I
nte
rna
t
io
na
l J
o
urna
l o
f
Adv
a
nces in Applie
d Science
s
(
I
J
AAS)
Vo
l.
14
,
No
.
4
,
Dec
em
b
er
20
25
,
p
p
.
1
4
2
0
~
1
4
3
2
I
SS
N:
2252
-
8
8
1
4
,
DOI
:
1
0
.
1
1
5
9
1
/ijaas
.
v14.
i
4
.
pp
1
4
2
0
-
1
4
3
2
1420
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
a
s
.
ia
esco
r
e.
co
m
Unv
eiling
a
no
ma
lies in ind
ustria
l c
o
ntrol sy
stems: a
kernel
SH
A
P
-
ba
sed a
pp
ro
a
ch wit
h
tempo
ra
l conv
o
lution a
utoenco
der
Sa
ng
ee
t
a
O
s
wa
l
1
,2
,
Su
bh
a
s
h
Sh
in
de
1
,
Vij
a
y
a
la
k
s
m
i
M
urli
2
1
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
En
g
i
n
e
e
r
i
n
g
,
Lo
k
ma
n
y
a
Ti
l
a
k
C
o
l
l
e
g
e
o
f
E
n
g
i
n
e
e
r
i
n
g
,
N
a
v
i
M
u
m
b
a
i
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
A
r
t
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
a
n
d
D
a
t
a
S
c
i
e
n
c
e
,
V
i
v
e
k
a
n
a
n
d
E
d
u
c
a
t
i
o
n
S
o
c
i
e
t
y
I
n
st
i
t
u
t
e
o
f
T
e
c
h
n
o
l
o
g
y
,
M
u
mb
a
i
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Ap
r
1
9
,
2
0
2
5
R
ev
is
ed
Oct
9
,
2025
Acc
ep
ted
No
v
4
,
2
0
2
5
I
n
d
u
strial
c
o
n
tr
o
l
sy
ste
m
s
(
ICS
)
a
re
o
ften
th
e
targ
e
t
o
f
c
y
b
e
r
-
a
tt
a
c
k
s
,
lea
d
in
g
to
u
n
d
e
sira
b
le
c
o
n
se
q
u
e
n
c
e
s.
ICS
s
o
p
e
ra
te
with
o
u
t
h
u
m
a
n
su
p
e
rv
isio
n
,
m
a
k
in
g
t
h
e
m
v
u
l
n
e
ra
b
le
to
a
d
v
e
rsa
ries
.
In
re
c
e
n
t
y
e
a
rs,
n
u
m
e
ro
u
s
d
e
e
p
lea
rn
in
g
-
b
a
se
d
so
lu
ti
o
n
s
h
a
v
e
d
e
m
o
n
stra
ted
th
e
ir
e
fficie
n
c
y
in
d
e
tec
ti
n
g
a
n
o
m
a
li
e
s
in
ICS
s
.
Ho
we
v
e
r,
th
e
r
e
is
a
la
c
k
o
f
a
b
il
it
y
t
o
p
i
n
p
o
in
t
t
h
e
se
n
so
rs
a
n
d
a
c
tu
a
to
rs
t
h
a
t
c
o
n
tri
b
u
ted
t
o
th
e
a
n
o
m
a
ly
.
I
n
t
h
is
re
se
a
rc
h
wo
rk
,
we
u
se
k
e
rn
e
l
S
h
a
p
ley
a
d
d
it
i
v
e
e
x
p
lan
a
ti
o
n
s
(
S
HA
P
)
t
o
e
x
p
lain
a
n
o
m
a
li
e
s
d
e
tec
ted
b
y
a
tem
p
o
ra
l
c
o
n
v
o
lu
ti
o
n
a
u
t
o
e
n
c
o
d
e
r
(
TCAE
)
.
Th
e
p
ro
p
o
se
d
T
C
AE
m
o
d
e
l
h
a
n
d
les
t
h
e
l
o
n
g
-
term
d
e
p
e
n
d
e
n
c
y
e
ffe
c
ti
v
e
ly
a
n
d
is
c
o
m
p
u
tatio
n
a
ll
y
e
ffe
c
ti
v
e
o
n
a
lar
g
e
d
a
tas
e
t.
A
c
o
m
p
re
h
e
n
si
v
e
e
x
p
la
n
a
ti
o
n
is
p
ro
v
id
e
d
,
fo
c
u
sin
g
o
n
th
e
fe
a
tu
re
th
a
t
c
o
n
t
rib
u
te
d
to
t
h
e
a
n
o
m
a
ly
f
o
r
e
a
c
h
id
e
n
ti
fie
d
a
tt
a
c
k
.
Th
e
S
HA
P
v
a
lu
e
s
a
re
e
x
tr
a
c
ted
fo
r
e
a
c
h
i
d
e
n
ti
f
ied
a
tt
a
c
k
a
n
d
v
isu
a
ll
y
d
e
p
ict
th
e
fe
a
t
u
re
t
h
a
t
c
o
n
tri
b
u
te
d
to
t
h
e
a
n
o
m
a
ly
fo
r
e
a
c
h
a
tt
a
c
k
,
h
e
lp
in
g
t
h
e
e
x
p
e
rt
t
o
h
a
n
d
le t
h
e
a
tt
a
c
k
a
n
d
b
u
il
d
u
se
r
tru
st.
K
ey
w
o
r
d
s
:
An
o
m
aly
d
etec
tio
n
Au
to
en
co
d
er
E
x
p
l
a
i
n
ab
le
AI
I
n
d
u
s
tr
ial
co
n
t
r
o
l sy
s
tem
T
em
p
o
r
al
c
o
n
v
o
lu
tio
n
n
etwo
r
k
Un
s
u
p
er
v
is
ed
lear
n
in
g
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
:
San
g
ee
ta
Oswal
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
E
n
g
i
n
ee
r
in
g
,
L
o
k
m
a
n
y
a
T
ilak
C
o
ll
eg
e
o
f
E
n
g
in
ee
r
i
n
g
,
M
u
m
b
ai
Un
iv
er
s
ity
Secto
r
-
4
,
Plo
t N
o.
17
an
d
1
8
,
Vik
as Na
g
ar
,
Ko
p
ar
k
h
air
an
e
,
Nav
i M
u
m
b
ai
,
I
n
d
ia
E
m
ail: san
g
ee
ta.
o
s
wal1
4
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
I
n
d
u
s
tr
y
4
.
0
[
1
]
in
co
r
p
o
r
ates
cy
b
er
-
p
h
y
s
ical
s
y
s
tem
s
,
I
o
T
,
an
d
cl
o
u
d
co
m
p
u
tin
g
in
t
o
in
d
u
s
tr
ial
co
n
tr
o
l
s
y
s
tem
(
I
C
S)
f
o
r
en
h
an
ce
d
au
to
n
o
m
y
a
n
d
d
e
p
en
d
ab
ilit
y
.
A
ty
p
ical
ex
am
p
le
o
f
an
I
C
S
is
a
s
m
ar
t
p
o
wer
g
r
id
,
a
wate
r
tr
ea
tm
en
t
p
lan
t,
o
r
a
wate
r
d
is
tr
ib
u
tio
n
s
y
s
tem
.
I
t
is
p
r
im
ar
ily
an
au
to
m
ated
s
y
s
tem
th
at
in
teg
r
ates
v
ar
io
u
s
ty
p
es
o
f
c
o
n
tr
o
l
s
y
s
tem
s
,
in
cl
u
d
in
g
s
u
p
er
v
is
o
r
y
co
n
t
r
o
l
a
n
d
d
ata
a
cq
u
is
itio
n
s
y
s
tem
s
(
SC
ADA)
,
d
is
tr
ib
u
ted
co
n
tr
o
l
s
y
s
tem
s
(
DC
S),
an
d
p
r
o
g
r
am
m
ab
le
lo
g
ic
co
n
tr
o
ller
s
(
PLCs
)
.
I
C
S
s
o
p
e
r
a
t
e
i
n
a
s
m
a
r
t
e
n
v
i
r
o
n
m
e
n
t
a
n
d
a
r
e
s
u
b
j
e
c
t
t
o
e
x
t
e
r
n
a
l
t
h
r
e
a
ts
.
A
n
o
m
a
ly
d
e
t
e
c
t
i
o
n
u
s
i
n
g
d
e
e
p
l
e
a
r
n
i
n
g
t
e
c
h
n
i
q
u
e
s
[
2
]
,
[
3
]
is
wi
d
e
l
y
u
s
e
d
i
n
I
C
S
f
o
r
f
a
u
l
t
d
e
t
e
ct
i
o
n
a
n
d
h
a
n
d
l
i
n
g
f
ai
l
u
r
e
s
o
f
s
e
n
s
o
r
s
a
n
d
a
c
t
u
at
o
r
s
.
N
u
m
e
r
o
u
s
a
n
o
m
a
l
y
d
e
t
e
c
t
i
o
n
t
e
c
h
n
i
q
u
es
h
a
v
e
b
e
e
n
c
r
e
a
t
e
d
t
o
i
d
e
n
t
i
f
y
c
y
b
e
r
a
t
t
a
c
k
s
i
n
I
C
S
.
Alth
o
u
g
h
p
r
io
r
r
esear
ch
o
n
I
C
S
d
o
m
ain
s
h
as
p
r
im
ar
ily
f
o
c
u
s
ed
o
n
im
p
r
o
v
in
g
m
o
d
el
p
er
f
o
r
m
an
ce
,
an
o
m
aly
d
etec
tio
n
m
o
d
els
h
av
e
b
ee
n
ch
allen
g
in
g
t
o
an
aly
ze
.
T
h
is
d
ef
icien
cy
ca
n
h
in
d
er
ef
f
o
r
ts
to
p
er
s
u
ad
e
ex
p
e
r
ts
to
tr
u
s
t
an
d
im
p
lem
en
t
p
o
ten
tially
ad
v
an
tag
e
o
u
s
an
o
m
aly
d
etec
tio
n
s
y
s
tem
s
.
T
h
e
r
esu
lts
o
f
th
e
an
o
m
aly
d
etec
tio
n
alg
o
r
ith
m
s
m
ay
r
ev
ea
l
an
o
m
alo
u
s
ca
s
es
p
r
ev
i
o
u
s
ly
u
n
k
n
o
w
n
to
th
e
d
o
m
ain
ex
p
er
t,
an
d
ex
p
lain
in
g
th
e
an
o
m
aly
en
h
an
ce
s
th
e
ex
p
er
t'
s
b
elief
in
th
e
alg
o
r
ith
m
.
A
lack
o
f
ex
p
lan
atio
n
h
in
d
er
s
d
ec
is
io
n
-
m
ak
er
s
an
d
d
o
m
ai
n
s
p
ec
ialis
ts
f
r
o
m
u
s
in
g
th
e
alg
o
r
ith
m
o
u
tp
u
t.
A
v
ia
b
le
ap
p
r
o
ac
h
to
ad
d
r
ess
th
is
p
r
o
b
lem
is
to
em
p
lo
y
e
x
p
lain
ab
le
ar
tific
ial
in
tellig
en
ce
(
XAI
)
m
o
d
els
[
4
]
in
co
n
ju
n
ctio
n
with
th
e
an
o
m
aly
d
etec
to
r
.
T
h
u
s
,
XAI
[
5
]
,
[
6
]
is
cr
u
cial
f
o
r
clar
if
y
in
g
a
n
d
u
n
d
er
s
tan
d
i
n
g
th
e
b
lack
-
b
o
x
m
o
d
els.
Nu
m
er
o
u
s
ex
is
tin
g
an
o
m
aly
d
etec
tio
n
m
eth
o
d
s
h
a
v
e
in
teg
r
ated
th
eir
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l Sci
I
SS
N:
2252
-
8
8
1
4
Un
ve
ilin
g
a
n
o
ma
lies
in
in
d
u
s
t
r
ia
l c
o
n
tr
o
l sys
tems:
a
ke
r
n
el
S
HA
P
-
b
a
s
ed
a
p
p
r
o
a
ch
…
(
S
a
n
g
ee
ta
Osw
a
l
)
1421
m
o
d
els
with
XAI
to
im
p
r
o
v
e
tr
an
s
p
ar
en
cy
an
d
r
eliab
ilit
y
.
Au
to
en
co
d
er
-
b
ased
f
r
am
ewo
r
k
s
,
f
o
r
en
d
-
to
-
en
d
wo
r
k
f
lo
ws
f
o
r
co
o
li
n
g
-
s
y
s
tem
an
o
m
aly
d
etec
tio
n
with
e
x
p
er
t
-
g
u
id
ed
lo
ca
lizatio
n
[
7
]
,
h
av
e
d
em
o
n
s
tr
ated
s
tr
o
n
g
p
er
f
o
r
m
an
ce
o
n
r
ea
l
-
wo
r
ld
m
ac
h
in
er
y
d
ata.
