T
E
L
K
O
M
N
I
K
A
T
elec
o
m
m
un
ica
t
io
n Co
m
pu
t
ing
E
lect
ro
nics
a
nd
Co
ntr
o
l
Vo
l.
24
,
No
.
5
,
Octo
b
er
20
26
,
p
p
.
1
5
9
7
~
1
6
1
0
I
SS
N:
1
6
9
3
-
6
9
3
0
,
DOI
: 1
0
.
1
2
9
2
8
/TE
L
KOM
NI
KA.
v
24
i
5
.
2
7
8
8
4
1597
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//telko
mn
ika
.
u
a
d
.
a
c.
i
d
Enha
ncing
SQ
L
i
nject
io
n
det
e
ction
accura
cy
wit
h
f
in
e
-
tuni
ng
LL
a
M
a
3.2
1B
usi
ng
lo
w
-
ra
nk ad
a
p
tatio
n
Dia
s
h F
irda
us
1
,
I
di S
um
a
rdi
2
,
M
uh
a
m
m
a
d Ichwa
n
1
,
M
a
ula
na
Seno
Aj
i
Yudh
a
nta
ra
1
1
F
a
c
u
l
t
y
o
f
I
n
f
o
r
m
a
t
i
c
s
a
n
d
Te
c
h
n
o
l
o
g
y
I
n
d
u
st
r
y
,
I
n
st
i
t
u
t
Te
k
n
o
l
o
g
i
N
a
s
i
o
n
a
l
B
a
n
d
u
n
g
(
I
TEN
A
S
)
,
B
a
n
d
u
n
g
,
I
n
d
o
n
e
s
i
a
2
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
c
s
E
n
g
i
n
e
e
r
i
n
g
,
Al
-
G
h
i
f
a
r
i
U
n
i
v
e
r
si
t
y
,
B
a
n
d
u
n
g
,
I
n
d
o
n
e
s
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Feb
20
,
2
0
2
6
R
ev
is
ed
J
u
n
11
,
2
0
2
6
Acc
ep
ted
J
u
n
30
,
2
0
2
6
S
tru
c
tu
re
d
q
u
e
ry
lan
g
u
a
g
e
i
n
jec
ti
o
n
(S
QL
i
n
jec
ti
o
n
/S
QLi
)
re
m
a
in
s
a
c
rit
ica
l
th
re
a
t
to
we
b
a
p
p
li
c
a
ti
o
n
se
c
u
rit
y
,
p
a
rti
c
u
larly
b
e
c
a
u
se
it
c
a
n
b
y
p
a
ss
trad
it
io
n
a
l
d
e
tec
ti
o
n
sy
ste
m
s
th
ro
u
g
h
o
b
f
u
sc
a
ti
o
n
.
Ex
isti
n
g
a
p
p
ro
a
c
h
e
s,
in
c
lu
d
in
g
ru
le
-
b
a
se
d
sy
ste
m
s
a
n
d
c
las
sic
a
l
m
a
c
h
in
e
lea
rn
in
g
(M
L)
m
o
d
e
ls,
o
ften
fa
il
to
c
a
p
tu
re
th
e
c
o
n
tex
tu
a
l
se
m
a
n
ti
c
s
o
f
c
o
m
0
p
lex
S
QL
q
u
e
ries
.
Th
is
stu
d
y
p
ro
p
o
se
s
a
li
g
h
twe
ig
h
t
a
n
d
c
o
n
tex
t
-
a
wa
re
S
QL
i
d
e
tec
ti
o
n
m
o
d
e
l
b
y
fin
e
-
tu
n
in
g
t
h
e
larg
e
lan
g
u
a
g
e
m
o
d
e
l
M
e
ta
a
rti
ficia
l
in
telli
g
e
n
c
e
(
LL
a
M
A
)
3
.
2
–
1
B
larg
e
lan
g
u
a
g
e
m
o
d
e
l
u
si
n
g
lo
w
-
ra
n
k
a
d
a
p
tatio
n
(Lo
RA).
Th
e
m
o
d
e
l
wa
s
train
e
d
u
sin
g
t
h
e
S
QL
-
I
n
jec
ti
o
n
-
E
x
ten
d
d
a
tas
e
t
o
b
tain
e
d
fr
o
m
Ka
g
g
le,
c
o
n
sistin
g
o
f
5
7
,
3
1
0
m
a
li
c
io
u
s
S
QL
i
sa
m
p
les
a
n
d
5
2
,
2
0
8
b
e
n
ig
n
u
se
r
-
in
p
u
t
sa
m
p
les
.
