I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
V
ol
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
, pp.
4778
~
4794
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
13
.i
4
.pp
4778
-
4794
4778
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
Fe
d
e
r
at
e
d
i
n
c
e
p
t
i
on
-
m
u
l
t
i
-
h
e
ad
at
t
e
n
t
i
on
m
od
e
l
s f
or
c
yb
e
r
-
at
t
ac
k
s d
e
t
e
c
t
i
on
I
m
ad
T
ar
e
q
AL
-
H
al
b
oos
i
1
, B
as
s
an
t
M
. E
lb
agou
r
y
1,
2
, S
al
s
ab
il
E
l
-
R
e
gai
ly
1
, E
l
-
S
aye
d
M
. E
l
-
H
or
b
at
y
1
1
D
e
pa
r
t
m
e
nt
of
C
om
put
e
r
S
c
i
e
nc
e
, F
a
c
ul
t
y of
C
om
put
e
r
a
nd
I
nf
or
m
a
t
i
on S
c
i
e
n
c
e
s
, A
i
n S
ha
m
s
U
ni
ve
r
s
i
t
y, C
a
i
r
o, E
gypt
2
F
a
c
ul
t
y of
C
om
put
e
r
S
c
i
e
nc
e
a
nd C
om
put
e
r
E
ngi
ne
e
r
i
ng
,
K
i
ng S
a
l
m
a
n I
nt
e
r
na
t
i
ona
l
U
ni
ve
r
s
i
t
y,
E
l
T
or
,
E
gypt
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 N
ov 24, 2023
R
e
vi
s
e
d F
e
b 20, 2024
A
c
c
e
pt
e
d M
a
r
21, 2024
With
the
proliferation
of
internet
of
things
(IoT)
devices,
ensuri
ng
the
security
of
these
interconnecte
d
systems
has
become
a
critical
co
ncern.
Cyberatt
acks ta
rgeting
IoT
devices pose significant threats to individu
als and
organizations
due
to
the
generation
of
vast
amounts
of
data
across
many
connected
devices,
which
traditi
onal
centralized
methods
cannot
solve.
Feder
ated
learn
ing
(FL)
could
be
a
promisin
g
solution
to
mitigate
privac
y
concerns
associated
with
centralized
approaches
and
address
cybersecuri
ty
concerns.
This
paper
uses
FL
and
deep
learning
(DL)
approac
hes
to
cybersecurit
y
in
IoT
applicati
ons.
The
goal
of
cyber
security
is
achieved
by
forming
a
federation
of
acquired
and
shared
models
at
the
head
of
the
various
participants.
We
use
inception
time
and
multi
-
head
attenti
on
(CNN)
algorit
hm
based
on
FL
to
detect
cyber
-
attacks
and
avoid
data
privacy
leaks
under
two
distribution
modes,
namely
IID
and
Non
-
IID.
In
contra
st,
the
FedAvg
and
FedMA
algorithms
aggregate
local
model
updates.
A
global
model
i
s
produced
after
several
communication
rounds
between
t
he
IoT
devices
and
the
model
parameter
server.
Cyber
threats
are
simulated
using
edge
-
IIoT
datasets.
Experiment
results
show
that
the
federated
inc
eption
model'
s
best
global
accuracy
was
93,
91%,
and
93,
49%
using
mult
i
-
head
attenti
on.
K
e
y
w
o
r
d
s
:
A
r
ti
f
ic
ia
l
in
te
ll
ig
e
nc
e
C
ybe
r
s
e
c
ur
it
y
D
e
e
p l
e
a
r
ni
ng
F
e
de
r
a
te
d l
e
a
r
ni
ng
I
nt
e
r
ne
t
of
t
hi
ngs
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
I
m
a
d T
a
r
e
q A
L
-
H
a
lb
oos
i
D
e
pa
r
tm
e
nt
of
C
om
put
e
r
S
c
ie
nc
e
, F
a
c
ul
ty
of
C
om
put
e
r
a
nd I
nf
or
m
a
ti
on S
c
ie
nc
e
s
, A
in
S
ha
m
s
U
ni
ve
r
s
it
y
El
-
K
ha
ly
f
a
E
l
-
M
a
m
oun S
tr
e
e
t
A
bba
s
ya
, C
a
ir
o 11566, E
gypt
E
m
a
il
:
e
m
a
dt
a
r
iq
1982@
gm
a
il
.c
om
1.
I
N
T
R
O
D
U
C
T
I
O
N
I
n
r
e
c
e
nt
ye
a
r
s
,
gove
r
nm
e
nt
s
,
a
c
a
de
m
ia
,
a
nd
in
dus
tr
y
ha
v
e
p
a
id
c
lo
s
e
a
tt
e
nt
io
n
to
th
e
in
te
r
ne
t
of
th
in
gs
(
I
oT
)
in
m
a
ny
f
ie
ld
s
,
s
uc
h
a
s
he
a
lt
hc
a
r
e
,
dr
one
s
,
s
m
a
r
t
c
it
ie
s
,
a
nd
c
ybe
r
-
phys
ic
a
l
s
ys
te
m
s
.
A
s
I
oT
de
vi
c
e
s
gr
ow
,
va
s
t
a
m
ount
s
of
da
ta
th
a
t
in
c
lu
d
e
us
e
r
s
'
pr
iv
a
te
i
nf
or
m
a
ti
on
w
il
l
be
ge
ne
r
a
te
d.
W
it
hout
r
e
li
a
bl
e
s
e
c
ur
it
y
s
ys
te
m
s
im
pl
e
m
e
nt
e
d
on
I
oT
de
vi
c
e
s
,
it
c
a
n
a
f
f
e
c
t
pe
r
s
ona
l
pr
iv
a
c
y
a
nd
m
a
y
be
e
xpos
e
d
to
m
a
ny
c
ybe
r
-
a
tt
a
c
ks
.
R
e
ga
r
dl
e
s
s
of
th
e
be
ne
f
it
s
pr
ovi
de
d
by
m
a
c
hi
ne
le
a
r
ni
ng
(
M
L
)
a
ppr
oa
c
he
s
,
it
s
c
a
pa
c
it
y
to
id
e
nt
if
y
th
r
e
a
ts
a
nd
e
xt
r
a
c
t
m
e
a
ni
ngf
ul
a
nd
c
om
pl
ic
a
te
d
d
a
ta
m
ode
ls
s
ta
nds
out
[
1]
.
T
r
a
di
ti
ona
l
ML
a
ppr
oa
c
he
s
c
ol
le
c
t
d
a
ta
on
a
c
e
nt
r
a
li
z
e
d
s
e
r
ve
r
w
hi
le
ig
nor
in
g
pr
iv
a
c
y
a
nd
s
e
c
ur
it
y
c
ons
id
e
r
a
ti
ons
.
B
e
c
a
u
s
e
of
th
e
vol
um
e
a
nd
s
e
n
s
it
iv
it
y
of
in
f
or
m
a
ti
on
tr
a
ns
m
it
te
d
be
twe
e
n
de
vi
c
e
s
in
I
oT
a
ppl
ic
a
ti
ons
,
th
is
pr
obl
e
m
m
ig
ht
be
a
m
pl
if
ie
d.
I
n
a
ddi
ti
on
to
p
r
ot
e
c
ti
ng
pr
iv
a
c
y,
one
of
t
he
pr
im
a
r
y
is
s
ue
s
w
it
h
I
oT
de
vi
c
e
s
is
th
a
t
I
oT
ne
twor
ks
ha
ve
s
e
t
ne
e
ds
f
or
pr
oc
e
s
s
in
g
r
e
s
our
c
e
s
,
w
hi
c
h
pos
e
s
a
s
ig
ni
f
ic
a
nt
ba
r
r
ie
r
.
A
s
a
r
e
s
ul
t,
one
s
ol
ut
io
n
is
t
o ha
ndl
e
da
ta
i
n a
de
c
e
nt
r
a
li
z
e
d m
a
nne
r
[
2]
, [
3]
.
F
e
de
r
a
te
d
le
a
r
ni
ng
(
F
L
)
is
th
e
m
os
t
pr
om
is
in
g
s
ol
ut
io
n
to
th
is
i
s
s
ue
.
I
n
FL
,
t
r
a
in
in
g
M
L
m
ode
ls
is
a
c
ol
la
bor
a
ti
ve
e
f
f
or
t
a
c
r
os
s
s
e
ve
r
a
l
c
li
e
nt
s
th
a
t
do
not
r
e
qui
r
e
th
e
lo
c
a
l
da
ta
to
be
t
r
a
ns
por
te
d
to
a
c
e
nt
r
a
li
z
e
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
F
e
de
r
at
e
d
in
c
e
pt
io
n
-
m
ul
ti
-
he
ad att
e
nt
io
n m
ode
ls
f
or
c
y
b
e
r
-
at
ta
c
k
s
de
te
c
ti
on
(
I
m
ad T
ar
e
q A
L
-
H
al
boos
i
)
4779
lo
c
a
ti
on,
r
e
duc
in
g
s
to
r
a
ge
c
o
s
ts
tr
a
n
s
m
is
s
io
n
w
hi
le
r
e
t
a
in
in
g
a
hi
gh
le
ve
l
of
us
e
r
pr
iv
a
c
y.
T
hi
s
a
ll
ow
s
bus
in
e
s
s
e
s
to
c
r
e
a
te
a
s
h
a
r
e
d
gl
oba
l
m
ode
l
w
it
hout
s
to
r
in
g
tr
a
in
in
g
da
ta
in
a
c
e
nt
r
a
li
z
e
d
lo
c
a
ti
on
[
4]
.
F
L
e
na
bl
e
s
m
ul
ti
pl
e
a
c
to
r
s
to
c
ol
la
bor
a
te
on
de
ve
lo
pi
ng
a
s
in
gl
e
,
r
obus
t
s
ys
te
m
w
it
hout
s
ha
r
in
g
da
ta
,
a
ddr
e
s
s
in
g
c
r
it
ic
a
l
is
s
ue
s
s
u
c
h
a
s
da
ta
pr
iv
a
c
y,
a
c
c
e
s
s
r
ig
ht
s
,
s
e
c
ur
it
y,
a
n
d
a
c
c
e
s
s
to
di
ve
r
s
e
da
ta
.
I
ns
te
a
d
of
a
pool
to
upda
te
t
he
de
vi
c
e
,
FL
obt
a
in
s
t
he
c
ur
r
e
nt
m
ode
l
a
nd c
om
put
e
s
a
m
odi
f
ie
d m
ode
l
on
t
he
s
ys
te
m
. T
he
s
e
l
oc
a
ll
y
tr
a
in
e
d
m
ode
ls
a
r
e
th
e
n
s
e
nt
f
r
om
th
e
m
a
c
hi
ne
s
to
th
e
c
e
nt
r
a
l
s
e
r
ve
r
a
nd
a
ggr
e
ga
te
d
be
f
or
e
be
in
g
di
s
tr
ib
ut
e
d
to
t
he
de
vi
c
e
s
a
s
a
uni
f
ie
d a
nd e
nha
n
c
e
d gl
oba
l
m
ode
l
[
5]
, [
6]
.
G
oogl
e
a
lr
e
a
dy
us
e
s
F
L
,
a
ll
ow
in
g
in
c
r
e
di
bl
e
pr
e
di
c
ti
ve
in
pu
t
f
e
a
tu
r
e
s
f
or
th
e
A
ndr
oi
d
ke
yboa
r
d,
on
-
de
vi
c
e
s
e
a
r
c
h
phone
s
,
a
nd
m
a
ny
ot
he
r
a
ppl
ic
a
ti
ons
.
R
e
c
e
nt
de
ve
lo
pm
e
nt
s
ha
ve
c
onc
e
nt
r
a
te
d
on
lo
w
e
r
in
g
s
ta
ti
s
ti
c
a
l
ba
r
r
ie
r
s
a
nd
im
pr
ovi
ng
F
L
s
e
c
ur
it
y
[
7]
,
[
8]
.
W
it
h
de
c
e
nt
r
a
li
z
e
d
tr
a
in
in
g
da
ta
,
to
p
s
e
r
vi
c
e
pr
ovi
de
r
s
ha
ve
us
e
d
F
L
a
ppr
oa
c
he
s
to
s
uppor
t
pr
iv
a
c
y
-
s
e
ns
it
iv
e
s
ys
t
e
m
s
.
