c
o
g
n
ize
d
su
rr
o
u
n
d
i
n
g
tex
t
a
n
d
re
c
o
g
n
ize
d
u
n
ifo
rm
re
so
u
r
c
e
lo
c
a
to
r
(
URL
)
stru
c
tu
re
,
a
s
we
ll
a
s n
e
two
rk
re
p
u
tatio
n
si
g
n
a
ls
in
a
sy
ste
m
.
A
tea
c
h
e
r
m
o
d
e
l
is
trai
n
e
d
o
n
2
0
5
,
4
8
8
re
a
l
-
wo
rld
QR
c
o
d
e
p
o
ste
r
sa
m
p
les
fro
m
4
5
c
o
u
n
tr
ies
a
n
d
2
3
lan
g
u
a
g
e
s,
a
n
d
th
e
k
n
o
wle
d
g
e
is
d
isti
ll
e
d
i
n
t
o
a
c
o
m
p
a
c
t
m
o
d
e
l
fo
r
t
h
e
stu
d
e
n
t
m
o
d
e
l,
in
o
rd
e
r
to
b
e
d
e
p
l
o
y
e
d
i
n
m
o
b
il
e
a
p
p
li
c
a
ti
o
n
s.
Th
e
stu
d
e
n
t
h
a
s
a
n
a
c
c
u
ra
c
y
o
f
9
5
.
5
5
%
a
n
d
a
re
c
a
ll
o
f
9
8
.
1
8
%
f
o
r
a
h
e
ld
-
o
u
t
tes
t
se
t,
with
a
n
i
n
fe
re
n
c
e
late
n
c
y
o
f
2
5
.
3
4
m
s
o
n
m
o
b
il
e
d
e
v
i
c
e
s
,
a
ll
o
f
wh
ich
a
re
a
t
o
r
b
e
l
o
w
t
h
e
d
e
p
lo
y
m
e
n
t
targ
e
ts.
R
o
b
u
stn
e
ss
o
f
9
5
.
0
0
%
wh
e
n
tes
ted
a
d
v
e
rsa
rially
o
n
fo
u
r
se
ts
o
f
a
tt
a
c
k
s.
Th
e
m
u
lt
i
-
m
o
d
a
l
fu
sio
n
a
p
p
r
o
a
c
h
e
n
h
a
n
c
e
s
p
e
rfo
rm
a
n
c
e
b
y
5
.
1
7
-
1
1
.
0
7
p
e
rc
e
n
tag
e
p
o
i
n
ts
o
v
e
r
t
h
e
u
n
imo
d
a
l
b
a
se
li
n
e
a
p
p
ro
a
c
h
e
s
(
p
<
0
.
0
0
1
).
Qu
ish
in
g
S
h
iel
d
,
to
o
u
r
b
e
st
k
n
o
wle
d
g
e
,
is
th
e
first
v
a
li
d
a
ted
m
u
lt
i
-
m
o
d
a
l
q
u
ish
i
n
g
d
e
tec
ti
o
n
s
y
ste
m
sa
ti
sfy
in
g
a
c
c
u
ra
c
y
,
sp
e
e
d
,
siz
e
a
n
d
p
ri
v
a
c
y
re
q
u
irem
e
n
t
fo
r
m
o
b
i
le d
e
p
l
o
y
m
e
n
t.
K
ey
w
o
r
d
s
:
C
r
o
s
s
-
m
o
d
al
atten
tio
n
Kn
o
wled
g
e
d
is
till
atio
n
Mo
b
ile
s
ec
u
r
ity
Mu
lti
-
m
o
d
al
d
ee
p
lear
n
in
g
Priv
ac
y
-
p
r
eser
v
i
n
g
in
f
er
en
ce
QR
co
d
e
p
h
is
h
in
g
Qu
is
h
in
g
d
etec
tio
n
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
:
Sh
ab
o
u
r
B
an
d
a
Dep
ar
tm
en
t o
f
Data
Scien
ce
an
d
An
aly
tics
,
Har
ar
e
I
n
s
titu
te
o
f
T
ec
h
n
o
lo
g
y
Nu
m
b
er
1
5
0
1
5
Gan
g
es R
o
ad
B
elv
ed
er
e
Har
ar
e,
Z
im
b
a
b
we
E
m
ail:
h
2
4
0
7
4
6
e@
h
it.a
c.
zw
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
b
lack
a
n
d
wh
ite
q
u
ick
r
esp
o
n
s
e
(
QR
)
co
d
e
is
n
o
lo
n
g
er
ju
s
t
a
p
ar
t
o
f
th
e
in
f
r
astru
ctu
r
e
;
it
i
s
n
o
w
u
s
ed
f
o
r
co
n
tactless
p
ay
m
e
n
ts
an
d
co
u
n
tles
s
to
u
ch
less
ac
tiv
ities
,
lik
e
ev
en
t
ch
ec
k
-
in
s
,
r
estau
r
an
t
m
en
u
s
,
h
ea
lth
ca
r
e
au
th
en
ticatio
n
,
an
d
m
o
r
e.
Acc
o
r
d
i
n
g
to
in
d
u
s
tr
y
s
tatis
t
ics,
th
e
n
u
m
b
er
o
f
s
ca
n
s
is
r
i
s
in
g
q
u
ick
ly
,
with
wo
r
ld
wid
e
s
ca
n
s
ex
ce
e
d
in
g
6
.
8
b
illi
o
n
in
2
0
2
4
[
1
]
.
T
h
is
u
b
iq
u
ity
,
h
o
wev
er
,
in
c
r
ea
s
es
ex
p
o
s
u
r
e
to
m
is
u
s
e
as
wid
e
s
p
r
ea
d
d
ep
lo
y
m
en
t
o
f
QR
co
d
es
h
as
ex
p
an
d
ed
th
e
attac
k
s
u
r
f
ac
e
an
d
m
a
d
e
th
em
an
attr
ac
tiv
e
v
ec
to
r
f
o
r
f
r
au
d
an
d
o
th
er
m
alicio
u
s
ac
tiv
ity
as
ev
er
y
QR
co
d
e
is
also
an
u
n
v
er
if
iab
le
h
y
p
e
r
lin
k
th
at
ca
n
r
ed
ir
ec
t
a
v
ictim
an
y
wh
er
e.
T
h
e
u
s
e
o
f
d
ee
p
lear
n
in
g
to
s
o
lv
e
th
is
ca
n
h
elp
p
r
ev
e
n
t
c
y
b
er
c
r
im
es
a
n
d
t
h
u
s
en
s
u
r
e
th
e
s
af
e
u
s
ag
e
o
f
QR
c
o
d
e
an
d
in
ter
n
et
s
er
v
ices [
2
]
.
QR
co
d
e
p
h
is
h
in
g
is
a
ty
p
e
o
f
v
is
u
al
in
n
o
cu
o
u
s
n
ess
-
b
ased
ex
p
lo
itatio
n
o
f
QR
co
d
es
th
at
lead
s
th
e
v
ictim
to
an
attac
k
er
’
s
in
f
r
astru
ctu
r
e.
B
y
th
e
e
n
d
o
f
2
0
2
3
,
QR
-
b
ased
p
h
is
h
in
g
ac
co
u
n
t
ed
f
o
r
2
2
%
o
f
all
p
h
is
h
in
g
attem
p
ts
,
as
r
ep
o
r
ted
in
[
3
]
,
n
o
tin
g
th
at
Qu
is
h
in
g
attac
k
s
s
u
r
g
ed
o
v
er
4
0
0
%
in
2
0
2
3
.
C
-
s
u
ite
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
3
2
2
1
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
,
Vo
l.
7
,
No
.
3
,
No
v
em
b
er
20
26
:
2
7
1
-
290
272
ex
ec
u
tiv
es
ar
e
4
2
tim
es
m
o
r
e
lik
ely
to
b
e
tar
g
ete
d
with
Qu
is
h
in
g
attac
k
s
th
an
n
o
n
-
ex
ec
u
tiv
es
an
d
s
u
f
f
er
lo
s
s
es o
f
m
o
r
e
th
an
$
1
m
illi
o
n
p
er
attac
k
,
ac
c
o
r
d
in
g
to
B
ar
r
a
cu
d
a
Netwo
r
k
s
[
4
]
.
T
h
e
th
r
ea
t
is
no
t sm
all
s
ca
le.
QR
-
c
o
d
e
p
h
is
h
in
g
(
Q
u
is
h
i
n
g
)
h
as
b
e
co
m
e
a
f
ast
-
g
r
o
wi
n
g
a
n
d
e
m
e
r
g
i
n
g
ex
p
l
o
i
te
d
m
ass
v
e
cto
r
,
wit
h
in
d
u
s
tr
y
r
e
p
o
r
ts
a
n
d
an
al
y
s
e
s
o
f
i
n
ci
d
e
n
ts
h
i
g
h
li
g
h
ti
n
g
s
ig
n
i
f
ic
a
n
t
in
cr
ea
s
es
in
m
ali
ci
o
u
s
QR
s
ca
n
s
a
n
d
a
n
u
m
b
e
r
o
f
h
i
g
h
-
p
r
o
f
i
le
p
er
s
o
n
al
a
n
d
c
o
r
p
o
r
ate
l
o
s
s
es
o
v
er
t
h
e
p
ast
y
ea
r
.
M
u
lt
i
p
le
r
e
p
o
r
t
e
d
ca
m
p
ai
g
n
s
r
e
v
e
al
h
o
w
att
ac
k
e
r
s
d
is
g
u
is
e
d
p
o
s
te
r
s
,
r
e
ce
i
p
ts
an
d
e
v
e
n
t
r
es
o
u
r
ce
s
to
p
r
o
v
i
d
e
a
u
t
h
e
n
t
ic
-
l
o
o
k
i
n
g
b
r
a
n
d
b
a
it
t
h
at
l
ea
d
s
to
c
r
e
d
en
tia
l
h
ar
v
esti
n
g
p
ag
e
s
a
n
d
p
a
y
m
en
t
f
r
a
u
d
t
r
a
f
f
ic
.
C
u
r
r
e
n
t
d
e
f
en
s
es
a
r
e
co
n
s
t
r
a
i
n
e
d
b
ec
a
u
s
e
s
o
m
e
em
a
il/
u
n
i
f
o
r
m
r
eso
u
r
c
e
l
o
ca
to
r
(
UR
L
)
f
i
lte
r
i
n
g
s
y
s
t
e
m
s
d
o
n
o
t
c
o
n
s
i
d
e
r
QR
i
m
a
g
es
as
m
e
d
ia
,
a
n
d
t
h
u
s
f
a
il
t
o
d
et
ec
t
a
n
y
m
a
lic
io
u
s
UR
L
s
e
m
b
e
d
d
ed
i
n
s
i
d
e
t
h
e
m
,
a
n
d
U
RL
-
o
n
l
y
s
y
s
te
m
s
ar
e
n
o
t
a
b
l
e
t
o
id
en
ti
f
y
s
o
ci
al
en
g
i
n
e
er
in
g
cu
es
t
h
r
o
u
g
h
v
is
u
al
a
n
d
c
o
n
te
x
t
c
o
n
te
n
t
t
h
a
t
is
u
s
ed
i
n
p
o
s
te
r
-
b
ase
d
att
ac
k
s
.
S
e
v
e
r
a
l
o
r
g
a
n
i
za
t
io
n
s
h
a
v
e
s
u
f
f
e
r
ed
l
o
s
s
es
t
h
r
o
u
g
h
t
h
is
m
ea
n
s
wh
ic
h
ci
r
c
u
m
v
en
te
d
tr
ad
iti
o
n
al
e
m
a
il
f
ilte
r
s
c
o
m
p
le
tel
y
b
ec
a
u
s
e
m
o
s
t
g
at
ew
ay
s
h
a
v
e
b
ee
n
u
n
ab
le
t
o
i
d
e
n
ti
f
y
QR
c
o
d
es
as
UR
L
s
to
b
e
f
o
ll
o
w
ed
[
5
]
,
[
6
]
.
Sh
a
r
e
v
s
k
i
e
t
a
l
.
[
6
]
co
n
d
u
ct
ed
a
n
at
u
r
al
is
ti
c
s
t
u
d
y
o
n
r
ea
l
s
c
a
n
an
d
c
o
m
p
r
o
m
is
e
e
v
e
n
ts
i
n
2
0
2
4
f
o
r
w
h
ic
h
e
x
i
s
tin
g
d
ef
e
n
s
es
we
r
e
u
n
a
b
le
t
o
s
t
o
p
.
Geis
le
r
a
n
d
P
ö
h
n
[
7
]
als
o
d
o
cu
m
e
n
t
ed
en
te
r
p
r
is
e
in
ci
d
e
n
ts
w
h
e
r
e
p
h
y
s
i
ca
l
-
w
o
r
l
d
QR
p
o
s
te
r
s
p
o
s
te
d
in
o
f
f
ic
e
lo
b
b
ies
a
n
d
c
o
n
f
er
en
ce
s
p
ac
es
w
er
e
r
e
p
l
ac
e
d
w
it
h
m
al
ici
o
u
s
s
u
b
s
tit
u
t
es.
Hau
n
s
ch
il
d
e
t
a
l
.
[
8
]
also
s
h
o
w
ed
th
at
m
u
lti
-
s
i
g
n
al
,
e
d
g
e
-
A
I
t
ec
h
n
i
q
u
es
a
r
e
r
eq
u
i
r
e
d
t
o
d
et
ec
t
n
ew
p
h
is
h
i
n
g
at
tac
k
s
i
n
r
e
al
-
ti
m
e
.
T
h
e
th
r
e
e
s
tr
u
c
tu
r
al
g
ap
s
t
h
a
t t
h
is
p
ap
e
r
aim
s
t
o
ta
ck
le
a
r
e
i
n
s
p
i
r
e
d
b
y
r
ea
l
-
w
o
r
l
d
ca
s
es
.