Th
e
b
e
n
ig
n
d
a
ta
c
o
m
p
r
ise
g
e
n
e
ra
l
u
se
r
-
g
e
n
e
ra
ted
tex
t
c
o
m
m
o
n
ly
su
b
m
it
ted
t
h
ro
u
g
h
we
b
a
p
p
li
c
a
ti
o
n
in
terfa
c
e
s.
Ex
p
e
rime
n
ta
l
re
su
lt
s
d
e
m
o
n
stra
te
th
a
t
th
e
p
ro
p
o
se
d
m
o
d
e
l
a
c
h
iev
e
s
a
n
a
c
c
u
ra
c
y
o
f
9
9
.
9
1
%
,
p
re
c
isio
n
o
f
9
9
.
9
6
%
,
re
c
a
ll
o
f
9
9
.
8
6
%
,
a
n
d
a
n
F
1
-
sc
o
re
o
f
9
9
.
9
1
%
,
o
u
t
p
e
rfo
rm
in
g
b
a
se
li
n
e
m
o
d
e
ls
in
c
lu
d
in
g
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
(CNN
),
lo
n
g
sh
o
rt
-
term
m
e
m
o
ry
(
LS
TM
),
a
n
d
n
o
n
-
fin
e
-
t
u
n
e
d
larg
e
lan
g
u
a
g
e
m
o
d
e
ls
(LL
M
s)
.
Th
e
se
fi
n
d
i
n
g
s
in
d
ica
te
th
a
t
L
o
RA
-
b
a
se
d
fi
n
e
-
tu
n
in
g
su
b
sta
n
ti
a
ll
y
e
n
h
a
n
c
e
s
d
e
tec
ti
o
n
p
e
rfo
rm
a
n
c
e
wh
il
e
m
a
in
tain
i
n
g
c
o
m
p
u
tati
o
n
a
l
e
fficie
n
c
y
,
m
a
k
i
n
g
th
e
a
p
p
ro
a
c
h
su
i
tab
le
fo
r
re
a
l
-
ti
m
e
d
e
p
lo
y
m
e
n
t
in
m
o
d
e
r
n
we
b
se
c
u
rit
y
s
y
ste
m
s.
F
u
r
th
e
rm
o
re
,
th
e
in
te
g
ra
ti
o
n
o
f
c
o
n
tex
t
-
a
wa
re
d
ig
i
tal
fo
re
n
sic
re
p
o
rti
n
g
s
u
p
p
o
rts
p
o
st
-
i
n
c
id
e
n
t
i
n
v
e
stig
a
ti
o
n
a
n
d
imp
r
o
v
e
s t
h
e
in
ter
p
re
tab
il
it
y
o
f
d
e
tec
ti
o
n
o
u
tco
m
e
s.
K
ey
w
o
r
d
s
:
Dig
ital f
o
r
en
s
ics
L
ar
g
e
lan
g
u
ag
e
m
o
d
el
L
L
aM
A
3
.
2
L
o
w
-
r
an
k
ad
ap
tatio
n
SQL
in
jectio
n
W
eb
ap
p
licatio
n
s
ec
u
r
ity
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
:
I
d
i Su
m
ar
d
i
Dep
ar
tm
en
t o
f
I
n
f
o
r
m
atics E
n
g
in
ee
r
in
g
,
Alg
h
if
ar
i U
n
i
v
er
s
ity
1
4
0
C
is
ar
an
ten
Ku
lo
n
Stre
et,
So
ek
ar
n
o
-
Hatta,
B
an
d
u
n
g
,
W
est J
av
a
4
0
2
9
3
,
I
n
d
o
n
esia
E
m
ail: id
is
u
m
ar
d
i@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
Stru
ctu
r
ed
q
u
er
y
lan
g
u
ag
e
in
j
ec
tio
n
(
SQL
in
jectio
n
/SQLi)
attac
k
s
ar
e
am
o
n
g
t
h
e
m
o
s
t
d
an
g
er
o
u
s
th
r
ea
ts
in
th
e
web
ap
p
licatio
n
ec
o
s
y
s
tem
d
u
e
to
th
eir
ab
ilit
y
to
d
ir
ec
tly
ex
p
lo
it
d
atab
ases
,
lead
in
g
to
f
atal
co
n
s
eq
u
en
ce
s
s
u
c
h
as
s
en
s
itiv
e
d
ata
(
PII
)
leak
ag
e,
u
n
au
th
o
r
ized
d
ata
m
o
d
i
f
icatio
n
,
an
d
e
v
en
s
er
v
e
r
ta
k
eo
v
er
th
r
o
u
g
h
r
e
m
o
te
c
o
d
e
ex
ec
u
tio
n
(
R
C
E
)
[
1
]
.
