I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
d
van
c
e
s
i
n
A
p
p
li
e
d
S
c
ie
n
c
e
s
(
I
JA
A
S
)
V
ol
.
14
, N
o.
4
,
D
e
c
e
m
be
r
20
25
, pp.
1263
~
1280
I
S
S
N
:
2252
-
8814
,
D
O
I
:
10.11591/
ij
a
a
s
.
v14.
i
4
.
pp1263
-
12
80
1263
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
aas
.i
ae
s
c
or
e
.c
om
S
t
ac
k
i
n
g ar
c
h
i
t
e
c
t
u
r
e
-
e
n
d
p
oi
n
t
d
e
t
e
c
t
i
on
:
a
h
yb
r
i
d
m
u
l
t
i
-
l
aye
r
e
d
ar
c
h
i
t
e
c
t
u
r
e
f
or
e
n
d
p
oi
n
t
t
h
r
e
at
d
e
t
e
c
t
i
on
A
b
d
R
ah
m
an
Wah
id
1
,
D
e
s
i
A
n
ggr
e
an
i
1
,
M
u
h
yi
d
d
in
A
.
M
.
H
ayat
1
,
A
e
d
ah
A
b
d
R
ah
m
an
2
,
M
u
h
am
m
ad
F
ai
s
al
1
1
D
e
pa
r
t
m
e
nt
of
I
nf
or
m
a
t
i
c
s
,
F
a
c
ul
t
y of
E
ngi
ne
e
r
i
ng,
M
uha
m
m
a
di
ya
h U
ni
ve
r
s
i
t
y of
M
a
ka
s
s
a
r
, M
a
ka
s
s
a
r
, I
ndone
s
i
a
2
S
c
hool
of
S
c
i
e
nc
e
a
nd T
e
c
hnol
ogy, A
s
i
a
e
U
ni
ve
r
s
i
t
y, K
ua
l
a
L
um
pur
, M
a
l
a
y
s
i
a
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
J
ul
15, 2025
R
e
vi
s
e
d
S
e
p 25, 2025
A
c
c
e
pt
e
d
N
ov 4, 2025
Modern
endpoint
threat
detection
systems
face
pe
rsistent
challen
ges
in
balancing
detection
accuracy,
resilience
against
zero
-
day
attacks,
a
nd
the
interpreta
bility
of
artificia
l
intelligence
(AI)
models.
Although
deep
le
arning
(DL)
approaches
often
achieve
high
accuracy
on
benchmark
dataset
s,
they
remain
vulnerable
to
adversarial
perturbations
and
operate
as
opaque
“black
boxes,”
thereby
reducing
trust
and
limiting
practical
adoption
in
critical
infrastruc
tures.
This
resear
ch
introduces
st
acking
architecture
-
en
dpoint
detection
(
STACK
-
ED
)
,
a
h
ybrid
m
ulti
-
l
ayered
a
rchitecture
for
e
n
dpoint
t
hreat
d
etection.
STACK
-
ED
integrates
three
complementary
para
digms
:
supervised
learning
for
known
attack
patterns,
self
-
supervised
F
graph
-
based
learning
for
structura
l
relationships
,
and
unsupervise
d
anomaly
detect
ion
for
emerging
or
unknown
thr
eats.
The
output
s
are
consoli
dat
ed
by
a
meta
-
learner
,
followed
by
a
post
-
hoc
correction
(PHC)
mechanism
to
mi
nimize
false
negatives.
The
framework
was
evaluated
on
a
combined
benc
hmark
dataset
(CSE
-
CIC
-
IDS2018
and
UNSW
-
NB15,
hereafter
referred
to
as
HIDS
-
Set).
Exper
imental
results
demonstrate
state
-
of
-
the
-
art
performance,
achieving
an F2
-
score of 98.89
% after hybr
id integration and a
ctive learning,
with
the
primary
optimization
objective
being
the
reduction
of
und
etected
attacks.
Furtherm
ore,
the
Shapley
additiv
e
explanati
ons
(SHAP)
method
enhances
interpret
abilit
y
by
revealing
feature
contribu
tions,
while
th
e
PHC
successfully
recovere
d
62.64%
of
missed
zero
-
day
candidates.
The
fi
ndings
position
STACK
-
ED
not
only
as
a
highly
accurate
detection
model
b
ut
also
as
an
ada
ptive,
resilient,
and
transparent
framework,
offering
pr
actical
implications
for
enterpris
e
-
grade
endpoint
defense
and
future
zer
o
-
trust
cybersecurit
y syst
ems.
K
e
y
w
o
r
d
s
:
A
dve
r
s
a
r
ia
l
r
obus
tn
e
s
s
A
r
ti
f
ic
ia
l
in
te
ll
ig
e
nc
e
E
ndpoint
de
te
c
ti
on
G
r
a
ph
ne
twor
k
H
ybr
id
l
e
a
r
ni
ng
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
A
bd R
a
hm
a
n W
a
hi
d
D
e
pa
r
tm
e
nt
of
I
nf
or
m
a
ti
c
s
, F
a
c
ul
ty
of
E
ngi
ne
e
r
in
g
,
M
uha
m
m
a
di
ya
h U
ni
ve
r
s
it
y of
M
a
ka
s
s
a
r
J
l.
S
ul
ta
n A
la
uddi
n, M
a
ka
s
s
a
r
, S
ul
a
w
e
s
i
S
e
la
ta
n, I
ndone
s
ia
E
m
a
il
:
105841116522
@
s
tu
de
nt
.uni
s
m
uh.a
c
.i
d
1.
I
N
T
R
O
D
U
C
T
I
O
N
T
he
di
gi
ta
l
s
e
c
ur
it
y
la
nds
c
a
pe
i
s
in
c
r
e
a
s
in
gl
y
s
h
a
pe
d
by
th
e
d
yna
m
ic
s
of
a
gl
oba
l
c
ybe
r
a
r
m
s
r
a
c
e
,
w
he
r
e
z
e
r
o
-
da
y
or
unknown
a
tt
a
c
ks
ha
v
e
be
c
om
e
r
out
in
e
th
r
e
a
ts
to
c
r
it
ic
a
l
in
f
r
a
s
tr
uc
tu
r
e
a
nd
e
nt
e
r
pr
is
e
e
nvi
r
onm
e
nt
s
[
1]
,
[
2]
.
I
n
I
ndone
s
ia
a
lo
ne
,
m
or
e
th
a
n
403
m
il
li
o
n
tr
a
f
f
ic
a
nom
a
li
e
s
w
e
r
e
r
e
c
or
de
d
in
2023
[
3]
,
th
r
e
a
te
ni
ng e
s
s
e
nt
ia
l
s
e
c
to
r
s
s
uc
h a
s
ba
nki
ng a
nd publi
c
s
e
r
vi
c
e
s
w
it
h both e
c
onomi
c
a
nd s
o
c
ia
l
r
e
pe
r
c
us
s
io
n
s
[
4]
.
C
onve
nt
io
na
l
s
ig
na
tu
r
e
-
ba
s
e
d
e
ndpoint
de
te
c
ti
on
(
E
D
)
ha
s
pr
ove
n
in
e
f
f
e
c
ti
ve
,
r
e
ly
in
g
on
da
ta
ba
s
e
s
of
known
th
r
e
a
ts
a
nd
f
a
il
in
g
to
r
e
s
pond
to
e
m
e
r
gi
ng
a
da
pt
iv
e
a
tt
a
c
ks
[
5]
.
T
o
ove
r
c
om
e
th
e
s
e
li
m
it
a
ti
ons
,
r
e
c
e
nt
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8814
I
nt
J
A
dv A
ppl
S
c
i
,
V
ol
.
14
, N
o.
4
,
D
e
c
e
m
be
r
20
25
:
126
3
-
12
80
1264
s
tu
di
e
s
ha
ve
e
xpl
or
e
d
a
r
ti
f
ic
ia
l
in
te
ll
ig
e
nc
e
(
A
I
)
-
ba
s
e
d
e
ndpoint
de
te
c
ti
on,
w
hi
c
h
pr
om
is
e
s
a
da
pt
iv
e
a
nom
a
ly
r
e
c
ogni
ti
on
a
nd
in
te
ll
ig
e
nt
de
c
is
io
n
-
m
a
ki
ng.
Y
e
t,
f
ir
s
t
-
ge
ne
r
a
ti
on
A
I
m
ode
ls
s
ti
ll
f
a
c
e
pa
r
a
doxe
s
de
e
p
le
a
r
ni
ng
(
D
L
)
a
c
hi
e
ve
s
hi
gh
a
c
c
ur
a
c
y
on
be
n
c
hm
a
r
ks
but
r
e
m
a
in
s
opa
qu
e
,
vul
ne
r
a
bl
e
to
a
dve
r
s
a
r
ia
l
pe
r
tu
r
ba
ti
ons
,
a
nd
in
c
ons
is
te
nt
in
de
te
c
ti
ng
z
e
r
o
-
da
y
th
r
e
a
ts
[
6]
,
[
7]
.
T
he
s
e
c
ons
tr
a
in
ts
hi
nde
r
a
dopt
io
n
in
m
is
s
io
n
-
c
r
it
ic
a
l
c
ybe
r
s
e
c
ur
it
y
a
nd
e
m
pha
s
iz
e
th
e
ne
e
d
f
or
s
ys
te
m
s
th
a
t
a
r
e
bot
h
r
obus
t
a
nd
e
xpl
a
in
a
bl
e
.
D
e
s
pi
te
not
a
bl
e
a
dva
n
c
e
s
s
uc
h
a
s
e
ns
e
m
bl
e
le
a
r
ni
ng,
gr
a
ph
ne
ur
a
l
ne
twor
ks
(
G
N
N
s
)
,
a
dve
r
s
a
r
ia
l
r
obus
tn
e
s
s
(
A
dvR
)
,
a
nd
e
xpl
a
in
a
bl
e
A
I
(
X
A
I
)
,
e
xi
s
ti
ng
w
or
ks
r
e
m
a
in
f
r
a
gm
e
nt
e
d,
f
oc
us
in
g
on
i
s
ol
a
te
d
te
c
hni
que
s
r
a
th
e
r
th
a
n
c
ohe
s
iv
e
in
te
gr
a
ti
on.
A
c
r
it
ic
a
l
r
e
s
e
a
r
c
h
ga
p
,
th
e
r
e
f
or
e
,
pe
r
s
is
ts
:
th
e
a
bs
e
n
c
e
of
a
f
r
a
m
e
w
or
k
th
a
t
uni
te
s
hybr
id
l
e
a
r
ni
ng, s
tr
uc
tu
r
a
l
r
e
pr
e
s
e
nt
a
ti
on, a
nom
a
ly
de
te
c
ti
on, i
nt
e
r
pr
e
ta
bi
li
ty
, a
nd a
dve
r
s
a
r
ia
l
r
e
s
il
ie
nc
e
.
T
hi
s
s
tu
dy
a
ddr
e
s
s
e
s
th
a
t
g
a
p
by
in
tr
oduc
in
g
s
ta
c
ki
ng
a
r
c
hi
te
c
tu
r
e
(
S
T
A
C
K
)
-
E
D
,
a
hybr
id
m
ul
ti
-
la
ye
r
e
d
a
r
c
hi
te
c
tu
r
e
[
8]
th
a
t
or
c
he
s
tr
a
te
s
m
ul
ti
pl
e
de
te
c
ti
on
pa
r
a
di
gm
s
in
a
uni
f
ie
d
f
r
a
m
e
w
or
k.
A
t
th
e
f
ir
s
t
la
ye
r
,
S
T
A
C
K
-
E
D
e
m
pl
oys
s
upe
r
vi
s
e
d
m
ode
l
s
,
na
m
e
ly
C
a
tB
oos
t
[
9]
a
nd
X
G
B
oos
t
[
10]
opt
im
iz
e
d
w
it
h
O
pt
una
[
11]
,
a
lo
ngs
id
e
s
e
lf
-
s
upe
r
vi
s
e
d
gr
a
ph
-
ba
s
e
d
c
om
pone
nt
s
(
G
I
N
E
C
onv
[
12
]
)
a
nd
uns
upe
r
vi
s
e
d
a
nom
a
ly
de
te
c
to
r
s
(
A
ut
oe
nc
ode
r
[
13]
,
I
s
ol
a
ti
on
F
or
e
s
t
[
14]
,
a
nd
one
-
c
la
s
s
s
uppor
t
v
e
c
to
r
m
a
c
hi
ne
(
S
V
M
)
[
15]
)
.
T
he
out
put
s
a
r
e
f
us
e
d
a
t
th
e
s
e
c
ond
la
ye
r
th
r
ough
a
m
e
ta
-
le
a
r
ne
r
opt
im
iz
e
d
vi
a
G
r
id
S
e
a
r
c
hC
V
[
16]
.
A
t
th
e
th
ir
d
la
ye
r
,
a
pos
t
-
hoc
c
or
r
e
c
ti
on
(
P
H
C
)
m
e
c
ha
ni
s
m
i
nt
e
gr
a
te
s
s
e
n
s
it
iv
e
out
li
e
r
de
te
c
to
r
s
,
in
c
lu
di
ng
M
a
ha
la
nobi
s
di
s
ta
nc
e
[
17]
,
hi
e
r
a
r
c
hi
c
a
l
de
ns
it
y
-
ba
s
e
d
s
pa
ti
a
l
c
lu
s
te
r
in
g
of
a
ppl
ic
a
ti
ons
(
H
D
B
S
C
A
N
)
[
18]
,
lo
c
a
l
out
li
e
r
f
a
c
to
r
(
L
O
F
)
[
19]
,
a
nd
c
os
in
e
s
im
il
a
r
it
y
[
20
]
to
r
e
c
ove
r
m
is
s
e
d
z
e
r
o
-
da
y
a
tt
a
c
ks
.