FL
out
pe
r
f
or
m
s
c
e
nt
r
a
li
z
e
d,
tr
a
di
ti
ona
l
s
ys
te
m
s
a
nd
th
e
c
os
t
s
a
nd
r
is
ks
a
s
s
o
c
ia
te
d
w
it
h
s
e
n
s
it
iv
e
da
ta
m
a
na
ge
m
e
nt
by
e
xc
e
ll
in
g
in
ba
ndw
id
th
a
nd
pow
e
r
-
c
ons
tr
a
in
e
d
c
ont
e
xt
s
a
nd
of
f
e
r
in
g
a
s
im
pl
e
,
e
f
f
ic
ie
nt
pl
a
tf
or
m
f
or
s
c
a
le
d
pe
r
s
ona
li
z
a
ti
on.
I
t
a
ls
o
gi
ve
s
us
e
r
s
ow
ne
r
s
hi
p
ove
r
th
e
ir
da
ta
,
a
ll
ow
in
g
de
ve
lo
pe
r
s
to
de
s
ig
n
c
r
e
a
ti
ve
a
pps
th
a
t
us
e
da
ta
in
s
ig
ht
s
[
9]
.
I
n
c
e
nt
r
a
li
z
e
d
le
a
r
ni
ng,
d
a
ta
f
or
le
a
r
ni
ng
m
ode
l
s
a
r
e
c
ol
le
c
te
d
f
r
om
m
ul
ti
pl
e
s
our
c
e
s
a
nd
th
e
n
c
onne
c
te
d
to
a
c
lo
ud
s
e
r
ve
r
to
c
r
e
a
te
a
s
ta
nda
r
d
m
ode
l
th
a
t
c
a
n
be
de
pl
oye
d a
n
d
s
ha
r
e
d a
c
r
os
s
de
vi
c
e
s
.
F
ig
ur
e
s
1
(
a
)
a
nd
1(
b)
s
how
s
th
e
di
f
f
e
r
e
nc
e
b
e
twe
e
n
c
e
nt
r
a
li
z
e
d
a
nd
FL
.
H
ow
e
v
e
r
,
th
e
r
e
a
r
e
va
r
io
us
li
m
it
a
ti
ons
w
h
e
n
a
ppl
yi
ng
F
L
in
I
oT
a
ppl
ic
a
ti
ons
,
in
c
lu
di
ng
th
e
r
e
li
a
bi
li
ty
of
th
e
le
a
r
ni
ng
m
ode
l.
I
f
th
e
gl
oba
l
m
ode
l
is
m
a
li
c
io
us
ly
m
odi
f
ie
d
or
br
oke
n,
th
e
upda
te
o
f
a
ll
lo
c
a
l
m
ode
ls
w
il
l
be
ne
ga
ti
ve
ly
a
f
f
e
c
te
d
[
10]
.
T
he
r
e
f
or
e
,
m
e
th
ods
f
o
r
us
in
g de
e
p
FL
a
r
e
f
unda
m
e
nt
a
l
in
t
hi
s
c
ont
e
xt
.
(
a
)
(
b)
F
ig
ur
e
1.
A
r
c
hi
te
c
tu
r
e
of
(
a
)
c
e
nt
r
a
li
z
e
d l
e
a
r
ni
ng a
nd (
b)
FL
pr
o
c
e
s
s
I
n
th
is
s
tu
dy,
w
e
us
e
two
m
ode
ls
of
de
e
p
le
a
r
ni
ng
(
D
L
)
a
ppr
oa
c
he
s
ba
s
e
d
on
F
L
to
de
te
c
t
c
ybe
r
-
a
tt
a
c
ks
(
hybr
id
in
c
e
pt
io
n
ti
m
e
a
nd
m
ul
ti
-
he
a
d
a
tt
e
nt
io
n
(
C
N
N
)
)
on
th
e
e
dge
-
in
dus
tr
ia
l
in
te
r
ne
t
of
th
in
gs
(
I
I
oT
)
da
ta
s
e
ts
.
U
s
in
g
two
ty
pe
s
of
di
s
tr
ib
ut
io
n:
non
-
in
de
pe
n
de
nt
a
nd
id
e
nt
ic
a
ll
y
di
s
tr
ib
ut
e
d
(
non
-
I
I
D
)
a
nd
in
de
pe
nde
nt
a
nd
id
e
nt
ic
a
ll
y
di
s
tr
ib
ut
e
d
(
I
I
D
)
.
T
he
f
e
de
r
a
te
d
a
ve
r
a
gi
ng
a
lg
or
it
hm
(
F
e
dA
vg
)
a
nd
f
e
de
r
a
te
d
m
a
tc
he
d a
ve
r
a
gi
ng (
F
e
dM
A
)
a
lg
or
it
hm
s
, on the
ot
he
r
ha
nd, a
ggr
e
ga
te
l
oc
a
l
m
ode
l
upda
te
s
.
T
he
f
ol
lo
w
in
g a
r
e
our
r
e
s
e
a
r
c
h c
ont
r
ib
ut
io
ns
:
‒
W
e
pr
ovi
de
te
s
t
s
e
va
lu
a
ti
ng
th
e
pr
opos
e
d
m
ode
l'
s
pe
r
f
or
m
a
nc
e
de
te
c
ti
ng
c
yb
e
r
-
a
tt
a
c
ks
on
I
o
T
de
vi
c
e
s
.
T
he
f
ol
lo
w
in
g
s
it
ua
ti
ons
w
e
r
e
c
om
pa
r
e
d
to
th
a
t
goa
l:
i)
a
c
e
nt
r
a
li
z
e
d
s
tr
a
te
gy
in
w
hi
c
h
a
ll
da
ta
i
s
s
h
a
r
e
d;
a
nd
ii
)
a
f
e
d
e
r
a
te
d
a
ppr
oa
c
h
in
w
hi
c
h
a
gl
oba
l
m
ode
l
is
d
e
ve
lo
pe
d
w
hi
le
lo
c
a
l
m
od
e
l
c
ha
ng
e
s
a
r
e
e
xc
ha
nge
d.
‒
D
e
te
c
ti
on
of
c
ybe
r
-
a
tt
a
c
ks
a
nd
I
oT
a
r
c
hi
te
c
tu
r
e
f
e
de
r
a
t
e
d
to
s
ol
ve
c
e
nt
r
a
li
z
e
d
pr
obl
e
m
s
,
s
uc
h
a
s
s
in
gl
e
poi
nt
s
of
f
a
il
u
r
e
,
a
nd
pr
e
s
e
r
ve
th
e
pr
iv
a
c
y
of
th
e
lo
c
a
ll
y
tr
a
in
e
d
da
ta
,
w
hi
c
h
is
r
e
qui
r
e
d
f
o
r
di
ve
r
s
e
a
ppl
ic
a
ti
ons
.
‒
T
o
our
knowle
dge
,
w
e
a
r
e
th
e
f
ir
s
t
to
us
e
th
e
hybr
id
in
c
e
pt
io
n
ti
m
e
a
nd
m
ul
ti
-
he
a
d
a
tt
e
nt
io
n
(
C
N
N
)
m
ode
l
w
it
h
FL
to
de
te
c
t
c
ybe
r
-
a
tt
a
c
ks
on I
oT
de
vi
c
e
s
.
‒
W
e
a
r
e
th
e
f
ir
s
t
to
c
om
pa
r
e
s
ynt
h
e
ti
c
m
in
or
it
y
ove
r
s
a
m
p
li
ng
m
e
th
od
(
S
M
O
T
)
a
nd
c
la
s
s
-
w
e
ig
ht
te
c
hni
que
s
ba
s
e
d on
FL
.
‒
W
e
us
e
t
he
F
e
dA
vg
a
nd F
e
dM
A
a
lg
or
it
hm
s
t
o a
ggr
e
ga
te
a
nd c
o
m
pa
r
e
l
oc
a
l
m
ode
l
upda
te
s
, not
e
t
ha
t
th
e
r
e
is
no pr
e
vi
ous
s
tu
dy t
ha
t
c
om
pa
r
e
s
t
hi
s
.
T
he
r
e
s
t
of
t
hi
s
w
or
k i
s
s
tr
uc
tu
r
e
d a
s
f
ol
lo
w
s
. T
he
r
e
la
te
d
w
or
k i
s
s
um
m
a
r
iz
e
d i
n s
e
c
ti
on 2. S
e
c
ti
on 3
di
s
c
us
s
e
s
th
e
m
e
th
od
s
.
S
e
c
ti
on
4
de
s
c
r
ib
e
s
th
e
s
y
s
te
m
de
s
ig
n.
S
e
c
ti
on
5
d
e
s
c
r
ib
e
s
th
e
e
xpe
r
im
e
nt
a
l
r
e
s
ul
t
s
.
F
in
a
ll
y, s
e
c
ti
on 6 c
onc
lu
s
io
n a
nd di
s
c
us
s
io
n.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
:
4778
-
4794
4780
2.
R
E
L
A
T
E
D
WORK
T
he
in
c
r
e
a
s
in
g
popula
r
it
y
of
FL
a
nd
th
e
us
e
of
th
e
in
te
r
ne
t
of
th
in
gs
ha
s
c
r
e
a
te
d
m
a
ny
in
te
r
e
s
ti
ng
r
e
s
e
a
r
c
h
pa
th
s
,
one
of
w
hi
c
h
is
th
e
di
s
c
ove
r
y
a
nd
c
la
s
s
if
ic
a
ti
on
of
c
ybe
r
-
a
tt
a
c
ks
in
I
oT
de
vi
c
e
s
,
a
nd
th
e
r
e
a
r
e
m
a
ny
r
e
s
e
a
r
c
h
w
or
ks
pr
opos
e
d
to
s
e
c
ur
e
I
oT
ne
twor
ks
f
r
o
m
m
a
li
c
io
us
a
tt
a
c
ks
.
I
n
th
is
pa
r
t,
w
e
w
il
l
go
th
r
ough the
m
os
t
r
e
c
e
nt
r
e
s
e
a
r
c
h.
T
a
bl
e
1 pr
ovi
de
s
a
n ov
e
r
vi
e
w
of
t
he
s
e
s
ugge
s
ti
ons
.
T
he
e
dge
-
I
I
oT
da
t
a
s
e
t,
pr
opos
e
d
by
F
e
r
r
a
g
e
t
al
.
[
11]
,
is
a
n
in
ve
s
ti
ga
ti
ona
l
r
e
s
our
c
e
th
a
t
ut
il
iz
e
s
c
e
nt
r
a
li
z
e
d
in
tr
us
io
n
de
te
c
ti
on
a
nd
f
e
d
e
r
a
te
d
DL
w
it
h
s
ta
nda
r
d
a
s
s
e
s
s
m
e
nt
c
r
it
e
r
ia
.
T
he
r
e
s
e
a
r
c
h
e
xpl
or
e
d
th
e
pr
e
di
c
ti
on
a
nd
de
te
c
ti
on
e
f
f
ic
a
c
y
of
di
f
f
e
r
e
nt
th
r
e
a
t
m
ode
ls
a
nd
c
ybe
r
-
a
tt
a
c
ks
us
in
g
bi
na
r
y,
s
ix
-
c
la
s
s
,
a
nd
f
if
te
e
n
-
c
la
s
s
c
a
te
gor
iz
a
ti
on.
T
he
be
s
t
r
e
s
ul
ts
f
or
th
e
15
-
c
la
s
s
a
r
e
a
c
hi
e
ve
d
w
it
h
I
I
D
a
c
c
ur
a
c
y
of
93.89%
a
nd
th
e
m
e
th
od
of
n
on
-
I
I
D
a
c
c
ur
a
c
y
of
91.45
%
.
T
a
ba
s
s
um
e
t
al
.
[
12]
pr
e
s
e
nt
F
E
D
G
A
N
-
I
D
S
,
a
FL
in
tr
us
io
n
de
te
c
ti
on
s
ys
t
e
m
(
I
D
S
)
ba
s
e
d
on
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
t
w
or
ks
(
G
A
N
)
f
or
de
te
c
ti
ng
c
ybe
r
-
a
tt
a
c
ks
in
in
te
ll
ig
e
nt
in
te
r
ne
t
of
th
in
gs
s
ys
te
m
s
s
uc
h
a
s
s
m
a
r
t
hom
e
s
,
s
m
a
r
t
c
it
ie
s
,
a
nd
in
te
ll
ig
e
nt
e
-
he
a
lt
hc
a
r
e
s
ys
te
m
s
.
E
xt
e
ns
iv
e
te
s
ti
ng
of
th
e
di
s
tr
ib
ut
e
d
in
tr
us
io
n
de
te
c
ti
on
m
od
e
l'
s
a
c
c
ur
a
c
y,
pe
r
f
or
m
a
nc
e
,
a
nd
c
onv
e
r
ge
nc
e
ut
il
iz
in
g t
hr
e
e
s
ta
nda
r
d da
ta
s
e
t
s
:
U
N
S
W
-
N
B
15, KD
D
99, a
nd
N
S
L
-
K
D
D
.
