I
nte
rna
t
io
na
l J
o
urna
l o
f
I
nfo
rm
a
t
ics a
nd
Co
m
m
un
ica
t
io
n T
ec
hn
o
lo
g
y
(
I
J
-
I
CT
)
Vo
l.
1
5
,
No
.
3
,
Sep
tem
b
er
20
2
6
,
p
p
.
1
1
5
4
~
1
1
6
6
I
SS
N:
2252
-
8
7
7
6
,
DOI
:
1
0
.
1
1
5
9
1
/iji
ct
.
v
1
5
i
3
.
pp
1
1
5
4
-
1
1
6
6
1154
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ict.
ia
esco
r
e.
co
m
Enha
nced ano
ma
ly
det
ec
tion in Io
T networks
via
f
e
a
ture
fusio
n and
learni
ng
-
ba
sed echo
sta
te
ne
tworks
P
.
P
a
lpa
nd
i
1
,
B
.
Sa
k
t
hiv
el
2
,
M.
P
o
nn
ra
j
a
k
um
a
ri
3
,
M
.
I
nd
ira
ni
4
,
S G
o
v
ind
a
ra
j
u
5
,
S.
Deiv
a
s
ig
a
m
a
ni
6
1
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
(
I
o
T)
,
S
e
t
h
u
I
n
st
i
t
u
t
e
o
f
Te
c
h
n
o
l
o
g
y
,
V
i
d
u
d
h
u
n
a
g
a
r
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
o
n
i
c
s a
n
d
C
o
m
mu
n
i
c
a
t
i
o
n
En
g
i
n
e
e
r
i
n
g
,
P
a
n
d
i
a
n
S
a
r
a
sw
a
t
h
i
Y
a
d
a
v
E
n
g
g
C
o
l
l
e
g
e
,
S
i
v
a
g
a
n
g
a
i
,
I
n
d
i
a
3
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
o
n
i
c
s a
n
d
C
o
m
mu
n
i
c
a
t
i
o
n
En
g
i
n
e
e
r
i
n
g
,
V
e
l
a
m
ma
l
E
n
g
i
n
e
e
r
i
n
g
C
o
l
l
e
g
e
,
C
h
e
n
n
a
i
,
I
n
d
i
a
4
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
B
u
s
i
n
e
ss S
y
st
e
ms,
V
e
l
Te
c
h
M
u
l
t
i
Te
c
h
D
r
.
R
a
n
g
a
r
a
j
a
n
D
r
.
S
a
k
u
n
t
h
a
l
a
En
g
i
n
e
e
r
i
n
g
C
o
l
l
e
g
e
,
A
v
a
d
i
,
I
n
d
i
a
5
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
S
c
h
o
o
l
o
f
S
c
i
e
n
c
e
a
n
d
H
u
m
a
n
i
t
i
e
s
,
S
t
.
J
o
se
p
h
U
n
i
v
e
r
si
t
y
,
C
h
e
n
n
a
i
,
I
n
d
i
a
6
F
a
c
u
l
t
y
o
f
En
g
i
n
e
e
r
i
n
g
,
T
e
c
h
n
o
l
o
g
y
&
B
u
i
l
t
En
v
i
r
o
n
me
n
t
,
U
C
S
I
U
n
i
v
e
r
s
i
t
y
,
K
u
a
l
a
Lu
mp
u
r
,
M
a
l
a
y
s
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Au
g
1
2
,
2
0
2
5
R
ev
is
ed
Ap
r
1
4
,
2
0
2
6
Acc
ep
ted
Ma
y
1
7
,
2
0
2
6
Th
e
fa
st
d
e
v
e
lo
p
m
e
n
t
o
f
in
ter
n
e
t
o
f
th
in
g
s
(I
o
T)
n
e
two
r
k
s
h
a
s
led
t
o
a
n
in
c
re
a
se
d
p
ro
b
a
b
i
li
ty
o
f
c
y
b
e
ra
tt
a
c
k
s.
In
tr
u
sio
n
d
e
tec
ti
o
n
sy
ste
m
s
(IDS)
a
re
n
e
e
d
e
d
fo
r
i
d
e
n
ti
fy
i
n
g
u
n
a
u
th
o
ri
se
d
a
c
c
e
ss
a
n
d
m
a
li
c
io
u
s
a
c
ti
v
it
i
e
s
in
su
c
h
d
y
n
a
m
ic
e
n
v
iro
n
m
e
n
ts.
H
o
we
v
e
r
,
e
x
isti
n
g
m
a
c
h
i
n
e
lea
rn
i
n
g
(
M
L
)
m
o
d
e
ls
fa
il
e
d
to
h
a
n
d
le
th
e
c
o
m
p
lex
it
y
a
n
d
v
a
riab
il
it
y
o
f
m
o
d
e
rn
c
y
b
e
r
th
re
a
ts.
In
th
is
w
o
rk
,
a
h
y
b
ri
d
d
e
e
p
lea
rn
in
g
(DL)
-
b
a
se
d
a
n
o
m
a
ly
d
e
tec
ti
o
n
m
o
d
e
l
is
p
re
se
n
ted
fo
r
I
o
T
c
y
b
e
rse
c
u
rit
y
.
Th
e
m
o
d
e
l
c
o
m
b
i
n
e
s th
re
e
ty
p
e
s o
f
fe
a
tu
re
s:
(
i
)
su
p
e
rv
ise
d
fe
a
tu
re
e
x
trac
ti
o
n
u
sin
g
li
n
e
a
r
d
isc
rimin
a
n
t
a
n
a
ly
sis
(LDA)
to
e
x
trac
t
th
e
m
o
st
d
isc
rimin
a
ti
v
e
f
e
a
tu
re
s,
(
ii
)
u
n
su
p
e
rv
ise
d
fe
a
tu
re
lea
rn
in
g
th
ro
u
g
h
a
u
t
o
e
n
c
o
d
e
rs
to
c
a
p
t
u
re
l
a
ten
t
re
p
re
se
n
tatio
n
s
o
f
th
e
i
n
p
u
t
d
a
ta,
a
n
d
(
iii
)
sta
ti
stica
l
fe
a
tu
re
s
su
c
h
a
s
m
e
a
n
,
v
a
rian
c
e
,
s
k
e
wn
e
ss
,
a
n
d
k
u
rt
o
sis
t
o
lea
rn
in
p
u
t
c
h
a
ra
c
teristics
.
Th
e
f
u
se
d
fe
a
tu
re
m
a
tri
x
is
fe
d
i
n
t
o
a
lea
rn
in
g
-
b
a
se
d
e
c
h
o
sta
te n
e
two
r
k
(LBE
S
N)
fo
r
fi
n
a
l
d
e
tec
ti
o
n
.
T
h
e
p
a
ra
m
e
ters
o
f
th
e
LBE
S
N
m
o
d
e
l
a
re
t
u
n
e
d
u
si
n
g
b
lac
k
e
a
g
le
o
p
ti
m
ize
r
(BEO).
E
x
p
e
rime
n
tal
re
su
lt
s
o
n
sta
n
d
a
rd
in
tr
u
sio
n
d
e
tec
ti
o
n
d
a
tas
e
ts
su
c
h
a
s
UN
S
W
-
NB1
5
,
KD
D9
9
,
a
n
d
I
n
S
DN
sh
o
w
t
h
a
t
th
e
p
ro
p
o
se
d
m
o
d
e
l
a
c
h
iev
e
s
su
p
e
ri
o
r
p
e
rfo
rm
a
n
c
e
in
term
s
o
f
a
c
c
u
ra
c
y
,
p
re
c
isio
n
,
re
c
a
ll
,
a
n
d
F1
-
sc
o
re
c
o
m
p
a
re
d
to
c
o
n
v
e
n
t
io
n
a
l
DL
tec
h
n
iq
u
e
s.
K
ey
w
o
r
d
s
:
B
lack
ea
g
le
o
p
tim
izer
I
n
tr
u
s
io
n
d
etec
tio
n
s
y
s
tem
s
I
n
ter
n
et
o
f
th
in
g
s
L
ea
r
n
in
g
-
b
ased
ec
h
o
s
tate
Netwo
r
k
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
P.
Palp
an
d
i
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
an
d
E
n
g
in
ee
r
in
g
(
I
o
T
)
,
Seth
u
I
n
s
titu
te
o
f
T
ec
h
n
o
lo
g
y
Vid
u
d
h
u
n
ag
ar
,
I
n
d
ia
E
m
ail:
p
alp
an
d
i@
s
eth
u
.
ac
.
in
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
in
ter
n
et
o
f
th
in
g
s
(
I
o
T
)
h
as
em
er
g
ed
as
a
tr
an
s
f
o
r
m
ativ
e
p
ar
ad
i
g
m
.
T
h
is
in
n
o
v
atio
n
in
ter
co
n
n
ec
ts
b
illi
o
n
s
o
f
p
h
y
s
ical
d
ev
ices,
s
en
s
o
r
s
,
an
d
in
t
ellig
en
t
s
y
s
tem
s
th
r
o
u
g
h
th
e
in
ter
n
et
[
1
]
.
I
o
T
is
u
s
ed
f
o
r
co
n
tin
u
o
u
s
d
ata
ex
ch
an
g
e
an
d
d
ec
is
io
n
-
m
ak
i
n
g
ac
r
o
s
s
d
if
f
er
en
t
d
o
m
ain
s
s
u
ch
as
s
m
ar
t
citie
s
,
h
ea
lth
ca
r
e,
a
g
r
icu
ltu
r
e
,
tr
a
n
s
p
o
r
tatio
n
,
an
d
in
d
u
s
tr
ial
au
to
m
atio
n
.
Ho
wev
e
r
,
t
h
e
p
e
r
f
o
r
m
a
n
ce
o
f
I
o
T
n
etwo
r
k
s
is
d
eg
r
ad
ed
b
y
v
a
r
io
u
s
cy
b
e
r
s
ec
u
r
ity
v
u
ln
er
a
b
ilit
ies
[
2
]
,
[
3
]
.
W
ith
th
e
g
r
o
wth
o
f
I
o
T
,
cy
b
er
s
ec
u
r
ity
h
as
b
ec
o
m
e
a
cr
itical
co
n
ce
r
n
.
I
t
in
v
o
lv
es
v
ar
io
u
s
attac
k
s
lik
e
d
is
tr
ib
u
ted
d
en
ial
o
f
s
er
v
ic
e
(
DDo
S)
attac
k
s
,
m
alwa
r
e
p
r
o
p
ag
atio
n
,
b
o
tn
et
r
ec
r
u
itm
e
n
t,
a
n
d
d
ata
e
x
f
il
tr
atio
n
.
C
o
m
p
ar
ed
to
co
n
v
e
n
tio
n
al
co
m
p
u
tin
g
e
n
v
ir
o
n
m
en
ts
,
I
o
T
n
etwo
r
k
s
in
v
o
l
v
e
m
ass
iv
e,
d
iv
er
s
e,
a
n
d
d
y
n
am
ic
d
ata
s
tr
ea
m
s
.
T
h
e
d
e
v
elo
p
m
e
n
t
o
f
r
ea
l
-
tim
e
an
d
ac
c
u
r
ate
in
tr
u
s
io
n
d
etec
tio
n
r
e
m
ain
s
a
d
if
f
ic
u
lt c
h
allen
g
e.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
E
n
h
a
n
ce
d
a
n
o
ma
ly
d
etec
tio
n
i
n
I
o
T n
etw
o
r
ks v
ia
fea
tu
r
e
fu
s
io
n
a
n
d
lea
r
n
in
g
-
b
a
s
ed
…
(
P
.