I
n
I
C
S
n
etwo
r
k
s
ec
u
r
ity
,
class
-
b
as
ed
SHAP
an
aly
s
e
s
p
r
o
v
id
e
s
em
i
-
g
lo
b
al
in
s
ig
h
ts
f
o
r
n
etwo
r
k
in
tr
u
s
io
n
d
etec
tio
n
s
y
s
tem
s
[
8
]
.
L
o
g
-
c
e
n
t
r
i
c
a
p
p
r
o
a
c
h
e
s
l
i
k
e
t
h
e
S
X
AD
f
r
a
m
e
w
o
r
k
t
r
a
n
s
f
o
r
m
b
l
a
c
k
-
b
o
x
l
o
g
d
e
t
e
c
t
o
r
s
i
n
t
o
tr
a
n
s
p
a
r
e
n
t
w
h
i
t
e
-
b
o
x
s
y
s
t
e
m
s
[
9
]
,
a
n
d
S
i
e
m
e
n
s
’
i
n
d
u
s
t
r
i
a
l
-
g
r
a
d
e
XA
I
w
h
i
t
e
p
a
p
e
r
o
u
t
l
i
n
e
s
e
s
s
e
n
t
i
a
l
a
r
c
h
i
t
e
c
t
u
r
es
a
n
d
r
e
q
u
i
r
e
m
e
n
t
s
f
o
r
e
x
p
l
a
i
n
a
b
l
e
m
a
n
u
f
a
c
t
u
r
i
n
g
a
r
t
i
f
ic
i
a
l
i
n
t
el
l
ig
e
n
c
e
(
A
I
)
[
1
0
]
.
Ho
wev
er
,
p
r
io
r
r
esear
ch
h
as
n
o
t
ev
alu
ated
th
e
d
ep
en
d
a
b
ilit
y
o
f
XAI
in
th
e
co
n
te
x
t o
f
I
C
S sec
u
r
ity
.
T
o
d
eliv
er
h
ig
h
-
q
u
ality
in
ter
p
r
etatio
n
f
o
r
a
n
o
m
aly
lo
ca
liz
atio
n
at
a
lo
w
co
m
p
u
tatio
n
al
co
s
t,
we
em
p
lo
y
m
o
d
e
r
n
m
eth
o
d
s
to
id
en
tify
an
o
m
al
o
u
s
d
ata
in
th
e
I
C
S
d
o
m
ai
n
.
Ad
d
itio
n
ally
,
we
d
eter
m
in
e
t
h
e
ef
f
ec
tiv
e
XAI
ap
p
r
o
ac
h
f
o
r
e
x
p
lain
in
g
th
e
a
n
o
m
alo
u
s
d
ata
p
o
in
t
an
d
b
u
ild
in
g
u
s
er
tr
u
s
t
in
d
ep
lo
y
in
g
th
e
m
o
d
el.
T
h
is
s
tu
d
y
u
s
es
th
e
s
ec
u
r
e
wate
r
tr
ea
tm
e
n
t
(
SW
aT
)
test
b
ed
d
ataset
[
1
1
]
to
tr
a
in
an
d
ev
alu
ate
t
h
e
an
o
m
aly
d
etec
tio
n
m
o
d
el,
te
m
p
o
r
al
c
o
n
v
o
lu
tio
n
a
u
to
en
c
o
d
er
(
T
C
AE
)
,
a
n
d
ass
ess
its
e
f
f
icac
y
in
l
o
ca
lizin
g
an
o
m
alies
th
r
o
u
g
h
XAI
.
W
e
p
r
o
p
o
s
e
a
T
C
AE
,
wh
ich
em
p
lo
y
s
a
tem
p
o
r
al
c
o
n
v
o
lu
ti
o
n
n
etwo
r
k
as
its
f
o
u
n
d
atio
n
.
T
C
AE
ex
ce
ls
in
h
an
d
lin
g
lo
n
g
-
ter
m
d
ep
en
d
en
cy
u
s
in
g
d
ilatio
n
co
n
v
o
lu
tio
n
to
ef
f
ec
tiv
ely
lear
n
lo
n
g
-
ter
m
i
n
tr
icate
tem
p
o
r
al
p
atter
n
s
.
T
h
e
d
e
n
s
ity
-
b
ased
s
p
atial
clu
s
ter
in
g
o
f
ap
p
lic
atio
n
s
with
n
o
is
e
(
D
B
SC
AN
)
al
g
o
r
i
t
h
m
is
u
s
e
d
t
o
f
l
a
g
t
h
e
a
t
t
ac
k
p
o
i
n
ts
,
a
n
d
w
e
p
r
o
p
o
s
e
Ke
r
n
e
l
S
h
a
p
l
e
y
ad
d
i
t
i
v
e
e
x
p
l
a
n
at
i
o
n
s
(
S
H
AP
)
[
1
2
]
,
s
t
r
a
t
e
g
ie
s
t
o
el
u
c
i
d
a
t
e
t
h
e
m
o
d
e
ls
'
d
e
c
is
i
o
n
s
b
y
c
a
l
c
u
l
a
t
i
n
g
t
h
e
S
H
AP
v
a
l
u
es
f
o
r
e
a
c
h
i
d
e
n
t
i
f
i
ed
a
t
t
a
c
k
.
T
h
e
u
s
e
o
f
a
lig
h
twei
g
h
t
T
C
AE
ar
ch
itectu
r
e
en
s
u
r
es
s
ca
lab
ilit
y
to
r
ea
l
I
C
S
en
v
ir
o
n
m
e
n
ts
,
u
n
lik
e
h
ea
v
ier
s
eq
u
en
ce
m
o
d
els
s
u
c
h
as
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
L
STM
)
o
r
T
r
an
s
f
o
r
m
er
s
n
o
r
m
al
d
ata
p
o
in
ts
.
T
h
e
T
C
AE
m
o
d
el
o
p
er
ates
o
n
th
e
p
r
em
is
e
th
at
an
o
m
alo
u
s
d
ata
p
o
in
ts
ex
h
ib
it
elev
ated
r
ec
o
n
s
tr
u
ctio
n
lo
s
s
,
as
th
e
m
o
d
el
is
tr
ain
ed
ex
clu
s
iv
ely
o
n
n
o
r
m
al
d
ata
p
o
in
ts
.
T
o
o
u
r
k
n
o
wled
g
e,
n
o
p
r
io
r
r
e
s
ea
r
ch
h
as
u
tili
ze
d
SHAP
to
o
f
f
er
b
lack
-
b
o
x
ex
p
lan
atio
n
s
f
o
r
ab
n
o
r
m
alities
id
en
tifie
d
b
y
an
a
u
to
en
c
o
d
er
o
n
th
e
SW
aT
d
ataset.
T
h
e
m
ain
co
n
tr
ib
u
tio
n
o
f
th
i
s
r
esear
ch
wo
r
k
is
:
i)
e
nd
-
to
-
en
d
f
u
s
io
n
o
f
a
TCAE
with
Ker
n
el
SHAP
f
o
r
an
o
m
aly
d
etec
tio
n
a
n
d
e
x
p
l
an
atio
n
in
I
C
S
;
ii)
p
r
o
v
id
es
f
in
e
-
g
r
ain
ed
ex
p
lan
atio
n
s
o
f
th
e
f
ea
t
u
r
es
th
at
co
n
tr
ib
u
te
m
o
s
t
to
an
o
m
alo
u
s
b
eh
av
io
r
;
iii)
t
h
e
u
s
e
o
f
a
lig
h
tweig
h
t
T
C
AE
ar
ch
itectu
r
e
en
s
u
r
es
s
ca
lab
ilit
y
to
r
ea
l
I
C
S
en
v
ir
o
n
m
e
n
ts
,
u
n
lik
e
h
ea
v
ier
s
eq
u
e
n
ce
m
o
d
els
s
u
ch
as
L
STM
s
o
r
T
r
an
s
f
o
r
m
er
s
;
an
d
iv
)
w
e
ev
alu
ate
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
o
n
th
e
SW
aT
d
ataset,
a
wid
ely
u
s
ed
I
C
S b
en
ch
m
ar
k
.
T
h
is
r
esear
ch
in
tr
o
d
u
ce
s
a
n
o
v
el
m
eth
o
d
u
tili
zin
g
SHAP
v
alu
es
to
elu
cid
ate
th
e
ab
n
o
r
m
alities
d
etec
ted
in
an
au
to
e
n
co
d
e
r
'
s
o
u
tp
u
t
f
o
r
ea
ch
id
e
n
tifie
d
att
ac
k
.
Ou
r
p
i
p
elin
e
f
lag
s
an
o
m
alies
at
in
ce
p
tio
n
,
p
in
p
o
in
ts
th
eir
r
o
o
t
ca
u
s
es
in
r
aw
m
ea
s
u
r
em
en
t
s
tr
ea
m
s
,
an
d
p
r
esen
ts
tr
an
s
p
ar
e
n
t,
h
u
m
an
-
r
ea
d
a
b
le
ex
p
lan
atio
n
s
th
at
b
u
ild
o
p
e
r
ato
r
tr
u
s
t
an
d
s
tr
ea
m
lin
e
d
ec
is
io
n
-
m
ak
in
g
.
T
h
e
T
C
AE
+SHAP
f
r
am
ewo
r
k
e
n
ab
les
ea
r
ly
an
o
m
al
y
d
etec
tio
n
b
y
ca
p
tu
r
in
g
s
u
b
tle
tem
p
o
r
al
d
e
v
iatio
n
s
in
I
C
S
d
ata,
wh
ile
SHAP
ex
p
lan
atio
n
s
p
in
p
o
in
t
r
o
o
t
c
au
s
es
b
y
attr
ib
u
tin
g
an
o
m
alies
to
s
p
ec
if
ic
s
e
n
s
o
r
s
.
Vis
u
al
to
o
ls
lik
e
f
o
r
ce
p
lo
ts
an
d
h
ea
t
m
ap
s
m
ak
e
th
ese
in
s
ig
h
ts
tr
an
s
p
ar
en
t,
h
elp
in
g
o
p
er
ato
r
s
u
n
d
er
s
tan
d
wh
en
,
wh
y
,
a
n
d
wh
e
r
e
an
o
m
alies
o
cc
u
r
,
b
u
ild
in
g
tr
u
s
t
in
AI
-
d
r
iv
e
n
d
e
cisi
o
n
s
.
T
o
g
eth
er
,
th
ese
ca
p
a
b
ilit
ies
en
s
u
r
e
ea
r
ly
war
n
in
g
,
r
ig
o
r
o
u
s
r
o
o
t
-
ca
u
s
e
an
aly
s
is
,
an
d
s
u
s
tain
ed
tr
u
s
t,
k
ey
p
r
er
e
q
u
is
ites
f
o
r
r
esil
ien
t,
m
is
s
io
n
-
cr
itical
I
C
S
o
p
er
atio
n
s
.
T
h
e
m
eth
o
d
will
b
e
ad
v
a
n
tag
eo
u
s
f
o
r
s
p
ec
iali
s
ts
n
ee
d
in
g
ju
s
tific
atio
n
an
d
v
is
ib
ilit
y
o
f
an
o
m
alies.
Do
m
ain
s
p
ec
ialis
ts
wh
o
u
tili
ze
d
th
e
ex
p
lan
atio
n
s
b
ased
o
n
r
ea
l
-
wo
r
ld
d
ata
o
f
f
er
ed
f
av
o
r
ab
le
c
o
m
m
e
n
ts
,
ass
er
tin
g
th
at
th
e
ex
p
lan
atio
n
s
f
ac
ilit
ated
th
eir
c
o
m
p
r
eh
e
n
s
io
n
an
d
ex
am
i
n
atio
n
o
f
th
e
ab
n
o
r
m
alities
.
T
h
e
s
u
b
s
eq
u
e
n
t
s
ec
tio
n
s
o
f
th
is
d
o
cu
m
e
n
t
ar
e
s
tr
u
ct
u
r
ed
as
f
o
llo
ws:
s
ec
tio
n
2
p
r
esen
ts
p
er
tin
en
t
r
esear
ch
o
n
a
n
o
m
aly
d
etec
tio
n
m
o
d
els
a
n
d
XAI
with
in
th
e
I
C
S
d
o
m
ain
.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
an
d
ex
p
lain
ab
le
AI
m
eth
o
d
o
lo
g
y
em
p
lo
y
ed
i
n
th
is
s
tu
d
y
ar
e
d
is
cu
s
s
ed
in
s
ec
tio
n
3
.
Sectio
n
4
c
o
n
tain
s
r
esu
lts
,
an
d
f
in
ally
,
t
h
e
co
n
clu
s
io
n
is
p
r
esen
ted
in
s
ec
tio
n
5
.
2.
RE
L
AT
E
D
WO
RK
2
.
1
.