T
h
r
ee
s
tr
u
ctu
r
al
g
a
p
s
in
cu
r
r
e
n
t
d
ef
en
s
es:
th
er
e
ar
e
a
lar
g
e
n
u
m
b
er
o
f
e
x
is
tin
g
d
ef
en
s
es,
b
u
t
all
o
f
th
em
h
av
e
th
r
ee
k
ey
wea
k
n
e
s
s
e
s
th
at
m
ak
e
th
eir
d
ef
en
s
es
s
tr
u
ctu
r
ally
u
n
f
it
ag
ain
s
t
Q
u
is
h
in
g
d
etec
tio
n
:
i)
i
n
s
p
ec
tio
n
g
ap
:
m
o
s
t
em
ail
s
ec
u
r
ity
g
atew
ay
s
r
ec
o
g
n
ize
QR
co
d
es
as
b
ein
g
an
o
p
a
q
u
e
im
ag
e.
T
h
ey
ca
n
n
o
t
d
etec
t th
e
h
id
d
e
n
lin
k
a
n
d
th
e
m
alicio
u
s
p
ay
lo
ad
c
r
ea
tes a
th
r
o
u
g
h
f
ar
e
th
at
g
o
es th
r
o
u
g
h
n
o
r
m
al
f
ilter
s
[
5
]
,
[
6
]
in
tact.
T
h
e
co
d
e
is
s
im
ilar
to
an
y
o
th
er
im
a
g
e
attac
h
m
en
t
;
i
i)
c
o
n
tex
tu
al
co
h
e
r
en
ce
g
a
p
:
q
u
is
h
in
g
attac
k
s
u
s
e
a
m
is
m
atch
,
wh
ich
is
a
QR
co
d
e
th
at
r
esem
b
les
a
b
r
an
d
an
d
a
UR
L
th
at
lead
s
to
attac
k
er
in
f
r
astru
ctu
r
e.
T
h
ese
s
ig
n
als
ar
e
p
r
o
ce
s
s
ed
in
d
ep
e
n
d
en
tly
in
e
x
is
tin
g
s
y
s
tem
s
with
o
u
t
co
n
s
id
er
atio
n
o
f
s
em
an
tic
in
co
n
s
is
ten
cies
wh
ich
co
n
s
titu
te
th
e
attac
k
v
e
cto
r
[
7
]
;
a
n
d
iii)
m
o
b
ile
d
ep
l
o
y
m
en
t
g
a
p
:
QR
co
d
es
ar
e
s
ca
n
n
ed
o
n
a
m
o
b
ile
d
ev
ice
b
u
t
th
e
d
etec
tio
n
o
cc
u
r
s
o
u
ts
id
e
o
f
th
e
m
o
b
ile
d
ev
ice
.
Ho
wev
er
,
cu
r
r
en
t
m
u
lti
-
m
o
d
al
p
h
is
h
in
g
m
o
d
els
ar
e
to
o
lar
g
e
(
1
0
0
MB+)
,
an
d
m
u
s
t
b
e
r
u
n
o
n
a
s
er
v
er
to
m
ak
e
in
f
er
en
ce
,
wh
ich
is
n
o
t
s
u
itab
le
f
o
r
o
n
-
d
ev
ice
d
ep
lo
y
m
e
n
t
[
8
]
.
M
o
s
t
r
ec
en
t
m
u
lti
-
alg
o
r
ith
m
en
s
em
b
le
a
p
p
r
o
ac
h
es
lik
e
r
an
d
o
m
f
o
r
est
,
C
atB
o
o
s
t,
Ad
aBo
o
s
t
an
d
m
u
ltil
ay
er
p
e
r
ce
p
tr
o
n
f
o
r
UR
L
-
o
n
ly
d
etec
tio
n
ac
h
iev
e
s
tr
o
n
g
ac
cu
r
ac
y
o
n
UR
L
d
atasets
[
9
]
b
u
t
ar
e
u
n
ab
le
to
b
e
d
ep
lo
y
ed
o
n
m
o
b
ile
s
ca
le
o
r
ar
e
o
b
liv
io
u
s
to
v
is
u
al
s
o
cial
en
g
in
ee
r
in
g
.
At
th
e
ex
ac
t m
o
m
en
t th
e
y
s
ca
n
,
u
s
er
s
ar
e
lef
t w
ith
o
u
t
p
r
o
tectio
n
[
1
0
]
.
T
h
is
p
ap
er
co
v
er
s
all
th
r
ee
g
a
p
s
an
d
ask
s
th
r
ee
s
p
ec
if
ic
q
u
esti
o
n
s
o
f
r
esear
ch
t
h
at
ca
n
b
e
m
ea
s
u
r
ed
:
−
D
o
es
m
u
lti
-
m
o
d
al
f
u
s
io
n
(
v
is
u
al+
tex
tu
al+
UR
L
+n
etwo
r
k
r
ep
u
tatio
n
)
p
r
o
d
u
ce
s
tatis
tic
ally
s
ig
n
if
ican
t
ac
cu
r
ac
y
im
p
r
o
v
e
m
en
ts
o
v
er
s
in
g
le
-
m
o
d
ality
a
n
d
d
u
al
-
m
o
d
a
lity
b
aselin
es o
n
a
lar
g
e
-
s
ca
le,
g
eo
g
r
a
p
h
ically
d
iv
er
s
e
Qu
is
h
in
g
d
ataset?
−
C
an
s
u
ch
a
m
u
lti
-
m
o
d
al
s
y
s
tem
b
e
co
m
p
r
ess
ed
to
u
n
d
er
5
0
MB
wh
ile
r
etain
in
g
g
r
e
ater
th
an
9
5
%
ac
cu
r
ac
y
an
d
r
u
n
n
in
g
with
in
3
0
0
m
s
o
n
m
id
-
r
a
n
g
e
An
d
r
o
id
h
ar
d
war
e?
−
W
h
ich
in
d
iv
id
u
al
m
o
d
ality
co
n
tr
ib
u
tes
m
o
s
t
to
d
etec
tio
n
ac
cu
r
ac
y
,
an
d
d
o
es
th
e
m
ar
g
in
al
g
ain
f
r
o
m
ea
ch
m
o
d
ality
d
if
f
er
s
ig
n
if
ican
tly
f
r
o
m
ze
r
o
?
T
h
is
p
ap
er
a
d
d
r
ess
es
all
th
r
ee
g
ap
s
th
r
o
u
g
h
th
e
f
o
llo
win
g
co
n
tr
i
b
u
tio
n
s
:
i)
Q
u
is
h
in
g
Sh
ield
,
a
f
o
u
r
-
m
o
d
ality
d
etec
tio
n
ar
c
h
itectu
r
e
th
at
in
teg
r
ates
v
is
u
al
,
tex
tu
al,
UR
L
-
s
tr
u
ctu
r
al,
an
d
n
etwo
r
k
r
e
p
u
tatio
n
en
co
d
er
s
th
r
o
u
g
h
c
r
o
s
s
-
m
o
d
a
l
atten
tio
n
f
u
s
io
n
.
T
r
ain
ed
o
n
2
0
5
,
4
8
8
QR
co
d
e
p
o
s
ter
s
am
p
les
ac
r
o
s
s
4
5
co
u
n
tr
ies
an
d
2
3
lan
g
u
ag
es,
t
h
e
teac
h
er
m
o
d
el
ac
h
iev
es
9
6
.
3
7
%
ac
cu
r
ac
y
a
n
d
9
8
.
1
0
%
r
ec
all
o
u
tp
er
f
o
r
m
in
g
th
e
n
ea
r
est
v
is
u
al+
UR
L
d
u
al
-
m
o
d
al
b
aselin
e
(
Ab
d
el
n
ab
i
et
a
l
.
[
1
1
]
:
9
3
.
4
0
%)
b
y
2
.
9
7
p
p
i
n
ac
cu
r
ac
y
a
n
d
5
.
9
0
p
p
in
r
ec
all
;
ii)
a
k
n
o
wled
g
e
d
is
till
atio
n
p
ip
elin
e
co
m
p
r
ess
in
g
th
e
3
5
2
MB
teac
h
er
in
to
a
2
9
.
6
2
MB
s
tu
d
en
t
m
o
d
el
with
o
n
l
y
0
.
8
2
p
e
r
ce
n
t
ag
e
p
o
in
t
ac
c
u
r
ac
y
l
o
s
s
,
a
co
m
p
r
ess
io
n
-
ac
cu
r
ac
y
r
eten
tio
n
r
ate
th
at
o
u
tp
e
r
f
o
r
m
s
p
u
b
lis
h
ed
b
en
ch
m
a
r
k
s
o
f
co
m
p
ar
ab
le
a
r
ch
itectu
r
es
[
1
2
]
,
[
1
3
]
,
in
w
h
ich
2
-
5
p
p
l
o
s
s
es
ar
e
ty
p
ical
at
co
m
p
ar
ab
le
co
m
p
r
e
s
s
io
n
r
atio
s
;
iii)
th
e
f
ir
s
t
p
u
b
lis
h
ed
m
u
lt
i
-
m
o
d
al
Q
u
is
h
in
g
d
etec
tio
n
s
y
s
tem
v
alid
ated
f
o
r
r
ea
l
-
wo
r
ld
A
n
d
r
o
i
d
d
e
p
lo
y
m
en
t,
s
im
u
ltan
eo
u
s
ly
s
atis
f
y
in
g
s
ize
(
<5
0
MB),
laten
c
y
(
<3
0
0
m
s
)
,
ac
cu
r
ac
y
(
>9
5
%),
r
o
b
u
s
tn
ess
(
>9
5
%),
an
d
f
u
ll
o
n
-
d
e
v
ice
p
r
iv
ac
y
co
n
s
tr
ain
ts
.
No
p
r
io
r
s
y
s
tem
ac
h
iev
es
all
f
iv
e
s
im
u
ltan
eo
u
s
ly
;
iv
)
a
g
eo
g
r
ap
h
ically
d
iv
er
s
e,
m
u
lti
-
lin
g
u
al
QR
co
d
e
p
h
is
h
in
g
d
ataset,
2
0
5
,
4
8
8
s
am
p
les,
4
5
co
u
n
tr
ies,
2
3
lan
g
u
a
g
es,
th
r
e
e
s
o
p
h
is
ticatio
n
tier
s
,
in
ter
-
an
n
o
tato
r
ag
r
ee
m
e
n
t
κ=
0
.
8
9
;
an
d
v
)
f
o
u
r
f
o
r
m
ally
s
p
ec
if
ied
alg
o
r
ith
m
s
(
Alg
o
r
it
h
m
s
1
-
4
)
an
d
a
co
m
p
lete
o
p
en
-
s
o
u
r
ce
r
ep
o
s
ito
r
y
f
o
r
r
ep
r
o
d
u
cib
ilit
y
.
T
a
b
le
1
s
u
m
m
ar
izes
all
o
p
en
-
s
o
u
r
ce
d
ep
en
d
en
cies
u
s
ed
in
th
e
Q
u
is
h
in
g
Sh
ield
im
p
lem
en
tatio
n
.
Sectio
n
2
p
r
esen
ts
th
e
Qu
is
h
in
g
Sh
ield
s
y
s
tem
m
et
h
o
d
o
lo
g
y
,
co
v
e
r
in
g
m
o
d
ality
s
elec
tio
n
r
atio
n
ale,
en
co
d
er
d
esig
n
,
cr
o
s
s
-
m
o
d
al
atten
tio
n
f
u
s
io
n
,
k
n
o
wled
g
e
d
is
till
atio
n
,
alg
o
r
ith
m
p
s
e
u
d
o
co
d
e,
a
n
d
p
r
iv
ac
y
-
p
r
ese
r
v
in
g
d
ep
lo
y
m
en
t
ar
ch
itectu
r
e.
I
n
s
ec
tio
n
3
,
th
e
e
x
p
er
im
en
tal
s
etu
p
is
d
escr
ib
e
d
th
at
co
n
tain
s
th
e
c
o
n
s
tr
u
ctio
n
o
f
th
e
d
ataset,
th
e
tr
ain
in
g
co
n
f
ig
u
r
atio
n
,
an
d
th
e
ev
alu
atio
n
p
r
o
t
o
co
l.
E
x
p
e
r
im
en
tal
r
esu
lts
ar
e
p
r
esen
ted
in
s
ec
tio
n
4
.
Fin
d
in
g
s
,
th
eo
r
y
im
p
licatio
n
s
an
d
lim
ita
tio
n
s
ar
e
d
is
cu
s
s
ed
in
s
ec
tio
n
5
.
Sectio
n
6
c
o
n
clu
d
es.
Evaluation Warning : The document was created with Spire.PDF for Python.
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
I
SS
N:
2722
-
3
2
2
1
Qu
is
h
in
g
S
h
ield
:
o
n
-
d
ev
ice
mu
lti
-
mo
d
a
l d
etec
tio
n
o
f q
u
ick
r
e
s
p
o
n
s
e
p
h
is
h
in
g
(
S
h
a
b
o
u
r
B
a
n
d
a
)
273
T
ab
le
1
.
Op
e
n
-
s
o
u
r
ce
lib
r
ar
ies an
d
to
o
ls
u
s
ed
in
th
e
Qu
is
h
in
g
Sh
ield
im
p
lem
en
tatio
n
C
o
m
p
o
n
e
n
t
Li
b
r
a
r
y
V
e
r
si
o
n
R
e
f
e
r
e
n
c
e
V
i
su
a
l
e
n
c
o
d
e
r
t
o
r
c
h
v
i
s
i
o
n
M
o
b
i
l
e
N
e
t
V
3
-
sma
l
l
0
.
1
5
.
2
H
o
w
a
r
d
e
t
a
l
.
[
14
]
Te
x
t
u
a
l
e
n
c
o
d
e
r
t
r
a
n
sf
o
r
mers
D
i
st
i
l
B
E
R
T
-
b
a
se
-
m
u
l
t
i
l
i
n
g
u
a
l
4
.
3
5
.
0
S
a
n
h
e
t
a
l
. [
15
]
U
R
L
e
n
c
o
d
e
r
C
u
s
t
o
m C
h
a
r
-
C
N
N
-
Zh
a
n
g
e
t
a
l
. [
16
]
A
t
t
e
n
t
i
o
n
f
u
si
o
n
t
o
r
c
h
.
n
n
.
M
u
l
t
i
h
e
a
d
A
t
t
e
n
t
i
o
n
2
.
0
.