P
a
lp
a
n
d
i
)
1155
T
o
o
v
er
co
m
e
th
ese
d
r
awb
ac
k
s
,
an
o
m
al
y
-
b
ased
in
tr
u
s
io
n
d
etec
tio
n
s
y
s
tem
s
(
A
-
I
DS)
h
av
e
g
ain
e
d
atten
tio
n
in
th
e
c
y
b
er
s
ec
u
r
ity
co
m
m
u
n
ity
[
4
]
.
T
h
ese
s
y
s
te
m
s
u
s
e
lear
n
in
g
alg
o
r
ith
m
s
to
id
en
tify
d
ev
iatio
n
s
f
r
o
m
n
o
r
m
al
b
eh
av
io
r
.
T
h
is
l
ea
r
n
in
g
a
p
p
r
o
ac
h
is
ap
p
lied
f
o
r
class
if
y
in
g
b
o
th
p
r
e
d
ef
in
e
d
an
d
n
o
v
el
attac
k
s
.
T
h
e
m
ac
h
in
e
lear
n
i
n
g
(
ML
)
an
d
d
ee
p
lear
n
in
g
(
DL
)
tech
n
iq
u
es
h
av
e
r
ec
en
tl
y
g
ain
ed
atten
tio
n
f
o
r
th
ei
r
ef
f
ec
tiv
en
ess
in
A
-
I
DS
f
o
r
th
e
class
if
icatio
n
o
f
an
o
m
alo
u
s
a
ctiv
ities
with
in
n
etwo
r
k
en
v
ir
o
n
m
en
ts
[
5
]
.
I
n
DL
ap
p
r
o
ac
h
es,
th
e
well
-
k
n
o
wn
m
o
d
els
lik
e
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
L
STM
)
n
etwo
r
k
s
,
co
n
v
o
lu
ti
o
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NN)
,
an
d
Au
to
e
n
co
d
er
s
ar
e
u
s
ed
to
h
an
d
le
c
o
m
p
lex
an
d
h
ig
h
-
d
im
en
s
io
n
al
d
ata.
T
h
ese
m
o
d
els
in
v
o
lv
e
d
if
f
er
e
n
t
lear
n
in
g
p
a
r
ad
ig
m
s
,
lik
e
s
u
p
e
r
v
is
ed
,
u
n
s
u
p
er
v
is
ed
,
an
d
s
em
i
-
s
u
p
er
v
is
ed
lear
n
in
g
.
T
h
is
p
r
o
v
id
es
f
lex
ib
ilit
y
in
a
d
ap
tin
g
to
v
ar
io
u
s
ty
p
es
o
f
c
y
b
er
s
ec
u
r
ity
ch
allen
g
es.
Ho
wev
er
,
co
n
v
en
tio
n
al
m
o
d
els
s
tr
u
g
g
le
to
ca
p
t
u
r
e
th
e
n
o
n
li
n
ea
r
r
elatio
n
s
h
ip
o
f
n
etwo
r
k
d
ata.
Als
o
,
th
e
tem
p
o
r
al
an
d
h
ig
h
-
d
im
en
s
io
n
a
l
p
atter
n
s
o
f
I
o
T
tr
af
f
ic
d
ata
c
an
n
o
t
b
e
ef
f
ec
ti
v
ely
p
r
o
c
ess
ed
b
y
ex
is
tin
g
m
o
d
e
ls
.
Fu
r
th
er
m
o
r
e,
th
e
m
o
d
els
b
ased
o
n
h
an
d
cr
a
f
ted
f
ea
tu
r
es
f
ailed
to
r
ep
r
esen
t
th
e
co
m
p
l
ex
ity
o
f
m
alicio
u
s
b
eh
av
io
r
s
ac
r
o
s
s
d
if
f
er
en
t
I
o
T
p
r
o
to
co
ls
an
d
en
v
ir
o
n
m
e
n
ts
.
T
o
f
ill
th
is
g
ap
,
a
h
y
b
r
id
DL
m
o
d
el
is
p
r
o
p
o
s
ed
f
o
r
I
o
T
c
y
b
er
s
ec
u
r
ity
ap
p
licatio
n
s
.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
in
teg
r
ates th
r
ee
d
if
f
er
e
n
t m
o
d
u
les f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
.
I
t in
v
o
lv
es
lin
ea
r
d
is
cr
im
i
n
an
t a
n
aly
s
is
(
L
DA)
an
d
Au
to
en
c
o
d
er
s
f
o
r
h
i
d
d
en
f
ea
tu
r
e
lear
n
i
n
g
.
Als
o
,
it
u
s
es
Stati
s
tical
Descr
ip
to
r
s
f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
.
T
h
ese
f
ea
tu
r
es
ar
e
co
m
b
in
ed
to
ca
p
tu
r
e
b
o
th
th
e
s
tatis
tical
b
eh
av
io
r
an
d
s
em
an
tic
p
atter
n
s
in
th
e
n
etwo
r
k
tr
af
f
ic.
T
h
e
n
,
t
h
e
f
u
s
ed
f
ea
tu
r
es
ar
e
p
r
o
ce
s
s
ed
b
y
a
lear
n
i
n
g
-
b
ased
ec
h
o
s
tate
n
etwo
r
k
(
L
B
E
SN)
f
o
r
f
in
al
d
etec
tio
n
.
2.
RE
L
AT
E
D
WO
RK
Desh
m
u
k
h
a
n
d
B
h
alad
h
ar
e
[
6
]
p
r
o
p
o
s
e
an
attac
k
er
class
if
icatio
n
m
o
d
el
f
o
r
th
e
n
etwo
r
k
u
s
i
n
g
a
d
ee
p
b
elief
n
etwo
r
k
(
DB
N)
.
T
o
i
m
p
r
o
v
e
ac
c
u
r
ac
y
,
t
h
e
DB
N
m
o
d
el
is
tr
ain
ed
u
s
in
g
tay
lo
r
-
s
p
id
er
m
o
n
k
ey
o
p
tim
izatio
n
(
T
a
y
lo
r
-
SMO)
.
E
x
p
er
im
en
tal
r
esu
lts
o
n
t
h
e
NSL
-
KDD
d
ataset
s
h
o
w
th
a
t
th
e
o
p
tim
izatio
n
p
r
o
ce
s
s
in
cr
ea
s
es
f
alse
p
o
s
itiv
e
r
ate
(
FP
R
)
an
d
f
alse
n
eg
a
tiv
e
r
ate
(
FNR
)
b
y
ch
o
o
s
in
g
th
e
m
o
s
t
r
elev
an
t
f
ea
tu
r
es.
Ma
n
im
u
r
u
g
an
et
a
l.
[
7
]
p
r
o
p
o
s
e
a
DB
N
-
b
ased
I
D
S
an
d
ev
alu
ate
it
u
s
in
g
th
e
C
I
C
I
DS
2
0
1
7
d
ataset.
I
t a
ttain
s
9
4
.
2
8
% a
cc
u
r
ac
y
f
o
r
th
e
n
o
r
m
al
class
,
with
o
th
e
r
at
tack
class
es v
ar
y
in
g
f
r
o
m
9
3
.
5
% to
9
5
.
8
%.
Ab
u
r
asain
[
8
]
p
r
o
p
o
s
e
b
lack
wid
o
w
o
p
tim
izatio
n
(
B
W
O)
b
ased
h
y
b
r
i
d
I
DS
f
o
r
attac
k
er
p
r
ev
en
tio
n
in
I
o
T
-
b
ased
p
r
ec
is
io
n
ag
r
ic
u
ltu
r
e
s
y
s
tem
s
.
T
h
e
d
ev
elo
p
ed
m
o
d
el
u
s
es
th
e
b
ald
ea
g
le
s
ea
r
ch
(
B
E
S)
o
p
tim
izer
f
o
r
f
ea
tu
r
e
s
elec
tio
n
an
d
B
W
O
f
o
r
p
ar
a
m
eter
t
u
n
in
g
to
ac
h
iev
e
h
ig
h
er
s
ec
u
r
ity
a
n
d
r
elia
b
ilit
y
.
A
g
en
e
r
ativ
e
DL
m
o
d
el
-
b
ased
I
DS
is
d
ev
elo
p
ed
b
y
Ab
d
alg
awa
d
et
a
l
.
[
9
]
.
T
h
e
h
y
b
r
id
m
o
d
el
o
f
ad
v
er
s
ar
ial
au
to
en
co
d
er
s
(
AAE
)
an
d
b
id
ir
ec
tio
n
al
GANs
(
B
iGAN)
is
u
s
ed
to
d
etec
t
n
etwo
r
k
attac
k
e
r
s
.
R
esu
lts
o
n
th
e
I
o
T
-
2
3
d
ataset
with
1
.
8
m
illi
o
n
n
etwo
r
k
f
lo
ws
s
h
o
w
th
at
t
h
e
AAE
with
t
h
e
B
iGAN
m
o
d
el
attain
ed
an
F1
-
r
ate
o
f
0
.
8
9
with
an
ac
cu
r
ac
y
o
f
0
.
9
3
.
Nay
a
k
a
n
d
B
h
attac
h
ar
y
y
a
[
1
0
]
p
r
o
p
o
s
e
an
I
DS
f
o
r
I
o
T
n
etwo
r
k
s
u
s
in
g
th
e
en
h
an
ce
d
s
h
u
f
f
le
b
id
ir
ec
tio
n
al
ch
an
n
el
a
tten
tio
n
-
b
ased
n
etwo
r
k
(
E
SB
C
A
-
Net)
.
I
t
u
s
es
r
ef
o
r
m
ed
h
is
to
g
r
am
eq
u
alis
atio
n
an
d
f
ea
tu
r
e
o
p
tim
izatio
n
v
ia
an
im
p
r
o
v
e
d
h
o
n
ey
b
ad
g
er
o
p
tim
is
er
to
ac
h
iev
e
h
ig
h
er
ac
cu
r
ac
y
co
m
p
ar
ed
to
ex
is
tin
g
m
eth
o
d
s
.
Ali
an
d
Yo
u
s
af
[
1
1
]
p
r
o
p
o
s
e
a
th
r
ee
-
tier
I
DS
f
o
r
in
tr
u
s
io
n
d
etec
tio
n
b
a
s
ed
o
n
DL
m
o
d
els.
T
h
e
DL
m
o
d
el,
c
o
m
b
in
e
d
wit
h
ty
p
e
-
I
I
f
u
zz
y
f
ilter
in
g
,
is
u
s
e
d
to
class
if
y
p
ac
k
ets
in
to
n
o
r
m
al,
s
u
s
p
icio
u
s
,
an
d
m
alicio
u
s
.
Fer
n
an
d
o
et
a
l.
[
1
2
]
p
r
o
p
o
s
ed
a
n
in
tr
u
d
er
p
r
ev
e
n
tio
n
m
o
d
el
u
s
in
g
e
n
s
em
b
le
le
ar
n
in
g
.
I
t
co
m
b
in
es
XGBo
o
s
t,
g
r
ad
ien
t
b
o
o
s
tin
g
,
an
d
d
ec
is
io
n
tr
ee
to
class
if
y
m
alicio
u
s
ac
ti
v
ities
.
R
esu
lts
o
n
th
e
B
o
t
-
I
o
T
d
ataset
s
h
o
w
th
at
t
h
e
e
n
s
em
b
le
m
o
d
el
ac
h
iev
es
a
n
ac
c
u
r
ac
y
o
f
u
p
to
9
3
%
f
o
r
b
i
n
ar
y
class
if
icatio
n
an
d
9
2
%
f
o
r
m
u
lticlas
s
ca
teg
o
r
is
atio
n
.
Alju
h
an
i
et
a
l.