T
hi
s
m
ul
ti
-
f
a
c
e
te
d
de
s
ig
n
s
ur
pa
s
s
e
s
s
in
gl
e
-
m
ode
l
a
nd
c
onve
nt
io
na
l
e
n
s
e
m
bl
e
a
ppr
oa
c
he
s
by
of
f
e
r
in
g
a
n
a
da
pt
iv
e
,
e
xpl
a
in
a
bl
e
,
a
nd
r
e
s
il
ie
nt
e
ndpoint
de
f
e
n
s
e
m
e
c
h
a
ni
s
m
.
T
h
e
m
a
in
c
ont
r
ib
ut
io
n
of
th
is
s
tu
dy
is
th
e
de
s
ig
n
of
a
hol
is
ti
c
,
hybr
id
,
a
nd
la
ye
r
e
d
f
r
a
m
e
w
or
k
th
a
t
in
te
gr
a
te
s
s
u
pe
r
vi
s
e
d,
s
e
lf
-
s
upe
r
vi
s
e
d,
a
nd
uns
upe
r
vi
s
e
d
pa
r
a
di
gm
s
f
or
e
ndpoint
th
r
e
a
t
de
te
c
ti
on,
th
e
in
te
gr
a
ti
on
of
pos
t
-
pr
oc
e
s
s
in
g
c
or
r
e
c
ti
on
m
e
c
h
a
ni
s
m
s
th
a
t
s
ig
ni
f
ic
a
nt
ly
im
pr
ove
z
e
r
o
-
da
y
de
te
c
ti
on
a
nd
r
e
duc
e
f
a
ls
e
ne
g
a
t
iv
e
s
,
th
e
us
e
of
S
ha
pl
e
y
a
ddi
ti
ve
e
xpl
a
n
a
ti
ons
(
S
H
A
P
)
to
im
pr
ove
th
e
t
r
a
ns
pa
r
e
nc
y
a
nd
r
e
li
a
bi
li
ty
of
th
e
m
ode
l'
s
de
c
is
io
n
-
m
a
ki
ng
pr
oc
e
s
s
,
a
nd
th
e
va
li
da
ti
on
of
th
e
pr
opos
e
d
f
r
a
m
e
w
or
k
on
th
e
H
I
D
S
-
S
e
t
hybr
id
be
nc
hm
a
r
k
da
ta
s
e
t
(
H
I
D
S
-
S
e
t)
,
a
c
hi
e
vi
ng
be
s
t
pe
r
f
or
m
a
nc
e
w
it
h a
n F
2 s
c
or
e
of
98.89%
w
hi
le
r
e
c
ove
r
in
g 62.64%
of
m
is
s
e
d z
e
r
o
-
da
y c
a
ndi
da
te
s
.
2.
L
I
T
E
R
A
T
U
R
R
I
V
I
E
W
E
D
is
a
n
e
vol
vi
ng
r
e
s
e
a
r
c
h
dom
a
in
th
a
t
m
us
t
a
d
a
pt
to
th
e
gr
ow
in
g
s
ophi
s
ti
c
a
ti
on
of
c
ybe
r
th
r
e
a
ts
[
21]
.
V
a
r
io
us
s
tu
di
e
s
ha
ve
e
xpl
or
e
d
e
ns
e
m
bl
e
le
a
r
n
in
g,
A
dvR
,
X
A
I
,
ye
t
m
o
s
t
r
e
m
a
in
is
ol
a
te
d,
a
ddr
e
s
s
in
g
onl
y
one
c
a
p
a
bi
li
ty
in
de
pt
h.
T
o
c
ont
e
xt
ua
li
z
e
th
is
r
e
s
e
a
r
c
h,
T
a
bl
e
1
c
om
pa
r
e
s
r
e
pr
e
s
e
nt
a
ti
ve
E
D
a
ppr
oa
c
he
s
a
c
r
os
s
f
iv
e
c
or
e
c
a
pa
bi
li
ti
e
s
:
hybr
id
a
r
c
hi
te
c
tu
r
e
, G
N
N
, A
dvR
, P
H
C
, a
nd X
A
I
.
T
a
bl
e
1. S
um
m
a
r
y a
nd c
om
pa
r
is
on of
r
e
la
te
d r
e
s
e
a
r
c
h w
or
ks
i
n
e
ndpoint
de
te
c
ti
on s
ys
te
m
s
S
t
udy (
A
ut
hor
,
ye
a
r
)
H
ybr
i
d
ar
ch
GNN
A
dvR
P
H
C
XAI
T
a
m
a
e
t
al
.
(
2019)
[
22]
✔
✘
✘
✘
✘
V
i
na
ya
kum
a
r
e
t
al
.
(
2019)
[
6]
✘
✘
✘
✘
✘
M
oha
m
e
d
e
t
al
.
(
2023)
[
23]
✔
✘
✘
✘
✘
G
hos
h (
2025)
[
24]
✔
✔
✘
✘
✘
M
a
goo a
nd G
a
r
g (
2021)
[
25]
✘
✘
✔
✘
⌀
K
ha
r
oubi
e
t
al
.
(
2025)
[
26]
✘
✘
✘
✘
✘
A
l
gha
z
a
l
i
a
nd H
a
noos
h (
2022)
[
27]
✔
✘
✘
✘
✘
V
i
s
hw
a
ka
r
m
a
a
nd K
e
s
s
w
a
ni
(
2025)
[
28]
✔
✘
✘
✘
✔
Z
hong
e
t
al
.
(
2024)
[
29]
✘
✔
✘
✘
⌀
H
e
e
t
al
.
(
2024)
[
30]
✘
✘
✔
✘
✘
A
r
r
e
c
he
e
t
al
.
(
2024)
[
31]
✘
✘
✘
✘
✔
R
os
ha
n a
nd Z
a
f
a
r
(
2024)
[
32]
✘
✘
✔
✘
✘
S
un
e
t
al
.
(
2024)
[
33]
✘
✘
✘
✔
⌀
T
hi
s
s
t
udy
(
2025)
✔
✔
✔
✔
✔
L
e
ge
nd:
✔
:
F
ul
l
y
di
s
c
us
s
e
d/
i
m
pl
e
m
e
nt
e
d
;
⌀
:
P
a
r
t
i
a
l
l
y
di
s
c
us
s
e
d/
i
m
pl
i
e
d
;
✘
:
N
ot
a
ddr
e
s
s
e
d
2.1. E
n
s
e
m
b
le
le
a
r
n
in
g b
as
e
d
ap
p
r
oac
h
E
ns
e
m
bl
e
le
a
r
ni
ng
ha
s
lo
ng
be
e
n
a
ppl
ie
d
to
e
nha
nc
e
E
D
a
c
c
ur
a
c
y
a
nd
r
obus
tn
e
s
s
by
le
ve
r
a
gi
ng
m
ul
ti
pl
e
c
la
s
s
if
ie
r
s
.
T
a
m
a
e
t
al
.
[
22]
in
tr
oduc
e
d
a
two
-
s
ta
ge
e
ns
e
m
bl
e
th
a
t
im
pr
ove
d
a
nom
a
ly
de
te
c
ti
on,
w
hi
le
M
oha
m
e
d
e
t
al
.
[
23]
a
dopt
e
d
e
ns
e
m
bl
e
vot
in
g
f
or
I
oT
c
ont
e
xt
s
.
A
lg
ha
z
a
li
a
nd
H
a
noos
h
[
27]
f
ur
th
e
r
in
te
gr
a
te
d
r
a
ndom
f
or
e
s
t
a
s
a
m
e
ta
-
le
a
r
ne
r
in
hybr
id
DL
.
D
e
s
pi
te
th
e
s
e
a
dva
nc
e
s
,
s
uc
h
m
ode
ls
of
te
n
de
pe
nd
on
ta
bul
a
r
f
e
a
tu
r
e
s
a
nd
ov
e
r
lo
ok
a
dve
r
s
a
r
ia
l
r
e
s
il
ie
nc
e
or
in
t
e
r
pr
e
ta
bi
li
ty
.
R
e
c
e
nt
e
xt
e
nde
d
d
e
te
c
ti
on
a
nd
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
dv A
ppl
S
c
i
I
S
S
N
:
2252
-
8814
St
ac
k
in
g ar
c
hi
te
c
tu
r
e
-
e
ndpoint
de
te
c
ti
on:
a hy
br
id
m
ul
ti
-
la
y
e
r
e
d ar
c
hi
te
c
tu
r
e
…
(
A
bd R
ahm
an W
ahi
d)
1265
r
e
s
pons
e
(
XDR
)
s
ur
ve
ys
unde
r
s
c
or
e
th
e
ne
e
d
f
or
m
ul
ti
-
la
ye
r
e
d
e
ns
e
m
bl
e
s
c
oupl
e
d
w
it
h
e
ndpoint
te
le
m
e
tr
y
[
29]
, but
f
e
w
f
r
a
m
e
w
or
ks
a
c
hi
e
ve
hol
is
ti
c
f
us
io
n a
c
r
os
s
t
h
e
s
e
c
a
pa
bi
li
ti
e
s
.
2.2. De
e
p
le
ar
n
in
g an
d
gr
ap
h
n
e
u
r
al
n
e
t
w
or
k
s
D
L
ha
s
r
e
vol
ut
io
ni
z
e
d
E
D
th
r
ough
a
ut
om
a
ti
c
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
f
r
om
ne
twor
k
tr
a
f
f
ic
.
V
in
a
ya
kum
a
r
e
t
al
.
[
6]
de
m
on
s
tr
a
te
d
D
L
’
s
f
e
a
s
ib
il
it
y
f
or
in
tr
us
io
n
de
te
c
ti
on,
a
nd
K
ha
r
oubi
e
t
al
.
[
26]
e
m
pl
oye
d
c
onvolut
io
na
l
n
e
ur
a
l
ne
twor
ks
(
C
N
N
s
)
f
or
I
oT
tr
a
f
f
ic
c
la
s
s
if
ic
a
ti
on.
Y
e
t,
D
L
r
e
m
a
in
s
opa
que
a
nd
s
e
ns
it
iv
e
to
a
dve
r
s
a
r
ia
l
pe
r
tu
r
ba
ti
ons
.
G
N
N
s
ha
ve
e
m
e
r
ge
d
a
s
pr
om
is
in
g
a
lt
e
r
na
ti
ve
s
.
G
hos
h
[
24]
ut
il
iz
e
d
gr
a
ph c
onvolut
io
na
l
ne
twor
ks
(
G
C
N
s
)
t
o
c
a
pt
ur
e
r
e
la
ti
ona
l
de
p
e
nde
nc
ie
s
, w
hi
le
Z
hong
e
t
al
.
[
29]
hi
gh
li
ght
e
d
th
e
ir
r
ol
e
in
s
ys
te
m
-
on
-
C
hi
p
(
S
oC
)
te
le
m
e
tr
y
pi
pe
li
ne
s
.
M
or
e
ove
r
,
li
nki
ng
G
N
N
-
ba
s
e
d
de
te
c
ti
on
w
it
h
f
r
a
m
e
w
or
ks
s
uc
h
a
s
M
I
T
R
E
A
T
T
&
C
K
pr
ovi
de
s
a
m
or
e
th
r
e
a
t
-
in
f
or
m
e
d
c
ont
e
xt
[
31]
.
S
ti
l
l,
m
os
t
G
N
N
-
ba
s
e
d
s
tu
di
e
s
f
oc
us
on
s
tr
uc
tu
r
a
l
m
ode
li
ng w
it
hout
i
nt
e
gr
a
ti
ng a
dve
r
s
a
r
ia
l
de
f
e
ns
e
or
P
H
C
.
2.3. Ad
ve
r
s
ar
ia
l
r
ob
u
s
t
n
e
s
s
an
d
e
xp
la
in
ab
le
A
I
A
dvR
r
e
m
a
in
s
a
ke
y
c
ha
ll
e
ng
e
,
a
s
a
tt
a
c
ke
r
s
e
xpl
oi
t
im
pe
r
c
e
pt
i
bl
e
pe
r
tu
r
ba
ti
ons
to
e
va
de
d
e
te
c
ti
on.
M
a
goo
a
nd
G
a
r
g
[
25]
a
na
ly
z
e
d
DL
vul
ne
r
a
bi
li
ty
to
f
a
s
t
gr
a
di
e
n
t
s
ig
n
m
e
th
od
(
F
G
S
M
)
a
nd
pr
oj
e
c
te
d
gr
a
di
e
nt
de
s
c
e
nt
(
P
G
D
)
a
tt
a
c
ks
,
w
hi
le
H
e
e
t
al
.
[
30]
a
nd
R
os
ha
n
a
nd
Z
a
f
a
r
[
32
]
pr
opos
e
d
ge
ne
r
a
li
z
e
d
a
dve
r
s
a
r
ia
l
de
f
e
ns
e
s
.
P
a
r
a
ll
e
ll
y,
th
e
d
e
m
a
nd
f
or
tr
a
ns
pa
r
e
n
c
y
ha
s
f
os
te
r
e
d
th
e
a
dopt
io
n
of
X
A
I
.
V
is
h
w
a
ka
r
m
a
a
nd
K
e
s
s
w
a
ni
[
28]
in
c
or
por
a
te
d
S
H
A
P
in
to
s
ta
c
ki
ng
e
ns
e
m
bl
e
s
,
a
nd
A
r
r
e
c
he
e
t
al
.
[
31]
de
s
ig
ne
d
e
xpl
a
in
a
bl
e
in
tr
us
io
n
de
te
c
ti
on
s
ys
te
m
(
I
D
S
)
f
r
a
m
e
w
or
ks
.
H
ow
e
ve
r
,
m
os
t
of
th
e
s
e
a
ppr
oa
c
he
s
s
ti
ll
tr
e
a
t
in
te
r
pr
e
ta
bi
li
ty
a
nd r
obus
tn
e
s
s
s
e
pa
r
a
te
ly
, l
a
c
ki
ng s
yne
r
gy i
n unif
ie
d a
r
c
hi
te
c
tu
r
e
s
.