1
V
a
sw
a
n
i
e
t
a
l
. [
17
]
Tr
a
i
n
i
n
g
P
y
T
o
r
c
h
2
.
0
+
C
U
D
A
1
1
.
8
2
.
0
.
1
-
I
N
T8
q
u
a
n
t
i
z
a
t
i
o
n
t
o
r
c
h
.
q
u
a
n
t
i
z
a
t
i
o
n
.
q
u
a
n
t
i
z
e
_
d
y
n
a
m
i
c
b
u
i
l
t
-
in
Jac
o
b
e
t
a
l
.
[
18
]
O
N
N
X
e
x
p
o
r
t
t
o
r
c
h
.
o
n
n
x
+
o
n
n
x
r
u
n
t
i
m
e
-
a
n
d
r
o
i
d
1
.
1
6
.
1
-
A
d
v
e
r
sari
a
l
t
e
s
t
i
n
g
Ty
p
o
S
q
u
a
t
t
e
r
(
o
p
e
n
-
so
u
r
c
e
)
1
.
2
.
0
-
S
t
a
t
i
st
i
c
a
l
t
e
s
t
i
n
g
sci
p
y
.
st
a
t
s
M
c
N
e
mar
1
.
1
1
.
0
-
D
a
t
a
s
e
t
a
n
n
o
t
a
t
i
o
n
La
b
e
l
st
u
d
i
o
1
.
8
.
0
-
Th
r
e
a
t
i
n
t
e
l
l
i
g
e
n
c
e
f
e
e
d
s
P
h
i
s
h
Ta
n
k
A
P
I
+
O
p
e
n
P
h
i
sh
f
r
e
e
f
e
e
d
-
-
R
a
n
d
o
m s
e
e
d
t
o
r
c
h
.
m
a
n
u
a
l
_
se
e
d
(
4
2
)
-
S
e
t
i
n
a
l
l
e
x
p
e
r
i
me
n
t
s
2.
M
E
T
H
O
D
2
.
1
.
Arc
hite
ct
ure
o
v
er
v
iew
Qu
is
h
in
g
Sh
ield
o
p
er
ates
ac
r
o
s
s
th
r
ee
f
u
n
ctio
n
al
lay
er
s
.
T
h
e
in
p
u
t
lay
er
ac
q
u
ir
es
a
n
d
p
r
e
-
p
r
o
ce
s
s
es
f
o
u
r
h
eter
o
g
en
e
o
u
s
s
ig
n
al
m
o
d
alities
f
r
o
m
a
s
in
g
le
QR
co
d
e
s
ca
n
ev
en
t:
a
2
2
4
×2
2
4
r
ed
,
g
r
ee
n
,
b
lu
e
(
R
GB
)
p
o
s
ter
im
ag
e,
o
p
tical
ch
ar
ac
te
r
r
ec
o
g
n
itio
n
(
OC
R
)
-
ex
t
r
ac
ted
s
u
r
r
o
u
n
d
in
g
tex
t
(
u
p
to
1
2
8
to
k
en
s
)
,
th
e
d
ec
o
d
e
d
UR
L
ch
ar
ac
ter
s
eq
u
en
ce
(
u
p
t
o
2
5
6
ch
ar
ac
ter
s
)
,
a
n
d
f
i
v
e
n
o
r
m
alize
d
n
etwo
r
k
r
ep
u
tatio
n
s
ca
lar
s
.
T
h
e
p
r
o
ce
s
s
lay
er
en
c
o
d
es
ea
c
h
m
o
d
ality
t
h
r
o
u
g
h
a
s
p
ec
ialized
n
eu
r
al
e
n
co
d
er
,
in
teg
r
ates
th
e
r
esu
ltin
g
2
5
6
-
d
im
e
n
s
io
n
al
em
b
ed
d
in
g
s
v
ia
cr
o
s
s
-
m
o
d
al
a
tten
tio
n
f
u
s
io
n
,
a
n
d
m
ap
s
t
h
e
5
1
2
-
d
im
en
s
io
n
al
j
o
in
t
r
ep
r
ese
n
tatio
n
to
p
r
o
d
u
ce
s
a
m
alicio
u
s
p
r
o
b
ab
ilit
y
s
co
r
e
th
r
o
u
g
h
a
f
o
u
r
-
s
tag
e
class
if
icatio
n
h
ea
d
.
T
h
e
o
u
tp
u
t
lay
er
ass
ig
n
s
th
e
s
co
r
e
to
o
n
e
o
f
th
r
ee
r
is
k
tier
s
,
SAFE,
W
AR
NI
NG,
o
r
B
L
OC
K,
an
d
m
ain
tai
n
s
a
p
r
iv
ac
y
-
p
r
eser
v
in
g
o
n
-
d
ev
ice
r
ep
u
tatio
n
ca
ch
e
to
av
o
id
r
e
d
u
n
d
an
t
i
n
f
er
en
ce
o
n
p
r
ev
io
u
s
ly
an
aly
ze
d
UR
L
s
.
Fig
u
r
e
1
illu
s
tr
ates
th
e
co
m
p
lete
in
f
er
en
ce
p
ip
elin
e
alo
n
g
s
id
e
t
h
e
k
n
o
wled
g
e
d
is
till
atio
n
p
r
o
c
ess
b
y
wh
ich
th
e
f
u
ll
teac
h
er
m
o
d
el
(
3
5
2
MB)
is
co
m
p
r
ess
ed
in
to
th
e
2
9
.
6
2
M
B
s
tu
d
en
t m
o
d
el
d
e
p
lo
y
ed
o
n
m
o
b
ile
h
ar
d
war
e.
Fig
u
r
e
1
.
Qu
is
h
in
g
d
etec
tio
n
in
f
er
en
ce
a
n
d
d
is
till
atio
n
p
ip
elin
e
s
h
o
win
g
f
o
u
r
in
p
u
t
m
o
d
alit
ies,
cr
o
s
s
-
m
o
d
al
atten
tio
n
f
u
s
io
n
,
k
n
o
wled
g
e
d
i
s
till
atio
n
,
an
d
o
n
-
d
ev
ice
d
e
p
lo
y
m
en
t f
lo
w
2
.
2
.
M
o
da
lity
s
elec
t
io
n r
a
t
io
na
le
Qu
is
h
in
g
’
s
attac
k
an
ato
m
y
is
th
e
b
asis
o
f
th
e
ch
o
ice
o
f
f
o
u
r
m
o
d
alities
:
v
is
u
al,
tex
tu
al,
UR
L
-
s
tr
u
ctu
r
al,
an
d
n
etwo
r
k
r
ep
u
tatio
n
,
ea
c
h
co
v
er
in
g
a
d
i
s
tin
ct
attac
k
s
u
r
f
ac
e
th
at
th
e
o
th
er
s
ca
n
n
o
t
d
etec
t
in
is
o
latio
n
.
E
v
en
if
th
e
UR
L
an
d
d
o
m
ai
n
ar
e
u
n
f
am
iliar
,
a
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN
)
ca
n
s
ee
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
3
2
2
1
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
,
Vo
l.
7
,
No
.
3
,
No
v
em
b
er
20
26
:
2
7
1
-
290
274
Ap
p
le
lo
g
o
o
n
a
p
h
is
h
in
g
p
o
s
t
er
,
wh
ich
in
d
icate
s
th
at
th
e
lo
g
o
is
d
etec
ted
b
y
v
is
u
al
s
ig
n
al
s
an
d
th
at
Ap
p
le
’
s
p
o
s
ter
lev
el
s
o
cial
en
g
in
ee
r
in
g
is
d
etec
ted
.
T
ex
tu
al
s
ig
n
als
f
r
o
m
OC
R
o
n
th
e
s
u
r
r
o
u
n
d
in
g
tex
t
id
e
n
tify
u
r
g
en
c
y
,
au
th
o
r
ity
im
p
er
s
o
n
at
io
n
an
d
p
s
y
c
h
o
lo
g
ical
m
an
ip
u
latio
n
th
at
Qu
is
h
in
g
p
o
s
ter
s
in
ten
tio
n
ally
u
s
e
in
o
r
d
er
to
lea
d
th
e
au
d
ie
n
ce
to
s
ca
n
with
o
u
t a
n
aly
zin
g
.
T
h
ese
t
wo
m
o
d
alit
ies ar
e
th
e
o
n
es th
a
t c
ap
tu
r
e
attac
k
s
in
wh
ich
ad
v
er
s
ar
ies
ar
e
u
s
in
g
v
alid
o
r
am
b
ig
u
o
u
s
UR
L
s
a
ca
s
e
ex
am
p
le
o
f
th
is
is
th
e
s
o
ci
al
s
ec
u
r
ity
n
u
m
b
er
s
(
SSN
)
s
u
s
p
en
d
ed
an
d
Ap
p
le
I
D
s
u
s
p
en
d
ed
ca
s
es
in
o
u
r
d
ep
lo
y
m
en
t
test
in
g
.
T
h
e
UR
L
-
s
tr
u
ctu
r
al
s
ig
n
als
ar
e
ab
le
to
u
n
co
v
er
s
y
n
tactic
an
o
m
alies,
I
P
-
ad
d
r
ess
h
o
s
ts
,
/b
in
.
s
h
p
ath
s
,
ty
p
o
s
q
u
attin
g
an
d
h
o
m
o
g
r
ap
h
s
u
b
s
titu
tio
n
s
th
at
wo
u
ld
r
em
ain
h
id
d
en
b
y
s
u
b
w
o
r
d
to
k
e
n
iz
atio
n
[
16
]
.
Netwo
r
k
r
ep
u
tati
o
n
s
ig
n
als
ar
e
b
ased
o
n
d
o
m
ain
i
n
f
r
astru
ctu
r
e
q
u
ality
,
n
ew
d
o
m
ain
s
,
r
e
v
o
k
e
d
s
ec
u
r
e
s
o
c
k
ets
lay
er
(
SS
L
)
ce
r
tific
ates
an
d
b
lack
lis
ted
h
o
s
ts
an
d
ar
e
n
o
t r
elate
d
to
v
is
u
al
an
d
te
x
tu
al
co
n
ten
t.
T
h
is
d
esig
n
is
s
u
p
p
o
r
ted
d
ir
e
ctly
b
y
th
e
d
ep
l
o
y
m
e
n
t
test
r
esu
lts
in
s
ec
tio
n
4
.
5
:
th
e
SS
N
s
u
s
p
en
d
ed
an
d
Ap
p
le
I
D
s
u
s
p
en
d
ed
att
ac
k
s
b
o
th
u
s
ed
leg
itima
te
-
ap
p
ea
r
in
g
UR
L
s
(
r
aw.
g
it
h
u
b
u
s
er
co
n
ten
t.c
o
m
an
d
www.
aa
p
.
o
r
g
)
,
wh
ich
wo
u
l
d
s
co
r
e
lo
w
s
u
s
p
icio
n
u
n
d
er
U
R
L
-
o
n
ly
o
r
n
etwo
r
k
a
n
aly
s
is
alo
n
e;
b
o
th
wer
e
co
r
r
ec
tly
class
if
ied
as
W
AR
N
I
NG
b
ec
au
s
e
th
e
v
is
u
al
en
co
d
er
ass
ig
n
ed
9
9
%
r
is
k
s
co
r
es
b
ased
o
n
b
r
an
d
im
p
er
s
o
n
atio
n
in
th
e
p
o
s
ter
im
ag
e.
On
ly
a
s
y
s
tem
th
at
p
r
o
ce
s
s
es
v
is
u
al
an
d
UR
L
s
ig
n
als
jo
i
n
tly
ca
n
ca
tch
th
is
class
o
f
attac
k
.
C
o
n
s
id
er
in
g
ad
d
itio
n
al
m
o
d
alities
s
u
ch
as
d
ev
ice
m
etad
ata
(
h
ar
d
wa
r
e
f
in
g
er
p
r
in
t)
an
d
g
eo
lo
ca
tio
n
at
s
ca
n
tim
e
wer
e
ex
clu
d
e
d
o
n
p
r
iv
ac
y
g
r
o
u
n
d
s
as
co
llectin
g
s
u
ch
i
d
en
tifi
er
s
co
n
tr
ad
icts
th
e
d
ata
-
m
in
im
izatio
n
ar
c
h
itectu
r
e
th
at
Qu
is
h
in
g
Sh
ield
is
b
u
ilt
ar
o
u
n
d
.
T
h
ese
m
a
y
b
e
in
c
o
r
p
o
r
ated
in
t
o
f
u
tu
r
e
wo
r
k
,
with
u
s
er
c
o
n
s
en
t.
2
.
3
.
M
ulti
-
m
o
da
l e
nco
der
de
s
ig
n
2
.
3
.
1
.
Vis
ua
l
enco
der
Mo
b
ileNetV3
-
s
m
all
[
14
]
is
a
p
r
e
-
tr
ain
e
d
C
NNs
th
at
p
r
o
ce
s
s
es
th
e
2
2
4
×
2
2
4
R
GB
im
ag
e,
y
ield
in
g
co
n
v
o
l
u
tio
n
al
f
ea
tu
r
es.
I
t
p
r
o
d
u
ce
s
a
5
7
6
-
d
im
e
n
s
io
n
al
f
ea
tu
r
e
v
ec
to
r
t
h
at
is
p
r
o
jecte
d
o
n
t
o
a
2
5
6
-
d
im
en
s
io
n
a
l
em
b
ed
d
in
g
s
p
ac
e
b
y
a
lin
ea
r
l
ay
er
f
o
llo
we
d
b
y
r
ec
tifie
d
lin
ea
r
u
n
it
(
R
eL
U
)
ac
tiv
atio
n
an
d
a
d
r
o
p
o
u
t
r
ate
o
f
0
.
3
.
Mo
b
ileNetV3
was
ch
o
s
en
b
ec
au
s
e
it
h
ad
a
g
o
o
d
b
alan
ce
o
f
ac
cu
r
ac
y
an
d
ef
f
icien
cy
o
n
m
o
b
ile
d
ev
ices,
with
an
i
n
f
er
en
ce
tim
e
o
f
less
th
an
5
0
m
s
a
n
d
2
.