[
1
3
]
p
r
o
p
o
s
e
a
p
r
iv
ate
b
l
o
ck
ch
ain
-
b
ased
s
ec
u
r
e
co
m
m
u
n
icatio
n
tech
n
iq
u
e
f
o
r
I
o
T
u
s
in
g
s
ess
io
n
-
b
ased
m
u
tu
al
v
er
if
icatio
n
an
d
k
ey
c
o
n
tr
ac
t.
I
n
ad
d
iti
o
n
,
atten
tio
n
-
b
ased
b
id
ir
ec
tio
n
al
L
STM
(
Ab
i
-
L
S
T
M)
n
etwo
r
k
s
ar
e
ap
p
lied
f
o
r
cy
b
er
-
attac
k
d
etec
tio
n
in
n
etwo
r
k
s
.
Ullah
a
n
d
Ma
h
m
o
u
d
[
1
4
]
p
r
o
p
o
s
ed
an
I
DS
m
o
d
el
u
s
in
g
a
m
u
lti
-
d
im
en
s
io
n
al
C
NN.
T
h
e
C
NN
m
o
d
el
is
f
o
r
m
ed
with
d
if
f
er
en
t
1
D,
2
D,
an
d
3
D
lay
e
r
s
to
lear
n
f
e
atu
r
es
d
ee
p
ly
.
T
h
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
ac
h
iev
es
an
ac
cu
r
ac
y
lev
el
o
f
9
3
.
8
5
%
co
m
p
ar
e
d
to
ex
i
s
tin
g
DL
m
o
d
els.
L
ah
asan
a
n
d
Sam
m
a
[
1
5
]
p
r
o
p
o
s
ed
a
t
wo
-
lay
er
o
p
tim
izer
m
o
d
el
f
o
r
in
tr
u
d
er
d
etec
tio
n
.
I
t
co
n
s
tr
u
cts
a
d
ee
p
au
to
en
co
d
er
m
o
d
el
b
ased
o
n
th
e
d
etec
t
io
n
ac
cu
r
ac
y
o
f
th
e
K
-
n
ea
r
est n
eig
h
b
o
r
(
KNN)
cla
s
s
if
ier
.
A
n
ew
ty
p
e
o
f
DL
m
o
d
el
-
b
ased
attac
k
class
if
icatio
n
m
o
d
el
ca
lled
d
ev
ice
-
b
ased
I
DS
(
DI
DS)
is
p
r
o
p
o
s
ed
b
y
Ma
d
h
u
et
a
l
.
[
1
6
]
.
T
o
im
p
r
o
v
e
th
e
class
if
icatio
n
ac
cu
r
ac
y
,
th
e
p
ar
a
m
e
ter
s
o
f
th
e
D
I
DS
ar
ch
itectu
r
e
ar
e
alter
ed
u
s
in
g
p
ar
ticle
s
war
m
o
p
tim
izatio
n
.
A
r
ec
u
r
r
en
t
n
e
u
r
al
n
etwo
r
k
m
o
d
el
o
f
L
STM
-
b
ased
in
tr
u
d
er
d
etec
tio
n
is
p
r
o
p
o
s
ed
b
y
Kesh
k
et
a
l
.
[
1
7
]
.
T
h
e
im
p
o
r
tan
t
f
ea
tu
r
es
f
r
o
m
th
e
n
et
wo
r
k
a
r
e
id
e
n
tifie
d
u
s
in
g
SP
I
P
(
s
h
ap
le
y
a
d
d
itiv
e
ex
p
lan
atio
n
s
,
p
er
m
u
tatio
n
f
ea
tu
r
e
im
p
o
r
tan
ce
,
in
d
i
v
id
u
al
c
o
n
d
itio
n
al
ex
p
ec
tatio
n
,
p
ar
tial
d
ep
en
d
en
c
e
p
lo
t
)
.
C
o
m
p
a
r
ed
to
ML
m
o
d
els,
th
e
L
STM
m
o
d
el
s
h
o
ws
h
ig
h
er
ac
c
u
r
ac
y
o
n
NSL
-
KDD,
UNS
W
-
NB
1
5
,
an
d
T
o
N
-
I
o
T
d
atasets
.
Ash
ik
u
an
d
Dag
li
[
1
8
]
p
r
esen
t
a
DL
-
b
ased
a
d
ap
tiv
e
I
DS
d
esig
n
ed
to
id
en
tify
b
o
th
k
n
o
wn
a
n
d
ze
r
o
-
d
a
y
n
etwo
r
k
attac
k
s
.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
u
s
es
C
NN
s
with
a
r
eg
u
lar
is
ed
m
u
lti
-
lay
er
p
er
ce
p
tr
o
n
(
ML
P)
to
in
cr
ea
s
e
d
etec
tio
n
f
lex
ib
ilit
y
co
m
p
ar
ed
to
tr
ad
itio
n
al
f
ee
d
-
f
o
r
war
d
n
etwo
r
k
s
.
I
n
in
tr
u
d
e
r
d
etec
tio
n
m
o
d
els,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
1
5
,
No
.
3
,
Sep
tem
b
er
20
2
6
:
1
1
5
4
-
1
1
6
6
1156
d
ee
p
tr
an
s
f
er
lear
n
in
g
is
an
a
d
v
an
ce
d
ML
tech
n
iq
u
e
t
h
at
u
s
es
k
n
o
wled
g
e
f
r
o
m
a
p
r
e
-
tr
a
in
ed
DL
m
o
d
el
to
d
etec
t
d
if
f
er
en
t
attac
k
er
p
atter
n
s
.
Ah
m
ad
et
a
l
.
[
1
9
]
p
r
o
p
o
s
e
an
an
o
m
aly
d
etec
tio
n
ap
p
r
o
ac
h
f
o
r
I
o
T
n
etwo
r
k
s
u
s
in
g
d
ee
p
tr
an
s
f
er
lear
n
in
g
.
T
h
e
E
d
g
eI
I
o
T
d
ataset
is
tr
an
s
f
o
r
m
ed
in
to
a
n
im
ag
e
f
o
r
m
at
f
o
r
C
NN
-
b
ased
p
r
o
ce
s
s
in
g
,
an
d
h
y
p
er
p
a
r
am
et
er
s
ar
e
tu
n
ed
u
s
in
g
R
an
d
o
m
Sear
ch
.
Hiza
l
et
a
l.
[
2
0
]
p
r
o
p
o
s
e
an
e
f
f
ec
tiv
e
I
DS
f
o
r
I
o
T
n
etwo
r
k
s
u
s
in
g
g
at
e
r
ec
u
r
r
e
n
t
u
n
it
(
G
R
U)
-
b
ased
DL
m
o
d
els.
T
h
e
p
r
ep
r
o
ce
s
s
in
g
tech
n
iq
u
es
o
f
f
ea
tu
r
e
s
elec
t
io
n
,
d
u
p
licatio
n
r
e
m
o
v
al,
an
d
n
o
r
m
aliza
tio
n
ar
e
ap
p
lied
to
in
cr
ea
s
e
d
etec
tio
n
a
cc
u
r
ac
y
.
Similar
ly
,
Nan
d
a
n
wa
r
an
d
Kata
r
y
a
[
2
1
]
r
ec
o
m
m
en
d
a
h
y
b
r
i
d
attac
k
er
d
etec
tio
n
m
o
d
el
u
s
in
g
GR
U.
At
f
ir
s
t,
th
e
f
ea
tu
r
es
ar
e
ca
p
tu
r
ed
u
s
in
g
a
C
NN.
Nex
t,
th
e
G
R
U
m
o
d
el
is
ap
p
lied
f
o
r
f
in
al
class
if
icatio
n
.
I
n
th
e
N_
B
aI
o
T
d
ataset,
th
e
C
NN+
GR
U
m
o
d
el
r
ea
ch
es
a
h
ig
h
test
in
g
ac
cu
r
ac
y
o
f
9
4
.
6
2
%
with
a
lo
s
s
o
f
0
.
0
0
6
3
.
Desh
m
u
k
h
an
d
R
av
u
la
k
o
llu
,
[
2
2
]
p
r
o
p
o
s
ed
a
m
eta
h
eu
r
is
tic
o
p
tim
izer
co
m
b
in
ed
d
etec
tio
n
m
o
d
el
ca
l
led
in
tellig
en
t
in
tr
u
s
io
n
d
etec
tio
n
n
etwo
r
k
(
I
I
DNe
t)
.
T
h
e
o
p
tim
izatio
n
s
tr
ateg
y
is
u
s
ed
f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
,
s
elec
tio
n
,
an
d
p
ar
a
m
eter
t
u
n
in
g
to
ac
h
iev
e
h
ig
h
er
attac
k
er
d
etec
tio
n
ac
cu
r
ac
y
.
R
es
u
lts
o
n
th
e
UNSW
-
N
B
1
5
d
ataset
s
h
o
w
th
at
I
I
DNe
t
ac
h
i
ev
es
th
e
h
ig
h
est
ac
cu
r
ac
y
o
f
9
4
.
8
9
%.
Alar
s
an
d
Ku
r
n
az
[
2
3
]
im
p
r
o
v
e
I
DS
p
er
f
o
r
m
a
n
ce
b
y
i
n
teg
r
atin
g
n
et
wo
r
k
an
d
h
o
s
t
tr
af
f
ic
d
ata
w
ith
DL
tech
n
iq
u
es.
Usi
n
g
a
m
ilit
ar
y
-
co
n
tex
t
n
etwo
r
k
in
tr
u
s
io
n
d
etec
tio
n
d
ata
s
et,
th
ey
ap
p
ly
a
C
NN
with
f
ea
tu
r
e
s
elec
tio
n
an
d
d
im
en
s
io
n
ality
r
ed
u
ctio
n
f
o
r
im
p
r
o
v
e
d
ac
cu
r
ac
y
.
An
a
u
to
en
co
d
er
c
o
m
b
in
e
d
with
a
M
L
P
-
b
ased
class
if
ier
m
o
d
el
is
p
r
o
p
o
s
ed
b
y
L
iu
et
a
l
.
[
2
4
]
.
I
t
u
s
es
tr
af
f
ic
f
ea
t
u
r
es
f
o
r
an
o
m
aly
id
e
n
tific
atio
n
,
clu
s
ter
in
g
,
a
n
d
class
if
icatio
n
.
Altu
n
ay
an
d
Al
b
ay
r
ak
[
2
5
]
p
r
o
p
o
s
e
th
r
ee
DL
m
o
d
els:
C
NN,
L
STM
,
an
d
a
h
y
b
r
id
C
NN+
L
STM
to
f
in
d
in
tr
u
s
io
n
s
in
I
o
T
n
etw
o
r
k
s
.
T
h
e
h
y
b
r
id
C
NN+
L
STM
m
o
d
el
ac
h
iev
ed
th
e
h
ig
h
est
ac
cu
r
ac
y
o
f
9
3
.
2
1
%
an
d
9
2
.
9
%
o
n
th
e
UNSW
-
NB
1
5
d
ataset
an
d
9
9
.
8
4
%
an
d
9
9
.
8
0
%
o
n
th
e
KDD
d
ata
s
et
f
o
r
b
in
ar
y
a
n
d
m
u
lti
-
class
clas
s
if
icatio
n
,
r
esp
ec
tiv
ely
.
L
ik
ewise,
th
e
au
th
o
r
s
[
2
6
]
,
[
2
7
]
p
r
o
p
o
s
e
th
e
g
r
ap
h
n
eu
r
al
n
etwo
r
k
s
(
GNN)
an
d
t
r
an
s
f
o
r
m
er
-
b
ased
m
o
d
els f
o
r
attac
k
er
d
etec
tio
n
in
n
etwo
r
k
s
.
3.