2.4. Re
s
e
ar
c
h
gap
F
r
om
th
e
a
bove
r
e
vi
e
w
,
it
is
e
vi
de
nt
th
a
t
a
lt
hough
e
ns
e
m
bl
e
le
a
r
ni
ng,
G
N
N
-
ba
s
e
d
de
te
c
ti
on,
a
dve
r
s
a
r
ia
l
r
e
s
il
ie
nc
e
,
a
nd
X
A
I
ha
ve
a
dva
nc
e
d
c
ons
id
e
r
a
bl
y
,
e
xi
s
ti
ng
w
or
ks
c
ont
in
ue
to
opt
im
iz
e
th
e
s
e
c
om
pone
nt
s
in
is
ol
a
ti
on.
P
o
s
t
-
e
xe
c
ut
io
n
c
or
r
e
c
ti
on
s
tr
a
te
gi
e
s
,
s
uc
h
a
s
th
o
s
e
pr
opos
e
d
by
S
un
e
t
al
.
[
33]
,
ha
ve
ye
t
to
be
a
ppl
ie
d
c
om
pr
e
he
ns
iv
e
ly
to
e
ndpoint
de
te
c
ti
on.
T
hi
s
s
tu
dy
in
tr
oduc
e
s
S
T
A
C
K
-
E
D
,
a
uni
f
ie
d
hybr
id
f
r
a
m
e
w
or
k
in
te
gr
a
ti
ng
s
upe
r
vi
s
e
d,
s
e
lf
-
s
upe
r
vi
s
e
d,
a
nd
un
s
upe
r
vi
s
e
d
pa
r
a
di
gm
s
w
it
h
a
dve
r
s
a
r
ia
l
a
d
a
pt
a
ti
on,
X
A
I
,
a
nd
P
H
C
.
T
h
e
f
r
a
m
e
w
or
k
br
id
ge
s
a
c
c
ur
a
c
y,
r
e
s
il
ie
n
c
e
,
a
nd
tr
a
ns
pa
r
e
nc
y
c
ha
r
a
c
te
r
is
ti
c
s
c
r
uc
ia
l
f
or
m
ode
r
n
S
oC
, e
nt
e
r
pr
is
e
c
ybe
r
s
e
c
ur
it
y, a
nd z
e
r
o
-
tr
us
t
e
nvi
r
onm
e
nt
s
.
3.
P
R
O
P
O
S
E
D
M
E
T
H
O
D
T
he
r
e
s
e
a
r
c
h
m
e
th
odol
ogy
w
a
s
de
s
ig
ne
d
to
bui
ld
th
e
S
T
A
C
K
-
E
D
a
r
c
hi
te
c
tu
r
e
.
I
t
in
vol
ve
d
s
e
ve
r
a
l
s
ta
ge
s
, be
gi
nni
ng w
it
h da
t
a
pr
e
pa
r
a
ti
on. T
he
pr
oc
e
s
s
c
ont
in
u
e
d
th
r
ough to t
he
f
in
a
l
m
ode
l
e
va
lu
a
ti
on.
3.1. Dat
a
p
ip
e
l
in
e
an
d
p
r
e
-
p
r
oc
e
s
s
in
g
T
he
pr
opos
e
d
da
ta
pi
pe
li
ne
,
il
lu
s
tr
a
te
d
in
F
ig
ur
e
1,
w
a
s
de
s
i
gne
d
to
gua
r
a
nt
e
e
th
e
in
te
gr
it
y
a
nd
r
e
pr
e
s
e
nt
a
ti
ve
ne
s
s
of
tr
a
f
f
ic
f
lo
w
s
pr
io
r
to
m
od
e
li
ng.
B
y
c
om
bi
ni
ng
m
ul
ti
pl
e
s
ta
g
e
s
of
c
le
a
n
s
in
g,
tr
a
ns
f
or
m
a
ti
on,
a
nd
di
m
e
ns
io
na
li
ty
r
e
duc
ti
on,
th
e
pi
pe
li
ne
a
i
m
s
to
pr
e
pa
r
e
da
ta
th
a
t
f
a
it
hf
ul
ly
r
e
f
le
c
ts
bot
h
nor
m
a
l
ope
r
a
ti
ons
a
nd ma
li
c
io
us
be
ha
vi
or
s
w
hi
le
m
in
im
iz
in
g n
oi
s
e
a
nd r
e
dunda
nc
y.
I
n
th
e
da
ta
s
e
ts
,
two
be
nc
hm
a
r
k
in
tr
us
io
n
de
te
c
ti
on
c
or
por
a
w
e
r
e
us
e
d
to
e
ns
ur
e
di
ve
r
s
it
y
of
a
tt
a
c
k
ty
pe
s
a
nd
ne
twor
k
c
ondi
ti
ons
.
T
he
U
N
S
W
-
N
B
15
da
ta
s
e
t
[
34]
pr
ovi
de
s
s
ta
ti
s
ti
c
a
l
f
lo
w
de
s
c
r
ip
to
r
s
,
w
id
e
ly
r
e
c
ogni
z
e
d
f
or
c
a
pt
ur
in
g
ba
c
kgr
ound
tr
a
f
f
ic
pa
tt
e
r
ns
.
C
om
pl
e
m
e
nt
a
r
il
y,
th
e
C
S
E
-
C
I
C
-
I
D
S
2018
da
ta
s
e
t
c
ont
a
in
s
s
e
s
s
io
n
-
le
ve
l
f
e
a
tu
r
e
s
(
e
.g.,
f
lo
w
in
te
r
-
a
r
r
iv
a
l
ti
m
e
,
pa
c
ke
t
le
ngt
h
m
a
xi
m
a
,
a
nd
f
or
w
a
r
d
pa
c
ke
t
r
a
te
)
th
a
t
e
m
pha
s
iz
e
be
ha
vi
or
a
l
s
ig
n
a
tu
r
e
s
of
m
ode
r
n
c
ybe
r
a
tt
a
c
ks
s
uc
h
a
s
di
s
tr
ib
ut
e
d
de
ni
a
l
of
s
e
r
vi
c
e
(
D
D
oS
)
,
br
ut
e
f
or
c
e
,
a
nd
in
f
il
tr
a
ti
on
a
tt
e
m
pt
s
.
S
a
m
pl
in
g
s
tr
a
te
gy.
T
o
m
a
in
ta
in
ba
la
nc
e
d
r
e
pr
e
s
e
nt
a
ti
on
a
c
r
os
s
c
l
a
s
s
e
s
,
a
s
tr
a
ti
f
ie
d
s
a
m
pl
in
g
of
10%
w
a
s
a
ppl
ie
d
to
e
a
c
h
d
a
ta
s
e
t.
T
hi
s
pr
oduc
e
d
a
c
om
bi
ne
d
c
or
pu
s
of
278,770
r
e
c
or
ds
,
e
qua
ll
y
di
s
tr
ib
ut
e
d
be
twe
e
n
be
ni
gn
a
nd
a
tt
a
c
k
in
s
ta
nc
e
s
.
T
he
s
tr
a
ti
f
ic
a
ti
on
e
ns
ur
e
d
th
a
t
r
a
r
e
but
c
r
it
ic
a
l
a
tt
a
c
k
c
a
te
gor
ie
s
w
e
r
e
r
e
ta
in
e
d
pr
opor
ti
ona
ll
y,
th
e
r
e
by
a
voi
di
ng
c
la
s
s
s
ke
w
.
L
a
be
l
s
t
a
nda
r
di
z
a
ti
on
a
nd
f
e
a
tu
r
e
r
e
f
in
e
m
e
nt
.
A
f
te
r
m
e
r
gi
ng,
a
ll
la
be
ls
w
e
r
e
s
ta
nda
r
di
z
e
d
in
to
a
bi
na
r
y
f
or
m
a
t
(
0=
nor
m
a
l
,
1=
a
tt
a
c
k
)
.
R
e
dunda
nt
id
e
nt
if
ie
r
s
w
e
r
e
r
e
m
ove
d,
c
a
te
gor
ic
a
l
va
r
ia
bl
e
s
w
e
r
e
e
nc
ode
d
us
in
g
one
-
hot
r
e
pr
e
s
e
nt
a
ti
on
[
35]
, a
nd nume
r
ic
va
r
ia
bl
e
s
w
e
r
e
s
c
a
le
d t
o t
he
i
n
te
r
va
l
[
0,1]
w
it
h M
in
M
a
x t
r
a
ns
f
or
m
a
ti
on
[
36]
.
T
he
s
e
s
te
p
s
pr
e
ve
nt
dom
in
a
nc
e
of
hi
gh
-
m
a
gni
tu
de
f
e
a
tu
r
e
s
a
nd
pr
om
ot
e
e
qua
l
c
ont
r
ib
ut
io
n
a
c
r
os
s
th
e
f
e
a
tu
r
e
s
pa
c
e
.
F
e
a
tu
r
e
e
ngi
ne
e
r
in
g
a
nd
s
e
le
c
ti
on.
T
o
e
nr
ic
h
di
s
c
r
im
in
a
to
r
y
pow
e
r
,
de
r
iv
e
d
in
di
c
a
to
r
s
w
e
r
e
c
r
e
a
te
d,
s
uc
h a
s
t
he
s
our
c
e
-
to
-
de
s
ti
na
ti
on byte
r
a
ti
o, pa
c
ke
t
f
r
e
que
nc
y, i
nt
e
r
-
a
r
r
iv
a
l
va
r
ia
nc
e
, a
nd t
r
a
f
f
ic
i
n
te
ns
it
y
[
37]
.
F
r
om
t
he
a
ugm
e
nt
e
d s
e
t,
50 pr
e
di
c
ti
ve
a
tt
r
ib
ut
e
s
w
e
r
e
s
e
le
c
te
d
vi
a
t
he
m
ut
ua
l
in
f
or
m
a
ti
on
(
M
I
)
c
r
it
e
r
io
n
[
38]
,
w
hi
c
h
qua
nt
if
ie
s
th
e
r
e
duc
ti
on
in
unc
e
r
ta
in
ty
of
th
e
ta
r
ge
t
la
be
l
w
he
n
a
gi
ve
n
f
e
a
tu
r
e
is
obs
e
r
ve
d.
F
e
a
tu
r
e
s
s
uc
h
a
s
de
s
ti
na
ti
on
por
t,
f
lo
w
dur
a
ti
on,
in
te
r
-
a
r
r
iv
a
l
s
ta
ti
s
ti
c
s
,
a
nd
f
or
w
a
r
d
pa
c
ke
t
r
a
te
s
e
xhi
bi
te
d
s
tr
ong
non
-
li
ne
a
r
a
s
s
oc
ia
ti
on
w
it
h
a
tt
a
c
k
pr
e
s
e
nc
e
a
nd
w
e
r
e
th
e
r
e
f
or
e
pr
io
r
i
ti
z
e
d.
B
a
la
nc
in
g
a
nd
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8814
I
nt
J
A
dv A
ppl
S
c
i
,
V
ol
.
14
, N
o.
4
,
D
e
c
e
m
be
r
20
25
:
126
3
-
12
80
1266
di
m
e
ns
io
na
li
ty
r
e
duc
ti
on.
T
o
a
ddr
e
s
s
r
e
s
id
ua
l
im
ba
la
nc
e
,
th
e
B
or
de
r
li
ne
S
M
O
T
E
te
c
hni
que
[
39]
w
a
s
e
m
pl
oye
d, s
ynt
he
s
iz
in
g
s
a
m
pl
e
s
a
lo
ng de
c
is
io
n bounda
r
ie
s
t
o e
nha
nc
e
th
e
de
te
c
ti
on of
m
in
or
it
y c
la
s
s
e
s
.
F
ig
ur
e
1. F
lo
w
da
ta
pi
pe
li
ne
a
nd pr
e
-
pr
oc
e
s
s
in
g
S
ubs
e
que
nt
ly
,
la
te
nt
r
e
pr
e
s
e
nt
a
ti
on
w
a
s
r
e
duc
e
d
to
10
di
m
e
ns
io
ns
us
in
g
uni
f
or
m
m
a
ni
f
ol
d
a
ppr
oxi
m
a
ti
on
a
nd
pr
oj
e
c
ti
on
(
U
M
A
P
)
[
40]
.
U
M
A
P
pr
e
s
e
r
ve
s
bot
h
lo
c
a
l
a
nd
gl
oba
l
s
tr
uc
tu
r
e
by
opt
i
m
iz
in
g
th
e
c
r
os
s
-
e
nt
r
opy obje
c
ti
ve
[
41]
:
∑
,
l
o
g
(
)
+
(
1
−
)
l
o
g
(
1
−
1
−
)
(
1)
w
he
r
e
a
nd
de
not
e
pa
ir
w
is
e
s
im
il
a
r
it
ie
s
in
th
e
or
ig
in
a
l
a
nd
pr
oj
e
c
te
d
s
pa
c
e
,
r
e
s
pe
c
ti
v
e
ly
.
T
hi
s
tr
a
ns
f
or
m
a
ti
on
c
a
pt
ur
e
s
hi
dde
n
to
pol
ogi
c
a
l
pa
tt
e
r
ns
,
e
na
bl
in
g
s
ubs
e
que
nt
m
ode
ls
to
ope
r
a
te
on
c
om
pa
c
t
ye
t
e
xpr
e
s
s
iv
e
m
a
ni
f
ol
ds
.
T
a
bl
e
2
pr
e
s
e
nt
s
e
xa
m
pl
e
s
of
th
e
f
in
a
l
p
r
e
-
pr
oc
e
s
s
e
d
f
e
a
tu
r
e
s
,
r
a
ngi
ng
f
r
om
s
ta
ti
s
ti
c
a
l
f
lo
w
m
e
a
s
ur
e
s
to
e
ngi
ne
e
r
e
d
r
a
ti
os
a
nd
e
nc
ode
d
c
a
te
gor
ic
a
l
va
r
ia
bl
e
s
.
T
he
s
e
s
tr
uc
tu
r
e
d
r
e
pr
e
s
e
nt
a
ti
ons
f
or
m
th
e
s
ta
nda
r
di
z
e
d i
nput
f
or
t
he
S
T
A
C
K
-
E
D
a
r
c
hi
te
c
tu
r
e
.