A
c
c
or
di
ng
to
C
a
m
pos
e
t
al
.
[
13]
,
a
FL
-
e
na
bl
e
d
I
D
S
te
c
hni
qu
e
ba
s
e
d
on
a
m
ul
ti
c
la
s
s
c
la
s
s
if
ie
r
is
e
va
lu
a
te
d
f
or
de
te
c
ti
ng
va
r
io
us
a
tt
a
c
k
s
in
a
n
I
oT
s
c
e
na
r
io
.
U
s
in
g
th
r
e
e
di
s
ti
nc
t
s
e
tt
in
gs
obt
a
in
e
d
by
pa
r
ti
ti
oni
ng
th
e
T
oN
-
I
oT
da
ta
s
e
t
a
c
c
or
di
ng
to
a
tt
a
c
k
ty
pe
a
nd
I
oT
de
vi
c
e
I
P
a
ddr
e
s
s
.
I
n
a
ddi
ti
on,
th
e
I
nt
e
r
na
ti
ona
l
B
us
in
e
s
s
M
a
c
hi
ne
s
(
I
B
M
)
f
e
de
r
a
te
d
le
a
r
ni
ng
f
r
a
m
e
w
or
k
is
us
e
d
to
im
pl
e
m
e
nt
a
nd
a
s
s
e
s
s
th
e
im
pa
c
t
of
va
r
io
us
a
ggr
e
ga
ti
on
f
unc
ti
ons
.
I
ns
te
a
d
of
th
e
na
ïv
e
F
e
dA
vg,
Z
ha
ng
e
t
al
.
[
14]
s
ugge
s
t
f
e
d
de
te
c
t,
a
FL
a
lg
or
it
hm
ic
f
r
a
m
e
w
or
k
f
o
r
on
-
de
vi
c
e
a
nom
a
ly
da
ta
de
te
c
ti
on
th
a
t
e
m
pl
oys
a
n
a
da
pt
iv
e
opt
im
iz
e
r
a
nd
a
c
r
os
s
-
r
ound
le
a
r
ni
ng
r
a
te
s
c
he
dul
e
r
.
U
s
e
th
e
N
-
B
a
I
oT
da
ta
s
e
t
to
a
s
s
e
s
s
th
e
F
e
d
I
oT
pl
a
tf
or
m
a
nd
th
e
f
e
d
de
te
c
t
a
lg
or
it
hm
r
e
ga
r
di
ng
m
ode
l
a
nd
s
ys
te
m
pe
r
f
or
m
a
nc
e
.
P
opoola
e
t
al
.
[
15]
pr
opos
e
a
FL
s
tr
a
te
gy
f
or
de
te
c
ti
ng
z
e
r
o
-
da
y
bot
ne
t
a
tt
a
c
ks
,
a
nd
a
n
a
ppr
opr
ia
te
de
e
p
ne
ur
a
l
ne
twor
k
a
r
c
hi
te
c
tu
r
e
is
us
e
d
f
or
ne
twor
k
tr
a
f
f
ic
c
la
s
s
if
ic
a
ti
on
in
I
oT
d
e
vi
c
e
s
to
li
m
it
da
ta
pr
iv
a
c
y
le
a
ka
ge
to
a
ggr
e
ga
te
lo
c
a
l
m
ode
l
upda
te
s
.
T
he
F
e
dA
vg
m
e
th
od
is
e
m
pl
oye
d.
M
ot
hukur
i
e
t
al
.
[
16]
s
ugge
s
te
d
a
th
r
e
a
t
de
te
c
ti
on
a
nd
c
la
s
s
if
ic
a
ti
on
te
c
hni
que
in
I
oT
ne
twor
ks
b
a
s
e
d
on
FL
.
T
h
e
a
ut
hor
s
c
om
bi
ne
ga
te
d
r
e
c
ur
r
e
nt
uni
ts
(
G
R
U
s
)
a
nd
r
a
ndom
f
or
e
s
t
(
R
F
)
m
ode
ls
to
f
or
m
th
e
ir
e
ns
e
m
bl
e
.
I
n
ot
he
r
w
or
ds
,
th
e
y
us
e
d
R
F
to
in
te
gr
a
te
th
e
G
R
U
m
ode
l
pr
e
di
c
ti
ons
to
im
pr
ove
th
e
c
la
s
s
if
ic
a
ti
on
pe
r
f
or
m
a
nc
e
.
F
e
d
-
I
I
oT
,
a
f
e
de
r
a
te
d
te
c
hni
que
f
or
id
e
nt
if
yi
ng
A
ndr
oi
d
m
a
lwa
r
e
in
I
I
oT
,
w
a
s
c
r
e
a
te
d
[
17]
.
F
e
d
-
I
I
oT
,
in
pa
r
ti
c
ul
a
r
,
c
r
e
a
te
s
a
nd
in
je
c
ts
hos
ti
le
m
a
te
r
ia
l
in
to
th
e
da
ta
s
e
t
us
in
g
two
G
A
N
m
ode
ls
.
T
he
s
e
r
ve
r
us
e
s
a
G
A
N
to
de
te
c
t
m
a
li
c
io
us
m
ode
ls
a
nd
r
e
je
c
t
pol
lu
te
d
da
ta
.
A
c
c
or
di
ng
to
N
ua
im
i
e
t
al
.
[
18]
,
m
ul
t
ip
le
I
D
S
s
a
r
e
de
ve
lo
pe
d
us
in
g
c
onve
nt
io
na
l
da
ta
a
na
ly
ti
c
s
m
e
th
ods
a
nd
th
e
ir
pe
r
f
or
m
a
nc
e
on
E
dge
-
I
I
oT
is
e
va
lu
a
te
d,
a
nd
I
D
S
s
a
r
e
c
om
pa
r
e
d
to
pr
e
vi
ous
w
or
k
to
il
lu
s
tr
a
te
th
a
t
hi
ghl
y
a
c
c
ur
a
te
bi
na
r
y
-
c
la
s
s
I
D
S
s
c
a
n be
pr
oduc
e
d, ho
w
e
ve
r
m
ul
ti
-
c
la
s
s
I
D
S
s
r
e
qui
r
e
c
a
r
e
f
ul
t
r
e
a
tm
e
nt
.
T
a
bl
e
1.
A
s
um
m
a
r
y of
pr
e
vi
ous
s
tu
di
e
s
A
ut
hor
s
Y
e
a
r
s
D
a
t
a
M
ode
l
D
e
s
c
r
i
pt
i
on
F
e
r
r
a
g
e
t
al
.
[1
1
]
2022
E
dge
-
I
I
o
T
s
e
t
D
e
e
p
FL
C
e
nt
r
a
l
i
z
e
d
a
nd
f
e
de
r
a
t
e
d
DL
i
s
us
e
d
t
o
i
nv
e
s
t
i
ga
t
e
t
he
de
t
e
c
t
i
on
e
f
f
i
c
a
c
y
a
nd
t
r
a
f
f
i
c
pr
e
di
c
t
a
bi
l
i
t
y
of
di
f
f
e
r
e
nt
c
ybe
r
-
a
t
t
a
c
ks
a
nd
t
hr
e
a
t
m
ode
l
s
.
T
a
ba
s
s
um
e
t
al
.
[1
2
]
2022
N
S
L
-
K
D
D
,
U
N
S
W
-
N
B
15, K
D
D
99.
F
e
dG
A
N
F
e
de
r
a
t
e
d
D
L
I
D
S
ba
s
e
d
on
G
A
N
f
or
de
t
e
c
t
i
ng
c
yb
e
r
t
hr
e
a
t
s
i
n
i
nt
e
l
l
i
ge
nt
I
oT
s
ys
t
e
m
s
.
C
a
m
pos
e
t
al
.
[1
3
]
2021
T
oN
-
I
oT
L
ogi
s
t
i
c
r
e
gr
e
s
s
i
on
A
n
F
L
-
e
na
bl
e
d
I
D
S
t
e
c
hni
que
ba
s
e
d
on
a
m
ul
t
i
c
l
a
s
s
c
l
a
s
s
i
f
i
e
r
i
s
a
s
s
e
s
s
e
d f
or
i
de
nt
i
f
yi
ng va
r
i
ous
t
hr
e
a
t
s
i
n a
n I
oT
c
ont
e
xt
.
Z
ha
ng
e
t
al
.
[1
4
]
2021
N
-
B
a
I
oT
D
e
e
p
A
ut
oe
nc
od
P
r
opos
e
f
e
d
de
t
e
c
t
,
a
n
F
L
a
l
gor
i
t
hm
i
c
f
r
a
m
e
w
or
k
f
or
on
-
de
vi
c
e
a
nom
a
l
y da
t
a
de
t
e
c
t
i
on.
P
opool
a
e
t
al
.
[1
5
]
2021
B
ot
-
I
oT
.
N
-
B
a
I
oT
DNN
T
he
f
e
de
r
a
t
e
d
D
L
a
ppr
oa
c
h
f
or
z
e
r
o
-
da
y
bot
ne
t
a
t
t
a
c
k
de
t
e
c
t
i
on
i
s
pr
opos
e
d t
o l
i
m
i
t
da
t
a
pr
i
va
c
y l
e
a
ka
ge
i
n I
oT
de
vi
c
e
s
.
M
ot
hukur
i
e
t
al
.
[1
6
]
2021
-
L
S
T
M
/
G
R
U
P
r
opos
e
d
a
n
e
ns
e
m
bl
e
f
e
de
r
a
t
e
d
-
ba
s
e
d
a
t
t
a
c
k
de
t
e
c
t
i
on
a
nd
c
l
a
s
s
i
f
i
c
a
t
i
on m
e
t
hod, c
om
bi
ni
ng R
F
a
nd G
R
U
s
m
ode
l
s
.
T
a
he
r
i
e
t
al
.
[1
7
]
2020
D
r
e
bi
n,
G
e
nom
e
,
C
ont
a
gi
o
F
e
dG
A
N
D
e
ve
l
ope
d F
e
d
-
I
I
oT
, a
n F
L
a
ppr
oa
c
h
f
or
de
t
e
c
t
i
ng A
ndr
oi
d m
a
l
w
a
r
e
i
n I
I
oT
.
T
he
r
e
i
s
no pr
e
vi
ous
s
tu
dy t
ha
t
us
e
s
t
he
I
nc
e
pt
io
n t
im
e
m
ode
l
w
it
h
FL
to
de
te
c
t
c
ybe
r
-
a
tt
a
c
ks
on I
oT
de
vi
c
e
s
.
H
ow
e
ve
r
,
our
m
ode
l
s
how
e
d
hi
gh
a
c
c
ur
a
c
y
in
d
e
te
c
ti
ng
a
tt
a
c
ks
c
om
pa
r
e
d
to
pr
e
vi
ous
s
tu
di
e
s
.
A
tt
e
nt
io
n
a
r
c
hi
te
c
tu
r
e
s
a
r
e
us
e
d
in
na
tu
r
a
l
la
ngua
g
e
pr
oc
e
s
s
i
ng,
a
nd
w
e
ha
ve
u
s
e
d
th
e
m
to
de
te
c
t
c
ybe
r
-
a
tt
a
c
ks
on
I
oT
de
vi
c
e
s
.
I
n
th
is
pa
pe
r
,
w
e
a
ppl
y
two
m
e
th
ods
to
de
te
c
t
m
ode
r
n
c
ybe
r
-
a
tt
a
c
ks
on
in
te
r
ne
t
of
th
in
gs
de
vi
c
e
s
.
T
he
f
ir
s
t
m
e
th
od
in
tr
oduc
e
s
F
L
w
it
h
D
L
a
ppr
oa
c
he
s
a
s
a
(
hybr
id
in
c
e
pt
io
n
ti
m
e
a
nd
m
ul
ti
-
he
a
d
a
tt
e
nt
io
n
(
C
N
N
)
a
lg
or
it
hm
)
to
de
te
c
t
c
ybe
r
-
a
tt
a
c
ks
in
I
oT
de
vi
c
e
s
us
in
g
th
e
E
dge
-
I
I
oT
da
ta
ba
s
e
to
a
ggr
e
ga
te
l
oc
a
l
m
ode
l
upda
te
s
i
n
FL
;
w
e
c
om
pa
r
e
t
he
F
e
dA
vg
a
nd F
e
dM
A
a
lg
or
it
hm
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
F
e
de
r
at
e
d
in
c
e
pt
io
n
-
m
ul
ti
-
he
ad att
e
nt
io
n m
ode
ls
f
or
c
y
b
e
r
-
at
ta
c
k
s
de
te
c
ti
on
(
I
m
ad T
ar
e
q A
L
-
H
al
boos
i
)
4781
3.