P
RO
P
O
SE
D
M
O
D
E
L
T
o
ef
f
ec
tiv
ely
s
o
lv
e
th
e
co
m
p
lex
ities
o
f
cy
b
er
attac
k
s
in
I
o
T
n
etwo
r
k
s
,
a
h
y
b
r
id
an
o
m
aly
-
b
ased
in
tr
u
s
io
n
d
etec
tio
n
m
o
d
el
is
p
r
o
p
o
s
ed
.
I
n
th
e
p
r
o
p
o
s
ed
d
e
tectio
n
ap
p
r
o
ac
h
,
a
th
r
ee
-
s
tag
e
f
ea
tu
r
e
ex
tr
ac
ti
o
n
p
r
o
ce
s
s
is
ca
r
r
ied
o
u
t.
T
h
e
f
e
atu
r
es
ar
e
ex
tr
ac
te
d
u
s
i
n
g
L
DA,
Au
to
en
co
d
er
a
n
d
s
tatis
tical
f
ea
tu
r
e
ex
t
r
ac
tio
n
.
T
h
e
ex
tr
ac
ted
f
ea
tu
r
es
a
r
e
p
r
o
ce
s
s
ed
u
s
in
g
th
e
L
B
E
SN
m
o
d
el
f
o
r
f
in
al
d
etec
tio
n
.
Als
o
,
th
e
L
B
E
SN
m
o
d
el
is
o
p
tim
is
ed
u
s
in
g
B
E
O
to
f
in
e
-
tu
n
e
its
p
ar
am
eter
s
f
o
r
o
p
tim
al
d
etec
tio
n
p
er
f
o
r
m
an
ce
.
T
h
e
wo
r
k
f
lo
w
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
is
s
h
o
wn
in
F
ig
u
r
e
1
.
Fig
u
r
e
1
.
W
o
r
k
f
lo
w
o
f
p
r
o
p
o
s
ed
d
etec
tio
n
a
p
p
r
o
ac
h
3
.
1
.
F
e
a
t
ure
p
re
pro
ce
s
s
ing
T
h
e
in
p
u
t
d
ata
c
o
llected
f
r
o
m
I
o
T
-
b
ased
n
etwo
r
k
s
c
o
n
tain
s
m
is
s
in
g
v
alu
es
an
d
i
n
co
n
s
is
ten
t
f
o
r
m
ats.
Als
o
,
it
in
clu
d
es
ca
teg
o
r
ical
an
d
n
u
m
er
ical
f
ield
s
.
I
n
th
is
s
tag
e,
th
e
d
ata
p
r
ep
r
o
ce
s
s
in
g
is
ca
r
r
ied
o
u
t
to
r
em
o
v
e
u
n
wan
ted
n
o
is
e.
I
n
itially
,
m
is
s
in
g
v
alu
es
ar
e
h
an
d
le
d
u
s
in
g
Simp
le
I
m
p
u
tatio
n
.
I
t
ca
lcu
lates
m
is
s
in
g
v
alu
es
u
s
in
g
th
e
m
ea
n
v
alu
e
o
f
n
u
m
er
ical
attr
ib
u
tes.
T
o
en
s
u
r
e
u
n
if
o
r
m
s
ca
lin
g
ac
r
o
s
s
f
ea
tu
r
es,
Z
-
s
co
r
e
n
o
r
m
alis
atio
n
is
ap
p
lied
.
T
h
i
s
p
r
o
ce
s
s
ce
n
tr
es
th
e
d
ata
ar
o
u
n
d
ze
r
o
an
d
s
ca
les
it
b
ased
o
n
th
e
s
tan
d
ar
d
d
ev
iatio
n
.
T
h
is
n
o
r
m
alis
atio
n
o
f
in
p
u
t d
ata
ca
n
b
e
co
m
p
u
ted
as f
o
llo
ws:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
E
n
h
a
n
ce
d
a
n
o
ma
ly
d
etec
tio
n
i
n
I
o
T n
etw
o
r
ks v
ia
fea
tu
r
e
fu
s
io
n
a
n
d
lea
r
n
in
g
-
b
a
s
ed
…
(
P
.
P
a
lp
a
n
d
i
)
1157
′
=
−
(
1
)
W
h
er
e
x
d
en
o
tes t
h
e
o
r
ig
in
al
i
n
p
u
t,
μ
&
σ
d
en
o
te
th
e
m
ea
n
a
n
d
s
tan
d
ar
d
d
ev
iatio
n
o
f
t
h
e
f
e
atu
r
e.
3
.
2
.
L
DA
L
DA
is
in
tr
o
d
u
ce
d
as
a
s
u
p
er
v
is
ed
lin
ea
r
f
ea
tu
r
e
e
x
tr
ac
tio
n
ap
p
r
o
ac
h
.
T
h
is
tech
n
i
q
u
e
p
r
o
jects
h
ig
h
-
d
im
en
s
io
n
al
in
p
u
t
in
to
a
lo
w
-
d
im
en
s
io
n
al
s
u
b
s
p
ac
e
an
d
p
r
e
s
er
v
es
th
e
class
s
ep
ar
ab
ilit
y
.
Un
lik
e
PC
A,
L
D
A
f
o
cu
s
es
o
n
m
a
x
im
izin
g
v
ar
ia
n
ce
with
o
u
t
class
in
f
o
r
m
at
io
n
.
T
h
is
is
n
ee
d
e
d
f
o
r
cy
b
er
s
ec
u
r
ity
ap
p
licatio
n
s
wh
er
e
d
is
tin
ct
attac
k
class
es
m
u
s
t
b
e
clea
r
ly
s
ep
a
r
ated
.
L
D
A
o
p
er
ates
b
y
ca
lcu
latin
g
t
h
e
with
in
-
class
s
ca
tter
m
atr
ix
an
d
th
e
b
etwe
en
-
class
s
ca
tter
m
atr
ix
:
=
∑
∑
(
−
)
(
−
)
,
∊
=
1
=
∑
∑
(
−
)
(
−
)
,
∊
=
1
(
2
)
T
h
en
,
L
DA
s
o
lv
es th
e
g
e
n
er
al
ized
eig
en
v
alu
e
p
r
o
b
lem
to
f
in
d
th
e
o
p
tim
al
p
r
o
jectio
n
m
atr
i
x
W
:
=
a
r
g
|
|
|
|
(
3
)
T
h
e
p
r
o
jecte
d
f
ea
tu
r
es
r
etain
m
ax
im
u
m
class
d
is
cr
im
in
ati
o
n
a
n
d
in
cr
ea
s
e
t
h
e
d
et
ec
tio
n
ca
p
a
b
ilit
y
o
f
r
ar
e
attac
k
ty
p
es in
I
o
T
n
etwo
r
k
s
.
3
.
3
.
Aut
o
enco
der
-
ba
s
ed
no
nli
nea
r
f
ea
t
ure
lea
rner
An
Au
to
en
c
o
d
er
is
u
s
ed
t
o
e
x
tr
ac
t
n
o
n
lin
ea
r
f
ea
tu
r
e
r
ep
r
e
s
en
tatio
n
s
f
r
o
m
th
e
p
r
ep
r
o
ce
s
s
ed
in
p
u
t
d
ata.
T
h
e
g
o
al
is
to
ca
p
tu
r
e
h
i
d
d
en
p
atter
n
s
,
r
ed
u
n
d
an
cies,
a
n
d
c
o
r
r
elatio
n
s
t
h
at
tr
ad
itio
n
al
s
tatis
tical
o
r
lin
ea
r
m
eth
o
d
s
m
ay
o
v
e
r
lo
o
k
.
I
t
co
n
s
is
t
s
o
f
two
co
m
p
o
n
en
ts
:
th
e
en
co
d
er
an
d
d
ec
o
d
er
.
T
h
e
f
u
n
c
tio
n
o
f
th
e
en
co
d
er
is
to
r
ed
u
ce
th
e
in
p
u
t
f
ea
tu
r
e
v
ec
to
r
in
to
a
lo
wer
-
d
im
e
n
s
io
n
al
laten
t
v
ec
to
r
u
s
in
g
a
n
o
n
lin
ea
r
tr
an
s
f
o
r
m
atio
n
.
T
h
e
m
ain
f
u
n
ctio
n
o
f
a
d
ec
o
d
er
is
to
r
ec
r
ea
te
th
e
o
r
ig
in
al
in
p
u
t
f
r
o
m
th
e
c
o
m
p
r
ess
e
d
f
o
r
m
at.
T
h
e
e
n
co
d
e
r
f
u
n
ctio
n
is
g
i
v
en
b
y
:
ℎ
=
(
)
=
(
+
)
,
̂
=
(
ℎ
)
=
(
ℎ
+
)
,
=
|
|
−
̂
|
|
2
(
4
)
Ov
er
all,
th
e
n
etwo
r
k
is
tr
ain
ed
to
r
ed
u
ce
t
h
e
r
ec
o
n
s
tr
u
ctio
n
er
r
o
r
.
Her
e,
h
is
th
e
late
n
t
f
ea
tu
r
es
ca
p
tu
r
in
g
m
ea
n
in
g
f
u
l
n
o
n
-
lin
ea
r
s
tr
u
ctu
r
es,
σ
is
th
e
ac
tiv
atio
n
f
u
n
ctio
n
(
R
eL
U
o
r
s
ig
m
o
id
)
a
n
d
̂
is
th
e
r
ec
o
n
s
tr
u
cted
o
u
t
p
u
t.
T
h
e
lat
en
t
v
ec
to
r
h
b
ec
o
m
es
a
co
m
p
ac
t
an
d
in
f
o
r
m
ativ
e
f
ea
tu
r
e
s
et
an
d
p
r
eser
v
es
cr
itical
b
eh
av
io
r
c
h
ar
ac
ter
is
tics
s
u
ch
as p
r
o
to
co
l a
n
o
m
alies o
r
ab
n
o
r
m
al
tr
af
f
ic
tim
i
n
g
.
3
.
4
.
St
a
t
is
t
ica
l
f
ea
t
ure
e
x
t
ra
ct
io
n m
o
du
le
T
h
ese
m
o
d
u
le
s
ca
p
tu
r
e
ess
en
tial
s
h
ap
e
an
d
d
is
tr
ib
u
tio
n
c
h
ar
ac
ter
is
tics
o
f
in
p
u
t
d
ata
f
o
r
an
oma
l
y
d
etec
tio
n
.
T
h
e
s
tatis
tical
f
ea
tu
r
es
ex
tr
ac
t
ed
f
r
o
m
ea
ch
in
p
u
t
v
ec
to
r
in
clu
d
e
m
ea
n
(
μ
)
,
s
tan
d
ar
d
d
ev
iatio
n
(
σ
)
,
sk
ewn
ess
(
γ
)
,
an
d
k
u
r
to
s
is
(
κ)
.
T
h
e
m
ea
n
v
alu
e
is
u
s
ed
to
m
e
asu
r
e
a
ce
n
tr
al
ten
d
e
n
cy
.
T
h
e
s
tan
d
ar
d
d
ev
iatio
n
q
u
an
tifie
s
v
ar
iab
ilit
y
.
T
h
e
s
k
ewn
ess
in
d
icate
s
th
e
asy
m
m
etr
y
.
L
ik
ewise,
k
u
r
to
s
i
s
ca
p
tu
r
es
tail
h
ea
v
in
ess
o
r
o
u
tlier
s
.
I
t c
an
b
e
f
r
am
e
d
as f
o
llo
ws:
=
1
∑
=
1
,
2
=
1
(
−
)
2
,
=
1
∑
(
−
)
3
=
1
,
=
1
∑
(
−
)
4
=
1
(
5
)
T
h
ese
s
tatis
tica
l m
etr
ics h
elp
id
en
tify
o
u
tlier
s
,
b
u
r
s
ts
,
an
d
o
t
h
er
an
o
m
alies c
o
m
m
o
n
in
cy
b
er
attac
k
s
ce
n
ar
io
s
.