5
4
M
p
a
r
a
m
eter
s
,
wh
ile
m
ain
tai
n
in
g
a
h
ig
h
r
e
p
r
esen
tatio
n
al
p
o
wer
.
I
m
ag
es
ar
e
n
o
r
m
alize
d
to
I
m
ag
eNe
t
s
tati
s
tics
(
m
ea
n
=[
0
.
4
8
5
,
0
.
4
5
6
,
0
.
4
0
6
]
,
s
td
=[
0
.
2
2
9
,
0
.
2
2
4
,
0
.
2
2
5
]
)
p
r
io
r
to
en
co
d
in
g
.
T
h
e
v
is
u
al
e
m
b
ed
d
in
g
is
:
_
=
(
(
(
_
·
3
(
_
)
+
_
)
)
)
w
h
er
e
,
_
∈
ℝ
^
(
576
×
256
)
is
th
e
p
r
o
jectio
n
m
atr
i
x
an
d
B
N
d
en
o
tes b
atch
n
o
r
m
aliza
tio
n
.
2
.
3
.
2
.
T
ex
t
ua
l
enco
d
er
Dis
til
B
E
R
T
-
b
ase
-
m
u
ltil
in
g
u
al
-
ca
s
ed
[
15
]
is
u
s
ed
to
p
r
o
ce
s
s
p
r
o
m
o
tio
n
al
tex
t
ar
o
u
n
d
th
e
QR
co
d
e
p
o
s
ter
s
ex
tr
ac
ted
f
r
o
m
th
e
O
C
R
im
ag
e.
T
h
e
m
u
ltil
in
g
u
al
v
ar
iatio
n
is
a
v
ailab
le
in
1
0
4
lan
g
u
a
g
es,
wh
ich
allo
ws
it
to
d
etec
t
an
y
o
f
th
e
2
3
-
lan
g
u
ag
e
d
ataset.
L
ay
er
s
1
-
3
ar
e
f
r
o
ze
n
in
o
r
d
er
to
av
o
id
ca
tast
r
o
p
h
ic
f
o
r
g
ettin
g
;
l
ay
er
s
4
-
6
ar
e
f
in
e
-
tu
n
ed
o
n
p
h
is
h
in
g
s
p
ec
if
ic
p
atter
n
s
.
T
h
e
[
C
L
S]
t
o
k
en
’
s
7
6
8
-
d
im
en
s
io
n
al
r
ep
r
esen
tatio
n
is
p
r
o
jecte
d
to
2
5
6
d
i
m
en
s
io
n
s
:
_
=
(
(
(
_
·
[
]
(
_
)
+
_
)
)
)
2
.
3
.
3
.
URL
enco
der
A
ch
ar
ac
ter
-
le
v
el
C
NN
[
16
]
p
r
o
ce
s
s
es
d
ec
o
d
ed
d
esti
n
atio
n
UR
L
s
ar
e
d
ec
o
d
e
d
as
a
s
eq
u
e
n
ce
o
f
u
p
to
2
5
6
ch
ar
ac
ter
s
th
at
ar
e
d
r
awn
f
r
o
m
a
v
o
ca
b
u
lar
y
o
f
7
9
ch
ar
a
cter
s
(
a
-
z,
A
-
Z
,
0
-
9
,
an
d
UR
L
s
p
ec
ial
ch
ar
ac
ter
s
)
an
d
r
ep
r
esen
ted
b
y
1
6
-
d
im
en
s
io
n
al
v
ec
to
r
s
p
er
ch
a
r
ac
ter
i
n
a
C
NN
[
15
]
.
T
h
r
ee
p
ar
alle
l
1
D
co
n
v
o
l
u
tio
n
s
(
k
er
n
el
s
izes
k
={
3
,
4
,
5
},
6
4
f
ilter
s
ea
ch
)
ex
tr
ac
t
m
u
lti
-
s
ca
le
n
-
g
r
a
m
p
atter
n
s
.
Glo
b
al
m
ax
-
p
o
o
lin
g
ag
g
r
eg
ates
s
alien
t f
ea
tu
r
es in
to
a
1
9
2
-
d
im
en
s
io
n
al
v
ec
to
r
p
r
o
jecte
d
t
o
2
5
6
d
im
e
n
s
io
n
s
:
_
=
(
(
(
_
·
(
₃
‖
₄
‖
₅
)
(
(
_
)
)
+
_
)
)
)
wh
er
e
‖
d
en
o
tes
co
n
ca
ten
atio
n
o
f
t
h
e
th
r
ee
p
ar
allel
co
n
v
o
lu
tio
n
o
u
tp
u
ts
.
C
h
ar
ac
ter
le
v
el
en
co
d
i
n
g
f
u
lly
ca
p
tu
r
es
ty
p
o
s
q
u
attin
g
,
h
o
m
o
g
r
ap
h
attac
k
s
,
an
d
s
u
b
d
o
m
ai
n
s
tack
in
g
an
o
m
alies
th
at
s
u
b
wo
r
d
to
k
e
n
izatio
n
in
tr
an
s
f
o
r
m
er
m
o
d
els s
y
s
tem
atica
lly
o
b
s
cu
r
es.
Evaluation Warning : The document was created with Spire.PDF for Python.
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
I
SS
N:
2722
-
3
2
2
1
Qu
is
h
in
g
S
h
ield
:
o
n
-
d
ev
ice
mu
lti
-
mo
d
a
l d
etec
tio
n
o
f q
u
ick
r
e
s
p
o
n
s
e
p
h
is
h
in
g
(
S
h
a
b
o
u
r
B
a
n
d
a
)
275
2
.
3
.
4
.
Net
wo
r
k
re
pu
t
a
t
io
n e
nco
der
A
two
-
lay
er
f
ee
d
f
o
r
war
d
n
e
two
r
k
(
5
→1
2
8
→2
5
6
)
with
b
atch
n
o
r
m
aliza
tio
n
an
d
0
.
3
d
r
o
p
o
u
t
p
r
o
ce
s
s
es
f
iv
e
n
o
r
m
alize
d
s
ca
lar
f
ea
tu
r
es
in
[
0
,
1
]
:
i
)
SS
L
ce
r
tific
ate
v
alid
ity
s
tatu
s
;
ii
)
d
o
m
ain
ag
e
in
n
o
r
m
alize
d
d
ay
s
;
iii
)
ce
r
tific
at
e
v
alid
atio
n
lev
el
(
DV/OV/E
V)
;
iv
)
r
ea
l
-
tim
e
b
lack
lis
t
m
e
m
b
er
s
h
ip
s
co
r
e,
a
n
d
v
)
ag
g
r
eg
ate
v
e
n
d
o
r
r
ep
u
tatio
n
s
co
r
e
f
r
o
m
th
r
ea
t i
n
tellig
en
c
e
API
s
.
_
=
(
(
(
_
2
·
(
(
(
_
1
·
_
+
_
1
)
)
)
+
_
2
)
)
)
2
.
4
.
Cro
s
s
-
mo
da
l a
t
t
ent
io
n f
us
io
n
T
h
e
f
o
u
r
2
5
6
-
d
i
m
e
n
s
i
o
n
al
e
n
c
o
d
e
r
o
u
t
p
u
ts
a
r
e
i
n
te
g
r
at
e
d
t
h
r
o
u
g
h
t
h
r
e
e
p
ai
r
wis
e
cr
o
s
s
-
m
o
d
al
atte
n
t
io
n
m
e
ch
a
n
is
m
s
[
1
7
]
,
[
19
]
.
V
is
u
al
-
te
x
t
att
e
n
ti
o
n
i
d
e
n
ti
f
ies
b
r
a
n
d
-
c
o
n
te
x
t
s
em
an
tic
m
is
m
a
tc
h
es
.
Vis
u
a
l
-
n
et
wo
r
k
att
e
n
ti
o
n
d
e
te
cts
h
i
g
h
-
q
u
alit
y
b
r
an
d
m
im
i
c
r
y
p
ai
r
e
d
wit
h
p
o
o
r
i
n
f
r
as
tr
u
ct
u
r
e
r
e
p
u
t
a
tio
n
.
T
e
x
t
-
n
et
wo
r
k
atte
n
t
io
n
c
o
r
r
elat
es
u
r
g
e
n
c
y
l
an
g
u
a
g
e
wit
h
s
u
s
p
i
ci
o
u
s
d
o
m
ain
c
h
ar
ac
t
er
is
ti
cs.
E
ac
h
m
e
c
h
a
n
is
m
u
s
es
m
u
l
ti
-
h
e
ad
a
tte
n
t
io
n
wit
h
h
=
8
h
e
a
d
s
an
d
h
ea
d
d
i
m
e
n
s
i
o
n
a
lit
y
d
_
k
=
3
2
,
c
o
m
p
u
t
i
n
g
s
ca
le
d
d
o
t
-
p
r
o
d
u
ct
a
tte
n
t
io
n
:
(
,
,
)
=
(
^
/
√
_
)
·
m
u
lti
-
h
ea
d
atten
tio
n
is
:
(
,
,
)
=
(
ℎ
₁
,
…
,
ℎ
_
ℎ
)
·
^
wh
er
e
,
ℎ
_
=
(
·
_
^
,
·
_
^
,
·
_
^
)
.
T
h
e
th
r
ee
cr
o
s
s
-
m
o
d
al
atte
n
tio
n
o
u
tp
u
ts
(
a_
v
t,
a_
v
n
,
a_
tn
)
to
g
eth
e
r
with
th
e
f
o
u
r
en
co
d
er
em
b
ed
d
in
g
s
ar
e
co
n
ca
ten
ate
d
(
7
×2
5
6
=1
,
7
9
2
d
im
en
s
io
n
s
)
an
d
p
r
o
jecte
d
t
o
5
1
2
d
im
en
s
io
n
s
v
ia
a
lin
ea
r
lay
e
r
with
lay
er
n
o
r
m
aliza
tio
n
an
d
0
.
5
d
r
o
p
o
u
t:
_
=
(
(
_
·
(
_
,
_
,
_
,
_
,
_
,
_
,
_
)
+
_
)
)
T
h
e
f
u
s
io
n
d
esig
n
d
elib
er
ately
in
clu
d
es
all
f
o
u
r
e
n
co
d
e
r
em
b
ed
d
in
g
s
alo
n
g
s
id
e
th
e
th
r
ee
cr
o
s
s
-
m
o
d
al
o
u
tp
u
ts
,
en
s
u
r
in
g
t
h
at
m
o
d
ality
-
s
p
ec
if
ic
in
f
o
r
m
ati
o
n
s
u
r
v
iv
es
in
to
th
e
class
if
icatio
n
h
ea
d
r
ath
er
th
an
b
ein
g
f
u
lly
co
llap
s
ed
b
y
atten
tio
n
p
o
o
lin
g
.
2
.
5
.
Cla
s
s
if
ica
t
io
n
net
wo
r
k
a
nd
k
no
wledg
e
dis
t
illa
t
io
n
T
h
e
5
1
2
-
d
im
en
s
io
n
al
f
u
s
ed
r
e
p
r
esen
tatio
n
is
f
ed
th
r
o
u
g
h
a
c
lass
if
icatio
n
h
ea
d
(
5
1
2
→2
5
6
→1
2
8
→2
)
.
A
lay
er
o
f
b
atch
n
o
r
m
aliza
tio
n
is
f
o
llo
wed
b
y
R
eL
U
an
d
0
.
5
d
r
o
p
o
u
t
f
o
r
ea
ch
h
id
d
e
n
lay
er
.
T
h
e
o
u
tp
u
t
lay
e
r
p
r
o
d
u
ce
s
class
p
r
o
b
ab
ilit
ies v
ia
So
f
tm
ax
:
(
|
)
=
(
_
·
ℎ
+
_
)
[
1
]
t
r
ain
in
g
u
s
es lab
el
-
s
m
o
o
th
e
d
c
r
o
s
s
-
en
tr
o
p
y
(
ε=
0
.
1
)
:
_
=
−
_
[
(
1
−
)
·
·
(
)
+
(
/
)
·
(
)
]
w
h
er
e
K=
2
.
T
h
e
A
d
am
o
p
tim
izer
[
20
]
with
lear
n
in
g
r
at
e
1
×1
0
⁻⁴
an
d
L2
weig
h
t
d
ec
ay
λ
=1
0
⁻⁵
[
21
]
,
ar
e
em
p
lo
y
ed
.
Gr
ad
ien
t c
lip
p
in
g
i
s
u
s
ed
to
co
n
tr
o
l f
o
r
ex
p
lo
d
in
g
g
r
ad
ien
ts
(
m
a
x
n
o
r
m
1
.
0
)
.
T
h
e
teac
h
er
m
o
d
el
is
d
is
till
e
d
to
th
e
s
tu
d
en
t
m
o
d
el
with
a
1
4
.
5
×
p
ar
am
eter
r
e
d
u
ctio
n
tech
n
iq
u
e
k
n
o
wn
as
k
n
o
wled
g
e
d
is
till
atio
n
[
1
7
]
,
[
19
]
(
120
.
9
M
p
a
r
am
eter
s
,
4
6
1
MB
FP
3
2
to
8
.
4
M
p
ar
am
ete
r
s
,
3
1
.
9
4
MB FP3
2
)
.
T
h
e
s
tu
d
en
t m
in
i
m
izes:
=
·
_
(
,
_
)
+
(
1
−
)
·
²
·
_
(
(
_
ℎ
/
)
,
(
_
/
)
)
w
h
er
e
α
=0
.
3
is
a
b
alan
ce
b
et
wee
n
h
ar
d
-
lab
el
an
d
s
o
f
t
-
lab
el
s
u
p
er
v
is
io
n
,
T
=
4
.
0
is
th
e
d
is
till
atio
n
tem
p
er
atu
r
e
to
m
ak
e
teac
h
er
p
r
o
b
ab
ilit
y
d
i
s
tr
ib
u
tio
n
s
s
o
f
ter
in
o
r
d
er
to
r
ev
ea
l
in
ter
-
class
s
im
ilar
ity
s
tr
u
ctu
r
e,
σ
is
So
f
tm
ax
ap
p
lied
at
tem
p
er
atu
r
e
T
,
an
d
L
_
KL
is
th
e
Ku
llb
ac
k
-
L
eib
ler
(
KL
)
d
iv
er
g
e
n
ce
b
etwe
en
th
e
s
tu
d
en
t a
n
d
teac
h
e
r
s
o
f
t
d
is
tr
ib
u
tio
n
s
.