T
H
E
P
R
O
P
O
S
E
D
M
E
T
H
O
D
S
3.1.
F
e
d
e
r
at
e
d
-
in
c
e
p
t
io
n
t
im
e
T
hi
s
s
e
c
ti
on
de
s
c
r
ib
e
s
th
e
a
r
c
hi
te
c
tu
r
e
of
th
e
e
nha
nc
e
I
nc
e
pt
io
n
ti
m
e
m
ode
l
(
in
c
e
pt
io
n
ti
m
e
v2
-
in
c
e
pt
io
n
ti
m
e
v3)
.
W
e
e
m
pl
oy
c
la
s
s
w
e
ig
ht
s
a
nd
s
m
ot
e
a
f
te
r
c
ol
le
c
ti
ng
a
nd
pr
oc
e
s
s
in
g
da
ta
.
T
o
a
voi
d
ove
r
f
it
ti
ng,
th
e
da
ta
s
e
t
is
di
vi
de
d
in
to
te
s
ti
ng
a
nd
tr
a
in
in
g.
T
h
e
A
da
m
a
nd
s
to
c
h
a
s
ti
c
gr
a
di
e
nt
d
e
s
c
e
nt
(
S
G
D
)
opt
im
iz
e
r
us
e
s
a
n
opt
im
iz
a
ti
on
te
c
hni
que
r
a
th
e
r
th
a
n
th
e
s
t
a
n
da
r
d
s
to
c
ha
s
ti
c
gr
a
di
e
nt
d
e
s
c
e
nt
a
ppr
oa
c
h
to
r
e
pe
a
te
dl
y
upda
te
n
e
twor
k
w
e
ig
ht
s
ba
s
e
d
on
tr
a
in
in
g
da
ta
.
W
it
h
a
de
pt
h
of
s
ix
a
nd
f
our
la
ye
r
s
a
nd
a
f
il
te
r
of
32,
w
it
h
a
s
hor
t
la
ye
r
ove
r
th
r
e
e
,
in
c
e
pt
io
n
ti
m
e
is
e
m
pl
oye
d.
T
hr
e
e
ke
r
ne
ls
w
il
l
be
ge
n
e
r
a
te
d
f
or
e
a
c
h
in
c
e
pt
io
n m
odul
e
f
r
om
40/
/(
2^
i)
a
s
i
i
nc
r
e
a
s
e
s
f
r
om
0
-
>
3.
W
e
e
m
pl
oye
d
a
n
in
c
e
pt
io
n
ti
m
e
m
ode
l
w
it
h
a
one
-
di
m
e
ns
i
ona
l
(
1D
)
in
put
ve
c
to
r
f
or
m
in
th
e
in
c
e
pt
io
n
ne
twor
k.
E
a
c
h
bl
oc
k
ha
s
th
r
e
e
in
c
e
pt
io
n
ti
m
e
m
odul
e
s
.
A
bot
tl
e
ne
c
k
la
ye
r
r
e
duc
e
s
th
e
num
be
r
of
pa
r
a
m
e
te
r
s
a
nd
pr
oc
e
s
s
in
g
c
os
t
s
by
lo
w
e
r
in
g
th
e
in
put
di
m
e
ns
io
ns
(
i.
e
.,
th
e
de
pt
h)
.
A
1D
s
li
di
ng
f
il
te
r
c
a
n
f
il
te
r
out
it
s
di
s
c
r
im
in
a
ti
ng
z
one
in
a
ti
m
e
s
e
r
ie
s
.
I
m
pr
ove
s
ge
ne
r
a
li
z
a
ti
on
a
nd
s
pe
e
ds
up
tr
a
in
in
g.
T
h
e
bot
tl
e
ne
c
k
out
put
f
e
e
ds
th
r
e
e
one
-
di
m
e
ns
io
na
l
c
onvolut
io
n
la
ye
r
s
w
it
h
10,
20,
a
nd
40
ke
r
ne
l
s
iz
e
s
.
F
ur
th
e
r
m
or
e
,
th
e
in
c
e
pt
io
n
ti
m
e
m
odu
le
'
s
in
put
s
pa
s
s
th
r
ough
th
e
f
ir
s
t
la
ye
r
c
ons
is
ti
ng
of
(
C
N
N
1D
,
f
il
te
r
s
=
32,
ke
r
ne
l=
10,
C
N
N
1D
,
f
il
te
r
s
=
32,
ke
r
ne
l
=
10.
M
a
x
-
P
ool
in
g
1D
,
pool
in
g
s
iz
e
=
20,
s
tr
id
e
=
1,
C
N
N
1D
,
f
il
te
r
s
=
32,
ke
r
ne
l=
10)
.
S
e
c
ond
l
a
ye
r
(
C
N
N
1D
,
f
il
te
r
s
=
32,
ke
r
ne
l=
20,
C
N
N
1D
,
f
il
te
r
s
=
32,
k
e
r
ne
l=
40,
C
N
N
1D
,
f
il
te
r
s
=
32,
k
e
r
ne
l=
10)
.
P
a
ddi
ng=
s
a
m
e
,
a
c
ti
va
ti
on=
r
e
c
ti
f
ie
d
li
ne
a
r
uni
t
(
R
e
L
U
)
e
a
c
h
a
ll
.
T
he
d
e
pt
h
c
onc
a
te
na
ti
on
la
ye
r
,
r
e
lu
a
c
ti
va
ti
on
la
ye
r
,
a
nd
ba
tc
h
nor
m
a
li
z
a
ti
on
la
ye
r
s
e
que
nc
e
th
e
out
put
s
of
th
e
f
our
c
onvolut
io
n
la
ye
r
s
a
lo
ng
th
e
de
pt
h
di
m
e
n
s
io
n.
A
ll
l
a
ye
r
s
, e
xc
e
pt
th
e
s
e
que
nc
e
la
ye
r
,
h
a
ve
th
e
s
a
m
e
s
tr
id
e
a
nd
pa
ddi
ng.
A
s
il
lu
s
tr
a
te
d
in
F
ig
ur
e
2,
a
ll
c
onvolut
io
n
la
y
e
r
s
h
a
ve
32
f
il
te
r
s
,
a
nd
r
e
s
id
ua
l
c
onne
c
ti
ons
a
r
e
e
m
pl
oye
d f
or
e
ve
r
y t
hi
r
d i
nc
e
pt
io
n t
im
e
m
odul
e
.
E
ve
r
y
th
r
e
e
m
odul
e
s
w
il
l
m
a
ke
a
r
e
s
id
ua
l
la
ye
r
c
ons
is
ti
ng
of
C
N
N
1D
(
f
il
te
r
=
128,
ke
r
ne
l=
1,
pa
ddi
ng=
s
a
m
e
)
ba
tc
h nor
m
a
li
z
a
ti
on;
a
f
te
r
t
ha
t,
w
e
w
il
l
a
dd
th
e
out
put
f
r
om
t
he
r
e
s
id
ua
l
a
nd t
he
l
a
s
t
in
c
e
pt
io
n
m
odul
e
w
it
h
a
dd
a
la
ye
r
.
T
he
n,
th
e
a
c
ti
va
ti
on
la
ye
r
w
it
h
r
e
lu
r
e
pe
a
ts
th
is
ope
r
a
ti
on
w
it
h
de
pt
h
num
be
r
.
U
lt
im
a
te
ly
,
w
e
a
dd
a
gl
oba
l
a
ve
r
a
ge
pool
in
g
la
ye
r
de
ns
e
la
ye
r
w
it
h
uni
ts
th
e
e
qua
l
num
be
r
of
c
la
s
s
e
s
w
it
h
a
c
ti
va
ti
on f
unc
ti
on s
of
tm
a
x. F
ig
ur
e
3 s
how
s
t
he
r
e
s
id
ua
l
la
y
e
r
.
F
ig
ur
e
2. T
he
c
us
to
m
i
nc
e
pt
io
n m
ode
l
3.2.
F
e
d
e
r
at
e
d
-
m
u
lt
i
-
h
e
ad
a
t
t
e
n
t
io
n
T
hi
s
s
e
c
ti
on
de
s
c
r
ib
e
s
th
e
a
r
c
hi
te
c
tu
r
e
of
th
e
m
ul
ti
-
he
a
d
a
tt
e
nt
io
n
(
C
N
N
)
m
ode
l.
T
h
e
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
e
s
ta
bl
is
he
s
a
w
e
ig
ht
e
d
a
ve
r
a
ge
of
(
s
e
que
nc
e
)
e
le
m
e
nt
s
,
th
e
w
e
ig
ht
s
of
w
hi
c
h
a
r
e
dyna
m
ic
a
ll
y
ge
ne
r
a
te
d
de
pe
ndi
ng
on
th
e
e
le
m
e
nt
s
'
ke
ys
a
nd
a
n
in
put
que
r
y.
T
he
goa
l
is
to
a
ve
r
a
ge
th
e
pr
ope
r
ti
e
s
of
s
e
ve
r
a
l
c
om
pone
nt
s
.
R
a
th
e
r
th
a
n
gi
vi
ng
e
a
c
h
e
le
m
e
nt
a
n
id
e
nt
ic
a
l
w
e
ig
ht
,
w
e
pr
e
f
e
r
to
w
e
ig
h
th
e
m
ba
s
e
d
on
th
e
ir
unde
r
ly
in
g
va
lu
e
s
.
I
n
ot
he
r
w
or
ds
,
w
e
ne
e
d
to
know
w
h
ic
h
in
put
s
w
e
w
a
nt
to
"
a
tt
e
nd"
to
m
or
e
th
a
n
ot
he
r
s
;
f
ur
th
e
r
i
nf
or
m
a
ti
on i
s
pr
ovi
de
d a
s
s
how
n i
n F
ig
ur
e
3.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
:
4778
-
4794
4782
F
ig
ur
e
3.
T
he
r
e
s
id
ua
l
la
ye
r
(
in
c
e
pt
io
n m
odul
e
)
3.2.1. At
t
e
n
t
io
n
T
he
a
tt
e
nt
io
n
m
e
c
h
a
ni
s
m
c
ov
e
r
s
a
nove
l
c
a
te
gor
y
of
la
ye
r
s
in
ne
ur
a
l
ne
twor
ks
th
a
t
ha
s
s
pa
r
ke
d
c
ons
id
e
r
a
bl
e
in
te
r
e
s
t
in
r
e
c
e
nt
ye
a
r
s
,
pa
r
ti
c
ul
a
r
ly
in
s
e
qu
e
nc
e
ta
s
ks
.
T
he
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
s
pe
c
if
ie
s
a
w
e
ig
ht
e
d
a
ve
r
a
ge
of
(
s
e
que
nc
e
)
it
e
m
s
,
th
e
w
e
ig
ht
s
of
w
hi
c
h
a
r
e
de
te
r
m
in
e
d
dyna
m
ic
a
ll
y
ba
s
e
d
on
th
e
in
put
que
r
y
a
nd
th
e
e
le
m
e
nt
s
'
ke
ys
.
H
is
a
tt
e
nt
io
n
c
ons
i
s
ts
of
pa
r
ts
(
que
r
y,
ke
ys
,
va
lu
e
s
,
s
c
or
e
f
unc
ti
on)
,
you
c
a
n
le
a
r
n
a
bout
it
in
m
or
e
de
ta
il
in
th
e
f
ol
lo
w
in
g
s
our
c
e
s
[
19]
–
[
23]
.
A
s
of
tm
a
x
c
a
lc
ul
a
te
s
th
e
a
ve
r
a
ge
w
e
ig
ht
s
a
c
r
os
s
a
ll
s
c
or
e
f
unc
ti
on
out
put
s
.
A
s
s
ig
n
m
or
e
s
ig
ni
f
ic
a
nt
im
po
r
ta
nc
e
to
th
os
e
va
lu
e
ve
c
to
r
s
w
ho
s
e
m
a
t
c
hi
ng
ke
y i
s
m
or
e
c
om
pa
r
a
bl
e
t
o t
he
que
r
y, w
hi
c
h c
a
n b
e
c
a
lc
ul
a
te
d a
s
(
1)
:
=
(
∫
(
,
)
)
∑
(
∫
(
,
)
)
,
=
∑
,
.
(
1)
M
os
t
a
tt
e
nt
io
n
m
e
th
ods
di
f
f
e
r
in
th
e
que
r
ie
s
ut
il
iz
e
d,
th
e
ke
y
a
nd
va
lu
e
ve
c
to
r
s
'
de
f
in
it
io
ns
,
a
nd
th
e
s
c
or
e
f
unc
ti
on
e
m
pl
oye
d.