Ano
m
a
ly
det
ec
t
io
n
An
o
m
aly
d
etec
tio
n
al
g
o
r
ith
m
s
ca
n
b
e
s
ep
ar
ated
in
to
two
ty
p
es:
d
is
tr
ib
u
tio
n
-
b
ased
,
wh
ich
m
o
d
els
th
e
d
is
tr
ib
u
tio
n
o
f
n
o
r
m
al
s
am
p
les,
an
d
r
ec
o
n
s
tr
u
ctio
n
-
b
as
ed
,
wh
ich
c
o
n
s
id
er
s
d
ata
p
o
in
ts
with
h
ig
h
r
ec
o
n
s
tr
u
ctio
n
er
r
o
r
to
b
e
a
n
o
m
alo
u
s
.
Sev
er
al
d
eep
-
lea
r
n
i
n
g
m
eth
o
d
s
[
1
3
]
h
av
e
b
ee
n
d
ep
lo
y
e
d
to
i
d
en
tify
an
o
m
alies
in
th
e
SW
aT
d
atas
et.
Me
th
o
d
s
u
tili
zin
g
au
to
en
c
o
d
er
s
[
1
4
]
,
g
en
er
ativ
e
a
d
v
er
s
a
r
ial
n
etwo
r
k
(
GAN
)
[
1
5
]
,
a
n
d
T
r
a
n
s
f
o
r
m
e
r
s
[
1
6
]
f
o
cu
s
o
n
e
n
h
an
cin
g
m
o
d
el
p
er
f
o
r
m
a
n
ce
wh
ile
p
lacin
g
les
s
em
p
h
asis
o
n
XAI
tech
n
iq
u
es.
Du
e
to
I
C
S's
in
tr
i
n
s
ic
d
ata,
d
ee
p
lear
n
in
g
is
p
r
ef
er
r
ed
o
v
er
m
ac
h
in
e
lear
n
i
n
g
m
o
d
els.
Seq
u
en
ce
lear
n
in
g
m
o
d
els
ar
e
em
p
lo
y
e
d
to
h
a
n
d
le
th
e
tem
p
o
r
al
d
ep
en
d
en
cy
o
f
t
h
e
d
ata.
Z
h
ao
et
a
l.
[
1
7
]
p
r
o
p
o
s
ed
a
g
ated
r
ec
u
r
r
e
n
t
u
n
it
(
GR
U
)
-
b
ased
an
o
m
aly
d
etec
to
r
f
o
r
SW
aT
,
s
h
o
win
g
im
p
r
o
v
ed
d
etec
tio
n
o
f
attac
k
s
co
m
p
ar
ed
to
C
NN
b
aselin
es.
Kim
et
a
l.
[
1
8
]
u
s
ed
s
tack
ed
L
STM
lay
er
s
with
atten
tio
n
m
e
ch
an
is
m
s
to
ca
p
tu
r
e
lo
n
g
-
ter
m
d
ep
en
d
en
cies
in
I
C
S
telem
etr
y
.
L
i
et
a
l.
[
1
9
]
p
r
o
p
o
s
ed
a
d
is
tan
g
le
tim
e
s
er
ies
to
s
o
lv
e
th
e
Ku
llb
ac
k
-
L
eib
ler
(
KL
)
v
a
n
is
h
in
g
p
r
o
b
lem
with
L
STM
.
Z
h
an
g
et
a
l.
[
2
0
]
p
r
o
p
o
s
ed
a
b
id
ir
ec
tio
n
al
L
STM
with
atten
tio
n
an
d
b
u
ilt
a
co
n
v
o
lu
t
io
n
au
to
e
n
co
d
er
m
o
d
el
(
C
AE
-
M)
to
ca
p
tu
r
e
tem
p
o
r
al
d
e
p
e
n
d
en
cy
in
th
e
tim
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
14
,
No
.
4
,
Dec
em
b
er
20
25
:
1
4
2
0
-
1
4
3
2
1422
s
er
ies.
Un
s
u
p
er
v
is
ed
An
o
m
al
y
d
etec
tio
n
is
f
av
o
r
ed
o
v
er
s
u
p
er
v
is
ed
tech
n
iq
u
es,
wh
ich
d
e
p
en
d
h
ea
v
ily
o
n
th
e
ass
u
m
p
tio
n
th
at
n
o
r
m
al
ca
s
es a
r
e
m
u
ch
m
o
r
e
c
o
m
m
o
n
th
an
an
o
m
alo
u
s
o
n
es.
Sp
atial
d
ep
en
d
e
n
cies
ar
e
b
ett
er
ca
p
tu
r
ed
i
n
m
o
d
els
lev
er
a
g
in
g
tr
a
n
s
f
o
r
m
e
r
s
an
d
g
r
ap
h
atten
tio
n
m
ec
h
an
is
m
s
.
T
u
li
et
a
l.
[
2
1
]
in
tr
o
d
u
ce
s
a
T
r
an
AD,
tr
an
s
f
o
r
m
er
-
b
ased
an
o
m
aly
d
etec
tio
n
f
r
am
ewo
r
k
th
at
lev
er
ag
es
s
elf
-
atten
tio
n
to
m
o
d
el
co
m
p
lex
d
ep
en
d
en
cies
in
I
C
S
d
ata
s
u
ch
as
SW
aT
an
d
wate
r
d
is
tr
ib
u
tio
n
(
W
ADI
)
,
o
u
tp
er
f
o
r
m
in
g
r
ec
u
r
r
en
t
ar
ch
itectu
r
es
in
s
ca
lab
ilit
y
.
Hy
b
r
id
m
o
d
els
lik
e
h
y
b
r
id
an
o
m
aly
d
etec
tio
n
with
m
u
lti
-
d
im
en
s
io
n
al
g
r
a
p
h
atten
tio
n
(
HAD
-
MD
GAT
)
[
2
2
]
co
m
b
i
n
e
T
C
N
with
g
r
ap
h
atten
tio
n
m
ec
h
an
is
m
s
to
in
te
g
r
ate
tem
p
o
r
al
an
d
r
elatio
n
al
d
e
p
en
d
e
n
c
ies
ac
r
o
s
s
p
r
o
ce
s
s
v
ar
ia
b
les.
R
ec
en
t
s
u
r
v
ey
s
[
2
]
,
[
1
6
]
f
u
r
th
er
h
ig
h
lig
h
t
th
e
g
r
o
win
g
ad
o
p
tio
n
o
f
T
r
an
s
f
o
r
m
e
r
an
d
T
C
N
v
ar
ian
ts
in
I
C
S
s
ec
u
r
ity
m
o
n
ito
r
i
n
g
,
r
ef
lectin
g
a
s
h
if
t
to
war
d
ar
ch
it
ec
tu
r
es
th
at
ca
n
s
ca
le
with
h
ig
h
-
d
im
en
s
io
n
al
in
d
u
s
tr
ial
tim
e
s
er
ies.
R
e
s
ea
r
ch
er
s
h
av
e
p
r
o
p
o
s
ed
a
m
o
d
el
to
r
e
d
u
ce
th
e
tr
ai
n
in
g
tim
e
f
o
r
q
u
ic
k
in
f
e
r
en
ce
.
Acr
o
s
s
all
th
ese
m
o
d
els,
th
e
ab
s
en
ce
o
f
a
p
r
o
p
er
ex
p
lan
atio
n
an
d
tr
an
s
p
ar
en
cy
to
s
h
e
d
lig
h
t
o
n
h
o
w
an
d
wh
y
a
m
o
d
el
h
as
m
ad
e
a
p
ar
ticu
lar
d
ec
is
io
n
is
a
p
r
o
b
lem
.
So
m
e
o
f
th
e
r
esear
ch
wo
r
k
o
n
ly
d
is
p
l
ay
s
to
p
k
f
ea
t
u
r
es
co
n
tr
ib
u
tin
g
to
all
th
e
an
o
m
aly
p
o
in
ts
,
wh
er
ea
s
we
h
av
e
an
aly
ze
d
ea
ch
attac
k
a
n
d
id
e
n
tifie
d
th
e
f
ea
tu
r
es f
o
r
ea
ch
attac
k
s
ep
ar
ately
.
2
.
2
.
Ex
pla
ina
ble A
I
a
pp
ro
a
ch
XAI
is
a
s
tu
d
y
d
o
m
ain
co
n
c
er
n
in
g
t
h
e
tr
an
s
p
a
r
en
cy
o
f
a
n
AI
-
b
ased
s
y
s
tem
[
2
3
]
.
E
x
p
lain
ab
ilit
y
r
ef
er
s
to
a
s
y
s
tem
'
s
ab
ilit
y
to
p
r
o
d
u
ce
a
s
et
o
f
f
ea
t
u
r
es
f
r
o
m
a
n
in
ter
p
r
etab
le
d
o
m
ai
n
th
at
im
p
ac
ted
th
e
d
ec
is
io
n
f
o
r
a
p
ar
ticu
lar
i
n
s
tan
ce
.
C
o
n
v
er
s
ely
,
ce
r
tain
r
esear
ch
u
s
es
th
e
ter
m
s
ex
p
lain
ab
ilit
y
an
d
in
ter
p
r
etab
ilit
y
in
ter
ch
a
n
g
ea
b
l
y
.
T
h
is
r
esear
ch
f
o
c
u
s
es
o
n
th
e
XAI
tech
n
iq
u
e
d
e
p
lo
y
ed
o
n
th
e
SW
aT
d
ataset.
I
n
th
is
r
esear
ch
p
ap
er
,
we
u
s
ed
th
e
ter
m
ex
p
lain
ab
ilit
y
to
id
en
tify
f
ea
tu
r
es
th
at
co
n
tr
ib
u
ted
to
a
r
ec
o
n
s
tr
u
ctio
n
er
r
o
r
.
T
ab
le
1
p
r
o
v
id
es a
co
n
ci
s
e
o
v
er
v
iew
o
f
r
ec
en
t r
esear
c
h
co
n
d
u
cted
.
XAI
tech
n
iq
u
es
s
u
ch
as
SHAP,
lo
ca
l
in
ter
p
r
etab
le
m
o
d
el
-
a
g
n
o
s
tic
ex
p
lan
atio
n
s
(
L
I
ME
)
,
in
teg
r
ated
g
r
ad
ien
ts
(
I
G)
,
an
d
ac
c
u
m
u
lat
ed
lo
ca
l
ef
f
ec
ts
(
AL
E
)
h
av
e
b
ee
n
ap
p
lied
to
tim
e
-
s
er
ies
an
o
m
aly
d
etec
tio
n
.
I
n
b
r
o
ad
e
r
tim
e
-
s
er
ies
co
n
tex
ts
,
R
o
jat
et
a
l.
[
6
]
d
em
o
n
s
tr
ated
th
e
ap
p
licatio
n
o
f
SHAP
f
o
r
r
ec
u
r
r
e
n
t
an
o
m
aly
d
etec
tio
n
m
o
d
els,
wh
ile
T
h
e
is
s
ler
et
a
l.
[
4
]
s
u
r
v
ey
e
d
in
t
er
p
r
etab
le
d
ee
p
an
o
m
aly
d
et
ec
tio
n
ap
p
r
o
ac
h
es
em
p
h
asizin
g
f
ea
tu
r
e
-
lev
el
attr
ib
u
tio
n
.
R
o
s
h
in
ta
an
d
Gab
o
r
[
2
4
]
ap
p
lied
SHAP
an
d
L
I
ME
to
m
u
ltiv
ar
iate
s
en
s
o
r
d
ata,
p
av
in
g
th
e
way
f
o
r
th
eir
ad
o
p
tio
n
in
I
C
S
d
o
m
ain
s
.
B
en
to
et
a
l.
[
2
5
]
in
tr
o
d
u
ce
d
T
im
eSHAP,
a
v
ar
ian
t
o
f
SHAP
tailo
r
ed
f
o
r
s
eq
u
en
tial
d
ata,
en
ab
lin
g
a
ttrib
u
tio
n
o
f
an
o
m
alies
to
s
p
ec
if
ic
tim
e
s
tep
s
.
Villan
i
et
a
l.
[
2
6
]
ap
p
lied
Ke
r
n
el
SHAP
to
L
STM
o
u
tp
u
ts
in
I
C
S,
s
h
o
win
g
h
o
w
f
ea
tu
r
e
im
p
o
r
tan
ce
v
a
r
ies
ac
r
o
s
s
attac
k
p
h
ases
.
Fu
n
g
et
a
l.
[
2
7
]
e
v
alu
ate
m
u
ltip
le
an
o
m
aly
d
etec
tio
n
m
o
d
els
o
n
I
C
S
d
atasets
lik
e
S
W
aT
an
d
W
ADI
,
with
em
p
h
asis
o
n
in
ter
p
r
etab
ilit
y
a
n
d
r
ec
o
n
s
tr
u
ctio
n
-
b
ased
m
et
h
o
d
s
.
J
u
ttle
et
a
l.