T
a
bl
e
2. S
a
m
pl
e
s
tr
uc
tu
r
e
of
f
in
a
l
f
e
a
tu
r
e
s
e
t
a
f
te
r
pr
e
-
pr
oc
e
s
s
in
g)
F
e
a
t
ur
e
s
S
a
m
pl
e
1
S
a
m
pl
e
2
S
a
m
pl
e
3
D
e
s
c
r
i
pt
i
on
F
l
ow
dur
a
t
i
on
0.0015
0.8920
0.0008
F
l
ow
dur
a
t
i
on (
s
e
c
onds
)
S
byt
e
s
0.0021
0.0005
0.0033
S
our
c
e
byt
e
P
kt
L
e
n M
a
x
0.125
0.045
0.150
M
a
xi
m
um
pa
c
ke
t
l
e
ngt
h
F
w
d P
kt
s
/
s
0.0035
0.8540
0.0042
M
a
xi
m
um
pa
c
ke
t
l
e
ngt
h
F
l
ow
I
A
T
M
e
a
n
0.0014
0.0120
0.0007
F
or
w
a
r
d pa
c
ke
t
f
r
e
que
nc
y
I
a
t
va
r
i
a
nc
e
0.0001
0.0023
0.0000
A
ve
r
a
ge
i
nt
e
r
-
a
r
r
i
va
l
S
byt
e
s
, dbyt
e
s
, r
a
t
i
o
0.4520
1.0000
0.3891
S
our
c
e
/
de
s
t
i
na
t
i
on byt
e
r
a
t
i
o
P
r
ot
okol
U
D
P
0
1
0
U
D
P
(
one
-
hot
)
P
r
ot
okol
T
C
P
1
0
1
T
C
P
(
one
-
hot
)
L
a
be
l
0
1
0
0=
nor
m
a
l
, 1=a
t
t
a
c
k
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
dv A
ppl
S
c
i
I
S
S
N
:
2252
-
8814
St
ac
k
in
g ar
c
hi
te
c
tu
r
e
-
e
ndpoint
de
te
c
ti
on:
a hy
br
id
m
ul
ti
-
la
y
e
r
e
d ar
c
hi
te
c
tu
r
e
…
(
A
bd R
ahm
an W
ahi
d)
1267
3.2. Ar
c
h
it
e
c
t
u
r
al
d
e
s
ig
n
T
he
S
T
A
C
K
-
E
D
a
r
c
hi
te
c
tu
r
e
w
a
s
c
onc
e
iv
e
d
a
s
a
la
ye
r
e
d
in
te
ll
ig
e
nc
e
s
ys
te
m
in
s
pi
r
e
d
by
th
e
w
or
kf
lo
w
of
hum
a
n
c
ybe
r
s
e
c
ur
it
y
a
na
ly
s
ts
. R
a
th
e
r
th
a
n
r
e
ly
in
g
on a
s
in
gl
e
de
te
c
ti
on me
c
ha
ni
s
m
, i
t
in
te
gr
a
te
s
m
ul
ti
pl
e
a
na
ly
ti
c
a
l
pe
r
s
pe
c
ti
ve
s
a
nd c
ons
ol
id
a
te
s
t
he
m
t
hr
ough a
hi
e
r
a
r
c
hi
c
a
l
de
c
is
io
n pr
oc
e
s
s
. A
s
de
pi
c
te
d i
n
F
ig
ur
e
2,
th
e
a
r
c
hi
te
c
tu
r
e
c
on
s
is
t
s
of
th
r
e
e
f
unc
ti
ona
l
l
e
ve
ls
:
i)
s
pe
c
ia
li
s
t
m
ode
l
c
om
pon
e
nt
s
(
ba
s
e
-
le
a
r
ne
r
s
)
,
ii
)
in
te
ll
ig
e
nc
e
f
us
io
n
by
a
m
e
ta
-
le
a
r
ne
r
,
a
nd
ii
i)
a
P
H
C
m
e
c
ha
ni
s
m
th
a
t
pr
ovi
de
s
r
e
s
il
ie
nc
e
a
ga
in
s
t
f
a
ls
e
ne
ga
ti
ve
s
a
nd z
e
r
o
-
da
y a
tt
a
c
ks
.
F
ig
ur
e
2. A
r
c
hi
te
c
tu
r
e
s
ta
c
ki
ng e
ns
e
m
bl
e
S
T
A
C
K
-
ED
3.2.1. L
e
ve
l
1
s
p
e
c
ia
li
s
t
m
od
e
l
c
o
m
p
on
e
n
t
s
(
b
as
e
-
le
ar
n
e
r
)
T
he
f
ir
s
t
le
ve
l
r
e
c
e
iv
e
s
th
e
10
-
di
m
e
ns
io
na
l
la
te
nt
da
ta
r
e
pr
e
s
e
nt
a
ti
on
a
nd
pr
oc
e
s
s
e
s
it
th
r
ough
th
r
e
e
c
om
pl
e
m
e
nt
a
r
y pe
r
s
pe
c
ti
ve
s
:
–
S
upe
r
vi
s
e
d
c
om
pone
nt
(
pa
tt
e
r
n
r
e
c
ogni
ti
on
e
xpe
r
t)
.
T
hi
s
c
om
p
one
nt
e
m
pl
oys
a
n
e
n
s
e
m
bl
e
of
C
a
tB
oo
s
t
a
nd
X
G
B
oos
t
c
la
s
s
if
ie
r
s
,
opt
im
iz
e
d
vi
a
B
a
ye
s
ia
n
hype
r
pa
r
a
m
e
te
r
s
e
a
r
c
h
(
O
pt
una
)
,
to
a
c
hi
e
ve
a
ba
la
nc
e
d
bi
a
s
-
va
r
ia
nc
e
tr
a
de
-
of
f
[
42]
.
I
t
s
pe
c
ia
li
z
e
s
in
id
e
nt
if
yi
ng
pa
tt
e
r
ns
c
ons
is
te
nt
w
it
h
pr
e
vi
ous
ly
known intr
us
io
ns
.
–
G
r
a
ph
-
ba
s
e
d
s
e
lf
-
s
upe
r
vi
s
e
d
c
om
pone
nt
(
ne
twor
k
s
tr
uc
tu
r
e
e
xpe
r
t)
.
L
e
ve
r
a
gi
ng
G
I
N
E
C
onv
w
it
hi
n
a
c
ont
r
a
s
ti
ve
le
a
r
ni
ng
f
r
a
m
e
w
or
k,
th
is
c
om
pone
nt
c
a
pt
ur
e
s
s
tr
uc
t
ur
a
l
de
pe
nde
nc
ie
s
a
m
ong
ne
twor
k
f
lo
w
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8814
I
nt
J
A
dv A
ppl
S
c
i
,
V
ol
.
14
, N
o.
4
,
D
e
c
e
m
be
r
20
25
:
126
3
-
12
80
1268
T
hr
ough
a
dve
r
s
a
r
ia
l
f
in
e
-
tu
ni
ng,
it
pr
oduc
e
s
r
obus
t
32
-
di
m
e
ns
io
na
l
e
m
be
ddi
ngs
c
a
pa
bl
e
of
di
s
c
r
im
in
a
ti
ng s
ubt
le
va
r
ia
ti
ons
i
n t
r
a
f
f
ic
r
e
la
ti
ons
.
–
U
ns
upe
r
vi
s
e
d
a
nom
a
ly
de
te
c
ti
on
c
om
pone
nt
(
be
h
a
vi
or
a
l
de
v
ia
ti
on
e
xpe
r
t)
.
O
pe
r
a
ti
ng
on
th
e
gr
a
ph
e
m
be
ddi
ngs
,
th
is
c
om
pone
nt
de
te
c
t
s
out
li
e
r
s
w
it
hout
r
e
qui
r
in
g
la
be
le
d
da
ta
.
I
t
in
te
gr
a
te
s
a
ut
oe
n
c
ode
r
s
,
I
s
ol
a
ti
on
F
or
e
s
t,
M
a
ha
la
nobi
s
di
s
ta
nc
e
,
a
nd
L
O
F
to
c
a
pt
ur
e
be
ha
vi
or
a
l
de
vi
a
ti
ons
in
di
c
a
ti
ve
of
nove
l
o
r
r
a
r
e
a
tt
a
c
ks
.
3.2.2. L
e
ve
l
-
2
in
t
e
ll
ig
e
n
c
e
f
u
s
io
n
b
y
m
e
t
a
-
le
ar
n
e
r
A
t
th
e
s
e
c
ond
le
ve
l,
th
e
s
y
s
te
m
f
unc
ti
ons
a
s
a
n
in
te
ll
ig
e
nc
e
f
us
io
n
hub.
H
e
r
e
,
out
put
s
f
r
om
a
ll
ba
s
e
-
le
a
r
ne
r
s
a
r
e
tr
a
ns
f
or
m
e
d
in
to
m
e
ta
-
f
e
a
tu
r
e
s
a
nd
a
ggr
e
ga
te
d
by
a
m
e
ta
-
le
a
r
ne
r
.
T
hi
s
m
e
c
ha
ni
s
m
pa
r
a
ll
e
ls
th
e
r
ol
e
of
a
s
e
c
ur
it
y ope
r
a
ti
ons
c
e
nt
e
r
, w
he
r
e
e
vi
de
nc
e
f
r
om
di
f
f
e
r
e
nt
a
na
ly
s
ts
i
s
w
e
ig
ht
e
d a
nd s
ynt
h
e
s
iz
e
d t
o
im
pr
ove
ove
r
a
ll
s
it
ua
ti
ona
l
a
w
a
r
e
ne
s
s
.
T
a
bl
e
3
s
um
m
a
r
iz
e
s
t
he
m
e
ta
-
f
e
a
tu
r
e
s
us
e
d
in
th
is
la
ye
r
,
in
c
lu
di
n
g
s
upe
r
vi
s
e
d pr
oba
bi
li
ti
e
s
, i
ni
ti
a
l
gr
a
ph pr
e
di
c
ti
ons
, a
nd a
nom
a
ly
s
c
or
e
s
. T
h
e
s
e
f
e
a
tu
r
e
s
a
r
e
t
he
n pr
oc
e
s
s
e
d by a
C
a
tB
oos
tC
l
a
s
s
if
ie
r
opt
im
iz
e
d
w
it
h
G
r
id
S
e
a
r
c
hC
V
to
m
a
xi
m
iz
e
th
e
F
2
-
s
c
or
e
,
w
hi
c
h
e
m
pha
s
iz
e
s
r
e
c
a
ll
in
s
e
c
ur
it
y
-
c
r
it
ic
a
l
c
ont
e
xt
s
.
=
(
)
w
he
r
e
=
[
ℎ
1
(
)
,
ℎ
2
(
)
,
…
,
ℎ
(
)
]
(
2)
H
e
r
e
,
is
th
e
pr
im
a
r
y
pr
e
di
c
ti
on
f
r
om
S
T
A
C
K
-
E
D
,
is
a
f
unc
ti
o
n
le
a
r
ne
d
by
th
e
m
e
ta
-
le
a
r
ne
r
,
X
is
th
e
or
ig
in
a
l
in
put
da
ta
ve
c
to
r
,
a
nd
ℎ
(
X
)
i
s
out
put
f
r
om
th
e
ba
s
e
-
le
a
r
ne
r
.
T
a
bl
e
3. M
e
ta
-
f
e
a
tu
r
e
s
a
s
i
nput
f
or
m
e
ta
-
le
a
r
n
e
r
s
M
e
t
a
-
f
e
a
t
ur
e
M
ode
l
de
s
c
r
i
pt
i
on a
nd s
our
c
e
E
ns
e
m
bl
e
pr
oba
A
t
t
a
c
k pr
oba
bi
l
i
t
y of
s
upe
r
vi
s
e
d c
om
pone
nt
(
C
a
t
B
oos
t
+X
G
B
oos
t
)
G
I
N
E
P
r
e
di
c
t
i
on
I
ni
t
i
a
l
bi
na
r
y pr
e
di
c
t
i
on of
t
he
gr
a
ph r
e
pr
e
s
e
nt
a
t
i
on c
om
pone
nt
A
ut
oe
nc
ode
r
s
c
or
e
A
nom
a
l
y s
c
or
e
(
r
e
c
ons
t
r
uc
t
i
on e
r
r
or
)
of
t
he
a
ut
oe
nc
od
er
M
a
ha
l
a
nobi
s
s
c
or
e
M
a
ha
l
a
nobi
s
di
s
t
a
nc
e
f
r
om
t
he
s
a
m
pl
e
t
o t
he
c
e
nt
e
r
of
t
he
“
nor
m
a
l
”
di
s
t
r
i
but
i
on
i
n G
N
N
s
pa
c
e
I
s
ol
a
t
i
on f
or
e
s
t
s
c
or
e
A
nom
a
l
y s
c
or
e
f
r
om
t
he
i
s
ol
a
t
i
on f
or
e
s
t
a
l
gor
i
t
hm
3.2.3. L
e
ve
l
3
p
os
t
-
h
oc
c
or
r
e
c
t
io
n
m
e
c
h
an
is
m
T
he
th
ir
d
le
ve
l
f
unc
ti
ons
a
s
a
s
a
f
e
gua
r
d
a
ga
in
s
t
unde
te
c
te
d
in
tr
us
io
ns
.
I
t
is
onl
y
a
c
ti
va
te
d
w
he
n
th
e
m
e
ta
-
le
a
r
ne
r
c
la
s
s
if
ie
s
a
s
a
m
pl
e
a
s
nor
m
a
l
(
=
0
)
.
I
n
s
uc
h
c
a
s
e
s
,
th
e
s
a
m
pl
e
unde
r
goe
s
a
s
e
c
onda
r
y
in
ve
s
ti
ga
ti
on
us
in
g
gr
a
ph
e
m
be
ddi
ng
s
a
nd
s
e
ns
it
iv
e
out
li
e
r
de
te
c
to
r
s
.