3
.
5
.
F
e
a
t
ure
f
us
io
n str
a
t
eg
y
I
n
s
tead
o
f
u
s
in
g
a
s
in
g
le
f
ea
tu
r
e
p
r
o
ce
s
s
in
g
m
eth
o
d
,
th
e
f
ea
tu
r
e
v
ec
t
o
r
s
f
r
o
m
L
DA,
au
t
o
en
co
d
er
,
an
d
s
tatis
t
ical
an
aly
s
is
ar
e
co
m
b
in
ed
to
in
cr
ea
s
e
d
etec
tio
n
ac
cu
r
a
cy
:
=
[
|
|
|
|
]
(
6
)
T
h
is
h
y
b
r
id
f
ea
tu
r
e
v
ec
to
r
u
s
e
s
th
e
d
is
cr
im
in
ativ
e
p
o
wer
o
f
L
DA,
th
e
n
o
n
lin
ea
r
r
ep
r
esen
tatio
n
f
r
o
m
th
e
au
to
en
c
o
d
er
,
an
d
th
e
b
e
h
av
io
r
al
in
s
ig
h
ts
f
r
o
m
s
tatis
tics
.
T
h
e
f
u
s
ed
v
ec
to
r
f
o
r
m
s
th
e
f
in
al
i
n
p
u
t
t
o
L
I
B
E
S
N.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
1
5
,
No
.
3
,
Sep
tem
b
er
20
2
6
:
1
1
5
4
-
1
1
6
6
1158
3
.
6
.
L
B
E
SN
c
la
s
s
if
ier
T
h
e
L
B
E
SN
m
o
d
el
is
a
p
o
wer
f
u
l
tim
e
-
s
er
ies
class
if
ier
b
ased
o
n
th
e
ec
h
o
s
tate
n
etwo
r
k
(
E
SN)
ar
ch
itectu
r
e.
E
SNs
ar
e
a
ty
p
e
o
f
r
eser
v
o
ir
co
m
p
u
tin
g
m
o
d
el
th
at
is
h
i
g
h
ly
ef
f
icien
t
f
o
r
s
e
q
u
en
tial
d
ata
d
u
e
to
th
eir
f
ast
tr
ain
i
n
g
an
d
d
y
n
a
m
ic
m
em
o
r
y
ca
p
ab
ilit
y
.
T
h
e
ar
ch
itectu
r
e
o
f
th
e
L
B
E
SN
m
o
d
el
is
s
h
o
wn
in
Fig
u
r
e
2
.
T
h
e
L
B
E
SN
co
n
tain
s
an
in
p
u
t
lay
er
t
h
at
m
a
p
s
f
ea
tu
r
e
v
ec
to
r
x
(
t)
to
t
h
e
r
eser
v
o
ir
,
an
d
th
e
r
eser
v
o
i
r
tr
an
s
f
o
r
m
s
th
e
in
p
u
t in
to
h
ig
h
-
d
im
en
s
io
n
al
d
y
n
am
ics.
I
t c
an
b
e
s
tated
as f
o
llo
ws:
(
+
1
)
=
(
1
−
)
.
(
)
+
.
ta
n
h
(
(
)
+
(
)
)
(
7
)
W
h
er
e,
r
(
t)
is
th
e
r
eser
v
o
ir
s
t
ate
at
tim
e
t,
is
th
e
in
p
u
t
w
eig
h
t
m
atr
ix
,
is
th
e
in
ter
n
al
co
n
n
ec
tio
n
weig
h
ts
,
α
is
th
e
leak
r
ate
(
c
o
n
tr
o
ls
tem
p
o
r
al
m
em
o
r
y
)
.
Un
lik
e
s
tan
d
ar
d
E
SNs
,
L
B
E
SN
in
co
r
p
o
r
ates
b
id
ir
ec
tio
n
al
r
eser
v
o
ir
s
an
d
in
tr
o
d
u
ce
s
tr
ain
ab
le
o
u
tp
u
t
lay
er
s
v
ia
s
u
p
er
v
is
ed
lear
n
in
g
.
I
t
is
u
s
ed
f
o
r
th
e
n
etwo
r
k
to
b
etter
ca
p
tu
r
e
d
ep
e
n
d
en
cies
f
r
o
m
b
o
th
p
ast
an
d
f
u
tu
r
e
co
n
tex
ts
in
a
s
eq
u
en
ce
.
L
B
E
SN
m
ain
tain
s
two
s
ep
ar
ate
s
tates
:
f
o
r
war
d
r
eser
v
o
ir
(
)
+
an
d
r
ev
e
r
s
e
r
eser
v
o
ir
(
)
,
b
o
th
ar
e
u
p
d
ated
u
s
in
g
th
e
s
am
e
r
eser
v
o
ir
d
y
n
am
i
cs
b
u
t
p
o
ten
tially
f
r
o
m
o
p
p
o
s
ite
en
d
s
o
f
th
e
s
eq
u
en
ce
.
(
+
1
)
=
(
1
−
)
.
(
)
+
.
ta
n
h
(
(
)
+
(
)
)
,
(
+
1
)
=
(
1
−
)
.
(
)
+
.
ta
n
h
(
(
)
+
(
)
)
(
8
)
At
th
e
en
d
o
f
th
e
s
eq
u
e
n
ce
,
b
o
th
s
tates a
r
e
co
n
ca
ten
ated
to
f
o
r
m
th
e
f
in
al
r
ep
r
esen
tatio
n
:
=
[
(
)
;
(
)
]
(
9
)
W
h
er
e
T
is
th
e
f
in
al
tim
e
s
te
p
.
T
h
e
o
u
tp
u
t
lay
er
is
a
lin
ea
r
class
if
ier
tr
ain
ed
u
s
in
g
r
id
g
e
r
eg
r
ess
io
n
o
n
th
e
r
eser
v
o
ir
s
tates:
=
min
|
|
−
|
|
2
+
|
|
|
|
2
(1
0
)
Fig
u
r
e
2
.
L
B
E
SN a
r
ch
itectu
r
e
3
.
7
.
B
la
c
k
ea
g
le
o
ptim
izer
(
B
E
O
)
T
h
e
B
E
O
is
a
b
io
-
in
s
p
ir
ed
m
e
tah
eu
r
is
tic
alg
o
r
ith
m
d
ev
elo
p
e
d
b
y
s
im
u
latin
g
th
e
h
u
n
tin
g
b
eh
av
io
r
o
f
b
lack
ea
g
les
[
2
8
]
.
I
t
ca
n
b
e
a
p
p
lied
to
co
m
p
lex
o
p
tim
izatio
n
p
r
o
b
lem
s
ch
ar
ac
ter
is
ed
b
y
n
o
n
-
lin
ea
r
ity
,
m
u
lti
-
m
o
d
ality
,
a
n
d
la
r
g
e
s
ea
r
c
h
s
p
ac
es.
T
h
e
in
s
p
ir
atio
n
co
m
es
f
r
o
m
th
e
way
b
lack
ea
g
les
co
o
p
er
ate,
s
talk
,
an
d
ca
p
tu
r
e
th
eir
p
r
ey
.
T
h
is
b
eh
av
io
r
is
em
u
lated
in
th
e
alg
o
r
ith
m
th
r
o
u
g
h
a
m
u
lti
-
p
h
ase
ap
p
r
o
ac
h
co
m
b
in
in
g
ex
p
lo
r
atio
n
(
s
ea
r
c
h
in
g
n
ew
r
eg
io
n
s
)
an
d
e
x
p
lo
itatio
n
(
r
ef
in
in
g
g
o
o
d
s
o
lu
tio
n
s
)
.
T
h
ese
p
h
ases
in
clu
d
e
Stalk
in
g
,
Attack
in
g
,
C
ap
tu
r
i
n
g
an
d
R
etu
r
n
in
g
/Ho
m
ec
o
m
i
n
g
.
I
n
itia
liz
a
tio
n
p
h
a
s
e
:
W
e
d
ef
in
e
an
o
p
tim
izatio
n
p
r
o
b
lem
as:
Min
im
ize
o
r
Ma
x
im
e
f
(
X)
=f
(
1
,
2
…
.
.
)
(1
1
)
W
h
er
e,
X
∈
ℝ
D
is
th
e
d
ec
is
io
n
v
ec
to
r
,
D
is
th
e
d
im
e
n
s
io
n
ality
,
x
j
∈
[
L
b
j
,
Ub
j
]
.
E
ac
h
v
ar
iab
le
lies
with
in
d
ef
in
ed
b
o
u
n
d
s
.
T
h
e
in
itial p
o
p
u
latio
n
(
s
et
o
f
b
lack
ea
g
les)
is
g
en
er
a
ted
r
an
d
o
m
ly
with
in
th
e
p
r
e
d
e
f
in
ed
s
ea
r
ch
s
p
ac
e:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
E
n
h
a
n
ce
d
a
n
o
ma
ly
d
etec
tio
n
i
n
I
o
T n
etw
o
r
ks v
ia
fea
tu
r
e
fu
s
io
n
a
n
d
lea
r
n
in
g
-
b
a
s
ed
…
(
P
.
P
a
lp
a
n
d
i
)
1159
,
=
+
(
−
)
(1
2
)
W
h
er
e,
∼
U(
0
,
1
)
,
,
is
th
e
p
o
s
iti
o
n
o
f
ea
g
le
i
in
d
im
en
s
io
n
j.
S
ta
lkin
g
I
n
n
atu
r
e,
b
lack
ea
g
les
f
ly
at
h
ig
h
altitu
d
es
an
d
s
ca
n
a
wi
d
e
ar
ea
f
o
r
p
o
ten
tial
p
r
ey
.
I
n
B
E
O,
th
is
tr
an
s
lates
to
a
g
lo
b
al
s
ea
r
ch
m
ec
h
an
is
m
wh
er
e
ea
ch
ea
g
le
u
p
d
ates
its
p
o
s
itio
n
b
ased
o
n
:
t
h
e
b
est
-
k
n
o
wn
s
o
lu
tio
n
(
g
lo
b
al
b
est),
r
a
n
d
o
m
in
f
lu
en
ce
to
m
ain
tain
d
iv
er
s
ity
.
I
t c
an
b
e
m
ath
em
atica
lly
s
ta
ted
as f
o
llo
ws:
,
(
+
1
)
=
,
(
)
+
1
.
1
.
(
,
(
)
−
,
(
)
)
+
2
.
2
.
(
−
)
.
(
2
.
−
1
)
(
1
3
)
W
h
er
e,
,
(
)
:
b
est
ea
g
le
in
d
im
en
s
io
n
j,
1
an
d
r
2
is
a
r
an
d
o
m
v
a
r
iab
le
wh
ich
v
ar
ies
f
r
o
m
ze
r
o
to
o
n
e.
1
is
th
e
c
o
n
tr
o
l
’
s
s
o
cial
attr
ac
tio
n
,
2
is
th
e
c
o
n
tr
o
l
’
s
r
a
n
d
o
m
ex
p
lo
r
ato
r
y
m
o
tio
n
.
T
h
is
m
o
v
e
m
en
t
h
elp
s
ea
g
les
co
n
v
er
g
e
to
g
o
o
d
ar
ea
s
wh
ile
s
till
s
ea
r
ch
in
g
elsewh
er
e.
A
tta
ck
in
g
On
ce
th
e
p
r
ey
is
s
p
o
tted
,
th
e
ea
g
le
b
eg
in
s
to
s
p
ir
al
d
o
wn
war
d
in
a
f
o
cu
s
ed
b
u
t
s
lig
h
tly
u
n
p
r
ed
ictab
le
p
ath
.