[
2
8
]
p
r
o
p
o
s
ed
a
C
-
SHAP
-
a
co
n
ce
p
t
-
b
ased
SHAP e
x
ten
s
io
n
f
o
r
tim
e
s
er
ies.
I
t e
n
ab
les I
C
S o
p
er
ato
r
s
to
in
ter
p
r
et
m
o
d
el
o
u
tp
u
ts
u
s
in
g
h
ig
h
-
lev
el
tem
p
o
r
al
p
att
er
n
s
an
d
s
y
s
tem
to
p
o
lo
g
y
o
v
e
r
lay
s
.
Ou
r
s
tu
d
y
co
n
tr
ib
u
tes
to
th
is
ev
o
lv
in
g
r
esear
ch
lan
d
s
ca
p
e
b
y
f
u
s
in
g
a
T
C
AE
with
Ker
n
el
SHAP
to
p
r
o
v
id
e
b
o
th
ac
c
u
r
ate
d
etec
ti
o
n
an
d
f
in
e
-
g
r
ain
ed
in
te
r
p
r
eta
b
ilit
y
f
o
r
I
C
S
an
o
m
alies.
Un
lik
e
p
r
io
r
w
o
r
k
s
th
at
p
r
im
ar
ily
f
o
cu
s
o
n
d
etec
tio
n
ac
cu
r
ac
y
,
o
u
r
ap
p
r
o
ac
h
h
ig
h
li
g
h
ts
th
e
s
p
ec
if
ic
s
en
s
o
r
s
an
d
co
n
tr
o
l
p
o
in
ts
m
o
s
t
r
esp
o
n
s
ib
le
f
o
r
an
o
m
al
o
u
s
b
eh
av
io
r
d
u
r
in
g
d
etec
ted
attac
k
w
in
d
o
ws.
T
ab
le
1
.
T
h
e
liter
atu
r
e
o
f
an
o
m
aly
d
etec
tio
n
u
s
in
g
XAI
M
o
d
e
l
Te
c
h
n
i
q
u
e
e
mp
l
o
y
e
d
X
A
I
e
mp
l
o
y
e
d
P
u
r
p
o
se
o
f
X
A
I
Y
e
a
r
U
S
A
D
[
1
4
]
A
u
t
o
e
n
c
o
d
e
r
s a
n
d
G
A
N
No
2
0
2
0
GAN
-
AD
[
2
9
]
G
e
n
e
r
a
t
i
v
e
a
d
v
e
r
sar
i
a
l
n
e
t
w
o
r
k
s
No
2
0
1
8
D
A
EM
O
N
[
3
0
]
A
d
v
e
r
sari
a
l
a
u
t
o
e
n
c
o
d
e
r
a
n
o
ma
l
y
d
e
t
e
c
t
i
o
n
i
n
t
e
r
p
r
e
t
a
t
i
o
n
Y
e
s
Th
e
t
o
p
-
k
d
i
me
n
si
o
n
s e
x
h
i
b
i
t
i
n
g
t
h
e
h
i
g
h
e
st
r
e
c
o
n
st
r
u
c
t
i
o
n
e
r
r
o
r
w
i
l
l
b
e
i
d
e
n
t
i
f
i
e
d
a
s
t
h
e
p
r
i
mary
s
o
u
r
c
e
o
f
t
h
e
a
n
o
ma
l
y
2
0
2
1
F
I
D
-
GAN
[
3
1
]
F
og
-
b
a
s
e
d
,
G
A
N
s
No
HAD
-
M
D
G
A
T
[
2
2
]
G
r
a
p
h
a
t
t
e
n
t
i
o
n
n
e
t
w
o
r
k
No
O
C
P
A
E
[
3
2
]
O
n
e
-
c
l
a
ss
p
r
e
d
i
c
t
i
v
e
a
u
t
o
e
n
c
o
d
e
r
No
2
0
2
2
M
A
D
_
G
A
N
[
3
3
]
GAN
-
LSTM
/
R
N
N
No
Tr
a
n
A
D
[
2
1
]
t
r
a
n
sf
o
r
mers
No
W
a
X
A
I
[
3
4
]
(
D
e
e
p
S
V
D
D
a
n
d
EC
O
D
)
Y
e
s
D
e
r
i
v
e
LI
M
E,
A
LE,
S
H
A
P
,
a
n
d
I
G
f
e
a
t
u
r
e
s
c
o
r
e
s
2
0
2
4
C
C
TA
K
[
3
5
]
V
A
E
w
i
t
h
T
C
N
a
n
d
K
A
N
Y
e
s
P
r
o
p
o
se
d
n
e
w
e
v
a
l
u
a
t
i
o
n
m
a
t
r
i
x
2
0
2
4
3.
M
E
T
H
O
D
3
.
1
.
E
x
pla
ina
ble
AI
Du
e
to
th
e
in
h
er
en
t
ca
p
ac
ity
o
f
b
lac
k
-
b
o
x
m
o
d
els
[
3
6
]
,
t
h
e
d
em
an
d
f
o
r
d
e
p
en
d
a
b
le
e
x
p
lan
atio
n
s
em
er
g
ed
.
T
h
ese
ex
p
lan
atio
n
s
f
o
s
ter
u
s
er
tr
u
s
t,
f
ac
ilit
ate
th
e
i
d
en
tific
atio
n
o
f
m
o
d
el
f
ailu
r
e
s
ites
,
an
d
elim
in
ate
o
b
s
tacle
s
to
th
e
im
p
lem
en
tatio
n
o
f
d
ee
p
n
eu
r
al
n
etwo
r
k
s
ac
r
o
s
s
v
ar
io
u
s
f
ield
s
.
B
y
d
ev
elo
p
in
g
m
o
r
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l Sci
I
SS
N:
2252
-
8
8
1
4
Un
ve
ilin
g
a
n
o
ma
lies
in
in
d
u
s
t
r
ia
l c
o
n
tr
o
l sys
tems:
a
ke
r
n
el
S
HA
P
-
b
a
s
ed
a
p
p
r
o
a
ch
…
(
S
a
n
g
ee
ta
Osw
a
l
)
1423
tr
an
s
p
ar
en
t
an
d
ex
p
licab
le
s
y
s
tem
s
,
u
s
er
s
will
h
av
e
a
b
ette
r
u
n
d
er
s
tan
d
in
g
an
d
,
co
n
s
eq
u
en
tly
,
m
o
r
e
tr
u
s
t
in
th
e
m
o
d
el.
No
ta
b
le
ex
am
p
les
o
f
tech
n
iq
u
es
u
tili
zin
g
ap
p
r
o
x
im
atio
n
s
in
clu
d
e
L
I
ME
[
3
7
]
,
a
m
o
d
el
-
a
g
n
o
s
tic
ap
p
r
o
ac
h
f
o
r
elu
cid
atin
g
p
r
e
d
ictio
n
s
v
ia
a
lo
ca
l
m
o
d
el,
an
d
Dee
p
L
I
FT
[
3
8
]
,
a
m
o
d
el
-
s
p
e
cif
ic
tech
n
iq
u
e
f
o
r
in
ter
p
r
etin
g
d
ee
p
lear
n
in
g
m
o
d
els
b
y
b
ac
k
p
r
o
p
ag
atin
g
th
e
c
o
n
tr
ib
u
tio
n
s
o
f
all
n
eu
r
o
n
s
to
th
e
in
p
u
t
f
ea
tu
r
es.
SHAP
[
1
2
]
in
teg
r
ates
p
r
io
r
m
eth
o
d
o
l
o
g
ies
f
o
r
elu
ci
d
atin
g
p
r
ed
ictio
n
s
b
y
q
u
an
tify
in
g
f
ea
tu
r
e
s
ig
n
if
ican
ce
,
em
p
lo
y
in
g
Sh
ap
ley
v
al
u
es
f
r
o
m
g
am
e
th
e
o
r
y
to
g
u
ar
a
n
tee
th
e
co
n
s
is
ten
cy
o
f
th
e
e
x
p
la
n
atio
n
s
.
T
h
e
SHAP
f
r
am
ewo
r
k
p
r
o
p
o
s
es
a
m
o
d
el
-
ag
n
o
s
tic
m
et
h
o
d
f
o
r
a
p
p
r
o
x
i
m
atin
g
SHAP
v
alu
es,
k
n
o
wn
as
Ker
n
el
SHAP.
Ker
n
el
SHAP
em
p
lo
y
s
lin
ea
r
L
I
ME
[
3
7
]
in
c
o
n
ju
n
ctio
n
with
Sh
ap
ley
v
alu
es
to
co
n
s
tr
u
ct
a
lo
ca
l
ex
p
lan
atio
n
m
o
d
el.
T
h
e
lo
ca
l
e
x
p
lan
atio
n
m
o
d
el
is
a
weig
h
ted
lin
ea
r
r
eg
r
ess
io
n
co
n
s
tr
u
cted
f
r
o
m
a
b
ac
k
g
r
o
u
n
d
d
ataset
an
d
a
s
am
p
le
o
f
p
o
ten
tial
f
ea
t
u
r
e
c
o
alitio
n
s
in
th
e
d
ata.
SH
AP
s
p
ec
if
y
ex
p
lan
atio
n
as
in
(
1
)
.
I
n
th
is
co
n
tex
t,
(
´
)
d
en
o
tes
th
e
ex
p
la
n
ato
r
y
m
o
d
el,
Φ
∈
r
ep
r
esen
ts
th
e
Sh
ap
ley
v
alu
es
f
o
r
f
ea
t
u
r
e
,
a
n
d
ea
ch
´
s
ig
n
if
ies
a
s
im
p
lifie
d
v
alu
e
o
f
th
e
in
p
u
t
f
ea
tu
r
es.
´
∈
{
0
,
1
}
r
ep
r
esen
ts
th
e
co
alitio
n
v
ec
t
o
r
o
f
th
e
m
ax
im
al
co
alitio
n
s
ize
.
I
n
th
is
co
n
tex
t,
´
=1
in
d
icate
s
th
e
p
r
esen
ce
o
f
f
ea
tu
r
e
in
s
id
e
th
e
co
alitio
n
,
wh
ile
its
b
in
ar
y
n
eg
atio
n
,
´
=0
,
s
ig
n
if
ies th
e
lack
o
f
f
ea
tu
r
e
.
T
h
e
Sh
ap
ley
v
alu
e
Φ
c
an
b
e
co
m
p
u
ted
in
(
2
)
.
(
z
´
)
=
∅
0
+
∑
∅
z
´
=
1
(
1
)
∅
=
∑
|
|
!
(
|
|
−
|
|
−
1
)
!
|
|
!
[
∪
{
}
⊑
(
∪
{
}
)
−
(
)
]
(
2
)
L
et
r
ep
r
esen
t
th
e
v
alu
es
o
f
th
e
in
p
u
t
f
ea
tu
r
es
with
in
th
e
f
ea
tu
r
e
s
u
b
s
ets
s
et
,
wh
er
e
all
ar
e
s
u
b
s
ets
o
f
,
with
d
en
o
tin
g
th
e
co
m
p
lete
s
et
o
f
f
ea
tu
r
e
s
.
A
m
o
d
el
∪
{
}
is
tr
ain
ed
with
th
e
f
ea
tu
r
e
in
clu
d
ed
,
wh
er
ea
s
a
s
ec
o
n
d
m
o
d
el
is
tr
ain
ed
with
o
u
t
th
e
f
ea
tu
r
e.
T
h
e
ex
p
r
ess
io
n
[
∪
{
}(
∪
{
})
−
(
)
]
s
er
v
es
to
co
m
p
a
r
e
th
e
p
r
ed
ictio
n
s
o
f
th
e
t
wo
m
o
d
els.
Sin
ce
th
e
im
p
a
ct
o
f
e
x
clu
d
in
g
a
ch
ar
ac
ter
is
tic
d
ep
en
d
s
o
n
o
th
e
r
f
ea
tu
r
es,
th
e
a
f
o
r
em
e
n
tio
n
ed
d
if
f
er
en
ce
s
ar
e
c
o
m
p
u
ted
f
o
r
a
ll v
iab
le
s
u
b
s
ets.
3
.
2
.
M
o
del
T
h
e
SW
aT
d
ataset
co
m
p
r
is
es
a
n
o
r
m
al
(
tr
ai
n
)
an
d
an
attac
k
(
test
)
d
ataset,
with
th
e
latter
co
n
tain
in
g
b
o
th
n
o
r
m
al
an
d
attac
k
p
o
in
ts
.
T
h
e
d
ataset
is
p
r
e
-
p
r
o
ce
s
s
ed
,
an
d
th
e
n
o
r
m
al
d
ataset
is
u
tili
ze
d
f
o
r
t
r
ain
in
g
th
e
T
C
AE
m
o
d
el.
T
h
e
attac
k
d
at
aset
is
u
s
ed
to
g
en
er
ate
p
r
e
d
i
ctio
n
s
.