T
e
c
hni
que
s
s
uc
h
a
s
M
a
ha
la
nobi
s
di
s
ta
nc
e
(
w
it
h
a
da
pt
iv
e
th
r
e
s
hol
ds
)
,
H
D
B
S
C
A
N
c
lu
s
te
r
in
g
(
n
oi
s
e
de
te
c
ti
on)
,
L
O
F
(
out
li
e
r
de
te
c
ti
on)
,
a
nd
c
os
in
e
s
im
il
a
r
it
y
a
r
e
a
ppl
ie
d
to
r
e
a
s
s
e
s
s
s
ubt
le
a
nom
a
li
e
s
.
I
f
a
ny
de
te
c
to
r
f
la
gs
th
e
s
a
m
pl
e
a
s
s
us
pi
c
io
us
,
th
e
f
in
a
l
pr
e
di
c
ti
on i
s
c
or
r
e
c
te
d t
o
a
tt
a
c
k
.
3.3. S
p
e
c
ia
li
s
t
c
om
p
on
e
n
t
s
T
hi
s
s
e
c
ti
on
e
l
a
bor
a
te
s
on
th
e
th
r
e
e
s
pe
c
i
a
li
s
t
c
om
pone
nt
s
th
a
t
c
ons
ti
tu
te
L
e
ve
l
1
of
th
e
S
T
A
C
K
-
E
D
a
r
c
hi
te
c
tu
r
e
.
E
a
c
h
c
om
pone
nt
r
e
pr
e
s
e
nt
s
a
di
s
ti
nc
t
a
na
ly
ti
c
a
l
pe
r
s
pe
c
ti
ve
,
c
ont
r
ib
ut
in
g
c
om
pl
e
m
e
nt
a
r
y
e
vi
de
nc
e
f
or
th
e
hi
ghe
r
-
le
ve
l
f
us
io
n
s
ta
ge
.
B
e
yond
th
e
ir
in
di
vi
dua
l
a
na
ly
ti
c
a
l
r
ol
e
s
,
th
e
s
e
s
p
e
c
ia
li
s
ts
a
r
e
de
s
ig
ne
d
to
r
e
f
le
c
t
how
r
e
a
l
-
w
or
ld
s
e
c
ur
it
y
te
a
m
s
di
s
tr
ib
ut
e
e
xpe
r
ti
s
e
a
c
r
os
s
di
f
f
e
r
e
nt
de
te
c
ti
on
m
oda
li
ti
e
s
.
E
a
c
h
c
om
pone
nt
c
a
pt
ur
e
s
a
uni
qu
e
f
a
c
e
t
of
e
ndpoint
be
ha
vi
or
pa
tt
e
r
n
r
e
c
ogni
ti
on
f
r
om
s
upe
r
vi
s
e
d
m
ode
ls
,
s
tr
uc
tu
r
a
l
c
ont
e
xt
f
r
om
gr
a
ph
-
ba
s
e
d
le
a
r
ni
ng,
a
nd
be
ha
vi
or
a
l
de
vi
a
ti
on
f
r
om
uns
upe
r
vi
s
e
d
te
c
hni
que
s
,
e
ns
ur
in
g
th
a
t
no
s
in
gl
e
vi
e
w
poi
nt
dom
in
a
te
s
th
e
de
c
is
i
on
-
m
a
ki
ng
pi
pe
li
ne
.
B
y
in
te
gr
a
ti
ng
th
e
s
e
he
te
r
oge
ne
ous
pe
r
s
pe
c
ti
ve
s
a
t
th
e
f
ounda
ti
ona
l
la
y
e
r
,
S
T
A
C
K
-
E
D
m
it
ig
a
te
s
m
ode
l
bi
a
s
,
r
e
duc
e
s
s
in
gl
e
-
poi
nt
f
a
il
ur
e
r
is
ks
,
a
nd
e
s
ta
bl
is
he
s
a
r
ic
he
r
e
vi
de
nc
e
b
a
s
e
f
or
th
e
m
e
ta
-
le
a
r
ne
r
to
s
ynt
he
s
iz
e
in
to
a
r
obus
t
f
in
a
l
ju
dgm
e
nt
.
3.3.1. S
u
p
e
r
vi
s
e
d
c
om
p
on
e
n
t
T
he
s
upe
r
vi
s
e
d
c
om
pone
nt
s
e
r
ve
s
a
s
a
pa
tt
e
r
n
r
e
c
ogni
ti
on
e
xpe
r
t
th
a
t
c
a
pt
ur
e
s
r
e
gul
a
r
it
ie
s
in
pr
e
vi
ous
ly
obs
e
r
ve
d
a
tt
a
c
k
be
h
a
vi
or
s
.
I
t
e
m
pl
oys
a
n
e
n
s
e
m
bl
e
of
gr
a
di
e
nt
boos
ti
ng
c
la
s
s
if
ie
r
s
(
C
a
tB
oo
s
t
a
nd
X
G
B
oos
t)
c
om
bi
ne
d
th
r
ough
a
s
of
t
vot
in
g
s
tr
a
te
gy,
w
he
r
e
pr
e
di
c
ti
on
pr
oba
bi
li
ti
e
s
a
r
e
a
ve
r
a
ge
d
to
f
or
m
a
c
ons
e
ns
u
s
.
H
ype
r
pa
r
a
m
e
te
r
tu
ni
ng
w
a
s
pe
r
f
or
m
e
d
us
in
g
O
pt
una
,
a
B
a
ye
s
ia
n
opt
im
iz
a
ti
on
f
r
a
m
e
w
or
k
th
a
t
s
ys
te
m
a
ti
c
a
ll
y
e
xpl
or
e
s
th
e
s
e
a
r
c
h
s
pa
c
e
to
m
in
im
iz
e
f
a
ls
e
ne
ga
ti
ve
s
.
T
hi
s
de
s
ig
n
e
ns
ur
e
s
a
r
obus
t
bi
a
s
-
va
r
ia
nc
e
tr
a
de
-
of
f
,
w
hi
c
h
is
c
r
it
ic
a
l
f
or
c
ybe
r
s
e
c
ur
it
y
c
ont
e
xt
s
w
he
r
e
unde
te
c
te
d
a
tt
a
c
ks
m
a
y
le
a
d
to
s
e
ve
r
e
c
ons
e
que
nc
e
s
.
T
a
bl
e
4
s
um
m
a
r
iz
e
s
th
e
hype
r
pa
r
a
m
e
te
r
r
a
nge
s
e
xpl
or
e
d
f
or
bot
h
C
a
tB
oos
t
a
nd
X
G
B
oos
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
dv A
ppl
S
c
i
I
S
S
N
:
2252
-
8814
St
ac
k
in
g ar
c
hi
te
c
tu
r
e
-
e
ndpoint
de
te
c
ti
on:
a hy
br
id
m
ul
ti
-
la
y
e
r
e
d ar
c
hi
te
c
tu
r
e
…
(
A
bd R
ahm
an W
ahi
d)
1269
dur
in
g
th
e
opt
im
iz
a
ti
on
pr
oc
e
s
s
.
T
h
e
pr
im
a
r
y
out
put
of
th
is
c
o
m
pone
nt
is
th
e
E
ns
e
m
bl
e
P
r
oba
,
r
e
pr
e
s
e
nt
in
g
th
e
pr
oba
bi
li
ty
of
a
tr
a
f
f
ic
f
lo
w
be
in
g
m
a
li
c
io
us
.
T
hi
s
out
put
f
or
m
s
a
ke
y
in
put
(
m
e
ta
-
f
e
a
tu
r
e
)
f
or
th
e
m
e
ta
-
le
a
r
ne
r
.
T
a
bl
e
4
.
H
ype
r
pa
r
a
m
e
te
r
s
e
a
r
c
h s
p
a
c
e
f
or
s
upe
r
vi
s
e
d
c
om
pone
nt
s
us
in
g optuna
M
ode
l
H
ype
r
pa
r
a
m
e
t
e
r
s
T
ype
R
a
nge
/
v
a
l
ue
C
a
t
B
oos
t
I
t
e
r
a
t
i
on
D
e
pt
h
L
e
a
r
ni
ng_r
a
t
e
12_l
e
a
f
_r
e
g
I
nt
e
ge
r
I
nt
e
ge
r
F
l
oa
t
F
l
oa
t
[
200.500]
[
4, 6]
[
0.01, 0.05]
[
1, 10]
X
G
B
oos
t
n_e
s
t
i
m
a
t
or
s
M
a
x_de
pt
h
L
e
a
r
ni
ng_r
a
t
e
r
e
g_l
a
m
bda
I
nt
e
ge
r
I
nt
e
ge
r
F
l
oa
t
F
l
oa
t
[
200. 500]
[
4, 6]
[
0.01, 0.05]
[
1, 10]
3.3.2. Gr
ap
h
-
b
as
e
d
s
e
lf
-
s
u
p
e
r
vi
s
e
d
c
o
m
p
on
e
n
t
s
T
he
s
e
lf
-
s
upe
r
vi
s
e
d
c
om
pone
nt
a
c
ts
a
s
a
ne
twor
k
s
tr
uc
tu
r
e
e
xp
e
r
t
.
N
e
twor
k
tr
a
f
f
ic
is
a
bs
tr
a
c
te
d
in
to
gr
a
ph
r
e
pr
e
s
e
nt
a
ti
ons
,
w
he
r
e
e
a
c
h
f
lo
w
is
m
ode
le
d
a
s
a
s
m
a
ll
g
r
a
ph
w
it
h
two
nod
e
s
(
e
ndpoint
s
)
c
onne
c
te
d
by
a
n
e
dge
th
a
t
c
a
r
r
ie
s
s
ta
ti
s
ti
c
a
l
a
tt
r
ib
ut
e
s
(
e
.g.,
byt
e
s
,
dur
a
ti
on,
a
nd
in
te
r
-
a
r
r
iv
a
l
ti
m
e
)
.
N
ode
f
e
a
tu
r
e
s
a
r
e
de
r
iv
e
d
f
r
om
U
M
A
P
e
m
be
ddi
ngs
,
e
ns
ur
in
g
a
c
om
pa
c
t
ye
t
di
s
c
r
im
in
a
ti
ve
la
te
nt
r
e
pr
e
s
e
nt
a
ti
on.
T
he
gr
a
ph
m
ode
l
e
m
pl
oye
d i
s
G
I
N
E
C
onv, tr
a
in
e
d t
hr
ough a
t
w
o
-
pha
s
e
s
tr
a
te
gy:
–
C
ont
r
a
s
ti
ve
l
e
a
r
ni
ng
.
T
r
a
in
in
g
be
gi
ns
w
i
th
th
e
I
nf
oN
C
E
lo
s
s
[
4
1]
,
de
s
ig
ne
d
to
m
a
x
im
iz
e
s
im
il
a
r
it
y
be
tw
e
e
n
e
m
be
d
di
ngs
o
f
no
r
m
a
l
-
n
or
m
a
l
pa
i
r
s
w
h
il
e
s
e
pa
r
a
t
in
g
n
o
r
m
a
l
-
a
t
ta
c
k
p
a
i
r
s
.
T
h
is
e
nc
o
ur
a
ge
s
th
e
f
o
r
m
a
t
io
n
o
f
e
m
be
dd
in
gs
th
a
t
a
r
e
bo
th
r
o
bus
t
a
nd
di
s
c
r
i
m
in
a
t
iv
e
[
43
]
,
w
he
r
e
r
e
pr
e
s
e
n
ta
t
io
ns
a
r
e
le
a
r
ne
d
by
d
is
t
in
g
ui
s
hi
ng
in
f
o
r
m
a
t
iv
e
pos
it
iv
e
s
a
m
p
le
s
f
r
om
n
oi
s
e
-
ba
s
e
d
ne
g
a
t
iv
e
s
th
r
o
ug
h
t
he
I
n
f
o
N
C
E
o
bj
e
c
ti
ve
.
=
[
e
x
p
(
(
,
)
e
x
p
(
(
,
)
)
+
∑
e
x
p
=
1
(
(
,
)
)
]
(
3)
W
he
r
e
a/
p
de
not
e
s
a
nc
hor
-
pos
it
iv
e
(
nor
m
a
l)
pa
ir
s
,
r
e
pr
e
s
e
nt
s
a
tt
a
c
k
e
m
be
ddi
ng
s
,
a
nd
τ
=
0.07
is
te
m
pe
r
a
tu
r
e
pa
r
a
m
e
te
r
.
–
A
dve
r
s
a
r
ia
l
f
in
e
-
tu
ni
ng
.
T
o
f
ur
th
e
r
e
nha
nc
e
r
obus
tn
e
s
s
,
a
dve
r
s
a
r
ia
l
s
a
m
pl
e
s
a
r
e
ge
n
e
r
a
te
d
us
in
g
th
e
f
a
s
t
gr
a
di
e
nt
m
e
th
od
(
F
G
M
)
[
42]
.
T
he
s
e
pe
r
tu
r
ba
ti
ons
a
r
e
in
tr
oduc
e
d
a
s
a
ddi
ti
ona
l
ne
ga
ti
ve
e
xa
m
pl
e
s
,
tr
a
in
in
g t
he
G
N
N
t
o r
e
s
is
t
e
va
s
io
n a
tt
e
m
pt
s
t
ha
t
s
ubt
ly
m
a
ni
pul
a
te
ne
twor
k f
e
a
tu
r
e
s
.
T
he
c
o
nc
e
p
tu
a
l
w
o
r
k
f
l
ow
o
f
t
hi
s
a
d
ve
r
s
a
r
ia
l
t
r
a
in
in
g
p
r
o
c
e
s
s
is
il
lu
s
tr
a
te
d
in
F
i
gu
r
e
3,
w
he
r
e
r
ob
us
t
e
m
be
dd
in
gs
a
r
e
opt
i
m
iz
e
d
i
te
r
a
t
iv
e
ly
th
r
o
ugh
c
on
tr
a
s
t
iv
e
a
nd
a
dve
r
s
a
r
ia
l
pa
ir
in
gs
.