S
e
lf
-
a
tt
e
nt
io
n
r
e
f
e
r
s
to
th
e
a
tt
e
nt
io
n
a
ppl
ie
d
w
it
hi
n
th
e
a
r
c
hi
te
c
tu
r
e
.
E
a
c
h
s
e
que
nc
e
e
le
m
e
nt
in
s
e
lf
-
a
tt
e
nt
io
n
pr
ovi
de
s
a
que
r
y,
va
lu
e
,
a
nd
ke
y.
W
e
c
onduc
t
a
n
a
tt
e
nt
io
n
la
ye
r
f
o
r
e
a
c
h
e
le
m
e
nt
,
c
he
c
ki
ng
th
e
s
im
il
a
r
it
y
of
a
ll
s
e
que
nc
e
m
e
m
be
r
s
'
ke
ys
ba
s
e
d
o
n
it
s
que
r
y
a
nd
r
e
tu
r
ni
ng
a
s
e
pa
r
a
te
,
a
ve
r
a
ge
d
va
lu
e
ve
c
to
r
f
or
e
a
c
h e
le
m
e
nt
.
3.2.2. S
c
al
e
d
-
d
ot
p
r
od
u
c
t
at
t
e
n
t
io
n
S
c
a
la
bl
e
-
dot
pr
oduc
t
a
tt
e
nt
io
n
is
th
e
f
unda
m
e
nt
a
l
id
e
a
of
s
e
lf
-
a
tt
e
nt
io
n.
W
e
w
a
nt
to
c
r
e
a
te
a
n
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
th
a
t
a
ll
ow
s
a
ny
e
le
m
e
nt
in
a
s
e
que
nc
e
to
pa
y
a
tt
e
nt
io
n
to
a
ny
ot
he
r
e
le
m
e
nt
w
hi
le
be
in
g
c
om
put
a
ti
ona
ll
y e
f
f
ic
ie
nt
. T
he
dot
-
pr
oduc
t
a
tt
e
nt
io
n i
s
f
e
d a
s
e
r
ie
s
of
i
nqui
r
ie
s
[
24]
.
Q
∈
R
T
×
d
k
,
ke
y
s
K
∈
R
T
×
d
k
a
n
d
v
a
l
ue
s
V
∈
R
T
×
d
v
,
T
is
th
e
s
e
que
nc
e
le
ngt
h,
a
nd
d
_
v
a
nd
d
_
k
de
not
e
th
e
hi
dde
n
di
m
e
ns
io
n
f
or
que
r
ie
s
V
va
lu
e
s
a
nd
K
ke
ys
,
r
e
s
pe
c
ti
ve
ly
.
T
he
s
im
il
a
r
it
y
of
th
e
ke
y
K
_
j
a
nd
que
r
y
Q
_
i
, us
in
g t
he
dot
-
p
r
oduc
t
a
s
t
he
s
im
il
a
r
it
y m
e
a
s
ur
e
, de
te
r
m
i
ne
s
t
he
a
tt
e
nt
io
n va
lu
e
f
r
om
j
to
i
a
s
s
how
n
in
F
ig
ur
e
4
.
(
,
.
)
=
(
√
)
(
2)
T
he
m
a
tr
ix
c
om
bi
na
ti
on
c
om
put
e
s
th
e
dot
pr
oduc
t
f
or
a
ll
c
onc
e
iv
a
bl
e
c
om
bi
na
ti
ons
of
ke
ys
a
nd
que
r
ie
s
,
r
e
s
ul
ti
ng
in
a
m
a
tr
ix
.
T
he
a
tt
e
nt
io
n
lo
gi
ts
f
or
a
s
in
gl
e
e
le
m
e
nt
to
a
ll
ot
he
r
c
om
pone
nt
s
in
th
e
s
e
que
nc
e
a
r
e
de
pi
c
te
d
in
e
a
c
h
r
ow
.
T
o
obt
a
in
a
w
e
ig
ht
e
d
m
e
a
n,
a
ppl
y
a
s
of
tm
a
x
a
nd
m
ul
ti
pl
y
by
th
e
va
lu
e
ve
c
to
r
[
25]
.
(
T
he
a
tt
e
nt
io
n de
te
r
m
in
e
s
t
he
w
e
ig
ht
s
.)
3.2.3.
M
u
lt
i
-
h
e
ad
at
t
e
n
t
io
n
M
ul
ti
-
he
a
d
a
tt
e
nt
io
n
is
a
m
odul
e
th
a
t
e
xe
c
ut
e
s
a
n
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
in
pa
r
a
ll
e
l.
A
ne
twor
k
c
a
n
a
tt
e
nd
to
a
s
e
qu
e
nc
e
u
s
in
g
s
c
a
l
e
d
-
dot
-
pr
oduc
t
a
tt
e
nt
io
n
[
26]
.
A
s
in
gl
e
w
e
ig
ht
e
d
a
v
e
r
a
ge
is
in
e
f
f
e
c
ti
ve
w
h
e
n
a
s
e
que
nc
e
e
le
m
e
nt
ha
s
to
a
ddr
e
s
s
m
a
ny
i
s
s
ue
s
.
A
s
a
r
e
s
ul
t,
w
e
br
oa
de
n
th
e
a
tt
e
nt
io
n
te
c
hni
que
s
to
in
c
lu
d
e
m
a
ny
he
a
ds
,
i.
e
.,
num
e
r
ous
s
e
a
r
c
he
s
on
th
e
s
a
m
e
f
e
a
tu
r
e
s
w
it
h
di
s
ti
nc
t
ke
y
-
va
lu
e
tr
ip
le
ts
.
W
e
pr
e
c
is
e
ly
r
e
s
tr
uc
tu
r
e
a
que
r
y,
va
lu
e
,
a
nd
ke
y
m
a
tr
ix
in
to
s
ub
-
que
r
ie
s
,
s
ub
-
va
lu
e
s
,
a
nd
s
ub
-
ke
y
s
th
a
t
a
r
e
th
e
n
in
di
vi
dua
ll
y
r
un
th
r
ough
th
e
s
c
a
le
d
-
dot
pr
oduc
t
a
tt
e
nt
io
n
[
27]
.
A
f
te
r
w
a
r
d,
th
e
he
a
ds
a
r
e
c
on
c
a
te
na
te
d
a
nd
m
ix
e
d w
it
h a
f
in
a
l
w
e
ig
ht
m
a
tr
ix
, A
s
s
how
n i
n F
ig
ur
e
5. T
hi
s
p
r
oc
e
dur
e
m
a
y be
e
xpr
e
s
s
e
d a
s
f
ol
lo
w
s
:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
F
e
de
r
at
e
d
in
c
e
pt
io
n
-
m
ul
ti
-
he
ad att
e
nt
io
n m
ode
ls
f
or
c
y
b
e
r
-
at
ta
c
k
s
de
te
c
ti
on
(
I
m
ad T
ar
e
q A
L
-
H
al
boos
i
)
4783
M
ul
ti
hue
d
(
,
,
)
=
(
ℎ
1
,
…
,
ℎ
ℎ
)
(
3)
W
he
r
e
=
a
tt
e
nt
io
n
(
,
,
)
th
a
t
is
r
e
f
e
r
r
e
d
to
a
s
th
e
m
ul
ti
-
he
a
d
a
tt
e
nt
io
n
la
ye
r
w
it
h
le
a
r
na
bl
e
pa
r
a
m
e
te
r
s
.
…
.
∈
×
,
…
.
∈
×
,
…
.
∈
×
∈
.
×
(
D
be
in
g
th
e
in
put
di
m
e
ns
io
na
li
ty
)
.
F
ig
ur
e
4. A
n
il
lu
s
tr
a
ti
on of
t
he
s
c
a
le
d
-
dot
-
pr
oduc
t
a
tt
e
nt
io
n. A
w
e
ig
ht
e
d s
um
va
lu
e
ge
n
e
r
a
te
s
t
he
r
e
s
ul
t,
w
it
h
w
e
ig
ht
s
de
f
in
e
d by the
dot
-
pr
oduc
t
to
que
r
y a
ll
ke
ys
[
25]
.
F
ig
ur
e
5.
T
he
m
ul
ti
-
he
a
d a
tt
e
nt
io
n. T
he
a
tt
e
nt
io
n m
e
c
h
a
ni
s
m
s
a
r
e
a
c
ti
va
te
d s
e
ve
r
a
l
ti
m
e
s
i
n
pa
r
a
ll
e
l
E
a
c
h
m
ul
ti
-
he
a
d
a
tt
e
nt
io
n
bl
oc
k
c
on
s
is
ts
of
f
our
s
uc
c
e
s
s
iv
e
le
v
e
ls
.
T
hr
e
e
C
N
N
(
de
ns
e
)
la
y
e
r
s
a
c
c
e
pt
ke
ys
,
que
r
ie
s
,
or
va
lu
e
s
in
th
e
f
ir
s
t
le
ve
l.
A
s
c
a
le
d
dot
-
pr
oduc
t
a
tt
e
nt
io
n
f
unc
ti
on
is
us
e
d
on
th
e
s
e
c
ond
le
ve
l.
T
he
pr
oc
e
s
s
e
s
on
th
e
f
ir
s
t
a
nd
s
e
c
ond
le
ve
l
s
a
r
e
r
e
p
e
a
te
d
h
ti
m
e
s
in
pa
r
a
ll
e
l,
de
pe
ndi
ng
on
th
e
num
be
r
of
he
a
ds
in
th
e
m
ul
ti
-
he
a
d
a
tt
e
nt
io
n
bl
oc
k.
O
n
th
e
th
ir
d
le
ve
l,
a
c
onc
a
te
na
ti
on
pr
oc
e
s
s
c
onne
c
ts
th
e
out
put
s
of
th
e
s
e
ve
r
a
l
he
a
d
s
.
A
f
in
a
l
li
ne
a
r
(
de
ns
e
)
la
ye
r
pr
oduc
e
s
th
e
out
put
on
th
e
f
our
th
le
ve
l.
T
he
f
ol
lo
w
in
g
a
r
e
th
e
pr
im
a
r
y c
om
pone
nt
s
of
a
tt
e
nt
io
n:
−
k a
nd q a
r
e
di
m
e
ns
io
n ve
c
to
r
s
, dk, c
a
r
r
yi
ng t
he
ke
ys
a
nd que
r
ie
s
.
−
V
de
not
e
s
a
di
m
e
ns
io
n ve
c
to
r
, dv, c
ont
a
in
in
g t
he
va
lu
e
s
.
−
K
, Q
, a
nd V
a
r
e
m
a
tr
ic
e
s
t
ha
t
bundle
t
oge
th
e
r
ke
ys
,
que
r
ie
s
, a
nd va
lu
e
s
.
−
W
K
, W
Q
, a
nd W
V
ge
ne
r
a
te
di
s
ti
nc
t
s
ubs
pa
c
e
r
e
pr
e
s
e
nt
a
ti
ons
of
t
he
ke
y, que
r
y, a
nd va
lu
e
m
a
tr
ic
e
s
.
−
W
O
, de
s
ig
na
ti
ng a
m
ul
ti
-
he
a
d pr
oj
e
c
ti
on ma
tr
ix
.
T
he
a
tt
e
nt
io
n
f
unc
ti
on
is
a
m
a
ppi
ng
f
r
o
m
a
que
r
y
to
a
s
e
t
of
k
e
y
-
va
lu
e
pa
ir
s
a
nd
th
e
n
to
a
n
out
put
.
T
he
r
e
s
ul
t
is
a
w
e
ig
ht
e
d
s
um
of
th
e
va
lu
e
s
,
w
it
h
th
e
w
e
ig
h
t
a
s
s
ig
ne
d
to
e
a
c
h
va
lu
e
de
te
r
m
in
e
d
by
th
e
c
om
pa
ti
bi
li
ty
f
unc
ti
on
of
th
e
que
r
y
w
it
h
th
e
r
e
le
v
a
nt
ke
y.
I
n
th
i
s
s
tu
dy, we
u
s
e
one
-
di
m
e
ns
io
n
a
l
c
onvolut
io
ns
w
it
h
m
ul
ti
-
he
a
d
C
N
N
s
,
w
he
r
e
th
e
di
m
e
n
s
io
n
di
c
ta
te
s
how
it
a
na
ly
z
e
s
in
put
da
t
a
.
T
o
ha
ndl
e
th
e
in
put
,
th
e
s
in
gl
e
-
c
ha
nne
l
C
N
N
e
m
pl
oys
a
s
in
gl
e
c
onvolut
io
na
l
h
e
a
d
w
it
h
a
s
in
gl
e
c
ha
nne
l.