T
h
ese
d
estru
cti
v
e
im
p
ac
ts
n
o
t
o
n
ly
d
is
r
u
p
t
t
ec
h
n
ical
o
p
er
atio
n
s
b
u
t
also
tr
ig
g
er
m
ass
iv
e
f
in
an
cial
lo
s
s
e
s
an
d
v
io
latio
n
s
o
f
s
tr
ict
d
ata
p
r
iv
ac
y
r
eg
u
latio
n
s
.
As
em
p
h
asized
in
r
ec
en
t
s
tu
d
ies,
SQLi
r
em
ain
s
a
p
r
im
ar
y
attac
k
v
ec
to
r
d
o
m
in
atin
g
th
e
cy
b
er
th
r
ea
t
la
n
d
s
ca
p
e
d
u
e
to
its
ef
f
ec
tiv
en
ess
in
p
en
etr
atin
g
w
ea
k
ap
p
licatio
n
d
ef
en
s
es
an
d
i
ts
o
n
g
o
in
g
ex
p
lo
itatio
n
b
y
s
ec
u
r
ity
r
esear
ch
er
s
in
p
u
r
s
u
it
o
f
im
p
r
o
v
ed
m
itig
ati
o
n
s
o
lu
tio
n
s
[
2
]
,
[
3
]
.
T
h
e
p
r
im
ar
y
is
s
u
e
n
o
lo
n
g
er
lies
in
s
im
p
le,
ea
s
ily
id
en
tifia
b
le
attac
k
s
b
u
t
r
ath
e
r
in
attac
k
er
s
’
ab
ilit
y
to
lev
er
ag
e
th
ese
tech
n
iq
u
es
to
m
an
ip
u
late
d
atab
ases
u
n
d
etec
ted
b
y
s
tan
d
ar
d
s
ec
u
r
ity
s
y
s
tem
s
.
T
h
is
th
r
ea
t
is
f
u
r
th
er
in
ten
s
if
ied
b
y
tec
h
n
o
lo
g
ical
ev
o
lu
ti
o
n
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
TEL
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l
C
o
n
tr
o
l
,
Vo
l.
24
,
No
.
5
,
Octo
b
e
r
20
26
:
1
5
9
7
-
1
6
1
0
1598
in
clu
d
in
g
t
h
e
ad
o
p
tio
n
o
f
lar
g
e
lan
g
u
a
g
e
m
o
d
els
(
L
L
Ms)
an
d
m
ac
h
i
n
e
lear
n
in
g
(
ML
)
alg
o
r
ith
m
s
th
at
ca
n
ass
is
t
attac
k
er
s
in
au
to
m
atic
ally
g
en
er
atin
g
d
i
v
er
s
e
p
ay
l
o
ad
v
ar
iatio
n
s
,
th
er
e
b
y
r
en
d
er
in
g
co
n
v
en
tio
n
al
d
etec
tio
n
m
eth
o
d
s
,
s
u
ch
as si
g
n
atu
r
e
-
b
ased
d
etec
tio
n
,
in
cr
ea
s
in
g
ly
in
ef
f
ec
tiv
e
[
4
]
,
[
5
]
.
Ov
er
th
e
p
ast
two
d
ec
ad
es,
s
ec
u
r
ity
r
esear
ch
er
s
an
d
p
r
ac
ti
tio
n
er
s
h
av
e
d
ev
elo
p
ed
v
ar
io
u
s
d
ef
en
s
e
m
ec
h
an
is
m
s
;
h
o
wev
er
,
ea
c
h
e
x
h
ib
its
f
u
n
d
a
m
en
tal
lim
itatio
n
s
wh
en
co
n
f
r
o
n
ted
with
m
o
d
e
r
n
attac
k
s
[
6
]
.