T
he
f
in
a
l
ou
tp
ut
of
th
is
c
om
p
one
nt
is
a
32
-
d
im
e
ns
i
ona
l
e
m
be
ddi
ng
th
a
t
c
a
pt
ur
e
s
bot
h
r
e
la
ti
ona
l
s
t
r
uc
t
ur
e
a
n
d
r
e
s
i
li
e
nc
e
to
a
dve
r
s
a
r
ia
l
pe
r
tu
r
ba
t
io
ns
.
3.3.3. Un
s
u
p
e
r
vi
s
e
d
an
om
al
y d
e
t
e
c
t
io
n
c
om
p
on
e
n
t
s
T
he
uns
upe
r
vi
s
e
d
c
om
pone
nt
f
unc
ti
ons
a
s
a
be
ha
vi
or
a
l
de
vi
a
ti
on
e
xpe
r
t
,
f
oc
us
in
g
on
th
e
de
te
c
ti
on
of
r
a
r
e
or
pr
e
vi
ous
ly
uns
e
e
n
a
tt
a
c
k
s
.
U
nl
ik
e
s
upe
r
vi
s
e
d
le
a
r
ni
ng,
it
ope
r
a
te
s
e
nt
ir
e
ly
on
th
e
32
-
di
m
e
n
s
io
na
l
G
N
N
e
m
be
ddi
ngs
a
nd
doe
s
not
r
e
ly
on
la
be
ls
.
I
ns
t
e
a
d
of
r
e
ly
in
g
on
a
s
in
gl
e
m
e
th
od, a
w
e
ig
ht
e
d
e
n
s
e
m
bl
e
of
a
nom
a
ly
de
te
c
to
r
s
i
s
c
ons
tr
uc
te
d,
in
c
lu
di
ng
A
ut
oe
nc
ode
r
r
e
c
ons
tr
uc
ti
on
e
r
r
or
,
M
a
ha
la
nobi
s
di
s
ta
nc
e
,
I
s
ol
a
ti
on
F
or
e
s
t,
a
nd
one
-
c
la
s
s
S
V
M
.
T
he
c
ont
r
ib
ut
io
n
of
e
a
c
h
de
te
c
to
r
is
w
e
ig
ht
e
d
a
c
c
or
di
ng
to
it
s
F
2
-
s
c
or
e
pe
r
f
or
m
a
nc
e
on
th
e
tr
a
in
in
g
da
ta
,
e
ns
ur
in
g
th
a
t
m
e
th
od
s
w
it
h
hi
ghe
r
di
s
c
r
im
in
a
ti
ve
c
a
pa
c
it
y
e
xe
r
t
gr
e
a
te
r
in
f
lu
e
nc
e
. F
or
m
a
ll
y, t
he
e
ns
e
m
bl
e
s
c
or
e
i
s
de
f
in
e
d a
s
(
4)
.
=
2
∑
2
=
1
;
=
∑
=
1
(
4)
W
he
r
e
is
th
e
nor
m
a
li
z
e
d
a
nom
a
ly
s
c
or
e
f
r
om
de
te
c
to
r
i
,
a
nd
k
is
th
e
to
ta
l
num
be
r
of
de
te
c
to
r
s
.
T
a
bl
e
5
pr
e
s
e
nt
s
th
e
r
e
la
ti
ve
w
e
ig
ht
s
a
s
s
ig
ne
d
to
e
a
c
h
d
e
te
c
to
r
in
th
e
f
in
a
l
e
n
s
e
m
bl
e
,
r
e
f
le
c
ti
ng
th
e
ir
pr
opor
ti
ona
l
c
ont
r
ib
ut
io
n.
T
he
out
c
om
e
of
th
is
c
om
pone
nt
is
a
s
e
t
of
a
nom
a
ly
s
c
or
e
s
th
a
t
f
or
m
th
e
uns
upe
r
vi
s
e
d
m
e
ta
-
f
e
a
tu
r
e
s
.
T
he
s
e
s
e
r
ve
a
s
a
c
r
it
ic
a
l
c
om
pl
e
m
e
nt
to
th
e
s
upe
r
vi
s
e
d
a
n
d
s
e
lf
-
s
upe
r
vi
s
e
d
pe
r
s
p
e
c
ti
ve
s
,
e
nr
ic
hi
ng
th
e
e
vi
de
nc
e
a
va
il
a
bl
e
t
o t
he
m
e
ta
-
le
a
r
ne
r
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8814
I
nt
J
A
dv A
ppl
S
c
i
,
V
ol
.
14
, N
o.
4
,
D
e
c
e
m
be
r
20
25
:
126
3
-
12
80
1270
F
ig
ur
e
3. W
or
kf
lo
w
of
a
dve
r
s
a
r
ia
l
f
in
e
-
tu
ni
ng
f
or
t
he
G
I
N
E
C
on
v gr
a
ph mode
l
T
a
bl
e
5. R
e
la
ti
ve
w
e
ig
ht
s
of
a
nom
a
ly
de
t
e
c
ti
on c
om
pone
nt
s
D
e
t
e
c
t
or
A
c
t
ua
l
w
e
i
ght
va
l
ue
G
N
N
(
i
ni
t
i
a
l
pr
e
di
c
t
i
on
)
0.409
M
a
ha
l
a
nobi
s
0.200
A
ut
oe
nc
ode
r
0.198
I
s
ol
a
t
i
on
f
or
e
st
0.109
O
ne
-
c
l
a
s
s
S
V
M
0.084
3.4.
I
n
t
e
ll
ig
e
n
c
e
f
u
s
io
n
b
y
m
e
t
a
-
le
ar
n
e
r
A
t
th
e
s
e
c
ond
le
ve
l
of
th
e
S
T
A
C
K
-
E
D
f
r
a
m
e
w
or
k,
in
te
ll
ig
e
nc
e
f
us
io
n
is
pe
r
f
or
m
e
d
by
a
m
e
ta
-
le
a
r
ne
r
th
a
t
in
te
gr
a
te
s
th
e
di
ve
r
s
e
e
vi
de
nc
e
pr
oduc
e
d
by
th
e
s
p
e
c
ia
li
s
t
c
om
pone
nt
s
.
T
hi
s
m
e
c
h
a
ni
s
m
e
m
ul
a
te
s
th
e
r
ol
e
of
a
n
in
te
ll
ig
e
nc
e
a
na
ly
s
is
c
e
nt
e
r
,
w
he
r
e
in
put
s
f
r
om
m
ul
ti
pl
e
e
xpe
r
ts
a
r
e
c
ons
ol
id
a
te
d
in
to
a
s
in
gl
e
,
m
or
e
r
e
li
a
bl
e
de
c
is
io
n.
T
he
m
e
ta
-
le
a
r
ne
r
is
im
p
le
m
e
nt
e
d
us
in
g
th
e
C
a
tB
oos
tC
la
s
s
if
ie
r
,
s
e
le
c
te
d
f
or
it
s
a
bi
li
t
y
to
ha
ndl
e
he
te
r
oge
ne
ou
s
f
e
a
tu
r
e
di
s
tr
ib
ut
io
ns
a
nd
f
or
it
s
r
obus
tn
e
s
s
to
ove
r
f
it
ti
ng.
I
nput
s
to
th
is
le
a
r
ne
r
a
r
e
th
e
m
e
ta
-
f
e
a
tu
r
e
s
de
r
iv
e
d
f
r
om
s
upe
r
vi
s
e
d,
s
e
lf
-
s
upe
r
vi
s
e
d,
a
nd
uns
upe
r
vi
s
e
d
c
om
pone
nt
s
(
s
e
e
T
a
bl
e
3)
,
in
c
lu
di
ng
pr
oba
bi
li
ti
e
s
,
a
nom
a
ly
s
c
or
e
s
,
a
nd
s
tr
uc
tu
r
a
l
pr
e
di
c
ti
ons
.
T
o
e
ns
ur
e
th
a
t
th
e
m
ode
l
ge
ne
r
a
li
z
e
s
w
e
ll
a
nd
e
m
pha
s
iz
e
s
r
e
c
a
ll
in
s
e
c
ur
it
y
-
s
e
ns
it
iv
e
s
c
e
na
r
io
s
,
th
e
op
ti
m
iz
a
ti
on
pr
oc
e
s
s
e
m
pl
oye
d
a
3
-
f
ol
d
c
r
os
s
-
va
li
da
ti
on
w
it
h
th
e
F
2
-
s
c
or
e
a
s
th
e
pr
im
a
r
y
obj
e
c
ti
ve
.
T
he
F
2
-
s
c
or
e
w
a
s
c
ho
s
e
n
b
e
c
a
us
e
it
pe
na
li
z
e
s
f
a
ls
e
ne
ga
ti
ve
s
m
or
e
he
a
vi
ly
,
r
e
f
le
c
ti
ng
th
e
hi
gh
c
os
t
o
f
unde
te
c
te
d
a
tt
a
c
ks
in
r
e
a
l
-
w
or
ld
a
ppl
ic
a
ti
ons
.
T
a
bl
e
6
s
um
m
a
r
iz
e
s
t
he
hype
r
pa
r
a
m
e
te
r
s
e
a
r
c
h s
p
a
c
e
e
xpl
or
e
d u
s
in
g G
r
id
S
e
a
r
c
hC
V
.
T
he
opt
im
iz
a
ti
on
pr
oc
e
dur
e
s
y
s
te
m
a
ti
c
a
ll
y
e
va
lu
a
te
d
a
ll
pa
r
a
m
e
te
r
c
om
bi
na
ti
ons
,
a
nd
th
e
c
onf
ig
ur
a
ti
on
th
a
t
m
a
xi
m
iz
e
d
th
e
F
2
-
s
c
or
e
dur
in
g
c
r
os
s
-
va
li
da
ti
on
w
a
s
s
e
le
c
te
d.
T
he
f
in
a
l
m
e
ta
-
le
a
r
ne
r
th
us
r
e
pr
e
s
e
nt
s
a
n
opt
im
iz
e
d
de
c
is
io
n
f
us
io
n
e
ngi
ne
,
d
e
s
ig
ne
d
not
onl
y
to
m
a
xi
m
iz
e
pr
e
di
c
ti
ve
a
c
c
ur
a
c
y
but
a
l
s
o
to
e
nha
nc
e
r
obus
tn
e
s
s
a
ga
in
s
t
unde
te
c
te
d
in
tr
us
io
ns
.
S
ubs
e
que
nt
va
li
da
ti
on
of
th
is
m
ode
l,
in
c
lu
di
ng
P
H
C
,
is
pr
e
s
e
nt
e
d i
n s
e
c
ti
on 4.
T
a
bl
e
6. H
ype
r
pa
r
a
m
e
te
r
s
e
a
r
c
h s
p
a
c
e
f
or
m
e
ta
-
le
a
r
ne
r
(
G
r
id
S
e
a
r
c
hC
V
)
H
ype
r
pa
r
a
m
e
t
e
r
s
T
ype
V
a
l
ue
e
xpl
or
e
d
I
t
e
r
a
t
i
ons
I
nt
e
ge
r
[
300, 500]
D
e
pt
h
I
nt
e
ge
r
[
4, 6, 8]
L
e
a
r
ni
ng_r
a
t
e
F
l
oa
t
[
0.01, 0.05, 0.1]
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
dv A
ppl
S
c
i
I
S
S
N
:
2252
-
8814
St
ac
k
in
g ar
c
hi
te
c
tu
r
e
-
e
ndpoint
de
te
c
ti
on:
a hy
br
id
m
ul
ti
-
la
y
e
r
e
d ar
c
hi
te
c
tu
r
e
…
(
A
bd R
ahm
an W
ahi
d)
1271
3.5.
I
n
t
e
r
p
r
e
t
ab
il
it
y an
d
p
os
t
-
h
oc
c
or
r
e
c
t
io
n
A
ke
y
de
s
ig
n
pr
in
c
ip
le
of
S
T
A
C
K
-
E
D
is
it
s
a
bi
li
ty
to
e
vol
ve
o
ve
r
ti
m
e
a
nd
r
e
m
a
in
tr
a
ns
pa
r
e
nt
in
it
s
de
c
is
io
n
-
m
a
ki
ng.
T
o
a
c
hi
e
ve
th
is
,
two
c
om
pl
e
m
e
nt
a
r
y
m
e
c
h
a
ni
s
m
s
a
r
e
in
tr
oduc
e
d:
in
te
r
pr
e
ta
bi
li
ty
th
r
ough
X
A
I
te
c
hni
que
s
a
nd
a
P
H
C
pr
oc
e
s
s
th
a
t
s
a
f
e
gu
a
r
ds
a
g
a
in
s
t
f
a
ls
e
ne
g
a
ti
ve
s
.
T
he
s
e
m
e
c
ha
ni
s
m
s
e
n
s
ur
e
th
a
t
S
T
A
C
K
-
E
D
doe
s
not
ope
r
a
te
a
s
a
s
ta
ti
c
or
opa
que
m
od
e
l,
but
r
a
th
e
r
a
s
a
s
y
s
te
m
c
a
p
a
bl
e
of
c
ont
in
uous
s
e
lf
-
a
s
s
e
s
s
m
e
nt
a
nd
r
e
f
in
e
m
e
nt
.
I
nt
e
r
pr
e
ta
bi
li
ty
e
na
bl
e
s
s
e
c
ur
it
y
a
na
ly
s
ts
to
tr
a
c
e
how
e
vi
de
nc
e
is
w
e
ig
ht
e
d,
r
e
duc
in
g
th
e
c
ogni
ti
ve
ba
r
r
ie
r
o
f
te
n
a
s
s
oc
ia
te
d
w
it
h
c
om
pl
e
x
m
a
c
hi
ne
le
a
r
ni
ng
pi
pe
li
ne
s
a
nd
s
tr
e
ngt
he
ni
ng
ope
r
a
ti
ona
l
tr
us
t.