T
h
e
u
n
d
er
ly
in
g
p
r
em
i
s
e
is
th
at
a
T
C
AE
tr
ain
ed
o
n
a
n
o
r
m
al
d
ataset
will
h
av
e
h
ig
h
er
r
ec
o
n
s
tr
u
ctio
n
l
o
s
s
f
o
r
an
o
m
alo
u
s
d
ata
p
o
i
n
ts
.
Su
b
s
eq
u
en
tly
,
we
ev
alu
ated
th
e
m
o
d
els'
p
r
ed
ictio
n
s
an
d
u
tili
ze
d
XAI
m
eth
o
d
s
to
clar
if
y
th
e
o
u
tc
o
m
es
r
elate
d
to
t
h
e
id
e
n
tifie
d
an
o
m
alies.
T
h
e
f
o
llo
win
g
s
ec
tio
n
s
o
u
tlin
e
th
e
d
ata
p
r
ep
r
o
ce
s
s
in
g
,
co
n
f
i
g
u
r
atio
n
o
f
th
e
an
o
m
al
y
d
etec
tio
n
m
o
d
el,
an
d
th
e
s
etu
p
o
f
XAI
.
3
.
2
.
1
.
Da
t
a
p
re
pro
ce
s
s
ing
T
h
is
s
tu
d
y
f
o
cu
s
ed
p
r
im
ar
il
y
o
n
attac
k
s
o
n
th
e
SW
aT
d
atasets
.
L
ab
els
ar
e
elim
i
n
ated
f
o
r
u
n
s
u
p
er
v
is
ed
p
r
o
ce
s
s
in
g
,
an
d
th
e
co
lu
m
n
s
ar
e
tr
a
n
s
f
o
r
m
ed
to
f
lo
ats
an
d
n
o
r
m
alize
d
with
a
m
in
-
m
ax
s
ca
ler
.
A
12
-
len
g
t
h
lo
ca
l
co
n
tex
tu
al
wi
n
d
o
w
is
u
s
ed
to
c
o
n
v
er
t
t
h
e
t
im
e
s
er
ies
to
a
s
lid
in
g
win
d
o
w
W
={
W
1
,
W
2
,
.
.
.
,
W
t}.
T
h
e
en
tire
tr
ain
in
g
win
d
o
w
s
ize
is
4
9
4
9
8
8
(
1
2
,
5
1
)
,
w
h
ile
th
e
test
win
d
o
w
s
ize
is
4
4
9
9
0
7
(
1
2
,
5
1
)
.
I
n
th
is
ca
s
e,
1
2
r
ep
r
esen
ts
th
e
wi
n
d
o
w
s
ize
,
an
d
5
1
r
ep
r
esen
ts
t
h
e
d
im
en
s
io
n
o
f
th
e
tim
e
s
er
ies.
3
.
2
.
2
.
T
em
po
ra
l c
o
nv
o
lutio
n
a
uto
enco
der
m
o
del
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
T
C
AE
s
h
o
wn
in
Fig
u
r
e
1
em
p
lo
y
s
a
T
C
N
au
to
en
co
d
er
to
ca
p
tu
r
e
n
o
r
m
al
tim
e
s
tam
p
s
an
d
u
s
es
th
is
r
ep
r
esen
tatio
n
to
id
en
tify
ab
n
o
r
m
al
p
a
tter
n
s
th
at
d
ev
iate
f
r
o
m
e
x
p
ec
ted
b
eh
av
io
u
r
.
T
h
e
m
o
d
el
em
p
l
o
y
s
d
ilated
c
o
n
v
o
lu
tio
n
al
lay
er
s
a
n
d
an
ex
p
a
n
s
iv
e
r
ec
ep
tiv
e
f
ield
to
ex
a
m
in
e
th
e
d
ata
ac
r
o
s
s
v
ar
io
u
s
tem
p
o
r
al
s
ca
les.
T
h
e
T
C
AE
m
o
d
el
f
ac
ilit
ates
th
e
co
n
cu
r
r
en
t
tr
ain
in
g
o
f
en
c
o
d
er
s
an
d
d
ec
o
d
er
s
.
E
n
co
d
er
s
ar
e
d
esig
n
ed
to
co
m
p
r
ess
in
p
u
t
tim
e
s
er
ies,
wh
er
ea
s
d
ec
o
d
er
s
ar
e
r
esp
o
n
s
ib
le
f
o
r
r
ec
o
n
s
tr
u
ctin
g
th
em
.
T
h
e
r
ec
o
n
s
tr
u
ctio
n
er
r
o
r
s
er
v
es
as
a
to
o
l
f
o
r
id
en
tify
in
g
an
o
m
al
o
u
s
b
eh
av
io
r
.
T
h
e
en
co
d
er
co
m
p
r
is
es
th
r
ee
tem
p
o
r
al
b
lo
c
k
s
,
ea
ch
u
s
in
g
ca
u
s
al,
d
ilated
1
D
co
n
v
o
lu
tio
n
s
with
a
d
o
u
b
lin
g
d
ilatio
n
s
ch
ed
u
le
q
∈
{1
,
2,
4,
8,
1
6
},
k
er
n
el
s
ize
k
=4
0
,
an
d
4
0
f
ilter
s
p
er
lay
e
r
,
f
o
llo
wed
b
y
r
esid
u
al
c
o
n
n
ec
tio
n
s
f
o
r
s
tab
ilit
y
.
E
ac
h
co
n
v
o
l
u
tio
n
is
f
o
llo
wed
b
y
R
eL
U
ac
tiv
atio
n
a
n
d
weig
h
t
n
o
r
m
aliza
tio
n
;
ch
an
n
el
co
m
p
r
ess
io
n
is
p
er
f
o
r
m
ed
v
i
a
a
1
×1
C
o
n
v
1
d
with
2
0
f
ilter
s
.
T
em
p
o
r
al
d
o
w
n
-
s
am
p
lin
g
u
s
es
av
er
ag
e
p
o
o
lin
g
with
a
s
tr
id
e
2
.
T
h
e
d
ec
o
d
e
r
m
ir
r
o
r
s
th
e
en
co
d
er
: u
p
-
s
am
p
l
in
g
(
s
tr
id
e
2
)
r
esto
r
es
tem
p
o
r
a
l
r
eso
lu
tio
n
,
f
o
llo
wed
b
y
d
ilat
ed
1
D
c
o
n
v
o
lu
tio
n
s
(
k
er
n
el
s
ize
4
0
,
4
0
f
ilter
s
)
an
d
R
eL
U
ac
tiv
atio
n
.
A
f
in
al
1
×1
C
o
n
v
1
d
p
r
o
jects b
ac
k
to
5
1
o
u
tp
u
t
c
h
an
n
els
with
lin
ea
r
ac
tiv
atio
n
to
r
ec
o
n
s
tr
u
ct
th
e
in
p
u
t
win
d
o
w.
W
e
o
p
tim
i
ze
m
ea
n
s
q
u
a
r
ed
er
r
o
r
(
MSE
)
r
ec
o
n
s
tr
u
ctio
n
lo
s
s
o
v
er
th
e
f
u
ll
win
d
o
w
(
1
2
×5
1
)
;
an
o
m
alies
ar
e
s
co
r
ed
as
p
er
-
win
d
o
w
MSE
b
etwe
en
in
p
u
t
an
d
r
ec
o
n
s
tr
u
ctio
n
.
T
h
e
ar
ch
itectu
r
al
d
etails
ar
e
p
r
o
v
id
ed
i
n
T
ab
le
2
.
T
h
e
T
C
A
E
m
o
d
el
is
tr
ain
ed
f
o
r
5
e
p
o
c
h
s
u
s
in
g
th
e
Ad
am
o
p
tim
izer
with
a
lear
n
in
g
r
ate
o
f
0
.
0
0
1
an
d
h
eld
o
u
t
1
0
%
o
f
th
e
n
o
r
m
al
tr
ain
in
g
win
d
o
ws
as
a
v
alid
atio
n
s
p
lit;
all
co
n
v
o
lu
tio
n
al
k
er
n
els
wer
e
in
itialized
with
Glo
r
o
t
n
o
r
m
al.
T
ab
le
3
s
u
m
m
a
r
izes
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
ag
ain
s
t e
x
is
tin
g
ap
p
r
o
ac
h
es.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
14
,
No
.
4
,
Dec
em
b
er
20
25
:
1
4
2
0
-
1
4
3
2
1424
Fig
u
r
e
1
.
T
h
e
p
r
o
p
o
s
ed
T
C
AE
m
o
d
el
T
ab
le
2
.
Su
m
m
a
r
y
o
f
T
C
AE
m
o
d
el
ar
ch
itectu
r
e
an
d
h
y
p
er
p
ar
am
eter
s
ettin
g
s
C
o
m
p
o
n
e
n
t
S
e
t
t
i
n
g
C
o
n
v
o
l
u
t
i
o
n
a
l
d
e
p
t
h
3
e
n
c
o
d
e
r
b
l
o
c
k
s
+
3
d
e
c
o
d
e
r
b
l
o
c
k
s
K
e
r
n
e
l
si
z
e
s
4
0
(
d
i
l
a
t
e
d
c
o
n
v
)
,
1
(
1
×
1
c
o
n
v
)
D
i
l
a
t
i
o
n
s
1
,
2
,
4
,
8
,
1
6
(
c
a
u
s
a
l
)
F
i
l
t
e
r
s
4
0
(
t
e
m
p
o
r
a
l
c
o
n
v
)
,
2
0
(
1
×
1
c
o
m
p
r
e
ss
i
o
n
)
,
5
1
o
u
t
p
u
t
A
c
t
i
v
a
t
i
o
n
s
R
e
LU
(
i
n
t
e
r
m
e
d
i
a
t
e
)
,
Li
n
e
a
r
(
o
u
t
p
u
t
)
P
o
o
l
i
n
g
/
u
p
-
s
a
mp
l
i
n
g
A
v
g
p
o
o
l
s
=
2
/
N
e
a
r
e
st
u
p
s
a
m
p
l
e
s=
2
N
o
r
mal
i
z
a
t
i
o
n
W
e
i
g
h
t
n
o
r
m
a
l
i
z
a
t
i
o
n
(
t
e
m
p
o
r
a
l
c
o
n
v
s)
R
e
si
d
u
a
l
s
P
e
r
-
b
l
o
c
k
i
d
e
n
t
i
t
y
sk
i
p
s
Lo
ss
M
S
E
o
v
e
r
(
1
2
×
5
1
)
w
i
n
d
o
w
T
ab
le
3
.
Mo
d
el
p
er
f
o
r
m
a
n
ce
M
o
d
e
l
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
D
A
G
M
M
0
.
4
6
9
5
0
.
6
6
5
9
0
.
5
5
0
7
LSTM
-
NDT
0
.
7
7
7
7
0
.
5
1
0
8
0
.
6
1
6
6
U
S
A
D
0
.
7
4
8
8
0
.
5
9
4
5
0
.
6
6
2
7
LSTM
-
AE
0
.
8
1
2
0
0
.
5
7
8
0
0
.
6
7
4
0
I
so
l
a
t
i
o
n
f
o
r
e
st
0
.
6
9
0
0
0
.
4
6
0
0
0
.
5
5
2
0
P
C
A
0
.
6
2
0
0
0
.
4
3
0
0
0
.
5
0
5
0
TC
A
E
(
O
u
r
s)
0
.
9
4
3
5
0
.
6
1
3
6
0
.
7
4
3
6
3
.
3
.
Ano
m
a
ly
d
et
ec
t
io
n
Fo
r
id
en
tif
y
in
g
an
o
m
al
o
u
s
a
ttack
p
o
in
ts
,
th
e
r
ec
o
n
s
tr
u
cti
o
n
lo
s
s
is
id
en
tifie
d
o
n
t
h
e
test
d
ata.
T
h
e
m
o
d
el
is
tr
ain
ed
ex
clu
s
iv
ely
o
n
n
o
r
m
al
tim
e
s
er
ies
d
u
r
in
g
th
e
tr
ain
in
g
p
h
ase.
T
h
e
T
C
AE
m
o
d
el
is
b
ased
o
n
th
e
r
atio
n
ale
th
at
th
e
m
o
d
e
l
tr
ain
ed
o
n
n
o
r
m
al
d
ata
p
o
in
t
s
will
h
av
e
lo
w
r
ec
o
n
s
tr
u
ctio
n
lo
s
s
,
b
ased
o
n
th
e
u
n
d
er
s
tan
d
i
n
g
th
at
t
h
e
au
to
e
n
co
d
er
s
h
o
u
ld
ac
c
u
r
ately
r
ec
o
n
s
tr
u
ct
n
o
r
m
al
d
ata
p
o
in
ts
in
th
e
tim
e
s
er
ies.