D
y
n
am
ic
q
u
an
tizatio
n
with
I
NT
8
[
18
]
,
[
2
2
]
is
t
h
en
u
s
ed
,
r
esu
ltin
g
in
a
to
tal
o
f
2
9
.
6
2
MB.
T
h
e
ex
ac
t q
u
an
tizatio
n
c
o
m
m
an
d
u
s
ed
was:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
3
2
2
1
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
,
Vo
l.
7
,
No
.
3
,
No
v
em
b
er
20
26
:
2
7
1
-
290
276
ℎ
.
.
_
(
_
,
{
.
,
.
}
,
=
ℎ
.
8
)
F
o
llo
wed
b
y
to
r
c
h
.
o
n
n
x
.
e
x
p
o
r
t(
)
with
o
p
s
et_
v
e
r
s
io
n
=1
3
.
T
h
e
s
y
s
tem
is
ex
p
o
r
ted
to
o
p
en
n
e
u
r
al
n
etwo
r
k
ex
ch
an
g
e
(
ONNX
)
f
o
r
m
at
in
o
r
d
er
to
b
e
d
ep
lo
y
ed
to
An
d
r
o
id
.
T
h
e
co
m
p
lete
tr
ain
i
n
g
a
n
d
d
is
till
atio
n
p
ip
elin
e
o
f
Qu
is
h
in
g
Sh
ield
is
p
r
o
v
id
ed
as
p
s
eu
d
o
co
d
e
in
th
e
p
r
o
ject
r
ep
o
s
ito
r
y
(
lin
k
to
b
e
p
r
o
v
id
e
d
wh
en
p
u
b
lis
h
ed
)
,
alo
n
g
with
p
r
e
-
tr
ain
ed
m
o
d
e
l
weig
h
ts
,
an
d
th
e
Qu
is
h
in
g
Sh
ield
An
d
r
o
i
d
APK
.
2
.
6
.
Alg
o
rit
hm
p
s
eudo
co
de
T
h
e
co
m
p
lete
Qu
is
h
in
g
Sh
ield
p
ip
elin
e
is
f
o
r
m
alize
d
i
n
Al
g
o
r
ith
m
s
1
-
4
.
Alg
o
r
ith
m
1
s
p
ec
if
ies
th
e
in
f
er
en
ce
f
o
r
wa
r
d
p
ass
;
Alg
o
r
ith
m
2
co
v
e
r
s
teac
h
er
tr
ain
in
g
;
Alg
o
r
ith
m
3
co
v
e
r
s
k
n
o
wl
ed
g
e
d
is
till
atio
n
an
d
q
u
an
tizatio
n
;
Alg
o
r
ith
m
4
d
escr
ib
es
th
e
p
r
iv
ac
y
-
p
r
eser
v
in
g
r
e
p
u
tatio
n
ca
ch
e.
All
h
y
p
er
p
ar
am
eter
s
a
n
d
en
v
ir
o
n
m
en
t v
e
r
s
io
n
s
ar
e
lis
ted
in
s
ec
tio
n
3
.
2
.
Alg
o
r
ith
m
1
.
Qu
is
h
in
g
Sh
ield
f
o
r
war
d
p
ass
(
in
f
e
r
en
ce
)
Input:
x_img
∈
ℝ
^(3×224×224)
-
QR poster image, ImageNet
-
normalized
x_text
∈
ℝ
^(1×128)
-
OCR
-
extracted token sequence (max 128 tokens)
x_url
∈
ℤ
^(1×256)
-
U
RL character index sequence, vocab size 79
x_net
∈
ℝ
^(1×5)
-
Normalized reputation scalar features
Output:
tier
∈
{SAFE, WARNING, BLOCK}
P_mal
∈
[0, 1]
—
Malicious probability
1.
e_v←Dropout (ReLU (BN (W_v · MobileNetV3(x_img)+b_v)))
// 576→256
-
dim visual embedding
2.
e_t←Dropout (ReLU (BN(W_t · DistilBERT_CLS(x_text)+b_t)))
// 768→256
-
dim; layers 1
–
3 frozen, layers 4
–
6 fine
-
tuned
3.
e_u←Dropout (ReLU (BN (W_u · MaxPool (Conv3
‖
Conv4
‖
Conv5(Embed(x_url)))+b_u)))
// 192→256
-
dim; three
parallel char
-
level convolutions
4.
e_n←Dropout (ReLU (BN (W_n2 · Dropout (ReLU (BN (W_n1 · x_net+b_n1)))+b_n2)))
// 5→128→256
-
dim network reputation embedding
5.
a_vt←MultiHeadAttn (Q=e_v, K=e_t, V=e_t, h=8, d_k=32) //
Brand
–
context mismatch
a_vn←MultiHeadAttn (Q=e_v, K=e_n, V=e_n, h=8, d_k=32) // Mimicry+poor infra
a_tn←MultiHeadAttn (Q=e_t, K=e_n, V=e_n, h=8, d_k=32) // Urgency+suspect domain
6.
z_fusion←Dropout (LayerNorm (W_f · Concat (e_v, e_t, e_u, e_n, a_vt,
a_vn,
a_tn)+b_f))
// 7×256=1792→512
-
dim fused representation
7.
h_1←Dropout (ReLU (BN(W_1·z_fusion+b_1))) // 512→256
h_2←Dropout (ReLU (BN(W_2·h_1+b_2)) // 256→128
logits←W_c·h_2+b_c // 128→2
8.
P_mal←Softmax(logits) [1]
9.
if P_mal<0.50 then tier←SAFE
elif P_mal<0.80 then tier←WARNING
else tie
r
←BLOCK
10.
Cache SHA256(url)→(P_mal, tier, TTL)
TTL=30 days if tier=BLOCK, else 7 days
Return:
tier, P_mal
Alg
o
r
ith
m
2
.
T
ea
ch
er
m
o
d
el
tr
ain
in
g
Input:
D={(x_img, x_text, x_url, x_net, y)}
—
labeled training set
T_max=50 epochs (early stopping δ=0.001, patience=10)
Hyperparams:
Adam optimizer [
16
], lr=1×10
⁻⁴
, L2 λ=10
⁻⁵
, batch=32
, ε=0.1
1.
Initialize all encoder, attention, and classification weights
2.
Freeze DistilBERT layers 1
–
3; set layers 4
–
6 trainable
3.
for epoch=1 to T_max:
for each mini
-
batch B
⊂
D:
with autocast(FP16):
P_teacher←ForwardPass(B) // Algorithm 1, steps 1
–
8
L←−Σ_y [(1−ε)·y·log(P) + (ε/K)·log(P)] // Label
-
smoothed CE
scaled_loss.backward()
clip_grad_norm_(params, max_norm=1.0) // Gr
adient clipping
optimizer.step() // FP32 parameter update
val_loss←Evaluate(D_val)
scheduler.step(val_loss)
//
ReduceL
ROnPlateau
(factor=0.1,
patience=5)
if |Δval_loss| <δ for patience consecutive epochs: break
4.
Save teacher_model.pt (120.9M params, 461 MB FP32)
Evaluation Warning : The document was created with Spire.PDF for Python.
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
I
SS
N:
2722
-
3
2
2
1
Qu
is
h
in
g
S
h
ield
:
o
n
-
d
ev
ice
mu
lti
-
mo
d
a
l d
etec
tio
n
o
f q
u
ick
r
e
s
p
o
n
s
e
p
h
is
h
in
g
(
S
h
a
b
o
u
r
B
a
n
d
a
)
277
Alg
o
r
ith
m
3
.
Kn
o
wled
g
e
d
is
till
atio
n
an
d
q
u
an
tizatio
n
Input:
teacher_model
-
pretrained, frozen (Algorithm 2)
D_train
-
training set
α = 0.3
-
hard
-
label vs. soft
-
label balance
T = 4.0
-
distillation temperature
1.
Initialize student with reduced encoder dimensions (14.5× fewer params than
teacher)
2.
Load frozen teacher_model; set requires_grad=False for all teacher params
3.
for epoch = 1 to 5:
for each mini
-
batch B
⊂
D_train:
P_student ← StudentForward(B)
L_CE ← CrossEntropy(y, P_student) // Hard
-
label
supervision
with torch.no_grad():
z_teacher ← teacher_model.get_logits(B)
z_student ← student_model.get_logits(B)
P_soft_T ← Softmax(z_teacher / T) // Soft teacher targets
P_soft_S ← Softmax(z_student / T)
L_KL ← KLDivergence(P_soft_T, P_soft_S) // Kullback
–
Leibler divergence
L ← α·L_CE + (1−α)·T²·L_KL // Combined distillation loss [
15
]
L.backward(); optimizer.step()
4.
Save student_model.pt (8.4M params, 31.94 MB
FP32)
5.
quantized ← torch.quantization.quantize_dynamic(student, {nn.Linear}, torch.qint8)
// Result: 29.62 MB [17
], [19
]
6.
torch.onnx.export(quantized, sample_input, 'quishingshield_v1.onnx',
opset_version=13)
// Also export: TorchScript, TFLite, CoreML, TF SavedModel
Alg
o
r
ith
m
4
.
Priv
ac
y
-
p
r
eser
v
i
n
g
r
ep
u
tatio
n
ca
ch
e
Data:
cache: HashMap<SHA256(url), (P_mal, tier, expiry)>
max_size: 10,000 entries |
eviction: LRU when full
function QueryOrScore(x_img, x_text, url, x_net):
1.
url_hash ← SHA256(url)
2.
if url_hash in cache AND cache[url_hash].expiry > now():
return cache[url_hash]
// Sub
-
millisecond cache hit; no inference
3.
tier, P_mal ← ForwardPass(x_img, x_text, url, x_net) // Algorithm 1
4.
TTL ← 30 days if tier=BLOCK else 7 days
cache[url_hash] ← (P_mal, tier, now() + TTL)
5.
if is_refresh_window(): // Background th
read, daily
for entry in cache where tier=BLOCK:
re_score against latest threat intelligence feed
6.
return tier, P_mal
// All operations on
-
device. Raw URL never transmitted.
// Satisfies GDPR data minimization and
Zimbabwe CDPA 2021, s.18 [
20]
,
[
2
1
]
2
.
7
.
Ris
k
cl
a
s
s
if
ica
t
io
n a
nd
priv
a
cy
-
preserv
ing
ca
che
T
h
e
s
ca
lar
p
r
o
b
ab
ilit
y
P
(
m
a
licio
u
s
|
x
)
ca
n
b
e
m
a
p
p
ed
to
a
th
r
ee
-
tier
r
is
k
-
ass
ess
m
en
t:
i)
SAFE:
P(m
alicio
u
s
)
<0
.
5
0
-
UR
L
clas
s
if
ied
b
en
ig
n
,
u
s
er
p
r
o
ce
ed
s
n
o
r
m
ally
;
ii)
W
AR
NI
N
G:
0
.
5
0
≤
P(m
alicio
u
s
)
<0
.
8
0
-
ele
v
ated
r
is
k
d
etec
ted
,
p
er
-
m
o
d
ality
r
is
k
b
r
ea
k
d
o
wn
d
is
p
lay
ed
,
u
s
er
ad
v
is
ed
to
c
h
ec
k
th
e
d
esti
n
atio
n
b
ef
o
r
e
g
o
in
g
ah
ea
d
;
iii)
B
L
OC
K:
P(m
alicio
u
s
)
≥
0
.
8
0
-
h
ig
h
co
n
f
id
en
ce
-
m
alicio
u
s
-
n
a
v
ig
atio
n
b
l
o
ck
ed
-
d
etaile
d
th
r
ea
t
ex
p
la
n
atio
n
d
is
p
lay
ed
.
T
h
e
o
n
-
d
ev
ice
p
r
iv
ac
y
-
p
r
e
s
er
v
in
g
r
e
p
u
tatio
n
ca
ch
e
s
to
r
es
u
p
to
1
0
,
0
0
0
s
ec
u
r
e
h
ash
alg
o
r
ith
m
(
SHA
)
-
256
(
UR
L
)
-
to
-
r
is
k
-
s
co
r
e
m
ap
p
in
g
s
with
d
if
f
er
en
tiated
t
im
e
-
to
-
liv
e
(
TTL
)
p
o
licies
(
3
0
d
ay
s
f
o
r
B
L
OC
K
class
if
icatio
n
s
,
7
d
ay
s
f
o
r
b
e
n
ig
n
)
.
All
in
f
er
e
n
ce
a
n
d
ca
c
h
i
n
g
r
u
n
ex
clu
s
iv
ely
on
-
d
e
v
ice,
s
atis
f
y
in
g
g
en
e
r
al
d
ata
p
r
o
tectio
n
r
eg
u
latio
n
(
G
DPR
)
d
ata
m
in
im
izatio
n
r
eq
u
ir
em
en
ts
[
2
3
]
a
n
d
Z
im
b
ab
we
’
s
cy
b
e
r
an
d
d
ata
p
r
o
tectio
n
ac
t
(
C
DPA)
[
2
4
]
.
3.
E
XP
E
R
I
M
E
N
T
A
L
SE
T
UP
3
.
1
.
Da
t
a
s
et
c
o
ns
t
ruct
io
n
Qu
is
h
in
g
Sh
ield
h
as 2
0
5
,
4
8
8
s
am
p
les o
f
QR
co
d
e
p
o
s
ter
s
co
l
lecte
d
f
r
o
m
J
u
ly
2
0
2
5
to
Feb
r
u
ar
y
2
0
2
6
.
T
o
o
u
r
k
n
o
wled
g
e,
it
is
th
e
lar
g
est
QR
co
d
e
d
ataset
p
u
b
lis
h
ed
s
o
f
ar
,
d
ed
icate
d
to
th
e
q
u
is
h
in
g
attac
k
.
T
a
b
le
2
g
iv
es su
m
m
ar
y
s
tatis
tics
.