First,
tr
ad
itio
n
al
ap
p
r
o
ac
h
es
th
at
r
ely
o
n
s
ig
n
atu
r
e
-
b
ased
s
y
s
tem
s
,
s
u
ch
as
clas
s
ical
web
a
p
p
licatio
n
f
ir
ewa
lls
(
W
AFs
)
,
o
p
er
ate
b
y
m
atch
in
g
tr
af
f
ic
ag
ai
n
s
t
d
atab
ases
o
f
k
n
o
wn
attac
k
p
atter
n
s
.
T
h
e
cr
itic
al
wea
k
n
ess
o
f
th
is
ap
p
r
o
ac
h
lies
in
it
s
r
ig
id
ity
:
ev
en
m
in
o
r
m
o
d
if
icatio
n
s
to
a
p
ay
lo
ad
ca
n
alter
its
s
ig
n
at
u
r
e,
allo
win
g
n
ew
attac
k
s
to
b
y
p
ass
d
etec
tio
n
an
d
r
esu
ltin
g
in
a
h
ig
h
f
alse
-
n
eg
ativ
e
r
ate
[
7
]
.
Nu
m
er
o
u
s
s
tu
d
ies
h
a
v
e
p
r
ev
i
o
u
s
ly
attem
p
ted
to
a
d
d
r
ess
SQLi
th
r
ea
ts
u
s
in
g
ML
an
d
d
e
ep
lear
n
in
g
(
DL
)
alg
o
r
ith
m
s
,
ac
h
iev
in
g
p
r
o
m
is
in
g
r
esu
lts
b
u
t
s
till
f
ac
i
n
g
s
tr
u
ctu
r
al
lim
itatio
n
s
.
R
ec
en
t
s
u
r
v
ey
s
in
d
icate
th
at
class
if
ier
s
lik
e
Naïv
e
B
ay
es
an
d
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
(
SVM)
ar
e
ef
f
ec
tiv
e
f
o
r
s
tan
d
ar
d
p
a
y
lo
ad
s
[
8
]
,
[
9
]
.
T
h
e
y
r
em
ain
v
u
l
n
er
ab
le
t
o
ad
v
er
s
ar
ial
ev
asio
n
tech
n
iq
u
es,
wh
er
e
attac
k
er
s
s
u
b
tly
m
o
d
if
y
q
u
er
y
s
y
n
tax
to
b
y
p
ass
d
etec
tio
n
[
1
0
]
.
A
p
r
e
v
io
u
s
s
tu
d
y
,
R
o
s
ca
et
a
l
.
[
5
]
in
tr
o
d
u
ce
d
a
ML
–
b
ased
f
r
a
m
ewo
r
k
f
o
r
SQLi
d
etec
tio
n
b
y
d
ev
el
o
p
in
g
a
tw
o
-
s
tag
e
p
r
o
ce
s
s
in
g
m
ec
h
an
is
m
th
at
in
teg
r
ates
SQL
q
u
er
y
n
o
r
m
aliza
tio
n
with
s
em
an
tic
f
ea
tu
r
e
ex
tr
ac
tio
n
.
As
its
p
r
im
ar
y
co
n
tr
ib
u
tio
n
,
th
e
ap
p
r
o
ac
h
ex
tr
ac
ts
eig
h
t
s
em
a
n
tic
ch
ar
ac
ter
is
tics
f
r
o
m
SQL
s
tatem
en
ts
af
ter
tr
an
s
f
o
r
m
in
g
r
aw
q
u
er
ies
in
to
g
en
er
alize
d
r
ep
r
esen
tatio
n
s
th
at
ar
e
m
o
r
e
s
u
itab
le
f
o
r
ML
ap
p
licatio
n
s
.
T
o
s
u
p
p
o
r
t
m
o
d
el
d
ev
elo
p
m
en
t,
th
e
a
u
th
o
r
s
co
m
p
iled
a
lar
g
e
-
s
ca
le
d
ataset
co
n
tain
in
g
9
0
,
0
0
0
SQL
q
u
er
ies,
c
o
m
p
r
is
in
g
1
7
,
6
9
5
leg
itima
te
q
u
er
ies
an
d
7
2
,
3
0
4
m
alicio
u
s
q
u
er
ie
s
.