W
h
en
T
C
AE
id
en
tifie
s
p
atter
n
s
th
at
d
ev
iate
n
o
tab
ly
f
r
o
m
th
e
n
o
r
m
,
we
ex
p
ec
t
to
s
ee
a
r
is
e
in
r
ec
o
n
s
tr
u
ctio
n
lo
s
s
.
T
o
id
e
n
tify
an
o
m
al
o
u
s
d
ata
p
o
i
n
ts
,
we
em
p
lo
y
e
d
a
m
u
lti
-
s
tep
ap
p
r
o
ac
h
th
at
co
m
b
in
es
r
ec
o
n
s
tr
u
ctio
n
lo
s
s
ca
lcu
lated
u
s
in
g
a
T
C
AE
an
d
d
e
n
s
ity
-
b
ased
clu
s
ter
in
g
t
h
r
o
u
g
h
DB
SC
AN.
T
o
ef
f
ec
tiv
ely
i
d
en
tify
an
o
m
a
lies
,
we
lev
er
ag
e
d
a
co
m
b
in
at
io
n
o
f
k
er
n
el
d
en
s
ity
esti
m
atio
n
(
KDE
)
an
d
DB
SC
A
N
clu
s
ter
in
g
.
First,
Gau
s
s
ian
_
k
d
e
was
ap
p
lied
to
th
e
r
ec
o
n
s
tr
u
ctio
n
lo
s
s
v
alu
es,
p
r
o
d
u
cin
g
a
d
en
s
ity
esti
m
ate
th
at
h
ig
h
lig
h
t
s
th
e
u
n
d
er
ly
in
g
d
is
tr
ib
u
tio
n
o
f
th
e
d
ata.
Su
b
s
eq
u
e
n
tly
,
th
e
DB
S
C
AN
alg
o
r
ith
m
was
em
p
lo
y
ed
to
g
r
o
u
p
d
ata
p
o
in
ts
b
ased
o
n
th
eir
d
en
s
ity
.
T
h
is
ap
p
r
o
ac
h
allo
wed
f
o
r
t
h
e
id
en
tific
atio
n
o
f
an
o
m
alies
as
p
o
in
ts
lab
eled
as
-
1
b
y
DB
SC
AN,
co
r
r
esp
o
n
d
in
g
to
s
p
ar
s
e
r
eg
io
n
s
in
t
h
e
d
ata
d
is
tr
ib
u
tio
n
.
A
v
is
u
aliza
tio
n
o
f
t
h
e
r
esu
lts
d
ep
icted
a
n
o
m
alies
in
r
e
d
a
n
d
n
o
n
-
an
o
m
alo
u
s
p
o
in
ts
i
n
b
lu
e,
p
r
o
v
id
in
g
a
clea
r
d
is
tin
ctio
n
b
etwe
en
cl
u
s
ter
s
an
d
o
u
tlier
s
,
as
s
h
o
wn
in
Fi
g
u
r
e
2
.
T
h
is
m
eth
o
d
n
o
t
o
n
l
y
id
en
tifie
d
o
u
tlier
s
ef
f
ec
tiv
ely
b
u
t
also
f
ac
ilit
ated
th
e
q
u
an
tific
atio
n
o
f
tim
esta
m
p
s
an
d
attac
k
p
o
in
ts
th
at
ar
e
m
ap
p
ed
to
d
etec
ted
an
o
m
alies
to
th
e
co
r
r
esp
o
n
d
i
n
g
attac
k
p
er
io
d
.
T
h
e
attac
k
s
in
th
e
d
ataset
wer
e
m
ap
p
ed
to
th
e
an
o
m
alies
f
lag
g
ed
b
y
th
e
m
o
d
el.
T
h
is
m
ap
p
in
g
allo
wed
u
s
to
m
e
asu
r
e
d
etec
tio
n
p
er
f
o
r
m
a
n
ce
b
y
co
m
p
a
r
in
g
th
e
id
en
tifie
d
an
o
m
alies
with
th
e
k
n
o
wn
attac
k
in
s
tan
ce
s
.
T
h
is
m
eth
o
d
o
l
o
g
y
en
ab
les
p
r
e
cise
ev
alu
atio
n
o
f
an
o
m
aly
d
etec
tio
n
m
o
d
els
b
y
p
r
o
v
id
in
g
in
s
ig
h
ts
in
to
m
o
d
el
p
er
f
o
r
m
an
ce
in
co
r
r
ec
tly
id
en
tify
in
g
attac
k
p
er
io
d
s
with
in
th
e
d
ata.
Fig
u
r
e
3
p
r
esen
ts
th
ese
o
u
tp
u
ts
an
d
h
ig
h
lig
h
ts
th
e
p
r
ec
is
io
n
wit
h
wh
ich
th
e
m
o
d
el
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l Sci
I
SS
N:
2252
-
8
8
1
4
Un
ve
ilin
g
a
n
o
ma
lies
in
in
d
u
s
t
r
ia
l c
o
n
tr
o
l sys
tems:
a
ke
r
n
el
S
HA
P
-
b
a
s
ed
a
p
p
r
o
a
ch
…
(
S
a
n
g
ee
ta
Osw
a
l
)
1425
id
en
tifie
s
attac
k
s
with
in
th
e
d
ataset.
B
y
p
in
p
o
in
tin
g
th
e
ex
a
ct
s
tar
t
an
d
en
d
in
d
ices
o
f
ea
ch
attac
k
,
it
b
ec
o
m
es
p
o
s
s
ib
le
to
an
aly
ze
th
e
ch
ar
a
cter
is
tics
o
f
th
e
d
etec
ted
an
o
m
alo
u
s
s
eg
m
en
ts
an
d
d
eter
m
in
e
th
e
u
n
d
e
r
ly
in
g
ca
u
s
es
o
f
th
e
attac
k
s
.
T
h
e
ab
ilit
y
to
m
ap
d
etec
ted
an
o
m
alies
b
ac
k
to
s
p
ec
if
ic
attac
k
p
o
in
ts
is
cr
u
cial
f
o
r
v
alid
atin
g
th
e
m
o
d
el'
s
ef
f
ec
tiv
en
ess
an
d
f
o
r
p
r
o
v
id
in
g
ac
tio
n
ab
le
in
s
ig
h
ts
f
o
r
s
y
s
tem
d
ef
e
n
s
e
s
tr
ateg
ies.
Fig
u
r
e
2
.
T
h
e
r
esu
lt o
f
DB
SC
AN
o
n
r
ec
o
n
s
tr
u
ctio
n
lo
s
s
Fig
u
r
e
3
.
A
s
n
ap
s
h
o
t o
f
d
etec
ted
attac
k
p
o
i
n
ts
3
.
4
.
E
x
pla
ina
ble A
I
us
ing
S
H
AP
W
h
ile
T
C
AE
h
as
b
ee
n
u
s
ed
f
o
r
a
n
o
m
aly
d
etec
tio
n
,
its
laten
t
o
u
tp
u
ts
ar
e
t
y
p
ically
o
p
aq
u
e
to
d
o
m
ain
ex
p
er
ts
.
T
o
ad
d
r
ess
th
is
,
we
in
teg
r
ate
Ker
n
el
SHAP p
o
s
t
h
o
c
to
ex
p
lain
an
o
m
aly
s
co
r
es b
y
attr
ib
u
tin
g
th
em
to
s
p
ec
if
ic
in
p
u
t
f
ea
tu
r
es
a
n
d
tim
e
s
tep
s
.
T
h
is
f
u
s
io
n
en
ab
les
te
m
p
o
r
al
in
ter
p
r
etab
ilit
y
,
allo
wi
n
g
an
aly
s
ts
to
tr
ac
e
an
o
m
alies
b
ac
k
to
co
n
tr
ib
u
tin
g
s
en
s
o
r
s
.
O
n
ce
th
e
an
o
m
aly
d
etec
tio
n
p
r
o
ce
s
s
id
en
tifie
d
at
tack
win
d
o
ws
an
d
th
eir
r
esp
ec
tiv
e
in
d
ices,
SHA
P
was
em
p
lo
y
ed
to
in
ter
p
r
et
th
e
f
ea
tu
r
es
co
n
tr
ib
u
tin
g
to
ea
ch
id
en
tifie
d
attac
k
.
T
o
s
tr
ea
m
lin
e
c
o
m
p
u
tatio
n
,
t
h
e
b
ac
k
g
r
o
u
n
d
d
ata
was
s
u
m
m
ar
ized
u
s
in
g
K
-
m
ea
n
s
clu
s
ter
in
g
,
r
ed
u
ci
n
g
t
h
e
d
ataset
to
1
0
0
r
ep
r
esen
tativ
e
clu
s
ter
s
(
K=
1
0
0
)
.
T
h
is
clu
s
ter
in
g
tech
n
iq
u
e
ef
f
ec
ti
v
ely
ca
p
t
u
r
ed
th
e
u
n
d
er
ly
in
g
f
ea
tu
r
e
s
p
ac
e
o
f
t
h
e
n
o
r
m
al
win
d
o
ws
(
win
d
o
ws_
n
o
r
m
al)
,
e
n
s
u
r
in
g
co
m
p
u
tatio
n
al
e
f
f
icie
n
cy
an
d
m
ea
n
in
g
f
u
l
b
aselin
e
co
m
p
ar
is
o
n
s
.
E
ac
h
attac
k
win
d
o
w
was
f
la
tten
ed
f
r
o
m
its
o
r
i
g
in
al
m
u
lt
id
im
en
s
io
n
al
f
o
r
m
at
(
1
2
-
tim
e
s
tep
s
×5
1
f
ea
tu
r
es)
in
to
a
s
in
g
le
-
d
im
en
s
io
n
al
in
p
u
t
f
o
r
SHAP
ca
lcu
latio
n
s
.
SHAP
v
alu
es
wer
e
co
m
p
u
ted
f
o
r
s
elec
ted
attac
k
win
d
o
ws
(
win
d
o
w
in
d
ices)
to
q
u
an
tif
y
th
e
f
ea
t
u
r
e
im
p
o
r
tan
ce
f
o
r
ea
ch
a
n
o
m
alo
u
s
in
s
tan
ce
.
Fo
r
in
ter
p
r
etab
ilit
y
,
th
e
SHAP
v
alu
es
wer
e
r
esh
ap
ed
b
ac
k
to
th
eir
o
r
ig
in
al
d
im
en
s
io
n
s
(
1
2
×5
1
)
,
an
d
th
e
m
ea
n
SHAP
v
alu
es
ac
r
o
s
s
all
an
aly
z
ed
win
d
o
ws
wer
e
ca
lcu
lated
t
o
h
ig
h
lig
h
t
f
ea
tu
r
es
with
s
ig
n
i
f
ican
t
co
n
tr
ib
u
tio
n
s
to
th
e
an
o
m
alies.
T
h
e
r
esu
lts
wer
e
v
is
u
alize
d
u
s
in
g
f
o
r
ce
p
l
o
ts
an
d
v
io
lin
ch
a
r
ts
,
p
r
o
v
id
in
g
a
d
etailed
r
an
k
in
g
o
f
f
ea
tu
r
es
b
ased
o
n
th
eir
m
e
an
SHAP
v
alu
es.
T
h
is
m
eth
o
d
p
in
p
o
in
ts
th
e
m
o
s
t
in
f
lu
en
ti
al
elem
en
ts
o
f
ea
c
h
attac
k
to
ex
p
lain
o
b
s
er
v
e
d
an
o
m
alies
an
d
im
p
r
o
v
e
t
h
e
an
o
m
aly
d
etec
tio
n
m
o
d
el's
in
ter
p
r
etab
ilit
y
.
Su
ch
in
s
ig
h
ts
f
ac
ilit
ate
a
d
e
ep
er
u
n
d
er
s
tan
d
in
g
o
f
s
y
s
tem
v
u
l
n
er
ab
ilit
ies
an
d
p
a
v
e
t
h
e
way
f
o
r
tar
g
eted
m
itig
atio
n
.
T
h
e
co
m
p
lete
wo
r
k
f
lo
w
u
s
ed
is
d
escr
ib
ed
in
Fig
u
r
e
4
.
Acc
o
r
d
in
g
to
Fig
u
r
e
4
,
th
e
p
r
o
ce
s
s
o
f
an
o
m
aly
id
e
n
tific
atio
n
an
d
th
e
u
s
e
o
f
XAI
in
t
h
e
id
en
tifie
d
attac
k
win
d
o
w
is
d
escr
ib
ed
i
n
d
etail.
T
o
ad
ap
t
Ker
n
el
SHAP
to
s
eq
u
en
tial
T
C
AE
em
b
ed
d
in
g
s
,
ea
ch
an
o
m
alo
u
s
win
d
o
w
(
1
2
×5
1
)
is
f
ir
s
t
f
latten
ed
in
to
a
6
1
2
-
d
im
e
n
s
io
n
al
v
ec
to
r
.