M
e
a
nw
hi
le
,
th
e
P
H
C
la
ye
r
a
c
ts
a
s
a
s
e
c
ond
a
r
y
a
na
ly
ti
c
a
l
s
w
e
e
p,
a
ll
ow
in
g
th
e
s
ys
t
e
m
to
r
e
vi
s
it
bor
de
r
li
ne
c
a
s
e
s
w
it
h
de
e
p
e
r
s
c
r
ut
in
y
by
le
ve
r
a
gi
ng
s
tr
uc
tu
r
a
l
e
m
be
ddi
ngs
a
nd
s
e
ns
it
iv
e
a
nom
a
ly
de
te
c
to
r
s
.
T
og
e
th
e
r
,
th
e
s
e
c
om
pone
nt
s
not
onl
y
e
nha
nc
e
tr
a
ns
pa
r
e
nc
y
a
nd
r
e
s
il
ie
nc
e
but
a
l
s
o
po
s
it
io
n
S
T
A
C
K
-
E
D
a
s
a
pr
a
c
ti
c
a
ll
y
de
pl
oya
bl
e
f
r
a
m
e
w
or
k
a
li
gne
d
w
it
h
m
ode
r
n
e
xpe
c
ta
ti
ons
f
or
a
c
c
ount
a
bl
e
a
nd
a
da
pt
iv
e
A
I
-
dr
iv
e
n s
e
c
ur
it
y s
ys
te
m
s
.
3.5.1. I
n
t
e
r
p
r
e
t
ab
il
it
y an
d
e
f
f
ic
ie
n
c
y i
m
p
r
ov
e
m
e
n
t
(
ac
t
iv
e
l
e
ar
n
in
g
)
I
nt
e
r
pr
e
ta
bi
li
ty
is
a
ddr
e
s
s
e
d
u
s
in
g
S
H
A
P
[
44]
,
w
hi
c
h
pr
ovi
de
s
a
th
e
or
e
ti
c
a
ll
y
gr
ounde
d
m
e
a
ns
to
qua
nt
if
y t
he
c
ont
r
ib
ut
io
n of
e
a
c
h m
e
ta
-
f
e
a
tu
r
e
t
o
th
e
f
in
a
l
de
c
is
io
n. B
y de
c
om
pos
in
g pr
e
di
c
ti
ons
i
nt
o a
ddi
ti
ve
f
e
a
tu
r
e
a
tt
r
ib
ut
io
ns
,
S
H
A
P
e
na
bl
e
s
hum
a
n
a
na
ly
s
ts
to
v
a
li
da
te
a
nd
unde
r
s
ta
nd
th
e
r
e
a
s
oni
ng
of
th
e
m
ode
l.
T
he
s
e
in
te
r
pr
e
ta
bi
li
ty
in
s
ig
ht
s
a
r
e
f
ur
th
e
r
ut
il
iz
e
d
to
s
uppor
t
a
n
A
c
ti
ve
L
e
a
r
ni
ng
c
y
c
le
[
45]
.
I
n
th
is
c
y
c
le
,
in
s
ta
nc
e
s
th
a
t
th
e
m
ode
l
i
s
m
o
s
t
unc
e
r
ta
in
a
bout
a
r
e
pr
io
r
it
iz
e
d
f
or
f
ur
th
e
r
in
s
pe
c
ti
on
a
nd
r
e
tr
a
in
in
g.
U
nc
e
r
ta
in
ty
is
m
e
a
s
ur
e
d
us
in
g
S
ha
nnon
e
nt
r
opy,
w
he
r
e
a
pr
ob
a
bi
li
ty
di
s
tr
ib
ut
io
n
c
lo
s
e
to
uni
f
or
m
in
d
ic
a
te
s
m
a
xi
m
um
c
onf
us
io
n
a
s
(
5)
.
(
)
=
−
2
(
)
−
(
1
−
)
2
(
1
−
)
(
5)
H
e
r
e
,
r
e
pr
e
s
e
nt
s
th
e
pr
e
di
c
te
d
pr
oba
bi
li
ty
of
th
e
pos
it
iv
e
c
l
a
s
s
.
W
h
e
n
≈0.5,
th
e
e
nt
r
opy
r
e
a
c
he
s
it
s
m
a
xi
m
um
,
s
ig
na
li
ng
th
a
t
th
e
s
a
m
pl
e
is
hi
ghl
y
a
m
bi
guous
.
B
y
i
te
r
a
ti
ve
ly
in
c
or
por
a
ti
ng
s
uc
h
hi
gh
-
unc
e
r
ta
in
ty
s
a
m
pl
e
s
in
to
th
e
tr
a
in
in
g
pr
oc
e
s
s
,
S
T
A
C
K
-
E
D
m
a
in
ta
in
s
c
ont
i
nuous
im
pr
ove
m
e
nt
a
nd
be
tt
e
r
ge
ne
r
a
li
z
a
ti
on
ove
r
t
im
e
.
3.5.2. P
os
t
-
h
oc
c
or
r
e
c
t
io
n
m
e
c
h
an
is
m
T
o
a
ddr
e
s
s
th
e
pr
obl
e
m
of
f
a
ls
e
ne
ga
ti
ve
s
c
a
s
e
s
w
he
r
e
a
tt
a
c
ks
a
r
e
in
c
or
r
e
c
tl
y
c
la
s
s
if
ie
d
a
s
nor
m
a
l,
S
T
A
C
K
-
E
D
in
tr
oduc
e
s
a
P
H
C
m
e
c
ha
ni
s
m
a
s
a
la
s
t
li
ne
of
de
f
e
ns
e
.
T
hi
s
m
e
c
h
a
ni
s
m
le
ve
r
a
ge
s
e
m
b
e
ddi
ngs
pr
oduc
e
d
by
th
e
gr
a
ph
-
ba
s
e
d
c
om
pone
nt
a
nd
s
ubj
e
c
ts
th
e
m
to
a
s
e
r
ie
s
of
hi
ghl
y
s
e
n
s
it
iv
e
a
nom
a
ly
de
te
c
to
r
s
.
T
he
pr
oc
e
s
s
be
gi
n
s
by
is
ol
a
ti
ng
s
a
m
pl
e
s
pr
e
di
c
te
d
a
s
nor
m
a
l
by
th
e
m
e
ta
-
le
a
r
ne
r
.
T
he
s
e
a
r
e
r
e
-
e
xa
m
in
e
d
in
th
e
G
N
N
la
te
nt
s
pa
c
e
(
32
-
di
m
e
ns
io
na
l
e
m
be
ddi
ngs
)
,
w
he
r
e
s
ubt
le
de
vi
a
ti
ons
f
r
om
nor
m
a
l
pa
tt
e
r
ns
c
a
n
be
m
or
e
e
a
s
il
y
de
te
c
te
d.
S
e
v
e
r
a
l
c
om
pl
e
m
e
nt
a
r
y
de
te
c
to
r
s
a
r
e
a
p
pl
ie
d
M
a
ha
la
nobi
s
di
s
ta
nc
e
,
w
hi
c
h
m
e
a
s
ur
e
s
s
ta
ti
s
ti
c
a
l
di
s
ta
nc
e
f
r
om
th
e
di
s
tr
ib
ut
io
na
l
c
e
nt
e
r
of
no
r
m
a
l
da
ta
.
H
D
B
S
C
A
N
,
a
de
n
s
it
y
-
ba
s
e
d
c
lu
s
t
e
r
in
g
a
lg
or
it
hm
c
a
pa
bl
e
of
la
be
li
ng
lo
w
-
de
ns
it
y
poi
nt
s
a
s
noi
s
e
.
L
O
F
,
w
hi
c
h
e
va
lu
a
te
s
th
e
lo
c
a
l
de
n
s
it
y
de
vi
a
ti
on
of
a
s
a
m
pl
e
r
e
la
ti
ve
t
o i
ts
ne
ig
hbor
s
,
a
nd
C
os
in
e
s
im
il
a
r
it
y, w
hi
c
h c
a
pt
ur
e
s
de
vi
a
ti
ons
i
n di
r
e
c
ti
ona
l
s
im
il
a
r
it
y
a
m
ong
e
m
be
ddi
ngs
.
I
f
a
ny
de
te
c
to
r
id
e
nt
if
ie
s
a
n
in
s
ta
nc
e
a
s
a
n
om
a
lo
us
,
th
e
pr
e
di
c
ti
on
i
s
c
or
r
e
c
te
d
to
a
tt
a
c
k.
T
hi
s
la
ye
r
e
d
c
or
r
e
c
ti
on
pr
oc
e
s
s
s
e
r
ve
s
a
s
a
s
a
f
e
gua
r
d
a
ga
in
s
t
e
va
s
iv
e
a
nd
z
e
r
o
-
da
y
a
tt
a
c
k
s
,
r
e
in
f
or
c
in
g
S
T
A
C
K
-
E
D
’
s
r
e
s
il
ie
nc
e
be
yond the
i
ni
ti
a
l
c
l
a
s
s
if
ic
a
ti
on
s
ta
ge
.
3.6.
A
d
ve
r
s
ar
ia
l
r
ob
u
s
t
n
e
s
s
e
val
u
at
io
n
B
e
yond
pr
e
di
c
ti
ve
a
c
c
ur
a
c
y,
a
n
e
s
s
e
nt
ia
l
di
m
e
ns
io
n
of
e
v
a
lu
a
ti
ng
in
tr
us
io
n
de
te
c
ti
on
m
ode
ls
li
e
s
in
th
e
ir
r
obus
tn
e
s
s
a
ga
in
s
t
a
dve
r
s
a
r
ia
l
m
a
ni
pul
a
ti
on.
I
n
r
e
a
l
-
w
or
ld
s
c
e
na
r
io
s
,
a
tt
a
c
ke
r
s
m
a
y
de
li
be
r
a
te
ly
in
tr
oduc
e
s
ubt
le
pe
r
tu
r
ba
ti
ons
to
n
e
twor
k
tr
a
f
f
ic
in
or
de
r
to
e
va
de
d
e
te
c
ti
on.
T
o
a
s
s
e
s
s
th
e
r
e
s
il
ie
nc
e
of
S
T
A
C
K
-
E
D
,
a
n
A
dvR
e
va
lu
a
ti
on
w
a
s
c
onduc
te
d.
T
he
pr
oc
e
dur
e
f
ol
lo
w
s
th
e
pr
in
c
ip
le
of
w
hi
te
-
box
a
dve
r
s
a
r
ia
l
te
s
ti
ng,
w
he
r
e
th
e
a
dve
r
s
a
r
y
is
a
s
s
um
e
d
to
ha
ve
knowle
dge
of
th
e
ta
r
ge
t
m
ode
l.
P
e
r
tu
r
ba
ti
ons
a
r
e
ge
ne
r
a
te
d
us
in
g
th
e
F
G
M
,
a
w
id
e
ly
a
dopt
e
d
e
va
s
io
n
te
c
hni
q
ue
th
a
t
a
ppl
ie
s
s
m
a
ll
but
s
tr
a
te
gi
c
a
ll
y
c
r
a
f
te
d
m
odi
f
ic
a
ti
ons
to
in
put
f
e
a
tu
r
e
s
.
F
or
e
a
c
h
pe
r
tu
r
ba
ti
on
s
tr
e
ngt
h
,
pa
r
a
m
e
te
r
iz
e
d
by
ϵ
,
a
dve
r
s
a
r
ia
l
s
a
m
pl
e
s
a
r
e
pr
oduc
e
d a
nd s
ubs
e
que
nt
ly
us
e
d t
o t
e
s
t
th
e
m
ode
l’
s
s
ta
bi
li
ty
.
T
w
o m
ode
ls
w
e
r
e
s
ubj
e
c
te
d t
o t
hi
s
e
va
lu
a
ti
on
.
3.6.1.
S
u
p
e
r
vi
s
e
d
e
n
s
e
m
b
le
(
m
od
e
l
b
as
e
li
n
e
)
T
he
ba
s
e
li
ne
m
ode
l
c
or
r
e
s
ponds
to
th
e
s
upe
r
vi
s
e
d
e
ns
e
m
b
le
c
om
pone
nt
(
C
a
tB
oos
t
+
X
G
B
oos
t
)
ope
r
a
ti
ng
w
it
hout
hi
ghe
r
-
le
ve
l
f
us
io
n
o
r
c
or
r
e
c
ti
on
m
e
c
ha
ni
s
m
s
.
T
hi
s
c
onf
ig
ur
a
ti
on
pr
ovi
de
s
a
be
nc
hm
a
r
k
to
qua
nt
if
y
th
e
in
he
r
e
nt
vul
ne
r
a
bi
li
ty
o
f
c
onve
nt
io
na
l
s
upe
r
vi
s
e
d
a
ppr
oa
c
he
s
unde
r
a
dve
r
s
a
r
ia
l
pe
r
tu
r
ba
ti
ons
,
pa
r
ti
c
ul
a
r
ly
in
de
te
c
ti
ng
s
ubt
le
or
e
v
a
s
iv
e
a
tt
a
c
k
p
a
tt
e
r
ns
.
B
y
is
ol
a
ti
ng
th
e
s
upe
r
vi
s
e
d
e
n
s
e
m
bl
e
f
r
om
th
e
a
ddi
ti
ona
l
r
e
s
il
ie
nc
e
la
ye
r
s
im
pl
e
m
e
nt
e
d
in
S
T
A
C
K
-
E
D
,
th
is
ba
s
e
li
ne
e
na
bl
e
s
a
c
le
a
r
a
s
s
e
s
s
m
e
nt
of
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8814
I
nt
J
A
dv A
ppl
S
c
i
,
V
ol
.
14
, N
o.
4
,
D
e
c
e
m
be
r
20
25
:
126
3
-
12
80
1272
in
he
r
e
nt
li
m
it
a
ti
ons
of
s
ta
nda
lo
ne
boos
ti
ng
-
ba
s
e
d
c
la
s
s
if
ie
r
s
a
n
d
hi
ghl
ig
ht
s
th
e
pe
r
f
o
r
m
a
nc
e
ga
in
s
in
tr
oduc
e
d
by t
he
m
ul
ti
-
la
ye
r
e
d a
r
c
hi
te
c
tu
r
e
.