I
n
o
p
tim
izatio
n
ter
m
s
,
t
h
is
m
ea
n
s
th
e
alg
o
r
ith
m
in
te
n
s
if
ies
its
lo
ca
l
s
ea
r
ch
ar
o
u
n
d
th
e
b
est
s
o
lu
tio
n
s
u
s
in
g
an
o
s
cillato
r
y
m
o
tio
n
.
I
t
im
p
r
o
v
es
th
e
ch
an
ce
s
o
f
f
in
d
in
g
b
etter
o
p
tim
a
in
th
at
r
eg
io
n
.
I
t
ca
n
b
e
m
at
h
em
atica
lly
s
tated
as f
o
llo
ws:
,
(
+
1
)
=
(
)
+
.
s
in
(
.
+
)
.
(
,
(
)
−
,
(
)
(
1
4
)
W
h
er
e,
A
is
th
e
s
p
ir
al
am
p
litu
d
e,
ω
is
th
e
f
r
e
q
u
en
c
y
(
co
n
tr
o
ls
o
s
cillatio
n
)
,
ϕ
is
th
e
p
h
ase
o
f
f
s
et
an
d
,
(
)
is
th
e
m
ea
n
lo
ca
tio
n
o
f
th
e
s
war
m
.
T
h
is
r
ep
r
esen
ts
n
o
n
-
lin
ea
r
an
d
o
s
cillato
r
y
m
o
tio
n
.
I
t
in
cr
ea
s
es
lo
ca
l sear
ch
ar
o
u
n
d
th
e
b
est ca
n
d
id
ate.
C
a
p
tu
r
in
g
I
n
n
atu
r
e,
th
e
ea
g
le
f
in
ally
m
a
k
es
a
s
h
ar
p
an
d
p
r
ec
is
e
m
o
v
e
m
en
t
to
g
r
ab
its
p
r
ey
.
T
h
is
p
h
ase
m
o
d
els
g
r
ee
d
y
s
ea
r
ch
,
wh
ich
s
u
p
p
o
r
ts
r
ap
id
c
o
n
v
e
r
g
en
ce
to
war
d
th
e
o
p
tim
al
s
o
lu
tio
n
.
I
t
ca
n
b
e
m
ath
em
atica
lly
ex
p
r
ess
ed
as f
o
llo
ws:
,
(
+
1
)
=
,
(
)
+
∊
.
(
,
(
)
−
,
(
)
(
1
5
)
W
h
er
e,
ϵ
∼
U(
−1
,
1
is
a
s
m
all
p
er
tu
r
b
atio
n
)
.
T
h
is
h
elp
s
f
in
e
-
t
u
n
e
s
o
lu
tio
n
s
n
ea
r
th
e
o
p
tim
u
m
an
d
en
s
u
r
es
q
u
ick
co
n
v
er
g
en
ce
.
R
etu
r
n
in
g
/Ho
mec
o
min
g
Af
ter
a
s
u
cc
ess
f
u
l
h
u
n
t,
ea
g
les
r
etu
r
n
t
o
th
eir
n
ests
.
I
n
B
E
O,
th
is
co
r
r
esp
o
n
d
s
to
c
h
ec
k
in
g
a
n
d
co
r
r
ec
tin
g
an
y
p
o
s
itio
n
s
th
at
h
av
e
g
o
n
e
o
u
t
o
f
b
o
u
n
d
s
.
T
h
i
s
en
s
u
r
es
th
at
all
s
o
lu
tio
n
s
r
e
m
ain
f
ea
s
ib
l
e
with
i
n
th
e
s
ea
r
ch
s
p
ac
e.
I
t c
a
n
b
e
s
tated
as f
o
llo
ws:
If
,
(
+
1
)
∊
[
,
]
,
a
p
p
ly,
,
(
+
1
)
=
{
,
(
+
1
)
<
,
(
+
1
)
>
(
1
6
)
Alter
n
ativ
ely
,
r
ef
lect
o
r
r
an
d
o
m
ly
r
ein
itialize
if
d
esire
d
:
,
(
+
1
)
=
+
.
(
−
)
(
1
7
)
T
h
e
B
E
O
ef
f
ec
tiv
ely
b
alan
ce
s
ex
p
lo
r
atio
n
a
n
d
ex
p
lo
itatio
n
ac
r
o
s
s
its
p
h
ases
.
B
y
m
im
ick
in
g
th
e
b
io
lo
g
ical
in
tellig
en
ce
o
f
b
lac
k
ea
g
les,
th
e
alg
o
r
ith
m
ca
n
n
av
i
g
ate
co
m
p
lex
lan
d
s
ca
p
es a
n
d
f
in
d
p
r
o
m
is
in
g
ar
ea
s
.
3
.
8
.
P
a
ra
m
et
er
t
un
ing
o
f
E
SN
m
o
del us
ing
BEO
T
h
e
B
E
O
is
u
s
ed
to
o
p
tim
ally
tu
n
e
th
e
h
y
p
er
p
ar
am
eter
s
o
f
th
e
L
B
E
SN.
T
h
ese
h
y
p
e
r
p
ar
am
eter
s
s
ig
n
if
ican
tly
im
p
ac
t
t
h
e
E
S
N
’
s
ab
ilit
y
to
m
o
d
el
tem
p
o
r
al
d
ep
en
d
en
cies
an
d
g
en
e
r
al
is
e
to
u
n
s
ee
n
d
ata.
T
h
e
o
p
tim
is
atio
n
p
r
o
b
lem
f
o
r
E
SN tu
n
in
g
is
as f
o
llo
ws:
min
(
(
)
)
=
1
−
(
1
8
)
W
h
er
e,
θ=
[
α
,
N
r
,
η
]
is
th
e
v
ec
to
r
o
f
E
SN
h
y
p
er
p
ar
a
m
eter
s
:
α
is
th
e
L
ea
k
R
ate
(
tem
p
o
r
al
m
e
m
o
r
y
f
ac
t
o
r
)
,
Nr
is
th
e
R
eser
v
o
ir
Size
(
n
u
m
b
er
o
f
in
ter
n
al
n
eu
r
o
n
s
)
,
a
n
d
η
is
t
h
e
L
ea
r
n
in
g
R
ate
(
o
u
tp
u
t
weig
h
t
ad
ju
s
tm
en
t
r
ate)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
1
5
,
No
.
3
,
Sep
tem
b
er
20
2
6
:
1
1
5
4
-
1
1
6
6
1160
E
ac
h
ea
g
le
in
th
e
B
E
O
al
g
o
r
i
th
m
r
e
p
r
esen
ts
a
ca
n
d
id
ate
θ
an
d
is
e
v
alu
ated
b
ased
o
n
its
ab
ilit
y
to
m
in
im
is
e
th
e
class
if
icatio
n
er
r
o
r
(
1
-
ac
c
u
r
ac
y
)
o
n
th
e
v
alid
atio
n
d
ataset.
T
h
e
B
E
O
is
u
s
ed
to
tu
n
e
L
B
E
SN
p
a
r
am
eter
s
b
y
in
itializin
g
a
p
o
p
u
latio
n
o
f
ca
n
d
i
d
ate
s
o
lu
tio
n
s
(
ea
g
les),
ea
ch
with
r
an
d
o
m
v
alu
es
f
o
r
leak
r
ate,
r
eser
v
o
ir
s
ize,
an
d
lear
n
in
g
r
ate.
E
ac
h
ea
g
le
’
s
f
itn
ess
is
ev
alu
ated
b
y
tr
ain
in
g
an
L
B
E
SN
m
o
d
el.
T
h
e
er
r
o
r
v
alu
e
is
ca
lcu
lated
o
n
v
alid
atio
n
d
ata.
T
h
e
b
est
-
p
e
r
f
o
r
m
in
g
ea
g
le
g
u
id
es
th
e
s
ea
r
ch
p
r
o
ce
s
s
th
r
o
u
g
h
f
o
u
r
p
h
ases
:
s
talk
in
g
(
ex
p
lo
r
atio
n
)
,
attac
k
i
n
g
(
lo
ca
l
s
ea
r
ch
)
,
ca
p
tu
r
in
g
(
g
r
ee
d
y
r
ef
i
n
em
en
t)
,
an
d
h
o
m
ec
o
m
in
g
(
b
o
u
n
d
ar
y
co
r
r
ec
tio
n
)
.
I
n
ea
c
h
iter
atio
n
,
ea
g
le
p
o
s
itio
n
s
ar
e
u
p
d
ated
an
d
f
itn
ess
r
e
-
e
v
alu
at
ed
to
id
en
tify
b
etter
s
o
lu
tio
n
s
.
Af
ter
s
ev
er
al
iter
atio
n
s
,
th
e
b
est
p
ar
am
eter
s
et
f
o
u
n
d
is
r
etu
r
n
e
d
.
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
e
L
B
E
SN
m
o
d
el
is
co
d
e
d
in
Py
th
o
n
an
d
v
alid
a
te
d
u
s
in
g
th
e
Go
o
g
le
Co
lab
en
v
ir
o
n
m
e
n
t
.
T
h
e
p
er
f
o
r
m
an
ce
o
f
th
e
m
o
d
el
is
as
s
es
s
ed
in
th
r
ee
d
if
f
er
en
t
d
atasets
o
f
KDDCu
p
99
,
UNSW
-
N
B
1
5
,
an
d
I
n
SDN
.
T
h
e
en
tire
d
ataset
is
s
ep
ar
ated
in
to
a
7
0
%
tr
ain
in
g
an
d
3
0
%
test
in
g
s
et
.
T
h
e
KDDCu
p
9
9
d
ataset
co
n
s
is
ts
o
f
2
2
d
is
tin
ct
attac
k
ty
p
es.
T
h
e
UNSW
-
NB
1
5
d
ataset
co
n
s
is
ts
o
f
1
0
d
if
f
e
r
en
t
attac
k
er
ca
teg
o
r
ies.
Du
r
in
g
p
r
ep
r
o
ce
s
s
in
g
,
t
h
e
s
y
n
th
etic
m
in
o
r
ity
o
v
er
-
s
am
p
lin
g
tech
n
iq
u
e
(
SMOT
E
)
is
ap
p
lied
to
s
o
lv
e
th
e
class
im
b
alan
ce
is
s
u
es.
T
h
e
p
er
f
o
r
m
an
ce
is
an
aly
s
ed
u
s
in
g
f
o
u
r
s
tan
d
ar
d
m
etr
ics:
p
r
ec
is
io
n
(
P),
r
ec
all
(
R
)
,
F1
-
s
co
r
e
(
F1
)
,
an
d
a
cc
u
r
ac
y
(
A)
.
T
o
o
p
tim
ize
th
e
p
er
f
o
r
m
an
ce
o
f
L
B
E
SN,
th
e
B
E
O
is
u
s
ed
with
1
0
0
iter
atio
n
s
an
d
a
p
o
p
u
latio
n
s
ize
o
f
2
0
ea
g
les.
T
h
e
leak
r
ate
(
α
)
v
ar
ies
f
r
o
m
0
.
1
to
1
.
0
.
T
h
e
r
eser
v
o
ir
s
ize
v
ar
ies
f
r
o
m
1
0
to
2
0
0
n
eu
r
o
n
s
.
T
h
e
lear
n
in
g
r
ate
is
s
et
to
0
.
0
1
.
T
h
e
o
p
tim
ized
v
alu
es a
r
e
g
iv
en
in
T
ab
le
1
.
T
h
e
f
itn
ess
cu
r
v
e
o
f
B
E
O
is
g
iv
en
in
Fig
u
r
e
3
.