W
e
u
s
e
th
e
K=
1
0
0
clu
s
ter
ce
n
tr
o
id
s
(
f
r
o
m
th
e
n
o
r
m
al
win
d
o
ws)
as
b
ac
k
g
r
o
u
n
d
s
am
p
les,
an
d
SHAP
in
ter
n
ally
g
en
er
ates
ap
p
r
o
x
im
ately
9
,
4
0
0
co
alitio
n
s
am
p
les
p
er
in
s
tan
ce
to
f
it
th
e
lo
ca
l
lin
ea
r
ex
p
lan
atio
n
m
o
d
el.
E
ac
h
SHAP
ca
ll
r
e
q
u
ir
es
o
n
e
f
o
r
war
d
p
ass
p
er
co
alitio
n
s
am
p
le,
s
o
ex
p
lain
in
g
a
s
in
g
le
win
d
o
w
e
n
tails
r
o
u
g
h
ly
9
,
4
0
0
m
o
d
el
e
v
alu
atio
n
s
.
On
o
u
r
h
ar
d
war
e
(
a
s
in
g
le
NVI
DI
A
A1
0
0
GPU)
,
co
m
p
u
tin
g
SHAP v
alu
es f
o
r
o
n
e
win
d
o
w
tak
es a
p
p
r
o
x
im
ately
4
0
s
ec
o
n
d
s
.
3
.
5
.
Deplo
y
m
ent
co
nte
x
t
I
n
its
cu
r
r
en
t
f
o
r
m
,
th
e
T
C
AE
+
k
er
n
el
SHAP
p
ip
elin
e
r
u
n
s
o
f
f
lin
e
in
a
Py
th
o
n
e
n
v
ir
o
n
m
en
t,
b
u
t
it
is
ar
ch
itected
as
a
s
tan
d
alo
n
e
in
f
er
en
ce
s
er
v
ice
f
o
r
I
C
S
n
etwo
r
k
s
.
I
n
a
p
r
o
d
u
ctio
n
s
ettin
g
,
t
h
e
tr
ain
ed
m
o
d
el
a
n
d
ex
p
lain
er
ar
e
p
ac
k
ag
e
d
(
in
a
D
o
ck
er
c
o
n
tain
er
)
an
d
d
ep
lo
y
ed
o
n
o
n
-
p
r
em
is
e
s
er
v
er
s
o
r
ed
g
e
g
atew
ay
s
th
at
alr
ea
d
y
r
ec
eiv
e
liv
e
s
en
s
o
r
an
d
ac
tu
ato
r
s
tr
ea
m
s
.
T
h
e
s
er
v
ice
co
n
tin
u
o
u
s
ly
in
g
ests
tim
estam
p
ed
m
ea
s
u
r
em
en
ts
,
co
m
p
u
tes
p
er
-
win
d
o
w
r
ec
o
n
s
tr
u
ctio
n
lo
s
s
es
to
f
lag
a
n
o
m
alies,
a
n
d
i
n
v
o
k
es
Ker
n
el
SHAP
to
p
r
o
d
u
ce
r
an
k
ed
,
tim
e
-
s
tep
–
lev
el
attr
ib
u
tio
n
s
.
Aler
ts
co
n
tain
i
n
g
th
e
a
n
o
m
aly
s
co
r
e
an
d
co
n
tr
ib
u
tin
g
v
ar
iab
les
ca
n
b
e
em
itted
as
J
SON
o
v
er
R
E
ST
API
s
o
r
m
ess
ag
e
q
u
eu
es
in
to
ex
is
tin
g
m
o
n
ito
r
in
g
d
ash
b
o
ar
d
s
o
r
alar
m
s
y
s
tem
s
.
Secu
r
ity
an
d
co
m
p
l
ian
ce
alig
n
with
NI
ST
SP
8
0
0
-
8
2
an
d
I
E
C
6
2
4
4
3
b
y
e
n
f
o
r
cin
g
en
cr
y
p
ted
telem
etr
y
ch
a
n
n
els,
r
o
le
-
b
ase
d
ac
ce
s
s
co
n
tr
o
ls
,
an
d
im
m
u
t
ab
le
au
d
it
lo
g
s
f
o
r
all
d
etec
ti
o
n
a
n
d
ex
p
lan
atio
n
ev
en
ts
.
Fu
tu
r
e
wo
r
k
will
v
alid
ate
liv
e
test
b
ed
in
teg
r
atio
n
,
b
e
n
ch
m
ar
k
en
d
-
to
-
e
n
d
laten
cy
a
n
d
th
r
o
u
g
h
p
u
t,
a
n
d
q
u
an
tify
t
h
e
im
p
ac
t o
f
ex
p
lain
ab
le
aler
ts
o
n
o
p
er
ato
r
d
ec
is
io
n
cy
cles th
r
o
u
g
h
c
o
n
tr
o
lled
u
s
er
s
tu
d
ies.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
14
,
No
.
4
,
Dec
em
b
er
20
25
:
1
4
2
0
-
1
4
3
2
1426
Fig
u
r
e
4
.
Flo
wch
ar
t
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
an
d
XAI
m
eth
o
d
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
I
n
th
is
s
ec
tio
n
,
th
e
r
esu
lts
ar
e
p
r
esen
ted
.
W
h
ile
th
e
p
r
o
p
o
s
ed
m
eth
o
d
o
f
f
er
s
h
ig
h
-
r
eso
lu
tio
n
attr
ib
u
tio
n
with
in
id
e
n
tifie
d
attac
k
win
d
o
ws,
its
in
ter
p
r
et
ab
ilit
y
is
lim
ited
b
y
th
e
tem
p
o
r
al
g
r
a
n
u
lar
ity
o
f
SHAP
ex
p
lan
atio
n
s
;
s
en
s
o
r
in
ter
ac
tio
n
s
o
u
ts
id
e
th
e
s
elec
ted
win
d
o
w
m
ay
b
e
u
n
d
er
r
ep
r
ese
n
ted
.
Ad
d
itio
n
ally
,
m
o
d
el
d
r
if
t
p
o
s
es
a
ch
allen
g
e
in
e
v
o
lv
in
g
I
C
S
en
v
ir
o
n
m
en
t
s
,
wh
er
e
s
h
if
ts
in
s
en
s
o
r
b
e
h
a
v
io
r
o
r
o
p
e
r
atio
n
al
p
atter
n
s
m
ay
d
eg
r
a
d
e
d
etec
tio
n
p
er
f
o
r
m
an
ce
o
v
er
tim
e.
W
e
d
is
cu
s
s
in
d
etail
two
at
ta
ck
s
d
etec
ted
b
y
o
u
r
s
y
s
tem
an
d
its
SHAP v
alu
es t
o
id
en
tify
t
h
e
f
ea
tu
r
es c
o
n
tr
ib
u
tin
g
to
th
e
an
o
m
alies.
4
.
1
.
Secure
wa
t
er
t
re
a
t
m
ent
d
a
t
a
s
et
T
h
e
r
esear
ch
wo
r
k
was
co
n
d
u
cted
o
n
SW
aT
,
a
wate
r
tr
ea
tm
en
t
p
lan
t
d
ev
elo
p
ed
b
y
iTr
u
s
t Sin
g
ap
o
r
e
,
to
ad
v
a
n
ce
r
esear
c
h
in
cy
b
er
-
p
h
y
s
ical
s
y
s
tem
s
[
3
9
]
.
T
h
e
S
W
aT
d
ataset
co
n
s
is
ts
o
f
6
s
tag
es,
P1
to
P6
,
with
v
ar
io
u
s
s
en
s
o
r
s
an
d
ac
tu
ato
r
s
as
d
escr
ib
ed
in
T
ab
le
4
.
I
t
in
clu
d
es
5
1
s
y
n
c
h
r
o
n
ized
v
ar
i
ab
les
(
2
5
s
en
s
o
r
s
,
2
6
ac
tu
ato
r
s
)
co
v
er
in
g
f
lo
w,
lev
el,
p
r
ess
u
r
e,
ch
em
ical
an
al
y
ze
r
s
,
p
u
m
p
s
,
m
o
to
r
ized
v
alv
es,
an
d
UV
u
n
its
,
r
ec
o
r
d
e
d
at
a
s
am
p
lin
g
r
ate
o
f
1
Hz.
A
s
er
ies o
f
attac
k
s
w
as
lau
n
ch
ed
o
n
SW
aT
to
d
is
tu
r
b
its
n
o
r
m
al
o
p
er
atio
n
.
T
h
e
attac
k
s
ca
r
r
ied
o
n
th
e
SW
aT
d
atasets
ar
e
d
escr
ib
ed
in
T
ab
le
5
a
n
d
class
if
ied
as
s
in
g
le
p
o
in
t
(
SP
)
an
d
m
u
lti
-
p
o
in
t
(
MP)
.
I
n
a
n
SP
attac
k
,
th
e
attac
k
er
m
an
ip
u
late
s
o
n
e
s
tate
v
ar
iab
le,
wh
er
ea
s
i
n
an
MP
attac
k
,
m
o
r
e
th
an
o
n
e
s
tate
v
ar
iab
les
ar
e
co
m
p
r
o
m
is
ed
an
d
th
e
co
r
r
esp
o
n
d
in
g
m
ea
s
u
r
e
m
en
ts
ar
e
s
p
o
o
f
ed
.
T
h
e
d
ataset
co
n
tain
s
4
1
d
o
cu
m
e
n
ted
attac
k
s
ce
n
ar
io
s
s
p
an
n
in
g
s
in
g
le
-
an
d
m
u
lti
-
s
tag
e
as
well
as
s
in
g
le
-
an
d
m
u
lti
-
p
o
i
n
t
m
an
ip
u
latio
n
s
;
o
u
r
p
r
o
p
o
s
e
d
m
o
d
el
co
r
r
ec
tly
d
etec
ted
3
1
o
f
th
ese
attac
k
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l Sci
I
SS
N:
2252
-
8
8
1
4
Un
ve
ilin
g
a
n
o
ma
lies
in
in
d
u
s
t
r
ia
l c
o
n
tr
o
l sys
tems:
a
ke
r
n
el
S
HA
P
-
b
a
s
ed
a
p
p
r
o
a
ch
…
(
S
a
n
g
ee
ta
Osw
a
l
)
1427
T
ab
le
4
.
T
h
e
SW
aT
d
ataset
s
e
n
s
o
r
s
an
d
ac
tu
ato
r
s
S
t
a
g
e
s
S
e
n
s
o
r
A
c
t
u
a
t
o
r
P1
LI
T
-
1
0
1
,
F
I
T
-
1
0
1
MV
-
1
0
1
,
P
1
0
1
P2
A
I
T
-
2
0
1
,
A
I
T
-
2
0
2
,
A
I
T
-
2
0
3
,
F
I
T
-
2
0
1
MV
-
2
0
1
,
P
-
2
0
1
,
P
-
2
0
2
,
P
-
2
0
3
,
P
-
2
0
4
,
P
-
2
0
5
,
P
-
2
0
6
P3
D
P
I
T
-
3
0
1
,
F
I
T
-
3
0
1
,
LI
T
-
3
0
1
MV
-
3
0
1
,
M
V
-
3
0
2
,
M
V
-
3
0
3
,
M
V
-
3
0
4
,
P
-
3
0
1
,
P
-
302
P4
A
I
T
-
4
0
1
,
A
I
T
-
4
0
2
,
F
I
T
-
4
0
1
,
LI
T
-
4
0
1
P
-
4
0
1
,
P
-
4
0
2
,
P
-
4
0
3
,
P
-
4
0
4
,
U
V
-
4
0
1
P5
A
I
T
-
5
0
1
,
A
I
T
-
5
0
2
,
A
I
T
-
5
0
3
,
A
I
T
-
5
0
4
,
F
I
T
-
5
0
1
,
F
I
T
-
5
0
2
,
F
I
T
-
5
0
3
,
F
I
T
-
5
0
4
,
P
I
T
-
5
0
1
,
P
I
T
-
5
0
2
,
P
I
T
-
503
P
-
5
0
1
,
P
-
5
0
2
P6
F
I
T
-
601
P
-
6
0
1
,
P
-
6
0
2
,
P
-
6
0
3
T
ab
le
5
.
Attack
s
o
n
SW
aT
d
ataset
Ty
p
e
o
f
a
t
t
a
c
k
N
u
mb
e
r
o
f
a
t
t
a
c
k
s
S
SSP
a
t
t
a
c
k
s
23
S
S
M
P
a
t
t
a
c
k
s
6
M
S
S
P
a
t
t
a
c
k
s
4
M
S
M
P
a
t
t
a
c
k
s
3
4
.
2
.
Ex
pla
ina
ble A
I
r
esu
lt
s
W
e
p
r
esen
t
th
e
XAI
r
esu
lt
o
n
th
e
attac
k
s
id
en
tifie
d
b
y
o
u
r
m
o
d
el
,
T
C
AE
.