T
h
e
d
ataset
was
g
en
er
ated
th
r
o
u
g
h
a
co
m
b
in
atio
n
o
f
p
u
b
licly
av
ailab
le
s
o
u
r
ce
s
an
d
s
y
n
th
etic
s
am
p
l
es
p
r
o
d
u
ce
d
u
s
in
g
g
en
er
ativ
e
p
r
e
-
tr
ain
ed
tr
an
s
f
o
r
m
er
4
o
(
GPT
-
4
o
)
,
aim
in
g
to
m
itig
ate
p
r
iv
ac
y
is
s
u
es
an
d
ad
d
r
ess
th
e
lim
ited
av
ailab
ilit
y
o
f
lab
eled
SQLi
d
ata.
Mu
ltip
le
ML
alg
o
r
ith
m
s
wer
e
s
u
b
s
eq
u
en
tly
ass
ess
ed
u
s
in
g
Azu
r
e
ML
Stu
d
io
u
n
d
er
v
a
r
io
u
s
s
am
p
lin
g
co
n
f
i
g
u
r
atio
n
s
.
Am
o
n
g
th
e
f
if
teen
ev
alu
ated
m
o
d
el
–
s
am
p
lin
g
co
m
b
in
atio
n
s
,
a
v
o
tin
g
en
s
em
b
le
m
o
d
el
in
co
r
p
o
r
atin
g
ex
t
r
em
e
g
r
ad
ien
t
b
o
o
s
tin
g
(
XGBo
o
s
t)
a
n
d
lig
h
t
g
r
ad
ie
n
t
b
o
o
s
tin
g
m
ac
h
in
e
(
L
i
g
h
tGB
M
)
d
em
o
n
s
tr
ated
th
e
h
i
g
h
est
p
er
f
o
r
m
a
n
c
e,
ac
h
iev
in
g
an
ac
cu
r
ac
y
o
f
9
6
.
8
6
%,
a
weig
h
ted
ar
ea
u
n
d
er
th
e
r
ec
eiv
e
r
o
p
er
ati
n
g
ch
a
r
ac
ter
is
tic
cu
r
v
e
(
AUC)
o
f
9
8
.
2
5
%,
a
weig
h
ted
F1
-
s
co
r
e
o
f
9
6
.
7
7
%,
an
d
a
Ma
tth
ews
co
r
r
elatio
n
co
ef
f
icien
t
(
MCC
)
o
f
8
9
.
8
9
%.
Nev
er
th
eless
,
d
esp
ite
th
e
p
r
o
m
is
in
g
p
r
ed
ictiv
e
p
er
f
o
r
m
an
ce
,
co
n
v
en
tio
n
al
ML
ap
p
r
o
ac
h
es
m
ay
s
till
h
av
e
d
if
f
icu
lty
ca
p
tu
r
in
g
th
e
co
n
tex
tu
al
an
d
s
em
an
tic
co
m
p
lex
ities
ass
o
ciate
d
with
s
o
p
h
is
ticated
an
d
o
b
f
u
s
ca
ted
SQLi
attac
k
s
.
Su
b
s
eq
u
en
tly
,
v
ar
io
u
s
s
tu
d
i
es
h
av
e
in
v
esti
g
ated
SQLi
d
etec
tio
n
tech
n
iq
u
es
u
s
in
g
d
iv
er
s
e
ap
p
r
o
ac
h
es,
r
an
g
in
g
f
r
o
m
p
a
tter
n
-
b
ased
m
eth
o
d
s
to
ML
an
d
DL
.
Falo
r
et
a
l
.
[
1
1
]
d
em
o
n
s
tr
ated
th
at
a
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN)
–
b
ased
ap
p
r
o
ac
h
ca
n
ac
h
iev
e
a
h
ig
h
ac
cu
r
ac
y
o
f
9
8
.
6
%,
h
i
g
h
lig
h
tin
g
C
NNs’
ef
f
ec
tiv
en
ess
in
lear
n
in
g
SQLi
attac
k
p
atter
n
s
,
alb
ei
t
at
th
e
co
s
t
o
f
h
i
g
h
co
m
p
u
tat
io
n
al
o
v
er
h
ea
d
.
I
n
co
n
tr
ast,
R
au
t
et
a
l
.