T
h
e
f
it
n
ess
cu
r
v
e
s
h
o
ws
th
at
th
e
B
E
O
is
ef
f
ec
tiv
e
in
r
ed
u
cin
g
t
h
e
lo
s
s
o
f
th
e
L
B
E
SN
m
o
d
el
f
o
r
attac
k
er
d
etec
tio
n
.
I
t
s
h
o
ws
an
in
itial
ex
p
lo
r
at
o
r
y
p
h
ase,
f
o
llo
wed
b
y
a
clea
r
co
n
v
er
g
e
n
ce
to
war
d
s
a
lo
wer
lo
s
s
v
alu
e.
I
t
is
o
b
s
er
v
ed
th
at
th
e
B
E
O
s
u
cc
ess
f
u
lly
o
p
tim
ized
th
e
L
B
E
SN
p
ar
am
eter
s
to
in
cr
ea
s
e
its
p
er
f
o
r
m
an
ce
in
d
etec
tin
g
attac
k
er
s
.
Stab
ilis
atio
n
at
th
e
en
d
s
u
g
g
ests
th
at
th
e
alg
o
r
ith
m
h
as
f
o
u
n
d
a
r
o
b
u
s
t
s
et
o
f
p
ar
am
eter
s
o
r
r
ea
ch
ed
th
e
lim
its
o
f
f
u
r
th
er
s
ig
n
i
f
ican
t
im
p
r
o
v
em
en
t
with
in
th
e
g
iv
en
iter
atio
n
s
.
Fig
u
r
e
3
.
B
E
O
f
itn
ess
cu
r
v
e
T
ab
le
1
.
Ad
ju
s
ted
v
alu
es o
f
t
h
e
L
B
E
SN m
o
d
el
P
a
r
a
me
t
e
r
D
e
scri
p
t
i
o
n
M
i
n
v
a
l
u
e
M
a
x
v
a
l
u
e
O
p
t
i
mi
z
e
d
v
a
l
u
e
Le
a
k
r
a
t
e
(
α)
C
o
n
t
r
o
l
s
mem
o
r
y
d
e
c
a
y
i
n
t
h
e
r
e
s
e
r
v
o
i
r
0
.
1
1
.
0
0
.
5
2
R
e
ser
v
o
i
r
s
i
z
e
N
u
mb
e
r
o
f
n
e
u
r
o
n
s i
n
t
h
e
r
e
serv
o
i
r
l
a
y
e
r
10
2
0
0
1
3
3
Le
a
r
n
i
n
g
r
a
t
e
Le
a
r
n
i
n
g
r
a
t
e
f
o
r
t
r
a
i
n
i
n
g
t
h
e
r
e
a
d
o
u
t
l
a
y
e
r
0
.
0
0
0
1
0
.
0
1
0
.
0
0
5
3
T
h
e
tr
ain
in
g
a
n
d
test
in
g
cu
r
v
es
o
f
ac
cu
r
ac
y
a
n
d
lo
s
s
f
o
r
t
h
r
ee
d
atasets
ar
e
g
iv
en
in
Fig
u
r
e
4
.
T
h
e
ac
cu
r
ac
y
c
u
r
v
es
s
h
o
w
a
s
m
o
o
th
an
d
co
n
s
is
ten
t
r
is
e
in
p
er
f
o
r
m
an
ce
with
m
in
im
al
d
ev
iatio
n
b
etwe
en
tr
ain
in
g
an
d
test
s
ets.
I
t
d
en
o
tes
th
at
s
tr
o
n
g
g
e
n
er
aliza
tio
n
with
o
u
t
o
v
er
f
itti
n
g
.
T
h
e
m
o
d
els
ac
h
i
ev
e
h
ig
h
ac
cu
r
ac
y
v
alu
es
with
in
r
elativ
ely
f
ew
e
p
o
ch
s
.
T
h
e
lo
s
s
cu
r
v
es,
in
p
ar
allel,
s
h
o
w
a
s
teep
d
ec
lin
e
in
t
h
e
ea
r
ly
ep
o
ch
s
an
d
f
latten
o
u
t
p
r
o
g
r
ess
iv
ely
.
T
h
e
co
n
f
u
s
io
n
m
atr
ices
p
r
o
v
e
th
e
m
o
d
el
’
s
s
tr
o
n
g
class
if
icatio
n
p
er
f
o
r
m
an
ce
ac
r
o
s
s
all
attac
k
ca
teg
o
r
ies,
as
s
h
o
wn
in
Fig
u
r
e
5
.
Fo
r
KDD9
9
,
t
h
e
m
o
d
el
attain
e
d
n
ea
r
-
p
er
f
e
ct
ac
cu
r
ac
y
ac
r
o
s
s
m
u
ltip
le
class
es
lik
e
n
ep
t
u
n
e
,
s
m
u
r
f
,
a
n
d
n
o
r
m
al
,
with
m
in
i
m
al
m
is
class
if
icatio
n
s
in
m
in
o
r
ity
class
es
s
u
ch
as
Per
l
an
d
phf
an
d
s
h
o
ws
a
r
o
b
u
s
t
g
en
er
aliza
tio
n
ca
p
ab
ilit
y
e
v
en
with
im
b
alan
ce
d
d
ata.
I
n
UNSW
-
N
B
1
5
,
th
e
m
o
d
el
ac
h
iev
es
h
ig
h
p
r
ec
is
io
n
in
class
if
y
in
g
m
ajo
r
attac
k
ty
p
es
s
u
ch
as
Gen
er
ic
,
Do
S
,
a
n
d
E
x
p
lo
its
.
Als
o
,
it
ac
h
iev
es
g
o
o
d
d
etec
tio
n
p
er
f
o
r
m
an
ce
f
o
r
less
f
r
eq
u
en
t
at
tack
s
lik
e
Sh
ellco
d
e
an
d
W
o
r
m
s
.
Fo
r
I
n
SDN
,
th
e
m
o
d
el
ef
f
ec
tiv
el
y
id
en
tifie
d
lar
g
e
-
s
ca
le
DDo
S
,
Pro
b
e
,
a
n
d
N
o
r
m
al
tr
af
f
ic
p
atter
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
E
n
h
a
n
ce
d
a
n
o
ma
ly
d
etec
tio
n
i
n
I
o
T n
etw
o
r
ks v
ia
fea
tu
r
e
fu
s
io
n
a
n
d
lea
r
n
in
g
-
b
a
s
ed
…
(
P
.
P
a
lp
a
n
d
i
)
1161
Fig
u
r
e
4
.
Mo
d
el
tr
ain
in
g
an
d
v
alid
atio
n
cu
r
v
e
f
o
r
th
r
ee
d
if
f
er
en
t d
atasets
Fig
u
r
e
5
.
C
o
n
f
u
s
io
n
m
atr
i
x
v
i
s
u
alis
atio
n
o
f
th
r
ee
d
if
f
er
en
t
d
ata
s
ets
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
1
5
,
No
.
3
,
Sep
tem
b
er
20
2
6
:
1
1
5
4
-
1
1
6
6
1162
T
h
e
p
er
f
o
r
m
a
n
ce
m
etr
ics
v
al
u
es
o
f
th
e
m
o
d
el
f
o
r
th
r
ee
d
if
f
er
en
t
d
atasets
ar
e
g
i
v
en
in
Fig
u
r
e
s
6
-
8
.
Fo
r
th
e
UNSW
-
NB
1
5
d
atase
t
,
th
e
m
o
d
el
r
ea
ch
es
an
o
v
er
a
ll
ac
cu
r
ac
y
o
f
9
6
.
0
%
with
a
h
ig
h
p
r
ec
is
io
n
an
d
r
ec
all
r
ate
ac
r
o
s
s
m
o
s
t
attac
k
ca
teg
o
r
ies.
I
t
ac
h
iev
es
F1
-
s
co
r
es
ab
o
v
e
0
.
9
7
f
o
r
th
e
attac
k
t
y
p
es
s
u
ch
as
B
FA
,
B
OT
NE
T
,
DDo
S
,
an
d
Do
S
.
I
t
d
en
o
tes
th
e
ac
c
u
r
ate
d
etec
tio
n
with
m
in
im
al
f
alse
p
o
s
itiv
es
an
d
n
eg
ativ
es
.
I
n
th
e
KDD9
9
d
ataset
,
th
e
m
o
d
el
ac
h
iev
ed
a
n
o
u
ts
tan
d
i
n
g
p
er
f
o
r
m
an
ce
with
an
o
v
er
all
ac
cu
r
ac
y
o
f
9
6
.
8
%
.
I
t
ac
h
iev
es
F1
-
s
co
r
es
a
b
o
v
e
0
.
9
6
f
o
r
th
e
attac
k
ty
p
es
s
u
ch
as
s
m
u
r
f
,
n
e
p
tu
n
e
,
p
o
r
t
s
wee
p
,
an
d
n
m
ap
.
I
t
p
r
o
v
es
th
e
m
o
d
el
’
s
ab
ilit
y
to
h
an
d
le
b
o
th
f
r
eq
u
e
n
t
an
d
in
f
r
eq
u
e
n
t
in
tr
u
s
io
n
ty
p
es
ef
f
ec
tiv
ely
.
T
h
is
s
tr
o
n
g
p
er
f
o
r
m
an
ce
ac
r
o
s
s
2
1
class
es
in
d
icate
s
th
e
m
o
d
el
’
s
r
o
b
u
s
tn
ess
in
h
ig
h
-
d
im
e
n
s
io
n
al
m
u
lti
class
en
v
ir
o
n
m
en
ts
.
Fo
r
th
e
co
m
p
lex
I
n
SDN
d
ataset
,
th
e
m
o
d
el
ac
h
iev
e
d
an
o
v
e
r
all
ac
cu
r
ac
y
o
f
9
4
.
1
%
.
I
t
s
h
o
ws
F1
-
s
co
r
es
ab
o
v
e
0
.
8
6
f
o
r
ca
teg
o
r
ies
lik
e
b
ac
k
d
o
o
r
an
d
wo
r
m
s
.
L
ik
ewise,
it
ac
h
iev
es
F1
-
s
co
r
es
ar
o
u
n
d
0
.
8
3
–
0
.
8
7
.
Ov
er
all,
th
e
B
E
O
-
tu
n
ed
L
B
E
SN
co
n
s
tan
tly
p
er
f
o
r
m
ed
well
ac
r
o
s
s
d
i
f
f
er
en
t
d
atasets
.
T
h
e
L
B
E
SN
m
o
d
el
ac
cu
r
ately
d
etec
ts
m
o
s
t k
n
o
wn
attac
k
t
y
p
es with
h
ig
h
r
eliab
il
ity
.
Fig
u
r
e
6
.
Per
f
o
r
m
an
c
e
f
o
r
UNSW
-
N
B
1
5
d
ataset
Fig
u
r
e
7
.
Per
f
o
r
m
an
c
e
f
o
r
th
e
KDD9
9
d
ataset
Fig
u
r
e
8
.
Per
f
o
r
m
an
c
e
f
o
r
I
n
S
DN
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
E
n
h
a
n
ce
d
a
n
o
ma
ly
d
etec
tio
n
i
n
I
o
T n
etw
o
r
ks v
ia
fea
tu
r
e
fu
s
io
n
a
n
d
lea
r
n
in
g
-
b
a
s
ed
…
(
P
.
P
a
lp
a
n
d
i
)
1163
T
h
e
ev
alu
atio
n
o
f
th
e
B
E
O
-
L
B
E
SN
wi
th
ex
is
tin
g
ar
ch
itectu
r
es
is
g
iv
en
in
T
ab
le
2
.
T
h
e
BEO
-
LBE
S
m
eth
o
d
ac
h
iev
es
th
e
h
ig
h
est
ac
cu
r
ac
y
o
f
9
5
.