I
t
pr
oc
e
s
s
e
s
e
a
c
h
pr
oc
e
s
s
va
r
ia
bl
e
in
di
vi
dua
ll
y
us
in
g
in
de
pe
nde
nt
s
in
gl
e
-
c
ha
nne
l
c
onvolut
io
na
l
he
a
ds
.
E
a
c
h
bl
oc
k
C
N
N
B
r
a
nc
h
w
il
l
in
c
lu
de
a
s
e
que
nt
ia
l
la
ye
r
of
C
onv1D
(
f
il
te
r
s
32,
ke
r
ne
l
s
iz
e
3)
,
C
onv1D
(
f
i
lt
e
r
s
16,
ke
r
ne
l
s
iz
e
5)
,
a
nd
C
onv1D
(
f
il
te
r
s
8,
ke
r
ne
l
s
iz
e
7)
.
T
he
hype
r
pa
r
a
m
e
te
r
s
a
r
e
ba
t
c
h
s
iz
e
64,
e
po
c
hs
50,
in
it
ia
l
le
a
r
ni
ng
r
a
te
0.5,
a
nd e
nd l
e
a
r
ni
ng r
a
te
0.008. T
he
m
ul
ti
-
he
a
d C
N
N
s
to
r
y i
s
s
e
e
n
in
F
ig
ur
e
6.
O
ne
a
dva
nc
e
d
us
e
of
m
ul
ti
-
he
a
d
a
tt
e
nt
io
n
is
in
th
e
c
ont
e
xt
of
s
e
lf
-
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
s
;
th
e
in
put
s
e
que
nc
e
i
s
pr
oc
e
s
s
e
d
by
a
tt
e
ndi
ng
to
di
f
f
e
r
e
nt
pa
r
ts
of
th
e
s
e
que
nc
e
to
c
om
put
e
a
w
e
ig
ht
e
d
s
um
of
it
s
e
le
m
e
nt
s
,
w
it
h
th
e
w
e
ig
ht
s
d
e
te
r
m
in
e
d
by
th
e
s
im
il
a
r
it
y
be
twe
e
n
th
e
e
le
m
e
nt
s
.
M
ul
ti
-
he
a
d
a
tt
e
nt
io
n
a
ll
ow
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
:
4778
-
4794
4784
th
e
ne
ur
a
l
ne
twor
k
to
gove
r
n
th
e
m
ix
in
g
of
in
f
or
m
a
ti
on
be
twe
e
n
s
e
gm
e
nt
s
of
a
n
in
put
s
e
que
nc
e
,
r
e
s
ul
ti
ng
in
r
ic
he
r
r
e
pr
e
s
e
nt
a
ti
ons
a
nd i
m
pr
ove
d pe
r
f
or
m
a
nc
e
on M
L
t
a
s
ks
.
F
ig
ur
e
6.
T
he
m
ul
ti
-
he
a
d C
N
N
s
tr
uc
tu
r
a
l
3.3. F
e
d
e
r
at
e
d
l
e
ar
n
in
g p
r
oc
e
s
s
T
h
e
FL
s
tr
u
c
tu
r
e
c
o
n
s
i
s
t
s
of
m
ul
ti
p
le
c
l
i
e
n
t
s
a
nd
o
n
e
s
e
r
v
e
r
;
m
a
in
l
y
, t
h
e
s
e
r
v
e
r
pr
o
vi
d
e
s
a
g
lo
b
a
l
s
h
a
r
e
d
m
o
d
e
l
.
I
n e
a
c
h c
o
m
m
u
ni
c
a
t
io
n b
e
t
w
e
e
n t
h
e
s
e
r
v
e
r
a
nd t
h
e
c
li
e
nt
, t
h
e
c
l
i
e
n
t
do
w
nl
oa
d
s
t
h
e
m
o
d
e
l
a
nd
tr
a
in
s
w
it
h
l
o
c
a
l
d
a
t
a
s
e
t
s
w
hi
l
e
up
d
a
t
in
g
m
o
de
l
pa
r
a
m
e
t
e
r
s
.
T
h
e
s
e
r
v
e
r
r
e
c
e
iv
e
s
t
h
e
c
ur
r
e
n
t
m
od
e
l
p
a
r
a
m
e
t
e
r
s
d
i
s
tr
ib
u
te
d
to
e
a
c
h
c
l
ie
nt
f
ol
l
ow
in
g
t
r
a
in
in
g
,
a
nd
t
h
e
n
t
h
e
u
pd
a
te
d
m
od
e
l
p
a
r
a
m
e
t
e
r
s
a
r
e
u
pl
o
a
de
d
t
o
th
e
s
e
r
ve
r
[
2
8]
.
W
e
te
s
te
d
two
m
e
th
ods
to
a
ggr
e
ga
te
c
li
e
nt
m
ode
l
pa
r
a
m
e
te
r
s
;
th
e
f
ir
s
t
us
e
d
th
e
F
e
dA
vg
m
e
th
od
[
29]
.
T
he
s
e
r
ve
r
is
pr
in
c
ip
a
ll
y
c
ont
r
ol
le
d
by
th
r
e
e
e
s
s
e
nt
ia
l
pa
r
a
m
e
te
r
s
:
C
,
th
e
pr
opor
ti
on
of
c
li
e
nt
s
;
B
,
th
e
lo
c
a
l
ba
tc
h
s
iz
e
f
or
c
li
e
nt
upd
a
te
s
;
a
nd
E
,
th
e
num
be
r
of
lo
c
a
l
ba
tc
h
upda
te
s
.
T
he
a
ggr
e
ga
ti
on
s
e
r
ve
r
in
it
ia
ll
y
c
hoos
e
s
a
C
p
e
r
c
e
nt
a
ge
of
K
c
li
e
nt
s
to
jo
in
th
e
F
L
pr
oc
e
s
s
a
nd
e
x
e
c
ut
e
R
\
FL
r
ounds
.
A
t
r
a
ndom,
th
e
a
ggr
e
ga
ti
on
s
e
r
ve
r
bui
ld
s
a
ge
ne
r
ic
m
ode
l
w
it
h
a
r
a
ndom
s
e
t
o
f
in
it
ia
l
w
e
ig
ht
s
w
.
T
he
ge
ne
r
ic
m
ode
l
is
th
e
n
obt
a
in
e
d
by
e
a
c
h
c
li
e
nt
k
f
r
om
th
e
a
ggr
e
ga
te
s
e
r
ve
r
.
E
ve
r
y
c
li
e
nt
r
e
tr
a
in
s
th
e
ge
ne
r
ic
m
ode
l
lo
c
a
ll
y
us
in
g
it
s
da
ta
a
nd
c
om
put
e
s
a
ne
w
s
e
t
of
w
e
ig
ht
s
f
or
th
e
ne
w
ly
c
r
e
a
te
d
l
oc
a
l
m
ode
l.
T
he
y
a
ppl
ie
d
th
e
f
ol
lo
w
in
g
(
4)
to
upda
te
t
he
m
ode
l
w
e
ig
ht
s
.
+
1
=
∑
=
1
+
1
(
4)
W
he
r
e
de
not
e
s
th
e
to
ta
l
s
iz
e
of
a
ll
c
li
e
nt
da
ta
s
e
t
s
a
nd
nk
th
e
s
i
z
e
of
e
a
c
h.
W
t+
1
I
t
is
th
e
it
e
r
a
ti
on'
s
upda
te
d
gl
oba
l
m
ode
l.
T
he
m
odi
f
ie
d
m
ode
l
ha
s
b
e
e
n
di
s
tr
ib
ut
e
d
to
th
e
c
li
e
nt
s
.
A
f
te
r
th
a
t,
th
e
s
e
r
ve
r
c
om
bi
ne
s
a
ll
c
li
e
nt
pa
r
a
m
e
te
r
s
(
k=
1)
.
T
he
ne
w
gl
oba
l
m
ode
l
is
s
ub
s
e
que
nt
l
y
s
e
nt
to
th
e
c
li
e
nt
s
,
w
ho
ut
il
iz
e
th
e
upda
te
d
pa
r
a
m
e
te
r
s
to
e
nh
a
nc
e
th
e
gl
oba
l
m
ode
l.
T
hi
s
pr
oc
e
dur
e
is
c
ont
in
ue
d
unt
il
th
e
m
ode
l
ha
s
r
e
a
c
he
d
c
onve
r
ge
nc
e
. T
h
e
pr
oc
e
dur
e
s
u
s
e
d t
o t
r
a
in
t
he
va
r
io
us
c
u
s
to
m
e
r
s
e
ts
a
r
e
de
pi
c
te
d i
n A
lg
or
it
hm
1.
I
n
th
e
s
e
c
ond
te
s
te
d
m
e
th
od,
w
e
us
e
d
th
e
F
e
dM
A
a
lg
or
it
hm
[
30]
.
T
he
F
e
dM
A
a
lg
or
it
hm
e
m
pl
oys
th
e
f
ol
lo
w
in
g
la
ye
r
-
w
is
e
m
a
tc
hi
ng
s
c
he
m
e
.
T
o
b
e
gi
n,
th
e
d
a
ta
c
e
nt
e
r
c
ol
le
c
ts
ju
s
t
th
e
w
e
ig
ht
s
of
th
e
f
ir
s
t
la
ye
r
s
f
r
om
th
e
c
li
e
nt
s
a
nd c
onduc
ts
on
e
-
la
ye
r
m
a
tc
hi
ng
to
ge
n
e
r
a
te
th
e
f
e
de
r
a
te
d
m
ode
l'
s
f
ir
s
t
-
la
ye
r
w
e
ig
ht
s
.
A
da
ta
c
e
nt
e
r
th
e
n
br
oa
dc
a
s
ts
th
e
s
e
w
e
ig
ht
s
to
c
li
e
nt
s
,
w
ho
s
ubs
e
que
nt
ly
tr
a
in
a
ll
s
ubs
e
que
nt
la
ye
r
s
on
th
e
ir
da
ta
s
e
ts
w
hi
le
f
r
e
e
z
in
g
th
e
m
a
tc
hi
ng
f
e
d
e
r
a
te
d
la
ye
r
s
.
T
hi
s
te
c
hni
que
c
ont
in
ue
s
unt
il
w
e
r
e
a
c
h
th
e
la
s
t
la
ye
r
,
w
he
r
e
w
e
do we
ig
ht
a
ve
r
a
gi
ng ba
s
e
d on the
c
la
s
s
pr
opor
ti
ons
of
da
ta
poi
nt
s
pe
r
c
li
e
nt
. T
he
F
e
dM
A
t
e
c
hni
que
ne
c
e
s
s
it
a
te
s
c
om
m
uni
c
a
ti
on
r
ounds
e
qua
l
to
th
e
num
be
r
of
ne
twor
k
la
ye
r
s
.
L
oc
a
l
c
li
e
nt
s
r
e
c
e
iv
e
th
e
m
a
tc
he
d
gl
oba
l
m
ode
l
a
t
th
e
s
ta
r
t
o
f
a
ne
w
r
ound
a
nd
r
e
c
ons
tr
uc
t
th
e
ir
l
oc
a
l
m
ode
ls
w
it
h
th
e
s
a
m
e
s
iz
e
a
s
th
e
or
ig
in
a
l
lo
c
a
l
m
ode
ls
b
a
s
e
d
on
th
e
pr
e
vi
ous
r
ound'
s
m
a
tc
hi
ng
r
e
s
ul
ts
.
I
n
c
ont
r
a
s
t
to
th
e
na
iv
e
m
e
th
od
of
us
in
g
a
w
hol
ly
m
a
tc
he
d gl
oba
l
m
ode
l
a
s
a
s
ta
r
ti
ng point
a
c
r
os
s
c
li
e
nt
s
on e
a
c
h c
yc
le
, t
hi
s
pr
oc
e
dur
e
a
ll
ow
s
us
t
o ke
e
p
th
e
s
iz
e
of
th
e
gl
oba
l
m
ode
l
s
hor
t.
F
e
dM
A
id
e
nt
if
ie
s
m
a
tc
he
d
s
e
ts
of
c
onvolut
io
na
l
f
il
te
r
s
a
nd
a
ve
r
a
ge
s
th
e
m
to
f
or
m
t
he
gl
oba
l
c
onvolut
io
na
l
f
il
te
r
s
. F
e
dM
A
'
s
m
a
tc
he
d a
nd globa
l
f
il
te
r
s
e
xt
r
a
c
t
th
e
s
a
m
e
f
e
a
tu
r
e
f
r
om
t
h
e
in
c
om
in
g
da
ta
.
T
h
e
F
e
dA
vg'
s
gl
oba
l
f
il
te
r
is
th
e
a
ve
r
a
ge
of
th
e
c
li
e
nt
'
s
f
il
te
r
.
F
ig
ur
e
7
s
how
s
th
e
a
r
ti
c
le
F
e
dM
A
a
lg
or
it
hm
s
.