3.6.2.
S
T
A
C
K
-
ED
T
he
c
om
pl
e
t
e
S
T
A
C
K
-
E
D
a
r
c
hi
te
c
tu
r
e
,
in
te
gr
a
ti
ng
s
upe
r
vi
s
e
d,
s
e
lf
-
s
upe
r
vi
s
e
d,
a
nd
un
s
upe
r
vi
s
e
d
c
om
pone
nt
s
a
lo
ng
w
it
h
th
e
m
e
ta
-
le
a
r
ne
r
,
w
a
s
te
s
te
d
a
ga
in
s
t
th
e
s
a
m
e
a
dve
r
s
a
r
ia
l
s
a
m
pl
e
s
.
F
or
e
a
c
h
pe
r
tu
r
be
d
in
s
ta
nc
e
,
th
e
m
e
ta
-
f
e
a
tu
r
e
s
w
e
r
e
r
e
c
on
s
tr
uc
te
d
f
r
om
th
e
e
m
b
e
d
di
ngs
a
nd a
nom
a
ly
s
c
or
e
s
of
th
e
ba
s
e
-
le
a
r
ne
r
s
,
e
ns
ur
in
g
th
a
t
th
e
f
u
s
io
n
pr
oc
e
s
s
r
e
m
a
in
e
d
c
ons
i
s
te
nt
unde
r
a
tt
a
c
k
c
ondi
ti
ons
.
P
e
r
f
or
m
a
nc
e
de
gr
a
da
ti
on
w
a
s
m
e
a
s
ur
e
d
us
in
g
th
e
F
2
-
s
c
or
e
,
c
ho
s
e
n
be
c
a
us
e
it
e
m
pha
s
iz
e
s
r
e
c
a
ll
a
nd
di
r
e
c
tl
y
pe
na
li
z
e
s
f
a
l
s
e
ne
ga
ti
ve
s
,
a
n
e
s
pe
c
ia
ll
y
c
r
it
ic
a
l
pr
ope
r
ty
in
c
ybe
r
s
e
c
ur
it
y,
w
he
r
e
unde
te
c
te
d
a
tt
a
c
ks
c
a
r
r
y
s
e
ve
r
e
im
pl
ic
a
ti
on
s
.
B
y
c
om
pa
r
in
g
th
e
r
a
te
of
F
2
-
s
c
or
e
r
e
duc
ti
on
a
c
r
os
s
in
c
r
e
a
s
in
g
va
lu
e
s
of
,
th
e
e
va
lu
a
ti
on
pr
ovi
de
s
in
s
ig
ht
in
t
o
th
e
r
e
la
ti
ve
r
e
s
il
ie
nc
e
of
S
T
A
C
K
-
E
D
ve
r
s
us
th
e
ba
s
e
li
ne
m
o
de
l.
I
n
s
um
m
a
r
y,
th
e
pr
opos
e
d
m
e
th
odol
ogy
in
te
gr
a
te
s
m
ul
ti
pl
e
pe
r
s
pe
c
ti
ve
s
in
to
a
uni
f
ie
d
S
T
A
C
K
-
E
D
f
r
a
m
e
w
or
k,
be
gi
nni
ng
w
it
h
a
s
tr
uc
tu
r
e
d
da
ta
pi
pe
li
ne
,
f
ol
lo
w
e
d
by
la
ye
r
e
d
a
r
c
hi
te
c
tu
r
e
de
s
ig
n,
a
nd
c
ul
m
in
a
ti
ng
in
in
te
ll
ig
e
nc
e
f
us
io
n,
in
te
r
pr
e
ta
bi
li
ty
,
a
nd
r
e
s
il
ie
nc
e
m
e
c
h
a
ni
s
m
s
.
E
a
c
h
m
e
th
odol
ogi
c
a
l
c
om
pone
nt
w
a
s
de
li
be
r
a
te
ly
s
e
le
c
te
d
to
a
ddr
e
s
s
th
e
li
m
it
a
ti
ons
of
c
onve
nt
io
na
l
de
te
c
ti
on
s
ys
te
m
s
:
s
upe
r
vi
s
e
d
le
a
r
ni
ng
f
or
known
pa
tt
e
r
ns
,
gr
a
ph
-
ba
s
e
d
s
e
lf
-
s
upe
r
vi
s
e
d
le
a
r
ni
ng
f
or
s
tr
uc
tu
r
a
l
in
s
ig
ht
s
,
un
s
upe
r
vi
s
e
d
a
nom
a
ly
de
t
e
c
ti
on
f
or
nove
l
th
r
e
a
ts
,
a
nd
P
H
C
a
s
a
s
a
f
e
gua
r
d
a
ga
in
s
t
f
a
ls
e
ne
ga
ti
ve
s
.
T
he
in
c
lu
s
io
n
of
e
xpl
a
in
a
bi
li
ty
(
vi
a
S
H
A
P
)
,
a
c
ti
ve
le
a
r
ni
ng,
a
nd
A
dvR
e
va
lu
a
ti
on
f
ur
th
e
r
e
ns
ur
e
s
th
a
t
S
T
A
C
K
-
E
D
not
onl
y
a
c
hi
e
v
e
s
hi
gh
pr
e
di
c
ti
ve
pe
r
f
or
m
a
nc
e
but
a
l
s
o
m
a
in
ta
in
s
tr
a
ns
pa
r
e
nc
y
a
nd
a
da
pt
a
bi
li
ty
in
a
dve
r
s
a
r
ia
l
e
nvi
r
onm
e
nt
s
.
H
a
vi
ng
e
s
ta
bl
i
s
he
d
th
is
m
e
th
odol
ogi
c
a
l
f
ounda
ti
on,
th
e
ne
xt
s
e
c
ti
on
pr
e
s
e
nt
s
th
e
e
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
a
nd
di
s
c
u
s
s
e
s
how
th
e
pr
opos
e
d
f
r
a
m
e
w
or
k
pe
r
f
or
m
s
i
n pr
a
c
ti
c
e
a
c
r
os
s
m
ul
ti
pl
e
e
va
lu
a
ti
on dim
e
ns
io
n
s
.
4.
R
E
S
U
L
T
T
hi
s
s
e
c
ti
on
pr
e
s
e
nt
s
th
e
r
e
s
ul
ts
of
a
s
e
r
ie
s
of
e
m
pi
r
ic
a
l
e
va
lu
a
ti
ons
de
s
ig
ne
d
to
va
li
da
te
th
e
pe
r
f
or
m
a
nc
e
of
S
T
A
C
K
-
E
D
a
c
r
os
s
m
ul
ti
pl
e
di
m
e
ns
io
ns
of
e
n
dpoi
nt
th
r
e
a
t
de
te
c
ti
on.
T
he
r
e
por
te
d
f
in
di
ngs
de
m
ons
tr
a
te
how
th
e
pr
opos
e
d
hybr
id
a
r
c
hi
te
c
tu
r
e
e
nha
nc
e
s
pr
e
di
c
ti
ve
a
c
c
ur
a
c
y,
s
tr
e
ngt
he
ns
r
e
s
il
ie
n
c
e
a
ga
in
s
t
bot
h c
onve
nt
io
na
l
a
nd
a
dve
r
s
a
r
ia
l
th
r
e
a
ts
, a
nd
im
pr
ove
s
de
c
is
io
n
r
e
li
a
bi
li
ty
th
r
ough
m
ul
ti
-
pe
r
s
pe
c
ti
ve
in
te
ll
ig
e
nc
e
f
us
io
n.
B
e
yond
num
e
r
ic
a
l
pe
r
f
or
m
a
nc
e
,
th
e
r
e
s
ul
ts
a
ls
o
hi
ghl
ig
ht
th
e
m
ode
l’
s
in
te
r
pr
e
ta
bi
li
ty
a
nd
a
da
pt
a
bi
li
ty
,
e
m
pha
s
iz
in
g
how
S
T
A
C
K
-
E
D
m
a
in
ta
in
s
s
ta
bl
e
pe
r
f
or
m
a
nc
e
in
dyna
m
ic
e
nvi
r
onm
e
nt
s
w
hi
le
of
f
e
r
in
g t
r
a
ns
pa
r
e
nt
i
ns
ig
ht
s
t
ha
t
s
uppor
t
pr
a
c
ti
c
a
l
c
ybe
r
s
e
c
ur
it
y de
c
is
io
n
-
m
a
ki
ng.
4.1.
O
ve
r
al
l
p
e
r
f
or
m
a
n
c
e
o
f
S
T
A
C
K
-
E
D
m
od
el
T
he
pe
r
f
or
m
a
nc
e
of
S
T
A
C
K
-
E
D
w
a
s
e
va
lu
a
te
d pr
ogr
e
s
s
iv
e
ly
a
c
r
os
s
i
ts
c
ons
ti
tu
e
nt
s
ta
ge
s
, r
e
f
le
c
ti
ng
th
e
la
ye
r
e
d
c
om
pl
e
xi
ty
of
th
e
a
r
c
hi
te
c
tu
r
e
.
T
a
bl
e
7
r
e
por
ts
th
e
r
e
s
ul
ts
in
te
r
m
s
of
A
c
c
ur
a
c
y,
F
2
-
s
c
or
e
,
P
r
e
c
is
io
n,
R
e
c
a
ll
,
a
nd
R
O
C
-
A
U
C
f
or
e
a
c
h
s
ta
ge
,
r
a
ngi
ng
f
r
o
m
in
di
vi
dua
l
c
om
pone
nt
s
to
th
e
f
in
a
l
hybr
id
m
ode
l
e
nha
nc
e
d by a
c
ti
v
e
l
e
a
r
ni
ng.
T
a
bl
e
7. S
um
m
a
r
y of
e
va
lu
a
ti
on me
tr
ic
pe
r
f
or
m
a
nc
e
f
or
e
a
c
h s
t
a
ge
of
t
he
S
T
A
C
K
-
E
D
m
ode
l
M
ode
l
s
t
a
ge
A
c
c
ur
a
c
y
F2
-
s
c
or
e
P
r
e
c
i
s
i
on
R
e
c
a
l
l
R
O
C
-
AUC
S
upe
r
vi
s
e
d (
T
3)
0.9626
0.9662
0.9572
0.9685
0.9626
S
e
l
f
-
s
upe
r
vi
s
e
d (
T
4)
0.8383
0.8298
0.8472
0.8256
0.8383
U
ns
upe
r
vi
s
e
d
-
t
r
a
i
ni
ng (
T
5)
0.8239
0.8226
0.8252
0.8219
0.8239
U
ns
upe
r
vi
s
e
d
-
va
l
i
da
t
i
on (
T
5)
0.8224
0.8224
0.8224
0.8224
0.8224
H
ybr
i
d t
r
a
i
ni
ng (
T
6)
0.9866
0.9911
0.9796
0.9940
0.9866
H
ybr
i
d va
l
i
da
t
i
on (
T
6
)
0.9857
0.9898
0.9793
0.9924
0.9857
H
ybr
i
d+A
L
(
T
7)
0.9829
0.9889
0.9733
0.9929
0.9829
A
c
le
a
r
tr
e
nd
of
pr
ogr
e
s
s
iv
e
p
e
r
f
or
m
a
nc
e
im
pr
ove
m
e
nt
c
a
n
b
e
obs
e
r
ve
d
a
c
r
os
s
th
e
la
ye
r
e
d
s
ta
ge
s
.
T
he
s
upe
r
vi
s
e
d
e
ns
e
m
bl
e
a
t
s
t
a
ge
T
3
s
e
r
ve
s
a
s
a
s
tr
ong
ba
s
e
li
ne
,
a
c
hi
e
vi
ng
a
n
F
2
-
s
c
or
e
of
0.9662
a
nd
a
c
c
ur
a
c
y
of
0.9626,
w
hi
c
h
de
m
on
s
tr
a
te
s
it
s
r
e
li
a
bi
li
ty
in
id
e
nt
if
yi
ng
known
a
tt
a
c
k
pa
tt
e
r
ns
.
T
he
gr
a
ph
-
ba
s
e
d
s
e
lf
-
s
upe
r
vi
s
e
d
c
om
pone
nt
(
T
4)
,
w
hi
le
not
a
f
in
a
l
c
la
s
s
if
ie
r
,
c
ont
r
ib
ut
e
s
s
ig
ni
f
ic
a
nt
ly
by
pr
oduc
in
g
e
m
be
ddi
ngs
th
a
t
c
a
pt
ur
e
r
e
la
ti
ona
l
s
tr
uc
tu
r
e
s
,
w
it
h
a
n
F
2
-
s
c
or
e
of
0.8298.
S
im
il
a
r
ly
,
th
e
uns
upe
r
vi
s
e
d
a
nom
a
ly
de
te
c
ti
on
c
om
pone
nt
(
T
5)
a
c
hi
e
ve
s
a
n
F
2
-
s
c
or
e
o
f
0.8224,
r
e
f
le
c
ti
ng
it
s
c
a
pa
c
it
y
to
id
e
nt
if
y
a
nom
a
li
e
s
w
it
hout
la
be
l
e
d
da
ta
.
T
he
m
o
s
t
s
ub
s
ta
nt
ia
l
pe
r
f
or
m
a
nc
e
ga
in
o
c
c
ur
s
a
t
s
ta
ge
T
6,
w
h
e
r
e
th
e
m
e
t
a
-
le
a
r
ne
r
in
te
gr
a
te
s
out
put
s
f
r
om
a
ll
ba
s
e
-
le
a
r
ne
r
s
.
T
he
hybr
id
f
us
io
n
a
c
hi
e
ve
s
a
n
F
2
-
s
c
or
e
of
0.9898
a
nd
a
c
c
ur
a
c
y
of
0.9857,
e
vi
de
nc
in
g
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
s
ta
c
ki
ng
in
s
ynt
he
s
iz
in
g
di
ve
r
s
e
f
or
m
s
of
e
vi
de
nc
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.