I
n
th
is
r
esear
c
h
wo
r
k
,
we
elab
o
r
ate
on
th
e
f
ea
t
u
r
e
co
n
tr
i
b
u
tio
n
to
ea
ch
id
e
n
tifie
d
attac
k
,
u
n
lik
e
p
r
ev
io
u
s
wo
r
k
,
wh
ic
h
p
r
esen
ts
o
n
ly
th
e
to
p
k
f
ea
tu
r
e
co
n
t
r
ib
u
tio
n
s
.
W
e
p
r
esen
t
in
d
etail
two
attac
k
s
,
Attack
n
u
m
b
er
6
,
wh
ich
is
a
s
in
g
le
-
s
tag
e
s
in
g
le
-
p
o
in
t
attac
k
(
SSSP
)
o
n
th
e
s
en
s
o
r
AI
T
2
0
2
,
a
n
d
a
ttack
n
u
m
b
er
2
2
,
wh
ich
is
a
m
u
lti
-
s
tag
e
m
u
lti
-
p
o
in
t
attac
k
(
MSM
P
)
o
n
s
en
s
o
r
s
an
d
ac
tu
a
to
r
s
in
s
tag
es 4
an
d
5
.
Attack
6
o
b
s
er
v
ed
an
an
o
m
aly
in
th
e
p
ar
am
eter
AI
T
-
2
0
2
,
wh
er
e
its
v
alu
e
ex
ce
ed
ed
7
.
0
5
.
A
co
r
r
ec
tiv
e
ac
tio
n
was
in
itiated
b
y
s
ettin
g
th
e
v
alu
e
o
f
AI
T
-
2
0
2
to
6
,
y
et
n
o
d
r
ain
ag
e
p
r
o
ce
s
s
was
tr
ig
g
er
ed
.
T
h
is
r
esu
lted
in
d
o
wn
s
tr
ea
m
e
f
f
ec
ts
s
u
ch
as
th
e
s
h
u
td
o
wn
o
f
P
-
2
0
3
an
d
a
s
u
b
s
eq
u
e
n
t
ch
an
g
e
in
wate
r
q
u
ality
.
T
o
u
n
d
er
s
tan
d
th
e
f
ea
t
u
r
es
c
o
n
tr
ib
u
tin
g
to
th
is
an
o
m
al
y
,
SHAP
v
alu
es
wer
e
co
m
p
u
ted
f
o
r
th
e
id
en
tifie
d
attac
k
win
d
o
w.
T
h
e
ca
lcu
lated
m
ea
n
SHAP
v
alu
es
r
ev
ea
le
d
th
at
AI
T
-
2
0
2
h
a
d
th
e
h
ig
h
es
t
co
n
tr
ib
u
tio
n
to
th
e
an
o
m
aly
,
with
a
m
ea
n
SHAP
v
alu
e
o
f
9
.
5
6
5
6
2
6
e
-
0
3
,
f
o
llo
wed
b
y
P
-
2
0
3
(
7
.
0
1
6
7
1
5
e
-
0
4
)
,
an
d
o
th
er
f
ea
tu
r
es
s
u
ch
as
P
-
6
0
2
,
FIT
-
6
0
1
,
an
d
P
-
2
0
5
,
wh
ich
ex
h
ib
ited
s
ig
n
if
ican
tly
s
m
aller
co
n
tr
ib
u
tio
n
s
.
T
h
is
s
u
g
g
ests
th
at
th
e
an
o
m
alo
u
s
b
eh
av
io
r
in
A
I
T
-
2
0
2
was
th
e
p
r
im
a
r
y
d
r
iv
er
o
f
th
is
attac
k
s
ce
n
ar
io
,
wh
ile
th
e
in
f
l
u
en
ce
o
f
o
th
er
f
ea
tu
r
es wa
s
n
eg
lig
ib
le
i
n
co
m
p
a
r
is
o
n
.
T
h
e
SHAP
an
aly
s
is
was
in
s
tr
u
m
en
tal
in
q
u
an
tify
i
n
g
t
h
e
f
ea
tu
r
e
im
p
o
r
ta
n
ce
f
o
r
t
h
e
d
etec
ted
an
o
m
aly
.
B
y
r
an
k
in
g
t
h
e
m
ea
n
SHAP
v
alu
es,
it
b
ec
am
e
ev
id
en
t
t
h
at
AI
T
-
2
0
2
'
s
d
ev
iatio
n
was
d
ir
ec
tl
y
co
r
r
elate
d
with
th
e
o
b
s
er
v
ed
im
p
ac
t
o
n
th
e
s
y
s
tem
.
Vis
u
aliza
tio
n
o
f
th
e
f
ea
t
u
r
e
co
n
tr
ib
u
tio
n
s
u
s
in
g
SHAP
f
o
r
ce
p
l
o
ts
in
Fig
u
r
e
5
an
d
v
i
o
lin
ch
ar
ts
in
Fig
u
r
e
6
p
r
o
v
id
ed
f
u
r
t
h
er
clar
ity
o
n
th
e
s
ig
n
i
f
ican
ce
o
f
A
I
T
-
2
0
2
an
d
its
r
elatio
n
s
h
ip
to
o
th
er
f
ea
tu
r
es
d
u
r
i
n
g
th
e
attac
k
.
T
h
is
in
ter
p
r
etab
ilit
y
h
ig
h
lig
h
ts
th
e
cr
itical
r
o
le
o
f
AI
T
-
2
0
2
in
th
e
attac
k
d
y
n
a
m
ics
an
d
u
n
d
er
s
co
r
es
th
e
n
e
ed
f
o
r
e
n
h
an
ce
d
m
o
n
ito
r
in
g
o
f
th
is
p
ar
am
eter
t
o
m
itig
ate
s
im
ilar
in
cid
en
ts
in
th
e
f
u
tu
r
e.
Fig
u
r
e
7
p
r
esen
ts
th
e
SHAP
v
alu
e
h
ea
tm
ap
an
d
s
en
s
o
r
-
wis
e
attr
ib
u
tio
n
p
lo
t,
v
is
u
ally
r
ein
f
o
r
cin
g
th
e
d
o
m
in
an
t r
o
le
o
f
AI
T
-
2
0
2
an
d
h
ig
h
lig
h
tin
g
th
e
r
el
ativ
e
in
s
ig
n
if
ican
ce
o
f
o
th
e
r
f
ea
tu
r
es d
u
r
in
g
a
ttack
6
.
Attack
2
3
in
v
o
lv
es
an
o
m
alies
in
t
h
e
p
ar
am
eter
s
UV
-
4
0
1
,
AI
T
-
5
0
2
,
a
n
d
P
-
5
0
1
.
T
h
e
s
ce
n
ar
io
was
ch
ar
ac
ter
ized
b
y
th
e
f
o
llo
win
g
co
n
d
itio
n
s
:
UV
-
4
0
1
was
ac
tiv
e,
AI
T
-
5
0
2
r
ec
o
r
d
ed
a
v
al
u
e
b
elo
w
1
5
0
,
a
n
d
P
-
5
0
1
r
em
ain
e
d
o
p
e
n
.
T
o
in
te
r
p
r
et
th
e
f
ea
tu
r
e
co
n
tr
ib
u
tio
n
s
f
o
r
th
is
an
o
m
aly
,
SHAP
v
alu
es
wer
e
co
m
p
u
ted
f
o
r
th
e
id
en
tifie
d
attac
k
win
d
o
w.
SHAP
s
u
m
m
ar
y
p
lo
ts
an
d
v
is
u
aliza
tio
n
s
p
r
o
v
id
e
d
a
clea
r
r
an
k
in
g
o
f
f
ea
tu
r
e
co
n
tr
ib
u
tio
n
s
,
o
f
f
er
in
g
in
s
ig
h
t
s
in
to
th
e
d
y
n
am
ics
o
f
th
e
a
n
o
m
aly
.
T
h
ese
f
in
d
in
g
s
u
n
d
er
s
c
o
r
e
th
e
n
ec
ess
ity
f
o
r
clo
s
ely
m
o
n
ito
r
in
g
UV
-
4
0
1
a
n
d
P
-
5
0
1
d
u
r
in
g
cr
itical
o
p
er
a
tio
n
s
to
p
r
ev
en
t
r
ec
u
r
r
e
n
ce
an
d
m
itig
ate
p
o
ten
tial
r
is
k
s
.
T
h
e
f
o
r
ce
p
lo
t
s
h
o
wn
in
Fig
u
r
e
8
p
r
esen
ts
th
e
f
ea
tu
r
e
co
n
tr
ib
u
tio
n
to
attac
k
2
3
,
a
n
d
th
e
v
io
lin
p
lo
t
in
Fig
u
r
e
9
s
h
o
ws
th
e
d
is
tr
ib
u
ti
o
n
o
f
SHAP
v
alu
es
f
o
r
ea
ch
f
ea
tu
r
e.
T
h
e
in
ter
p
r
eta
b
ilit
y
p
r
o
v
id
e
d
b
y
SHAP
en
h
an
ce
s
th
e
u
n
d
er
s
tan
d
in
g
o
f
s
y
s
tem
v
u
ln
e
r
ab
ilit
ies
an
d
s
u
p
p
o
r
ts
th
e
d
ev
elo
p
m
en
t
o
f
tar
g
eted
co
u
n
ter
m
ea
s
u
r
es f
o
r
s
im
ilar
s
ce
n
ar
io
s
.
Fig
u
r
e
1
0
s
h
o
ws
t
h
e
SHAP
h
ea
tm
ap
f
o
r
a
ttack
2
3
,
h
ig
h
lig
h
tin
g
th
e
tem
p
o
r
al
im
p
o
r
tan
ce
o
f
f
ea
t
u
r
es
lik
e
AI
T
-
5
0
2
,
P
-
5
0
1
,
an
d
UV
-
4
0
1
.
Fig
u
r
e
1
1
r
an
k
s
s
en
s
o
r
s
b
y
th
eir
m
ea
n
SHAP
v
alu
es,
c
o
n
f
ir
m
in
g
th
ese
as
th
e
to
p
co
n
t
r
ib
u
to
r
s
to
t
h
e
an
o
m
aly
.
U
n
lik
e
e
x
is
tin
g
ap
p
r
o
ac
h
es
th
at
m
e
r
ely
p
r
esen
t
a
r
an
k
ed
lis
t
o
f
t
o
p
-
k
co
n
tr
ib
u
tin
g
f
ea
tu
r
es
with
o
u
t
d
etailed
co
n
tex
t,
o
u
r
m
eth
o
d
o
lo
g
y
p
r
o
v
i
d
es
g
r
an
u
lar
in
s
ig
h
ts
in
to
in
d
iv
id
u
al
attac
k
win
d
o
ws.
T
h
is
allo
ws
f
o
r
tar
g
ete
d
an
aly
s
is
an
d
t
ailo
r
ed
m
itig
atio
n
s
tr
ateg
ies
f
o
r
ea
c
h
a
n
o
m
aly
.
B
y
f
o
cu
s
in
g
o
n
f
ea
tu
r
e
co
n
t
r
ib
u
tio
n
s
s
p
ec
if
ic
to
ea
ch
attac
k
,
we
en
s
u
r
e
a
co
m
p
r
eh
e
n
s
iv
e
u
n
d
er
s
tan
d
in
g
o
f
t
h
e
r
o
o
t c
au
s
es.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
14
,
No
.
4
,
Dec
em
b
er
20
25
:
1
4
2
0
-
1
4
3
2
1428
Fig
u
r
e
5
.
Fo
r
ce
p
lo
t f
o
r
a
ttack
n
u
m
b
er
6
Fig
u
r
e
6
.
Vio
lin
Plo
t f
o
r
f
e
atu
r
e
co
n
tr
ib
u
tio
n
f
o
r
attac
k
n
u
m
b
er
6
Fig
u
r
e
7
.
Hea
tm
ap
an
d
s
en
s
o
r
-
wis
e
attr
ib
u
tio
n
f
o
r
attac
k
n
o
6
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l Sci
I
SS
N:
2252
-
8
8
1
4
Un
ve
ilin
g
a
n
o
ma
lies
in
in
d
u
s
t
r
ia
l c
o
n
tr
o
l sys
tems:
a
ke
r
n
el
S
HA
P
-
b
a
s
ed
a
p
p
r
o
a
ch
…
(
S
a
n
g
ee
ta
Osw
a
l
)
1429
Fig
u
r
e
8
.
Fo
r
ce
p
lo
t f
o
r
attac
k
n
u
m
b
er
2
3
Fig
u
r
e
9
.
Vio
lin
Plo
t f
o
r
SHA
P v
alu
es f
o
r
attac
k
n
u
m
b
e
r
2
3
Fig
u
r
e
10
.
Hea
tm
ap
f
o
r
SHAP v
alu
es
attac
k
2
3
Evaluation Warning : The document was created with Spire.PDF for Python.