A
lg
or
it
hm
1:
F
e
dA
vg,
c
on
s
id
e
r
in
g
K
c
li
e
n
ts
B
r
e
pr
e
s
e
nt
s
th
e
lo
c
a
l
m
in
i
-
ba
tc
h
s
iz
e
,
E
r
e
pr
e
s
e
nt
s
th
e
num
be
r
of
lo
c
a
l
e
poc
h
s
,
th
e
num
be
r
of
gl
oba
l
r
ounds
is
,
r
e
pr
e
s
e
nt
s
th
e
pr
opor
ti
on
of
c
li
e
nt
s
a
nd i
s
t
he
l
e
a
r
ni
ng
.
A
lg
or
it
hm
1:
F
e
de
r
a
te
d l
e
a
r
ni
ng (
F
e
dA
vg)
[
29
]
1.
S
e
r
ve
r
(
K
,C
,R
)
2.
1
←
g
e
n
e
r
ic
m
o
d
e
l
()
3.
F
o
r
e
a
c
h
r
o
un
d
=
1
,…,do
4.
←m
a
x(
C
.K
,1)
5.
←(
r
a
n
d
o
m
s
e
t
of
m
c
l
ie
n
t
s
)
6.
F
o
r
e
a
c
h
c
l
ie
n
t
K
ϵ
St
in
P
a
r
a
l
l
e
l
, do
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
F
e
de
r
at
e
d
in
c
e
pt
io
n
-
m
ul
ti
-
he
ad att
e
nt
io
n m
ode
ls
f
or
c
y
b
e
r
-
at
ta
c
k
s
de
te
c
ti
on
(
I
m
ad T
ar
e
q A
L
-
H
al
boos
i
)
4785
7.
_
(
+
1
)
^
c
l
ie
n
t
up
d
a
t
e
(
wt
.
K
)
8.
E
nd
9.
w
t
+
1
←
∑
nk
n
k
k
=
1
w
t
+
1
k
10.
E
nd
1.
C
li
e
nt
(
w
,k)
2.
←
(
s
p
l
it
P
in
t
o
b
a
t
c
he
s
of
s
iz
e
B
)
3.
F
o
r
e
a
c
h
l
o
c
a
l
e
p
o
c
h
i
fr
o
m
1
to
E
,
do
4.
F
or
ba
tc
h
∈
, do
5.
←
−
ⴙ
∆
ꝭ
(
,
)
6.
E
nd
7.
E
nd
8.
S
e
n
d
w
to
s
e
r
v
e
r
F
ig
ur
e
7.
T
he
F
e
dM
A
a
lg
or
it
hm
s
tr
uc
tu
r
e
I
n
th
is
w
or
k,
C
N
N
m
a
k
e
s
u
s
e
of
one
-
di
m
e
ns
io
na
l
c
onvolut
io
ns
,
w
he
r
e
th
e
di
m
e
ns
io
n
di
c
ta
te
s
how
it
a
na
ly
z
e
s
in
put
da
ta
.
T
he
s
iz
e
of
th
e
pa
r
a
m
e
te
r
s
us
e
d
in
our
e
x
pe
r
im
e
nt
,
ba
tc
h_s
iz
e
=
512,
le
a
r
ni
ng
r
a
te
=
0.01,
r
e
tr
a
in
_l
r
=
0.01,
f
in
e
_t
une
_l
r
=
0.01,
e
poc
hs
=
5,
r
e
tr
a
in
_e
poc
hs
=
10,
f
in
e
_t
une
_e
poc
h
s
=
10,
pa
r
ti
ti
on_s
te
p_s
iz
e
=
6,
lo
c
a
l_
poi
nt
s
=
5000,
n_ne
ts
=
10
(
num
be
r
of
c
li
e
nt
s
)
.
us
e
d
opt
im
iz
e
r
S
G
D
.
T
he
la
te
s
t
a
lg
or
it
hm
on
FL
w
a
s
done
on
F
e
d
M
A
in
2020,
s
o
w
e
tr
ie
d
to
e
xpe
r
im
e
nt
.
A
lg
or
it
hm
2
s
um
m
a
r
iz
e
s
our
F
e
dM
A
.
A
lg
or
it
hm
2:
F
e
de
r
a
te
d l
e
a
r
ni
ng (
F
e
dM
A
)
[
30]
I
nput
:
lo
c
a
l
w
e
ig
ht
s
of
N
-
la
ye
r
a
r
c
hi
te
c
tu
r
e
s
{
(
1
)
,
(
2
)
,
…
,
(
)
}
=
1
O
ut
put
:
gl
oba
l
w
e
ig
ht
s
{
(
1
)
,
(
2
)
,
…
,
(
)
}
=
1
;
w
hi
le
≤
do
if
<
th
e
n
{
−
1
}
=
1
=
−
(
{
(
)
}
=
1
)
;
(
)
=
1
∑
(
)
−
1
;
e
ls
e
(
)
=
∑
∑
=
1
(
)
w
he
r
e
is
a
f
r
a
c
ti
on of
da
ta
poi
nt
s
w
it
h l
a
be
ls
on t
he
w
or
ke
r
;
E
nd
f
or
∈
{
1
,
2
.
.
.
,
}
do
(
+
1
)
←
(
)
(
+
1
)
;
//
pe
r
m
ut
a
te
t
he
ne
xt
-
la
ye
r
w
e
ig
ht
s
T
r
a
in
{
(
+
1
)
,
.
.
.
(
)
}
w
it
h
(
)
f
r
oz
e
n;
E
nd
=
+
1
;
E
nd
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
:
4778
-
4794
4786
O
ur
f
in
a
l
a
lg
or
it
hm
w
it
h
f
e
de
r
a
te
d
a
ve
r
a
ge
a
nd
m
ode
l
s
a
r
c
h
pl
us
s
m
ot
e
w
il
l
be
.
T
he
s
um
m
a
r
y
of
th
e
pr
opos
e
d A
lg
or
it
hm
3, i
n s
te
ps
:
−
S
te
p 1:
W
e
s
pl
it
th
e
I
oT
da
ta
s
e
t
in
to
t
w
o gr
oups
, t
r
a
in
(
X
_t
r
a
in
, y_t
r
a
in
)
a
nd
va
li
da
ti
on (
X
_va
l,
y_va
l)
s
e
ts
w
it
h a
r
a
ti
o of
20%
f
or
va
li
da
ti
on a
nd r
e
m
a
in
f
or
tr
a
in
in
g w
it
h s
tr
a
ti
f
y
opt
io
n.
−
S
te
p2:
N
or
m
a
li
z
e
X
_t
r
a
in
, X
_va
l
w
it
h M
in
M
a
x nor
m
a
li
z
a
ti
on t
e
c
hni
que
s
−
S
te
p3:
W
e
ha
ve
t
w
o opti
ons
t
o ove
r
c
om
e
t
he
i
m
ba
la
nc
e
pr
obl
e
m
us
in
g c
la
s
s
w
e
ig
ht
s
or
s
m
ot
e
t
e
c
hni
que
s
−
S
te
p 4:
S
a
m
pl
in
g
our
t
r
a
in
da
ta
s
e
t
be
twe
e
n n c
li
e
nt
s
w
it
h t
w
o t
e
c
hni
que
s
:
I
I
D
or
non
-
IID
−
S
te
p 5:
B
a
tc
h s
iz
e
e
a
c
h n c
li
e
nt
da
ta
s
e
t
to
ba
tc
he
s
w
it
h s
iz
e
ba
tc
h s
iz
e
−
S
te
p 6:
W
e
r
e
duc
e
t
he
opt
im
iz
e
r
-
le
a
r
ni
ng r
a
te
e
ve
r
y f
iv
e
e
poc
hs
w
it
h 0.5
−
S
te
p 7:
W
e
ke
e
p t
r
a
c
ki
ng t
he
gl
oba
l
m
ode
l
a
c
c
ur
a
c
y t
o ge
t
th
e
be
s
t
m
ode
l
w
e
ig
ht
s
.
A
lg
or
it
hm
3:
P
r
oc
e
s
s
f
e
de
r
a
te
d l
e
a
r
ni
ng
I
nput
:
I
oT
D
a
ta
s
e
t,
tr
a
in
D
a
ta
M
ode
,
n
C
li
e
nt
s
,
s
a
m
pl
in
g
T
e
c
hni
que
,
ba
tc
h
-
s
iz
e
,
c
om
m
on
R
ounds
,
in
it
L
e
a
r
ni
ng R
a
te
O
ut
put
:
M
ode
l
W
e
ig
ht
s
1.
D
iv
id
e
I
oT
D
a
ta
s
e
t
in
to
t
r
a
in
in
g a
nd va
li
da
ti
on s
e
t
2.
N
or
m
a
li
z
e
X
_t
r
a
in
, X
_va
l
3.
C
he
c
k t
he
i
m
ba
la
nc
e
da
ta
s
e
t
to
s
ol
ve
o
I
f
tr
a
in
D
a
ta
M
ode
=
S
M
O
T
E
:
▪
U
ps
a
m
pl
in
g t
he
X
_t
r
a
in
D
a
ta
s
e
t
▪
S
e
t
c
la
s
s
-
W
e
ig
ht
s
<
-
N
one
o
E
ls
e
▪
C
la
s
s
-
W
e
ig
ht
s
<
-
G
e
t
C
la
s
s
W
e
ig
ht
s
f
or
y_t
r
ia
n
4.
C
he
c
k t
he
s
a
m
pl
in
g t
e
c
hni
que
t
o s
pl
it
t
he
da
ta
be
twe
e
n n C
li
e
nt
s
o
I
f
s
a
m
pl
in
g T
e
c
hni
que
=
=
I
I
D
:
▪
S
e
t
c
li
e
nt
s
da
ta
s
e
t
<
-
S
pl
it
tr
a
in
da
ta
s
e
t
be
twe
e
n n C
li
e
nt
s
w
it
h I
de
nt
ic
a
ll
y D
is
tr
ib
ut
e
d
o
E
ls
e
▪
S
e
t
c
li
e
nt
s
da
ta
s
e
t
<
-
S
pl
it
t
r
a
in
da
ta
s
e
t
be
twe
e
n n C
li
e
nt
s
w
it
h N
on
-
I
de
nt
ic
a
ll
y D
is
tr
ib
ut
e
d
5.
S
e
t
c
li
e
nt
s
B
a
tc
he
d <
-
S
pl
it
e
a
c
h n C
li
e
nt
s
i
n c
li
e
nt
’
s
da
ta
s
e
t
to
ba
tc
he
s
w
it
h s
iz
e
ba
tc
h S
iz
e
6.
I
ni
ti
a
li
z
e
t
he
gl
oba
l
m
ode
l
7.
S
e
t
be
s
t
G
lo
ba
l
M
ode
l
W
e
ig
ht
s
<
-
gl
oba
l
M
ode
l
G
e
t
W
e
ig
ht
s
(
)
8.
S
e
t
be
s
t
A
c
c
ur
a
c
y R
e
s
ul
t
<
-
0
9.
S
e
t
le
a
r
ni
ng r
a
te
<
-
in
it
L
e
a
r
ni
ng
R
a
te
10.
I
ni
ti
a
li
z
e
opt
im
iz
e
r
w
it
h
le
a
r
ni
ng R
a
te
11.
S
e
t
lo
s
s
F
unc
<
-
lo
s
s
f
unc
ti
on t
o be
C
a
te
gor
ic
a
l
C
r
os
s
-
e
nt
r
opy
12.
S
e
t
e
va
lu
a
ti
on M
e
tr
ic
s
<
-
E
va
lu
a
ti
on me
tr
ic
s
(
a
c
c
ur
a
c
y, pr
e
c
is
io
n, r
e
c
a
ll
, f
1
-
s
c
or
e
)
13.