[
1
2
]
ad
o
p
ted
th
e
Ah
o
–
C
o
r
asick
p
atter
n
-
m
atch
in
g
al
g
o
r
ith
m
,
w
h
ich
is
co
m
p
u
tatio
n
ally
lig
h
tweig
h
t
b
u
t
ac
h
ie
v
ed
o
n
ly
9
1
.
3
%
ac
cu
r
ac
y
an
d
ten
d
s
to
f
ail
in
d
etec
tin
g
a
d
v
an
ce
d
o
r
o
b
f
u
s
ca
ted
attac
k
s
.
A
m
o
r
e
co
m
p
le
x
DL
ap
p
r
o
ac
h
was
in
tr
o
d
u
ce
d
b
y
[
1
3
]
th
r
o
u
g
h
a
h
y
b
r
id
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
L
STM
)
–
C
NN
m
o
d
el,
wh
ich
r
ep
o
r
ted
th
e
h
ig
h
est
ac
c
u
r
ac
y
o
f
9
9
.
9
6
%,
th
o
u
g
h
it
r
eq
u
ir
es
s
u
b
s
tan
tial
co
m
p
u
tatio
n
a
l
r
eso
u
r
ce
s
.
C
o
n
v
e
n
t
i
o
n
a
l
M
L
a
p
p
r
o
a
c
h
es
c
o
n
t
i
n
u
e
t
o
d
e
m
o
n
s
t
r
a
t
e
c
o
m
p
e
t
i
t
i
v
e
p
e
r
f
o
r
m
a
n
c
e
.
Pa
t
il
e
t
a
l
.
[
1
4
]
a
p
p
l
i
e
d
d
e
c
i
s
i
o
n
t
r
e
e
s
(
D
T
)
,
r
a
n
d
o
m
f
o
r
e
s
t
(
R
F
)
,
a
n
d
SV
M
,
a
c
h
i
e
v
i
n
g
a
c
c
u
r
a
c
ie
s
o
f
u
p
t
o
9
6
%
,
w
h
i
l
e
e
m
p
h
a
s
i
z
i
n
g
t
h
e
i
m
p
o
r
ta
n
c
e
o
f
m
o
d
e
l
e
x
p
l
a
i
n
a
b
i
l
it
y
i
n
e
n
h
a
n
c
i
n
g
t
r
u
s
t
i
n
d
e
t
ec
t
i
o
n
s
y
s
t
e
m
s
,
d
e
s
p
i
t
e
c
h
a
ll
e
n
g
es
i
n
r
e
a
l
-
ti
m
e
d
e
p
l
o
y
m
e
n
t
.
Si
m
i
la
r
l
y
,
M
u
s
t
a
p
h
a
et
a
l
.
[
1
5
]
s
h
o
w
e
d
t
h
a
t
R
F
a
n
d
a
r
t
i
f
i
c
i
al
n
e
u
r
a
l
n
e
tw
o
r
k
s
(
AN
N
)
c
a
n
r
e
a
c
h
9
7
%
a
cc
u
r
a
c
y
;
h
o
we
v
e
r
,
e
v
a
l
u
a
ti
o
n
o
n
a
s
i
n
g
le
d
a
t
a
s
et
l
i
m
i
ts
t
h
e
g
e
n
e
r
al
i
z
a
b
il
it
y
o
f
t
h
e
i
r
m
o
d
e
ls
.
F
u
r
t
h
e
r
m
o
r
e
,
s
e
v
e
r
al
s
t
u
d
i
es
h
a
v
e
p
r
o
p
o
s
e
d
h
y
b
r
i
d
a
n
d
d
o
m
a
i
n
-
s
p
e
c
i
f
ic
a
p
p
r
o
a
c
h
e
s
.
W
a
n
g
e
t
a
l
.
[
1
6
]
u
tili
ze
d
a
co
m
b
in
atio
n
o
f
C
NN
an
d
b
id
i
r
ec
tio
n
al
lo
n
g
s
h
o
r
t
-
ter
m
m
e
m
o
r
y
(
B
iLST
M
)
,
ac
h
ie
v
in
g
a
n
ac
cu
r
ac
y
o
f
9
9
.
3
%,
alth
o
u
g
h
th
ei
r
wo
r
k
p
r
im
ar
il
y
f
o
cu
s
ed
o
n
in
p
u
t
in
jectio
n
r
ath
er
th
an
SQLi
s
p
ec
if
ically
.