C
o
llectio
n
p
r
o
ce
s
s
:
m
alicio
u
s
s
am
p
les
wer
e
s
o
u
r
ce
d
v
ia
th
r
ee
ch
an
n
els:
i)
au
to
m
ated
cr
awlin
g
o
f
Ph
is
h
T
an
k
an
d
Op
en
Ph
is
h
f
ee
d
s
to
r
etr
iev
e
co
n
f
ir
m
ed
p
h
is
h
in
g
UR
L
s
,
f
o
llo
wed
b
y
g
en
er
atio
n
o
f
QR
co
d
es
em
b
ed
d
e
d
in
c
o
n
tex
tu
ally
r
e
n
d
er
ed
p
o
s
ter
tem
p
lates
(
b
an
k
in
g
,
p
ar
ce
l
d
eli
v
er
y
,
tech
n
o
l
o
g
y
im
p
er
s
o
n
atio
n
p
atter
n
s
)
;
ii)
m
an
u
al
co
llect
io
n
o
f
co
n
f
ir
m
e
d
Qu
is
h
in
g
ca
m
p
aig
n
ar
tifa
cts
p
r
o
v
id
e
d
b
y
cy
b
er
s
ec
u
r
ity
in
tellig
en
ce
p
ar
tn
er
s
u
n
d
e
r
s
ig
n
ed
d
ata
-
s
h
ar
i
n
g
ag
r
ee
m
e
n
t
s
;
an
d
iii)
p
o
s
t
-
p
r
o
ce
s
s
in
g
o
f
p
u
b
licly
av
ailab
le
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
3
2
2
1
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
,
Vo
l.
7
,
No
.
3
,
No
v
em
b
er
20
26
:
2
7
1
-
290
278
Kag
g
le
p
h
is
h
in
g
UR
L
d
at
asets
,
f
r
o
m
wh
ich
p
o
s
ter
-
b
ased
s
y
n
th
etic
QR
ar
tifa
cts
wer
e
g
en
er
ated
to
ex
ten
d
co
v
er
ag
e
o
f
s
o
p
h
is
ticatio
n
tier
s
.
B
en
ig
n
s
am
p
les
wer
e
co
llected
f
r
o
m
o
f
f
icial
o
r
g
an
i
za
tio
n
al
web
s
ites
,
v
er
if
ied
co
m
m
er
cial
QR
ca
m
p
aig
n
s
,
an
d
p
u
b
lic
ev
en
t
m
at
er
ials
ac
r
o
s
s
4
5
co
u
n
t
r
ies,
with
a
s
u
b
s
et
s
o
u
r
ce
d
f
r
o
m
Kag
g
le
b
en
ig
n
UR
L
d
ata
s
ets.
E
ac
h
b
en
ig
n
QR
co
d
e
wa
s
v
alid
ated
b
y
r
eso
lv
i
n
g
its
UR
L
an
d
co
n
f
ir
m
in
g
n
etwo
r
k
r
e
p
u
tatio
n
s
co
r
es a
b
o
v
e
th
r
esh
o
ld
.
B
en
ig
n
s
am
p
les
wer
e
tak
e
n
f
r
o
m
th
e
leg
itima
te
co
m
m
er
cial
QR
ca
m
p
aig
n
s
a
n
d
f
r
o
m
t
h
e
m
ater
ials
g
ath
er
ed
f
r
o
m
o
f
f
icial
o
r
g
an
iz
atio
n
web
s
ites
ac
r
o
s
s
4
5
co
u
n
tr
ies
in
r
eg
ar
d
to
p
u
b
lic
ev
en
t
s
.
A
s
m
all
n
u
m
b
er
o
f
s
am
p
les
wer
e
also
o
b
tain
ed
f
r
o
m
t
h
e
a
v
ailab
le
p
u
b
lic
d
atasets
o
n
Kag
g
le
t
o
e
n
h
a
n
ce
co
v
er
ag
e
an
d
v
ar
iatio
n
o
f
b
e
n
ig
n
s
am
p
les.
QR
co
d
e
v
alid
atio
n
was
p
er
f
o
r
m
ed
s
y
s
tem
atica
lly
,
wh
er
eb
y
QR
co
d
es
wer
e
r
eso
lv
ed
an
d
m
atch
e
d
with
t
h
e
r
esp
ec
tiv
e
UR
L
o
f
t
h
e
le
g
itima
te
lan
d
in
g
p
ag
e,
an
d
t
h
e
web
s
ite
was
th
en
ch
ec
k
ed
f
o
r
n
etwo
r
k
r
ep
u
tatio
n
to
ce
r
tify
th
e
QR
co
d
e.
An
n
o
tatio
n
g
u
id
elin
es
a
n
d
v
a
lid
atio
n
:
th
r
ee
i
n
d
ep
e
n
d
en
t
a
n
n
o
tato
r
s
with
c
y
b
er
s
ec
u
r
ity
an
d
d
i
g
ital
f
o
r
en
s
ics
b
ac
k
g
r
o
u
n
d
s
,
lab
ele
d
all
th
e
s
am
p
les
f
r
o
m
v
is
u
al,
tex
tu
al
an
d
UR
L
asp
ec
t
s
.
All
th
e
s
am
p
les
wer
e
an
n
o
tated
h
o
lis
tically
b
y
th
r
ee
in
d
ep
en
d
en
t
an
n
o
tato
r
s
,
b
o
th
with
cy
b
er
s
ec
u
r
ity
a
n
d
d
ig
ital
f
o
r
en
s
ics
b
ac
k
g
r
o
u
n
d
s
.
T
h
e
an
n
o
tatio
n
p
r
o
to
co
l
r
e
q
u
ir
e
d
an
n
o
tato
r
s
to
d
o
cu
m
e
n
t
th
e
p
r
im
ar
y
d
ec
e
p
tio
n
v
ec
to
r
(
v
is
u
al
b
r
an
d
im
p
e
r
s
o
n
atio
n
,
u
r
g
e
n
cy
lan
g
u
ag
e,
UR
L
an
o
m
aly
,
o
r
i
n
f
r
astru
ctu
r
e
q
u
ality
)
f
o
r
ea
ch
m
alicio
u
s
s
am
p
le.
A
f
o
u
r
th
ex
p
er
t
r
eso
lv
ed
all
d
i
s
ag
r
ee
m
en
ts
b
y
ad
ju
d
icatio
n
.
T
h
e
r
esu
ltin
g
in
ter
-
an
n
o
tato
r
a
g
r
ee
m
en
t
(
κ=
0
.
8
9
)
s
u
b
s
tan
tially
ex
ce
ed
s
th
e
κ=
0
.
7
0
ac
ce
p
tab
ilit
y
th
r
esh
o
ld
[
2
5
]
,
co
n
f
ir
m
in
g
lab
el
r
eliab
ilit
y
.
E
th
ical
ap
p
r
o
v
als
an
d
d
ata
g
o
v
er
n
an
ce
:
m
alicio
u
s
s
am
p
le
c
o
llectio
n
f
r
o
m
Ph
is
h
T
a
n
k
a
n
d
Op
en
Ph
is
h
was
co
n
d
u
cted
u
n
d
e
r
th
eir
r
e
s
p
ec
tiv
e
o
p
en
-
d
ata
ter
m
s
.
C
o
llectio
n
f
r
o
m
cy
b
e
r
s
ec
u
r
ity
i
n
tellig
en
ce
p
ar
tn
e
r
s
was
co
n
d
u
cted
u
n
d
er
s
ig
n
e
d
in
s
titu
tio
n
al
d
ata
-
s
h
ar
in
g
ag
r
ee
m
en
ts
th
at
p
r
o
h
ib
it
d
is
clo
s
u
r
e
o
f
p
a
r
tn
e
r
id
en
titi
es.
No
p
er
s
o
n
al
d
ata
wer
e
co
llected
an
d
all
UR
L
s
ar
e
p
u
b
licly
r
ep
o
r
ted
th
r
ea
t
in
d
ic
ato
r
s
.
B
en
ig
n
UR
L
cr
awlin
g
co
m
p
lied
with
ea
ch
tar
g
et
s
ite
’
s
r
o
b
o
ts
.
tx
t.
An
an
o
n
y
m
ized
d
ataset
m
a
n
if
est
will
b
e
r
elea
s
ed
with
th
e
r
ep
o
s
ito
r
y
.
Up
o
n
p
u
b
licatio
n
,
th
e
d
ataset
will
b
e
av
ail
ab
le
at
h
ttp
s
://d
o
i.o
r
g
/1
0
.
x
x
x
x
/x
x
x
x
(
DOI
to
b
e
ass
ig
n
ed
)
.
Data
f
ilter
in
g
a
n
d
q
u
ality
co
n
tr
o
l:
d
u
p
licate
UR
L
s
r
em
o
v
ed
b
y
SHA
-
2
5
6
h
ash
co
m
p
a
r
is
o
n
,
d
ata
f
ilter
in
g
an
d
q
u
ality
co
n
tr
o
l.
Sam
p
les
with
Q
R
co
d
e
d
ec
o
d
in
g
f
ai
lu
r
e
r
ate
>5
%
wer
e
ex
clu
d
ed
f
r
o
m
th
r
ee
d
ec
o
d
in
g
lib
r
ar
ies.
B
elo
w
a
c
o
n
f
id
en
ce
th
r
esh
o
l
d
o
f
0
.
7
0
O
C
R
-
ex
tr
ac
ted
tex
t
was
m
ar
k
ed
as
lo
w
-
co
n
f
i
d
en
c
e
an
d
was n
o
t u
s
ed
f
o
r
t
r
ain
in
g
t
h
e
tex
tu
al
en
c
o
d
er
,
o
n
ly
UR
L
an
d
v
is
u
al
s
ig
n
als we
r
e
u
s
ed
f
o
r
s
u
ch
s
am
p
les.
Kn
o
wn
b
iases
an
d
m
itig
atio
n
p
lan
:
two
lim
itatio
n
s
a
p
p
ly
t
o
g
en
e
r
aliza
b
ilit
y
.
First,
th
er
e
ar
e
ab
o
u
t
7
0
%
o
f
m
alicio
u
s
s
am
p
les
th
at
ar
e
p
o
s
ter
-
r
en
d
e
r
ed
in
s
tead
o
f
ca
p
tu
r
ed
in
au
t
h
en
tic
d
ep
lo
y
m
en
t
en
v
ir
o
n
m
en
ts
(
p
h
y
s
ical
s
ig
n
ag
e
an
d
em
b
e
d
d
ed
em
ails
)
.
T
h
e
m
o
d
el
m
ig
h
t
b
e
o
p
tim
is
tically
ca
lib
r
ated
f
o
r
clea
n
d
ig
ital
im
ag
es,
wh
er
ea
s
QR
co
d
e
q
u
ality
an
d
lig
h
tin
g
m
ig
h
t
v
ar
y
m
o
r
e
i
n
r
ea
l
wo
r
l
d
q
u
is
h
in
g
p
o
s
ter
s
.
Plan
n
e
d
m
itig
atio
n
:
ex
p
an
d
t
h
e
d
at
aset
with
p
h
y
s
ically
p
h
o
to
g
r
ap
h
e
d
Qu
is
h
in
g
s
am
p
les
in
a
s
u
b
s
eq
u
en
t
d
ata
-
co
llect
io
n
ca
m
p
aig
n
.
Sec
o
n
d
,
alth
o
u
g
h
t
h
e
d
ata
p
r
o
v
id
ed
co
v
e
r
s
4
5
co
u
n
tr
ies,
E
n
g
lis
h
-
lan
g
u
a
g
e
p
o
s
ter
s
ac
co
u
n
t
f
o
r
6
7
%
o
f
th
e
s
am
p
les,
s
o
th
e
d
etec
tio
n
ca
p
ab
i
liti
es
ar
e
n
o
t
as
g
o
o
d
f
o
r
th
e
o
th
er
lan
g
u
ag
es,
esp
ec
ially
th
e
lo
w
r
eso
u
r
ce
l
an
g
u
ag
es
w
h
er
e
t
h
er
e
a
r
e
f
e
we
r
p
h
is
h
in
g
s
am
p
les
in
tellig
en
ce
f
ee
d
s
m
ay
b
e
s
o
m
ewh
at
wea
k
er
th
a
n
th
e
a
g
g
r
eg
ate
m
etr
ics
s
u
g
g
est.
T
h
e
b
iases
ar
e
b
o
th
r
ea
s
o
n
ab
le
p
o
t
en
tial
ca
n
d
id
ates
f
o
r
ex
p
an
s
io
n
o
f
th
e
d
ata
s
ets.
Plan
n
ed
m
itig
atio
n
:
co
llab
o
r
ate
with
r
eg
io
n
al
cy
b
e
r
s
ec
u
r
ity
p
ar
tn
er
s
to
g
et
lan
g
u
ag
e
s
am
p
les
f
o
r
u
n
d
er
-
r
ep
r
esen
ted
lan
g
u
ag
es.
C
lass
b
alan
ce
:
th
e
d
ataset
is
alm
o
s
t
p
er
f
ec
tly
b
alan
ce
d
(
5
1
.
2
%
m
alicio
u
s
,
4
8
.
8
%
b
e
n
ig
n
)
,
r
eso
lv
in
g
th
e
p
r
o
b
lem
o
f
class
im
b
alan
ce
,
wh
ic
h
n
eg
ativ
ely
im
p
ac
ts
u
n
im
o
d
al
UR
L
class
if
ier
s
u
s
in
g
asy
m
m
etr
ic
d
ataset
s
[
2
6
]
.
3
.
2
.
T
ra
ini
ng
co
nfig
ura
t
io
n
All
m
o
d
els
wer
e
tr
ain
e
d
o
n
G
o
o
g
le
C
o
lab
Pro
with
Py
T
o
r
ch
2
.
0
C
UDA
1
1
.
8
o
n
a
NVI
DI
A
T
esla
T
4
GPU
(
1
5
.
6
4
GB
VR
AM
)
.
T
h
e
r
a
n
d
o
m
s
ee
d
w
as
f
ix
ed
at
4
2
(
to
r
ch
.
m
an
u
al_
s
ee
d
(
4
2
)
;
n
u
m
p
y
.
r
an
d
o
m
.
s
ee
d
(
4
2
)
;
r
an
d
o
m
.
s
ee
d
(
4
2
)
)
f
o
r
all
e
x
p
er
im
en
ts
.