F
or
e
a
c
h r
ound in c
om
m
on R
ounds
:
o
S
e
t
gl
oba
l
W
e
ig
ht
s
<
-
gl
oba
l
M
ode
l
G
e
t
W
e
ig
ht
s
(
)
o
S
e
t
c
li
e
nt
N
a
m
e
s
<
-
ge
t
c
li
e
nt
na
m
e
f
r
om
c
li
e
nt
s
B
a
tc
he
d i
n a
l
is
t
of
c
li
e
nt
s
ba
tc
he
d
o
S
huf
f
le
c
li
e
nt
N
a
m
e
s
o
S
e
t
le
a
r
ni
ng r
a
te
<
-
lr
S
c
he
dul
e
r
(
le
a
r
ni
ng
R
a
te
, r
ound)
o
S
e
t
opt
im
iz
e
r
l
e
a
r
ni
ng r
a
te
<
-
L
e
a
r
ni
ng
R
a
te
o
S
e
t
e
m
pt
y r
ound M
ode
l
W
e
ig
ht
s
l
is
t
o
F
or
a
c
li
e
nt
i
n c
li
e
nt
N
a
m
e
s
:
▪
I
ni
ti
a
li
z
e
t
he
l
oc
a
l
m
ode
l
▪
B
ui
ld
l
oc
a
l
m
ode
l
w
it
h (
in
put
s
ha
pe
, num C
la
s
s
e
s
)
▪
C
om
pi
le
l
oc
a
l
M
ode
l
W
it
h (
opt
im
iz
e
r
=
opt
im
iz
e
r
, l
os
s
=
lo
s
s
F
unc
, m
e
tr
ic
s
=
e
va
lu
a
ti
on M
e
tr
ic
s
)
▪
S
e
t
L
oc
a
l
m
ode
l
<
-
L
oc
a
l
m
ode
l
s
e
t
w
e
ig
ht
s
(
gl
oba
l
W
e
ig
ht
s
)
▪
F
it
l
oc
a
l
M
ode
l
w
it
h (
c
li
e
nt
s
B
a
tc
he
d[
c
li
e
nt
]
, c
la
s
s
W
e
ig
ht
s
)
▪
S
e
t
lo
c
a
l
M
ode
l
W
e
ig
ht
s
<
-
ge
t
lo
c
a
l
M
ode
l
G
e
t
W
e
ig
ht
s
(
)
▪
A
ppe
nd l
oc
a
l
M
ode
l
W
e
ig
ht
s
t
o r
ound M
ode
l
W
e
ig
ht
s
o
S
e
t
a
ve
r
a
ge
W
e
ig
ht
s
<
-
a
ve
r
a
ge
t
he
r
ound M
ode
l
W
e
ig
ht
s
l
is
t
o
S
e
t
gl
oba
l
m
ode
l
<
-
gl
oba
l
m
ode
l.
s
e
t
w
e
ig
ht
s
(
a
ve
r
a
ge
W
e
ig
ht
s
)
14.
S
e
t
c
ur
r
e
nt
G
lo
ba
l
M
ode
l
A
c
c
ur
a
c
y <
-
E
va
lu
a
te
t
he
gl
oba
l
M
ode
l
w
it
h (
X
_va
l,
y_va
l)
15.
C
he
c
k f
or
t
he
be
s
t
gl
oba
l
m
ode
l
w
e
ig
ht
o
I
f
c
ur
r
e
nt
G
lo
ba
l
M
ode
l
A
c
c
ur
a
c
y >
be
s
t
A
c
c
ur
a
c
y R
e
s
ul
t:
▪
S
e
t
be
s
t
G
lo
ba
l
M
ode
l
W
e
ig
ht
s
<
-
gl
oba
l
M
ode
l
G
e
t
W
e
ig
ht
s
(
)
16.
R
e
tu
r
n be
s
t
G
lo
ba
l
M
ode
l
W
e
ig
ht
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
F
e
de
r
at
e
d
in
c
e
pt
io
n
-
m
ul
ti
-
he
ad att
e
nt
io
n m
ode
ls
f
or
c
y
b
e
r
-
at
ta
c
k
s
de
te
c
ti
on
(
I
m
ad T
ar
e
q A
L
-
H
al
boos
i
)
4787
3.4
.
M
u
lt
i
-
s
t
age
f
r
am
e
w
o
r
k
of
t
r
u
s
t
e
d
f
e
d
e
r
at
e
d
l
e
ar
n
in
g
T
he
m
ul
ti
-
s
ta
ge
a
r
c
hi
te
c
tu
r
e
of
th
e
FL
e
xe
c
ut
io
n,
a
s
il
lu
s
tr
a
te
d
in
F
ig
ur
e
8,
m
a
y
be
s
e
pa
r
a
te
d
in
to
two s
ta
ge
s
:
tr
a
in
in
g a
nd pr
e
di
c
ti
on. At
e
a
c
h s
ta
ge
of
FL
e
xe
c
ut
i
on, t
he
m
ode
l
e
nc
ount
e
r
s
s
e
c
ur
it
y a
nd pr
iv
a
c
y
c
onc
e
r
ns
. F
L
ne
e
ds
a
ppr
opr
ia
te
s
e
c
ur
it
y a
nd pr
iv
a
c
y s
a
f
e
gu
a
r
ds
a
t
e
a
c
h s
ta
g
e
.
F
ig
ur
e
8. T
he
m
ode
ls
of
t
he
m
ul
ti
-
s
ta
ge
F
L
, i
nc
lu
di
ng t
he
t
r
a
in
in
g m
ode
ls
3.4.1.
T
r
ai
n
in
g s
t
age
F
L
ne
e
ds
t
he
c
ol
la
bor
a
ti
on of
nume
r
ous
l
oc
a
l
I
oT
de
vi
c
e
s
t
o t
r
a
in
a
gl
oba
l
m
ode
l.
M
a
li
c
io
us
I
oT
c
a
n
m
odi
f
y
it
s
da
ta
,
m
ode
l
gr
a
di
e
nt
s
,
a
nd
pa
r
a
m
e
te
r
s
dur
in
g
th
e
m
ode
l
-
tr
a
in
in
g
s
te
p.
A
s
a
r
e
s
ul
t,
if
a
dve
r
s
a
r
ie
s
pe
ne
tr
a
te
th
e
I
oT
de
vi
c
e
s
,
th
e
y
c
a
n
di
s
r
upt
th
e
in
te
gr
it
y
of
th
e
tr
a
in
in
g
da
ta
s
e
t
or
m
ode
l,
im
pa
ir
in
g
th
e
pe
r
f
or
m
a
nc
e
of
th
e
gl
oba
l
m
ode
l.
F
ur
th
e
r
m
or
e
,
th
e
s
e
r
ve
r
c
a
n
pe
r
f
or
m
pa
s
s
iv
e
or
a
ggr
e
s
s
iv
e
in
f
e
r
e
nc
e
a
s
s
a
ul
t
s
.
F
ur
th
e
r
m
or
e
,
dur
in
g
th
e
upl
oa
d
a
nd
dow
nl
oa
d
of
m
ode
l
upda
te
s
,
in
te
r
m
e
di
a
te
s
in
th
e
c
om
m
uni
c
a
ti
on
r
out
e
m
a
y
a
tt
a
c
k
th
e
m
ode
l
s
,
r
e
s
ul
ti
ng
in
ta
m
p
e
r
e
d
w
it
h
or
s
to
le
n
m
ode
l
upda
te
s
.
A
s
a
r
e
s
ul
t,
s
e
c
ur
in
g t
he
c
om
m
uni
c
a
ti
on of
m
ode
l
upda
te
s
b
e
twe
e
n I
oT
d
e
v
ic
e
s
a
nd t
he
s
e
r
ve
r
i
s
c
r
it
ic
a
l.
3.4.2.
P
r
e
d
ic
t
in
g
s
t
age
O
nc
e
th
e
m
ode
l
ha
s
b
e
e
n
tr
a
in
e
d,
th
e
gl
oba
l
m
ode
l
i
s
de
li
ve
r
e
d
to
a
ll
I
oT
de
vi
c
e
s
,
r
e
g
a
r
dl
e
s
s
of
w
he
th
e
r
th
e
y
to
ok
pa
r
t
in
th
e
tr
a
in
in
g
or
no
t.
C
ybe
r
s
e
c
ur
it
y
t
hr
e
a
ts
a
r
e
c
om
m
on
dur
in
g
th
is
s
ta
ge
.
A
tt
a
c
ks
ty
pi
c
a
ll
y
do
not
a
lt
e
r
th
e
ta
r
ge
t
m
ode
l
but
t
r
ic
k
it
in
to
g
iv
i
ng
in
c
or
r
e
c
t
pr
e
di
c
ti
ons
.
A
f
te
r
obt
a
in
in
g
th
e
opt
im
a
l
gl
oba
l
m
ode
l
w
e
ig
ht
s
,
w
e
c
a
n
br
oa
d
c
a
s
t
to
c
li
e
nt
s
,
m
a
ke
pr
e
di
c
ti
ons
f
or
th
e
te
s
t
da
ta
s
e
t,
a
nd
th
e
n
a
s
s
e
s
s
t
he
out
c
om
e
s
us
in
g
a
c
c
ur
a
c
y, pr
e
c
i
s
io
n, F
1
-
s
c
or
e
, r
e
c
a
ll
,
a
nd l
os
s
.
4.
S
Y
S
T
E
M
D
E
S
I
G
N
4.1.
D
at
as
e
t
s
T
he
c
ybe
r
s
e
c
ur
it
y
E
dge
-
I
I
oT
da
ta
s
e
t
f
or
I
I
oT
a
nd
I
oT
a
ppl
ic
a
ti
ons
is
ut
il
is
e
d
in
ML
-
ba
s
e
d
I
D
S
to
a
s
s
e
s
s
t
he
m
ode
l
s
i
n t
hi
s
pa
pe
r
. T
he
E
dge
-
I
I
oT
da
ta
s
e
t
c
ove
r
s
f
our
te
e
n di
s
ti
nc
t
f
or
m
s
of
a
tt
a
c
ks
s
e
pa
r
a
te
d i
nt
o
f
iv
e
c
la
s
s
if
ic
a
ti
ons
(
di
s
tr
ib
ut
e
d
de
ni
a
l
-
of
-
s
e
r
vi
c
e
(
D
D
oS
)
a
nd
de
ni
a
l
-
of
-
s
e
r
vi
c
e
(
D
D
oS
)
a
s
s
a
ul
ts
,
in
je
c
ti
on
a
tt
a
c
ks
,
in
f
or
m
a
ti
on
ga
th
e
r
in
g,
m
a
lwa
r
e
a
tt
a
c
ks
,
a
nd
m
a
n
in
th
e
m
id
dl
e
(
M
I
T
M
)
a
tt
a
c
ks
)
.
E
dge
I
I
oT
[
11]
ha
s
20,952,648
ty
pi
c
a
l
a
tt
a
c
k
s
ta
ti
s
ti
c
s
,
w
hi
c
h
c
om
pr
is
e
11,223,940
nor
m
a
l
a
nd
9,728,708
a
tt
a
c
ks
.
W
e
di
vi
de
d
th
is
da
ta
s
e
t
in
to
20%
f
or
te
s
ti
ng
a
nd
80%
f
or
tr
a
in
in
g,
w
i
th
th
e
opt
io
n
of
s
tr
a
ti
f
yi
ng
to
m
a
in
ta
in
th
e
pe
r
c
e
nt
a
ge
s
c
on
s
ta
nt
a
c
r
os
s
a
ll
c
la
s
s
e
s
.
T
he
da
ta
s
e
t
yi
e
ld
e
d 61 f
e
a
tu
r
e
s
a
nd 1,909,671 s
a
m
pl
e
s
, 1,527,736 a
nd
381,935 da
ta
s
e
ts
a
r
e
a
va
il
a
bl
e
f
or
t
r
a
in
in
g a
nd
t
e
s
ti
ng, r
e
s
pe
c
ti
ve
ly
, w
hi
c
h w
e
r
e
di
vi
de
d i
nt
o 15
c
a
te
gor
ie
s
a
s
s
how
n i
n T
a
bl
e
2
. W
e
us
e
d c
la
s
s
w
e
ig
ht
s
s
in
c
e
t
he
da
ta
w
e
r
e
i
m
ba
la
nc
e
d. F
ol
lo
w
in
g va
r
io
us
a
tt
e
m
pt
s
t
o f
in
e
-
tu
ne
th
e
hype
r
-
pa
r
a
m
e
te
r
s
,
th
e
ba
tc
h
s
iz
e
w
a
s
s
e
t
a
t
64,
w
hi
c
h
is
a
ppr
opr
ia
te
f
or
m
e
m
or
y
th
a
t
m
a
in
ta
in
s
th
e
tr
a
in
in
g
pr
oc
e
s
s
w
it
hout
f
lu
c
tu
a
ti
on
a
nd
c
onve
r
ge
s
w
it
h
ge
ne
r
a
l
pe
r
f
or
m
a
nc
e
.
T
he
de
pt
h
w
a
s
lo
w
e
r
e
d
to
f
our
or
s
ix
f
or
c
e
r
ta
in
E
dge
-
I
I
oT
da
ta
be
c
a
us
e
s
om
e
da
ta
s
e
ts
r
e
qui
r
e
a
m
or
e
s
ophi
s
ti
c
a
te
d
m
ode
l
to
c
onve
r
ge
to
a
Evaluation Warning : The document was created with Spire.PDF for Python.