6
%
.
T
h
e
e
x
is
tin
g
m
o
d
els
lik
e
AAE
+
B
iG
AN
an
d
AB
iLST
M
s
h
o
w
ac
cu
r
ac
ies
o
f
9
4
.
1
%
an
d
9
4
.
2
%
,
r
esp
ec
tiv
ely
.
T
h
e
C
NN
+
GR
U
an
d
L
STM
s
h
o
w
th
e
ac
cu
r
ac
ies
o
f
9
4
.
5
%
an
d
9
2
.
8
%
,
r
esp
ec
tiv
el
y
.
T
h
e
m
o
d
els
DB
N
an
d
MLP
s
h
o
w
lo
wer
ac
cu
r
ac
ies
ar
o
u
n
d
9
1
.
0
%
to
9
1
.
8
%
.
Ov
er
all,
th
e
p
r
o
p
o
s
ed
B
E
O
-
L
B
E
SN
m
o
d
el
o
u
tp
er
f
o
r
m
e
d
all
ex
is
tin
g
m
eth
o
d
s
with
an
ac
cu
r
ac
y
o
f
9
5
.
6%
.
T
h
e
p
o
wer
f
u
l
co
m
b
in
atio
n
o
f
E
SN
’
s
d
y
n
am
ic
r
eser
v
o
ir
co
m
p
u
tin
g
with
B
E
O
in
cr
ea
s
es
lear
n
in
g
ac
cu
r
ac
y
ac
r
o
s
s
v
ar
io
u
s
n
etwo
r
k
tr
af
f
ic
class
es.
T
o
v
a
l
i
d
a
t
e
t
h
e
p
e
r
f
o
r
m
a
n
c
e
o
f
t
h
e
m
o
d
e
l
f
o
r
r
e
a
l
-
w
o
r
l
d
I
o
T
t
r
a
f
f
i
c
d
a
t
a
s
e
ts
,
t
h
e
m
o
d
e
l
i
s
a
p
p
l
i
e
d
t
o
I
o
T
-
2
3
a
n
d
E
d
g
e
-
I
I
o
T
d
a
t
a
s
e
ts
.
T
h
i
s
d
a
t
as
e
t
c
o
n
s
is
ts
o
f
d
at
a
c
o
l
l
e
c
t
e
d
f
r
o
m
e
d
g
e
d
e
v
i
ce
s
i
n
c
l
u
d
i
n
g
a
t
t
ac
k
s
l
i
k
e
B
o
t
n
e
t
,
D
D
o
S
,
D
o
S
,
a
n
d
W
o
r
m
s
.
T
h
e
r
e
s
u
lt
s
o
f
B
E
O
-
L
B
E
SN
o
n
t
h
e
s
e
r
e
al
-
w
o
r
l
d
d
a
t
as
e
ts
a
r
e
g
i
v
e
n
i
n
T
a
b
l
e
3
.
T
h
ese
r
esu
lts
s
h
o
w
th
at
th
e
B
E
O
-
L
B
E
SN
m
o
d
el
ca
n
e
f
f
e
ctiv
ely
class
if
y
I
o
T
-
b
ased
att
ac
k
tr
af
f
ic
ev
en
in
d
y
n
a
m
ic
an
d
ev
o
l
v
in
g
I
o
T
en
v
ir
o
n
m
e
n
ts
.
T
o
ev
alu
ate
th
e
ef
f
icien
cy
o
f
th
e
B
E
O
-
L
B
E
SN
m
o
d
el
in
ter
m
s
o
f
r
u
n
tim
e
an
d
m
em
o
r
y
u
s
ag
e,
th
e
test
s
ar
e
p
er
f
o
r
m
ed
o
n
b
o
th
b
en
ch
m
a
r
k
d
atasets
an
d
r
ea
l
-
wo
r
ld
I
o
T
tr
af
f
ic
d
atasets
.
T
h
e
co
m
p
u
tatio
n
al
an
aly
s
is
r
esu
lts
ar
e
g
iv
en
in
T
ab
le
4
.
I
t
is
o
b
s
er
v
e
d
th
at
B
E
O
-
L
B
E
SN
ac
h
iev
es
s
tr
o
n
g
d
etec
tio
n
p
er
f
o
r
m
a
n
ce
with
a
m
o
d
e
r
ate
co
m
p
u
tatio
n
al
o
v
er
h
ea
d
f
o
r
b
o
th
tr
ain
in
g
tim
e
an
d
m
em
o
r
y
u
s
ag
e.
T
h
ese
o
v
er
h
ea
d
s
ar
e
m
an
ag
ea
b
le
i
n
ed
g
e
co
m
p
u
tin
g
en
v
ir
o
n
m
en
ts
.
Fu
r
t
h
er
,
to
an
aly
ze
th
e
e
f
f
icien
cy
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
f
o
r
r
e
a
l
-
tim
e
ad
ap
tab
ilit
y
,
th
e
p
er
f
o
r
m
an
ce
is
test
ed
f
o
r
ze
r
o
-
d
a
y
attac
k
s
.
T
o
s
im
u
late
ze
r
o
-
d
a
y
attac
k
s
,
a
n
o
v
el
attac
k
ty
p
e
is
g
en
er
ate
d
in
b
o
th
th
e
KDDCu
p
9
9
an
d
I
o
T
-
2
3
d
atasets
.
T
h
e
p
e
r
f
o
r
m
an
c
e
o
f
th
e
B
E
O
-
L
B
E
SN m
o
d
el
o
n
ze
r
o
-
d
ay
attac
k
s
is
g
iv
en
in
T
ab
le
5
.
T
h
e
r
esu
l
ts
s
h
o
w
th
at
th
e
B
E
O
-
L
B
E
S
N
m
o
d
el
ca
n
id
en
tify
ze
r
o
-
d
a
y
attac
k
s
ef
f
ec
tiv
ely
with
th
e
h
ig
h
est ac
cu
r
ac
y
an
d
F1
-
s
co
r
e
r
ates.
T
ab
le
2
.
C
o
m
p
a
r
is
o
n
with
o
th
er
m
o
d
els
A
u
t
h
o
r
(
s)
Te
c
h
n
i
q
u
e
A
c
c
u
r
a
c
y
(
%)
M
.
S
.
D
e
s
h
m
u
k
h
e
t
a
l
.
D
B
N
w
i
t
h
Ta
y
l
o
r
-
S
M
O
9
1
.
8
S
.
M
a
n
i
m
u
r
u
g
a
n
D
B
N
9
2
.
3
R
.
Y
.
A
b
u
r
a
s
a
i
n
e
t
a
l
.
B
W
O
-
b
a
s
e
d
h
y
b
r
i
d
m
o
d
e
l
9
3
.
2
A
b
d
a
l
g
a
w
a
d
e
t
a
l
.
A
A
E
+
B
i
G
A
N
9
4
.
1
N
.
K
.
S
.
N
a
y
a
k
e
t
a
l
.
ESBC
A
-
N
e
t
9
3
.
5
M
a
r
w
a
K
e
sh
k
e
t
a
l
.
LSTM
9
2
.
8
H
i
ma
n
sh
u
N
a
n
d
a
n
w
a
r
e
t
a
l
.
C
N
N
+
G
R
U
9
4
.
5
Q
i
g
a
n
g
L
i
u
e
t
a
l
.
M
LP
9
1
.
0
U
.
C
.
A
k
u
t
h
o
t
a
e
t
a
l
.
Tr
a
n
sf
o
r
mer m
o
d
e
l
9
3
.
9
S
a
r
i
k
a
S
a
x
e
n
a
e
t
a
l
.
GNN
9
4
.
6
P
r
o
p
o
se
d
m
e
t
h
o
d
(
B
EO
-
L
B
ESN
)
O
p
t
i
mi
z
e
d
ESN
u
si
n
g
B
EO
9
5
.
6
T
ab
le
3
.
Per
f
o
r
m
an
ce
o
n
r
ea
l
-
wo
r
ld
I
o
T
tr
af
f
ic
d
atasets
D
a
t
a
s
e
t
A
c
c
u
r
a
c
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
R
e
c
a
l
l
(
%)
F1
-
sc
o
r
e
(
%)
Tr
a
i
n
i
n
g
t
i
me
(
sec
o
n
d
s)
I
n
f
e
r
e
n
c
e
t
i
me
(
sec
o
n
d
s/
s
a
mp
l
e
)
I
o
T
-
23
9
5
.
2
9
4
.
5
9
6
.
0
9
5
.
3
2
4
5
0
.
4
2
Ed
g
e
-
I
I
o
T
9
6
.
0
9
5
.
8
9
6
.
4
9
6
.
1
2
9
0
0
.
5
5
T
ab
le
4
.
An
aly
s
is
o
f
c
o
m
p
u
tatio
n
al
o
v
er
h
ea
d
M
e
t
r
i
c
U
N
S
W
-
N
B
1
5
K
D
D
C
u
p
9
9
I
o
T
-
23
Ed
g
e
-
I
I
o
T
Tr
a
i
n
i
n
g
t
i
me
(
sec
o
n
d
s)
3
5
0
3
1
0
2
4
5
2
9
0
I
n
f
e
r
e
n
c
e
t
i
me
(
sec
o
n
d
s
/
sam
p
l
e
)
0
.
5
0
0
.
4
5
0
.
4
2
0
.
5
5
M
e
m
o
r
y
u
s
a
g
e
(
M
B
)
1
5
3
0
1
4
5
0
1
7
8
0
1
8
2
0
M
o
d
e
l
s
i
z
e
(
M
B
)
1
2
.
5
1
1
.
2
1
4
.
0
1
5
.
6
T
ab
le
5
.
Per
f
o
r
m
an
ce
an
aly
s
is
f
o
r
ze
r
o
-
d
a
y
attac
k
s
D
a
t
a
s
e
t
Ze
r
o
-
d
a
y
a
t
t
a
c
k
t
y
p
e
A
c
c
u
r
a
c
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
R
e
c
a
l
l
(
%)
F1
-
sc
o
r
e
(
%)
K
D
D
C
u
p
9
9
N
e
w
D
o
S
9
3
.
5
9
4
.
2
9
1
.
8
9
3
.
0
I
o
T
-
23
B
o
t
n
e
t
(
v
a
r
i
a
n
t
)
9
4
.
8
9
4
.
6
9
5
.
4
9
5
.
0
5.
CO
NCLU
SI
O
N
I
n
th
is
wo
r
k
,
a
n
o
v
el
h
y
b
r
id
D
L
s
y
s
tem
is
s
u
g
g
ested
to
en
h
a
n
ce
an
o
m
aly
d
etec
tio
n
i
n
I
o
T
n
etwo
r
k
s
.
B
y
in
teg
r
atin
g
d
if
f
er
e
n
t
f
ea
tu
r
e
ex
tr
ac
tio
n
tech
n
iq
u
es,
th
e
m
o
d
el
ef
f
ec
tiv
ely
ca
p
tu
r
es
b
o
th
ex
p
licit
an
d
laten
t
ch
ar
ac
ter
is
tics
o
f
n
etwo
r
k
tr
af
f
ic.
T
h
e
f
u
s
ed
f
ea
tu
r
es
ar
e
cla
s
s
if
ied
u
s
in
g
a
L
B
E
SN.
I
t
u
s
e
s
d
y
n
am
ic
m
em
o
r
y
f
o
r
r
eser
v
o
ir
co
m
p
u
tin
g
.
I
n
ad
d
itio
n
,
th
e
p
ar
am
eter
s
ar
e
f
in
e
-
tu
n
e
d
u
s
in
g
th
e
B
E
O
to
en
s
u
r
e
o
p
tim
al
Evaluation Warning : The document was created with Spire.PDF for Python.