T
h
e
Ad
am
o
p
tim
izer
[
20
]
was
u
s
ed
with
a
lear
n
in
g
r
ate
o
f
α
=1
×
1
0
⁻⁴,
L
2
weig
h
t
d
ec
ay
o
f
λ
=1
0
⁻⁵
[
21
]
,
an
d
R
ed
u
ce
L
R
On
Plateau
s
ch
ed
u
lin
g
(
f
ac
to
r
0
.
1
,
p
atien
ce
5
ep
o
ch
s
)
.
B
atch
s
ize
was
3
2
.
Mix
ed
-
p
r
ec
is
io
n
tr
ain
in
g
(
FP
1
6
f
o
r
war
d
/b
ac
k
wa
r
d
,
FP
3
2
p
ar
am
eter
u
p
d
ates)
s
av
ed
a
b
o
u
t
5
0
% o
f
m
em
o
r
y
u
s
ag
e
an
d
in
cr
ea
s
ed
t
h
e
tr
ain
in
g
s
p
ee
d
b
y
ar
o
u
n
d
2
x
.
T
h
e
s
ix
r
e
g
u
lar
izatio
n
s
tr
ateg
ies
u
s
ed
wer
e:
d
r
o
p
o
u
t
(
r
ate
0
.
3
f
o
r
en
co
d
er
s
,
0
.
5
f
o
r
th
e
f
u
s
io
n
an
d
class
if
icatio
n
lay
er
s
)
,
b
atch
n
o
r
m
ali
za
tio
n
af
ter
ea
ch
lin
ea
r
p
r
o
jectio
n
,
L
2
weig
h
t
d
ec
ay
,
lab
el
s
m
o
o
th
in
g
(
ε=
0
.
1
)
,
g
r
a
d
ien
t
clip
p
in
g
(
m
ax
n
o
r
m
1
.
0
)
an
d
ea
r
ly
s
to
p
p
in
g
(
δ=0
.
0
0
1
,
p
atien
ce
1
0
e
p
o
ch
s
)
.
Kn
o
wled
g
e
d
is
till
atio
n
was
p
er
f
o
r
m
ed
f
o
r
5
ep
o
c
h
s
with
α
=0
.
3
a
n
d
T
=4
.
0
.
I
NT
8
d
y
n
am
ic
q
u
a
n
tizat
io
n
was
ap
p
lied
p
o
s
t
-
d
is
till
atio
n
u
s
in
g
to
r
ch
.
q
u
an
tizatio
n
.
q
u
an
tize_
d
y
n
am
ic
(
Py
T
o
r
ch
2
.
0
b
u
ilt
-
in
API
)
.
Fu
ll
h
y
p
er
p
ar
am
ete
r
s
ea
r
ch
r
an
g
es a
n
d
s
en
s
itiv
ity
an
aly
s
is
ar
e
p
r
o
v
i
d
ed
in
t
h
e
s
u
p
p
lem
en
tar
y
m
ate
r
ials
.
Evaluation Warning : The document was created with Spire.PDF for Python.
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
I
SS
N:
2722
-
3
2
2
1
Qu
is
h
in
g
S
h
ield
:
o
n
-
d
ev
ice
mu
lti
-
mo
d
a
l d
etec
tio
n
o
f q
u
ick
r
e
s
p
o
n
s
e
p
h
is
h
in
g
(
S
h
a
b
o
u
r
B
a
n
d
a
)
279
T
ab
le
2
.
Qu
is
h
in
g
Sh
ield
d
ataset
s
u
m
m
ar
y
in
clu
d
in
g
s
am
p
le
co
u
n
ts
,
class
b
alan
ce
,
g
eo
g
r
ap
h
ic
d
iv
er
s
ity
,
lan
g
u
ag
e
c
o
v
er
a
g
e,
an
d
an
n
o
t
atio
n
ag
r
ee
m
e
n
t
A
t
t
r
i
b
u
t
e
V
a
l
u
e
To
t
a
l
(
%)
N
o
t
e
s
To
t
a
l
sam
p
l
e
s
2
0
5
,
4
8
8
1
0
0
Ju
l
y
2
0
2
5
-
F
e
b
2
0
2
6
M
a
l
i
c
i
o
u
s
1
0
5
,
2
5
9
5
1
.
2
P
h
i
s
h
i
n
g
/
Q
u
i
s
h
i
n
g
B
e
n
i
g
n
1
0
0
,
2
2
9
4
8
.
8
V
e
r
i
f
i
e
d
l
e
g
i
t
i
m
a
t
e
Tr
a
i
n
i
n
g
sp
l
i
t
1
4
3
,
8
4
1
70
S
t
r
a
t
i
f
i
e
d
r
a
n
d
o
m
V
a
l
i
d
a
t
i
o
n
s
p
l
i
t
3
0
,
8
2
3
15
H
y
p
e
r
p
a
r
a
me
t
e
r
t
u
n
i
n
g
Te
st
sp
l
i
t
3
0
,
8
2
4
15
Te
mp
o
r
a
l
h
o
l
d
o
u
t
G
e
o
g
r
a
p
h
i
c
d
i
v
e
r
si
t
y
4
5
c
o
u
n
t
r
i
e
s
-
6
c
o
n
t
i
n
e
n
t
s
La
n
g
u
a
g
e
c
o
v
e
r
a
g
e
2
3
l
a
n
g
u
a
g
e
s
-
6
7
%
E
n
g
l
i
sh
I
n
t
e
r
-
a
n
n
o
t
a
t
o
r
a
g
mt
.
κ
=
0
.
8
9
-
3
i
n
d
e
p
e
n
d
e
n
t
a
n
n
o
t
a
t
o
r
s
C
r
a
w
l
e
d
U
R
Ls
(
mal
i
c
i
o
u
s)
~
6
2
,
0
0
0
-
P
h
i
s
h
Ta
n
k
/
O
p
e
n
P
h
i
s
h
f
e
e
d
s
P
o
st
e
r
-
g
e
n
e
r
a
t
e
d
(
m
a
l
.
)
~
4
3
,
2
5
9
-
S
y
n
t
h
e
t
i
c
c
o
n
t
e
x
t
u
a
l
i
ma
g
e
s
P
o
st
e
r
-
g
e
n
e
r
a
t
e
d
(
b
e
n
.
)
~
1
0
0
,
2
2
9
-
O
f
f
i
c
i
a
l
a
n
d
c
o
mm
e
r
c
i
a
l
so
u
r
c
e
s
S
o
p
h
i
s
t
i
c
a
t
i
o
n
:
b
a
s
i
c
1
8
,
0
5
1
18
U
R
L
s
h
o
r
t
e
n
e
r
s,
u
r
g
e
n
c
y
c
u
e
s
S
o
p
h
i
s
t
i
c
a
t
i
o
n
:
m
o
d
e
r
a
t
e
5
2
,
1
4
8
52
Ty
p
o
s
q
u
a
t
t
i
n
g
,
h
o
m
o
g
r
a
p
h
s
S
o
p
h
i
s
t
i
c
a
t
i
o
n
:
a
d
v
a
n
c
e
d
3
0
,
0
8
6
30
C
o
m
p
r
o
mi
s
e
d
si
t
e
s,
S
S
L
c
e
r
t
s
3
.
3
.
E
v
a
lua
t
i
o
n
pr
o
t
o
co
l
C
las
s
if
icatio
n
m
etr
ics:
a
cc
u
r
a
cy
,
p
r
ec
is
io
n
,
r
ec
all,
F1
s
co
r
e
an
d
s
p
ec
if
icity
ar
e
r
ep
o
r
ted
o
n
h
eld
-
o
u
t
tem
p
o
r
al
test
s
et
(
n
=
3
0
,
8
2
4
)
.
R
ec
all
is
weig
h
ted
ab
o
v
e
p
r
ec
is
io
n
in
th
is
co
n
tex
t
b
ec
au
s
e
f
alse
n
eg
ativ
es,
m
is
s
ed
p
h
is
h
in
g
attac
k
s
,
ca
r
r
y
m
ater
ially
h
ig
h
e
r
o
p
e
r
atio
n
al
co
s
t
th
an
f
alse
alar
m
s
.
A
s
y
s
t
em
th
at
lets
2
%
o
f
attac
k
s
th
r
o
u
g
h
ca
n
n
o
t b
e
c
o
n
s
id
er
ed
s
ec
u
r
e
r
eg
a
r
d
less
o
f
its
p
r
ec
is
io
n
.
S
t
a
t
i
s
t
i
c
a
l
v
a
l
i
d
a
t
i
o
n
:
b
o
o
t
s
t
r
a
p
p
i
n
g
(
1
0
,
0
0
0
t
i
m
e
s
r
e
s
a
m
p
l
i
n
g
)
i
s
d
o
n
e
f
o
r
a
l
l
m
a
j
o
r
m
e
t
r
i
c
s
,
a
s
s
h
o
w
n
i
n
T
a
b
l
e
3
.
T
h
e
p
a
i
r
w
i
s
e
m
o
d
e
l
c
o
m
p
a
r
i
s
o
n
(
p
<
0
.
0
0
1
t
h
r
e
s
h
o
l
d
)
i
s
c
a
r
r
i
e
d
o
u
t
u
s
i
n
g
M
c
N
e
m
a
r
’
s
t
e
s
t
.
T
h
r
o
u
g
h
o
u
t
t
h
e
d
o
c
u
m
e
n
t
e
f
f
e
c
t
s
i
z
e
s
(
C
o
h
e
n
’
s
h
f
o
r
p
r
o
p
o
r
t
i
o
n
s
)
a
r
e
r
e
p
o
r
t
e
d
a
l
o
n
g
w
i
t
h
p
-
v
a
l
u
e
s
.
N
o
m
u
l
t
i
p
l
e
-
c
o
m
p
a
r
i
s
o
n
c
o
r
r
e
c
t
i
o
n
i
s
a
p
p
l
i
e
d
,
a
s
e
a
c
h
c
o
m
p
a
r
i
s
o
n
a
d
d
r
e
s
s
e
s
a
d
i
s
t
i
n
c
t
p
r
e
-
r
e
g
i
s
t
e
r
e
d
r
e
s
e
a
r
c
h
q
u
e
s
t
i
o
n
.
A
d
v
e
r
s
a
r
i
a
l
r
o
b
u
s
t
n
e
s
s
t
e
s
t
i
n
g
p
r
o
t
o
c
o
l
:
f
o
u
r
c
a
t
e
g
o
r
i
e
s
o
f
a
t
t
a
c
k
s
w
e
r
e
t
e
s
t
e
d
:
i
)
U
R
L
o
b
f
u
s
c
a
t
i
o
n
u
s
i
n
g
s
h
o
r
t
e
n
e
r
s
(
b
i
t
.
l
y
,
t
i
n
y
u
r
l
.
c
o
m
)
;
i
i
)
h
o
m
o
g
r
a
p
h
s
u
b
s
t
i
t
u
t
i
o
n
(
C
y
r
i
l
l
i
c
/
G
r
e
e
k
c
h
a
r
a
c
t
e
r
s
r
e
p
l
a
c
i
n
g
A
m
e
r
i
c
a
n
s
t
a
n
d
a
r
d
c
o
d
e
f
o
r
i
n
f
o
r
m
a
t
i
o
n
i
n
t
e
r
c
h
a
n
g
e
(
A
S
C
I
I
)
l
o
o
k
a
l
i
k
e
s
i
n
d
o
m
a
i
n
n
a
m
e
s
)
;
i
i
i
)
A
I
-
g
e
n
e
r
a
t
e
d
p
e
r
s
u
a
s
i
v
e
p
o
s
t
e
r
t
e
x
t
d
e
s
i
g
n
e
d
t
o
m
a
x
i
m
i
z
e
u
r
g
e
n
c
y
s
i
g
n
a
l
s
w
h
i
l
e
e
v
a
d
i
n
g
k
e
y
w
o
r
d
-
b
a
s
e
d
f
i
l
t
e
r
s
;
a
n
d
i
v
)
c
o
m
b
i
n
e
d
a
t
t
a
c
k
s
u
s
i
n
g
a
l
l
t
h
r
e
e
e
v
a
s
i
o
n
s
t
r
a
t
e
g
i
e
s
s
i
m
u
l
t
a
n
e
o
u
s
l
y
.
T
h
e
a
d
v
e
r
s
a
r
i
a
l
s
a
m
p
l
e
s
(
n
=
2
,
0
0
0
p
e
r
c
a
t
e
g
o
r
y
)
w
e
r
e
c
r
e
a
t
e
d
t
h
r
o
u
g
h
a
u
t
o
m
a
t
e
d
m
e
a
n
s
:
T
y
p
o
S
q
u
a
t
t
e
r
(
d
e
f
a
u
l
t
s
e
t
t
i
n
g
s
,
h
o
m
o
g
r
a
p
h
s
u
b
s
t
i
t
u
t
i
o
n
w
i
t
h
C
y
r
i
l
l
i
c
a
n
d
G
r
e
e
k
l
o
o
k
a
l
i
k
e
s
)
f
o
r
h
o
m
o
g
r
a
p
h
i
c
a
t
t
a
c
k
s
,
a
n
d
G
P
T
-
4
(
t
e
m
p
e
r
a
t
u
r
e
=
0
.
7
,
1
0
p
r
o
m
p
t
i
n
g
t
e
m
p
l
a
t
e
s
p
e
r
c
a
t
e
g
o
r
y
)
f
o
r
A
I
-
g
e
n
e
r
a
t
e
d
pe
r
s
u
a
s
i
v
e
p
o
s
t
e
r
t
e
x
t
.
R
o
b
u
s
t
n
e
s
s
i
s
m
e
a
s
u
r
e
d
a
s
a
c
c
u
r
a
c
y
o
n
a
d
v
e
r
s
a
r
i
a
l
s
a
m
p
l
e
s
,
a
n
d
t
h
e
r
e
s
u
l
t
s
a
r
e
s
u
b
s
t
a
n
t
i
a
t
e
d
b
y
t
h
e
M
c
N
e
m
a
r
t
e
s
t
t
o
s
h
o
w
t
h
a
t
t
h
e
a
c
c
u
r
a
c
y
d
i
f
f
e
r
s
f
r
o
m
t
h
e
r
a
n
d
o
m
b
a
s
e
l
i
n
e
.
B
aseli
n
e
s
ele
cti
o
n
r
ati
o
n
al
e:
c
o
m
p
ar
ati
v
e
b
as
eli
n
es
r
e
p
r
es
en
t
t
h
e
f
u
ll
d
et
ec
t
io
n
p
a
r
a
d
i
g
m
s
p
ac
e
:
UR
L
b
la
c
k
lis
t
(
r
e
ac
ti
v
e,
u
n
i
m
o
d
a
l)
,
UR
L
-
o
n
l
y
m
u
l
ti
-
al
g
o
r
it
h
m
m
ac
h
i
n
e
l
ea
r
n
i
n
g
(
M
L
)
[
9
]
,
UR
L
-
o
n
l
y
t
r
a
n
s
f
o
r
m
e
r
-
b
as
ed
(
UR
L
T
r
a
n
[
2
7
]
)
,
v
is
u
a
l
QR
d
e
te
cti
o
n
[
2
8
]
,
a
n
d
t
h
e
t
wo
c
lo
s
est
m
u
lti
-
m
o
d
a
l
b
asel
i
n
es
i
n
t
h
e
lit
er
at
u
r
e
(V
is
u
a
lPh
is
h
Net
[
1
1
]
,
[
2
9
]
)
.
T
r
an
s
f
o
r
m
e
r
-
b
ase
d
m
u
l
ti
-
m
o
d
al
m
o
d
els
at
s
c
ale
(
c
o
n
t
r
asti
v
e
la
n
g
u
a
g
e
-
i
m
a
g
e
p
r
e
-
tr
ai
n
i
n
g
(
C
L
I
P
)
-
b
as
ed
p
h
i
s
h
i
n
g
d
et
ec
t
o
r
s
)
we
r
e
c
o
n
s
id
er
ed
b
u
t
e
x
cl
u
d
e
d
b
e
ca
u
s
e
n
o
p
u
b
l
is
h
e
d
s
y
s
te
m
ev
al
u
at
es
t
h
e
m
u
n
d
er
m
o
b
il
e
d
ep
l
o
y
m
e
n
t
co
n
s
t
r
a
in
ts
,
o
m
is
s
i
o
n
is
s
t
ate
d
e
x
p
lici
tl
y
a
n
d
t
h
e
co
m
p
a
r
is
o
n
g
ap
h
as
b
e
en
w
ell
d
o
cu
m
e
n
t
e
d
i
n
t
h
is
p
a
p
e
r
.
Mo
b
ile
d
ep
lo
y
m
en
t
m
ea
s
u
r
e
m
en
t
p
r
o
to
co
l:
in
f
e
r
en
ce
late
n
cy
was
m
ea
s
u
r
ed
o
n
a
Sam
s
u
n
g
Gala
x
y
A3
2
(
Qu
alco
m
m
Sn
a
p
d
r
a
g
o
n
6
8
0
,
4
GB
R
AM
)
u
s
in
g
An
d
r
o
id
p
r
o
f
iler
with
ONN
X
R
u
n
tim
e
1
.
1
6
.
1
.
2
0
war
m
-
u
p
r
u
n
s
(
d
is
ca
r
d
e
d
)
+1
0
0
tim
ed
r
u
n
s
(
1
0
0
-
s
ec
o
n
d
s
id
l
e
b
ef
o
r
e
ea
c
h
r
u
n
)
.
Me
an
late
n
cy
an
d
t
h
e
s
tan
d
ar
d
d
ev
iatio
n
ar
e
p
r
esen
ted
.
Peak
R
AM
wa
s
m
ea
s
u
r
ed
u
s
in
g
An
d
r
o
id
’
s
Deb
u
g
.
Me
m
o
r
y
I
n
f
o
A
PI.
Var
ian
ce
in
th
e
d
ev
ice
lev
el
was m
ea
s
u
r
ed
b
y
r
ep
licatin
g
ex
p
e
r
im
en
ts
o
n
Xiao
m
i Red
m
i 9
T
an
d
Sam
s
u
n
g
Gala
x
y
S2
2
.
T
ab
le
3
.
T
r
ai
n
in
g
c
o
n
f
ig
u
r
atio
n
an
d
k
ey
h
y
p
er
p
ar
am
eter
s
u
s
ed
f
o
r
teac
h
er
tr
ain
i
n
g
an
d
s
tu
d
en
t d
is
till
atio
n
P
a
r
a
me
t
e
r
V
a
l
u
e
F
r
a
mew
o
r
k
P
y
T
o
r
c
h
2
.
0
+
C
U
D
A
1
1
.
8
H
a
r
d
w
a
r
e
N
V
I
D
I
A
Te
sl
a
T4
(
1
5
.
6
4
G
B
V
R
A
M
)
O
p
t
i
mi
z
e
r
A
d
a
m [
17
]
,
l
r
=
1
e
-
4
,
L
2
w
d
=
1
e
-
5
[
20
]
LR
s
c
h
e
d
u
l
e
R
e
d
u
c
e
L
R
O
n
P
l
a
t
e
a
u
(
f
a
c
t
o
r
=
0
.
1
,
p
a
t
i
e
n
c
e
=
5
)
B
a
t
c
h
si
z
e
32
M
i
x
e
d
p
r
e
c
i
s
i
o
n
F
P
1
6
f
w
d
/
b
w
d
,
F
p
3
2
u
p
d
a
t
e
s
D
r
o
p
o
u
t
0
.
3
(
e
n
c
o
d
e
r
s)
,
0
.
5
(
f
u
si
o
n
/
c
l
a
ss
i
f
i
c
a
t
i
o
n
)
La
b
e
l
smo
o
t
h
i
n
g
e
p
s
=
0
.
1
G
r
a
d
i
e
n
t
c
l
i
p
p
i
n
g
max
_
n
o
r
m
=
1
.
0
Ea
r
l
y
st
o
p
p
i
n
g
d
e
l
t
a
=
0
.
0
0
1
,
p
a
t
i
e
n
c
e
=
1
0
e
p
o
c
h
s
K
D
t
e
m
p
e
r
a
t
u
r
e
T=
4
.
0
,
a
l
p
h
a
=
0
.
3
R
a
n
d
o
m s
e
e
d
4
2
(
t
o
r
c
h
,
n
u
mp
y
,
r
a
n
d
o
m)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
3
2
2
1
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
,
Vo
l.
7
,
No
.
3
,
No
v
em
b
er
20
26
:
2
7
1
-
290
280
4.
RE
SU
L
T
S
4
.
1
.
Cla
s
s
if
ica
t
io
n
p
er
f
o
r
m
a
nce
T
h
e
class
if
icatio
n
r
esu
lt
o
f
th
e
teac
h
er
an
d
s
tu
d
en
t
m
o
d
el
o
v
er
th
e
test
s
et
(
n
=3
0
,
8
2
4
)
is
r
ep
o
r
ted
in
T
ab
le
4
.
T
h
e
teac
h
er
ac
h
iev
es
9
6
.
3
7
%
ac
cu
r
ac
y
(
9
5
%
C
I
:
9
6
.
1
1
-
9
6
.
6
2
%)
an
d
9
8
.
1
0
%
r
ec
all
(
9
5
%
C
I
:
9
7
.
8
9
-
9
8
.
3
0
%),
ex
ce
e
d
in
g
all
f
iv
e
r
esear
ch
tar
g
et
th
r
esh
o
l
d
s
.
T
h
e
s
tu
d
en
t
’
s
a
cc
u
r
ac
y
is
9
5
.
5
5
%
(
9
5
%
C
I
:
9
5
.
2
8
-
9
5
.
8
2
%)
an
d
th
e
s
tu
d
en
t
’
s
r
ec
all
is
9
8
.
1
8
%
(
9
5
%
C
I
: 9
7
.
9
7
-
9
8
.
3
8
%)
an
d
is
s
tatis
tic
ally
g
r
ea
ter
th
an
th
e
teac
h
er
’
s
r
ec
all,
as
d
is
cu
s
s
ed
i
n
th
eo
r
y
in
s
ec
tio
n
5
.
2
.
I
t
h
as
a
tr
ad
e
-
o
f
f
o
f
0
.
8
2
p
p
at
9
1
.
6
%
o
f
s
ize
r
ed
u
ctio
n
(
3
5
2
MB
to
2
9
.
6
2
MB)
wh
ich
is
b
etter
th
an
p
u
b
lis
h
ed
co
m
p
r
ess
io
n
b
en
c
h
m
ar
k
s
f
o
r
co
m
p
ar
a
b
le
ar
ch
itectu
r
es
[
1
2
]
,
[
1
3
]
.
Fo
r
al
l
d
if
f
er
e
n
ce
s
r
elativ
e
to
th
e
U
R
L
b
lack
lis
t
b
aselin
e,
th
e
C
o
h
en
’
s
h
is
>0
.
1
4
a
n
d
th
e
χ²
is
>4
5
.
2
,
p
<0
.
0
0
1
(
Mc
N
em
ar
’
s
test
)
.
Mu
lti
-
p
an
el
v
is
u
al
co
m
p
a
r
is
o
n
s
ar
e
o
f
f
er
e
d
in
Fig
u
r
e
2
.
Fig
u
r
e
2
(
a
)
d
is
p
lay
s
b
ar
ch
a
r
ts
o
f
th
e
class
if
icatio
n
m
etr
ics
with
co
n
f
id
en
ce
in
ter
v
als,
wh
ile
Fig
u
r
es
2
(
b
)
a
n
d
2
(
c
)
s
h
o
w
th
e
c
o
n
f
u
s
io
n
m
atr
ices
an
d
m
o
d
el
s
izes,
r
esp
ec
tiv
ely
,
f
o
r
b
o
th
m
o
d
els.
Fig
u
r
e
2
(
d
)
co
m
p
ar
es
th
e
in
f
er
en
ce
s
p
ee
d
s
f
o
r
b
o
th
m
o
d
els
an
d
Fig
u
r
e
2
(
e
)
d
is
p
lay
s
th
e
I
NT
8
co
m
p
r
ess
io
n
r
atio
.
Fig
u
r
e
2
p
r
o
v
id
es
ad
d
itio
n
al
in
f
o
r
m
ati
o
n
ab
o
u
t
th
e
tr
en
d
s
an
d
p
atter
n
s
o
f
er
r
o
r
s
in
T
a
b
le
4
th
at
is
n
o
t c
o
n
v
ey
e
d
b
y
a
n
u
m
er
ical
s
u
m
m
ar
y.
T
ab
le
4
.
C
lass
if
icatio
n
p
er
f
o
r
m
an
ce
o
f
t
h
e
teac
h
er
a
n
d
s
tu
d
en
t m
o
d
els o
n
th
e
h
eld
-
o
u
t te
s
t set
M
e
t
r
i
c
Te
a
c
h
e
r
[
9
5
%
C
I
]
S
t
u
d
e
n
t
[
9
5
%
C
I
]
Δ
Ta
r
g
e
t
A
c
c
u
r
a
c
y
(
%)
9
6
.
3
7
[
9
6
.
1
1
-
9
6
.
6
2
]
9
5
.
5
5
[
9
5
.
2
8
-
9
5
.
8
2
]
-
0
.
8
2
>
9
5
%
✓
P
r
e
c
i
s
i
o
n
(
%)
9
5
.
3
0
[
9
4
.
9
9
-
9
5
.
6
1
]
9
5
.
0
0
[
9
4
.
6
8
-
9
5
.
3
1
]
-
0
.
3
0
>
9
0
%
✓
R
e
c
a
l
l
(
%)
9
8
.
1
0
[
9
7
.
8
9
-
9
8
.
3
0
]
9
8
.
1
8
[
9
7
.
9
7
-
9
8
.
3
8
]
+
0
.
0
8
>
9
5
%
✓
F1
-
s
c
o
r
e
(
%)
9
6
.
6
8
[
9
6
.
4
3
-
9
6
.
9
3
]
9
6
.
5
6
[
9
6
.
3
1
-
9
6
.
8
1
]
-
0
.
1
2
>
9
2
%
✓
S
p
e
c
i
f
i
c
i
t
y
(
%)
9
4
.
9
3
[
9
4
.
6
1
-
9
5
.
2
5
]
9
4
.
5
6
[
9
4
.
2
3
-
9
4
.
8
8
]
-
0
.
3
7
>
9
0
%
✓
A
d
v
.
r
o
b
u
st
n
e
ss
(
%)
9
8
.
7
5
[
9
8
.
5
5
-
9
8
.
9
4
]
9
5
.
0
0
[
9
4
.
7
1
-
9
5
.
2
9
]
-
3
.
7
5
>
9
5
%
✓
M
o
d
e
l
s
i
z
e
(
M
B
)
3
5
2
.
4
3
2
9
.
6
2
-
9
1
.
6
%
✓
<
5
0
M
B
✓
I
n
f
e
r
e
n
c
e
l
a
t
e
n
c
y
(
ms)
7
7
.
7
4
±
3
.
1
2
5
.
3
4
±
1
.
7
-
6
7
.
4
%
<
3
0
0
ms
✓
(
a)
(
b
)
(
c)
(
d
)
(
e)
Fig
u
r
e
2
.
Mu
lti
-
p
an
el
p
e
r
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
: te
ac
h
er
v
s
.
s
tu
d
en
t m
o
d
els
:
(
a)
class
if
icati
o
n
m
etr
ics with
b
o
o
ts
tr
ap
co
n
f
id
en
ce
in
ter
v
als
,
(
b
)
c
o
n
f
u
s
io
n
m
atr
ices
,
(
c)
m
o
d
el
s
ize
co
m
p
ar
is
o
n
,
(
d
)
in
f
er
en
ce
s
p
ee
d
,
an
d
(
e)
I
NT
8
q
u
an
tizatio
n
co
m
p
r
e
s
s
io
n
r
atio
an
d
r
esu
ltin
g
s
ize
r
ed
u
ctio
n
Evaluation Warning : The document was created with Spire.PDF for Python.