Alth
o
u
g
h
th
ese
s
tu
d
ies
r
ep
o
r
t
h
ig
h
d
etec
tio
n
ac
cu
r
ac
y
,
t
h
ey
s
h
ar
e
c
o
m
m
o
n
lim
itatio
n
s
.
T
h
e
p
r
im
ar
y
lim
itatio
n
lies
in
th
e
s
em
an
tic
g
ap
,
n
am
ely
th
e
i
n
a
b
ilit
y
o
f
th
ese
m
o
d
els
to
u
n
d
er
s
tan
d
th
e
c
o
n
tex
t
an
d
in
ten
t
b
eh
in
d
SQL
q
u
er
y
s
tr
u
ctu
r
es,
p
ar
ticu
lar
ly
w
h
en
a
ttack
s
ar
e
m
o
d
if
ied
u
s
in
g
s
o
p
h
is
ticated
s
y
n
tax
o
b
f
u
s
ca
tio
n
te
ch
n
iq
u
es.
T
h
e
p
r
e
v
io
u
s
ly
id
en
tifie
d
s
em
an
tic
g
ap
u
n
d
er
s
co
r
es
th
e
u
r
g
en
c
y
o
f
ad
o
p
tin
g
L
L
M
s
.
Un
lik
e
tr
ad
itio
n
al
ML
o
r
DL
ap
p
r
o
a
ch
es,
L
L
Ms
ar
e
tr
ain
e
d
o
n
m
ass
iv
e
co
r
p
o
r
a
o
f
s
o
u
r
ce
co
d
e
an
d
i
n
ter
n
et
te
x
t,
wh
ich
en
d
o
ws
th
em
with
d
ee
p
co
n
tex
tu
al
an
d
s
em
an
tic
u
n
d
er
s
tan
d
in
g
.
R
ath
er
th
an
m
er
ely
m
atch
in
g
s
tatis
t
ical
p
atter
n
s
,
L
L
Ms
ar
e
ca
p
ab
le
o
f
“r
ea
d
i
n
g
”
SQL
q
u
er
ies
an
d
d
ete
r
m
in
in
g
wh
et
h
er
th
e
u
n
d
er
l
y
in
g
lo
g
ic
is
m
alicio
u
s
,
ev
en
wh
en
th
e
s
y
n
tax
h
as
b
ee
n
m
a
n
ip
u
la
ted
th
r
o
u
g
h
s
o
p
h
is
ticated
o
b
f
u
s
ca
tio
n
tech
n
iq
u
es
[
1
7
]
.
Ho
wev
e
r
,
d
ep
l
o
y
in
g
l
ar
g
e
-
s
ca
le
L
L
Ms
s
u
ch
as
G
PT
-
4
o
r
lar
g
e
lan
g
u
ag
e
m
o
d
el
Me
ta
ar
tific
ial
in
tellig
en
ce
(
L
L
aM
A)
7
0
B
f
o
r
r
ea
l
-
tim
e
i
n
tr
u
s
io
n
d
etec
tio
n
is
g
en
er
ally
im
p
r
ac
tical
d
u
e
to
h
ig
h
laten
c
y
an
d
s
u
b
s
tan
tial o
p
er
atio
n
al
co
s
ts
[
1
8
]
.
As
an
in
n
o
v
ativ
e
s
o
lu
tio
n
,
L
L
aM
A
3
.
2
1
B
em
er
g
es
as
a
g
am
e
-
ch
an
g
er
in
t
h
e
ca
te
g
o
r
y
o
f
s
m
all
lan
g
u
ag
e
m
o
d
els
(
SLM
s
)
.
W
ith
o
n
ly
o
n
e
b
illi
o
n
p
a
r
am
ete
r
s
,
th
is
m
o
d
el
is
s
p
ec
if
ically
d
esig
n
ed
f
o
r
h
ig
h
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
E
n
h
a
n
cin
g
S
QL in
jectio
n
d
ete
ctio
n
a
cc
u
r
a
cy
w
ith
fin
e
-
tu
n
in
g
LL
a
Ma
3
.
2
1
B
u
s
in
g
…
(
Dia
s
h
F
ir
d
a
u
s
)
1599
ef
f
icien
cy
,
m
ak
in
g
it
lig
h
tweig
h
t
en
o
u
g
h
to
r
u
n
o
n
e
d
g
e
d
ev
ices
s
u
ch
as
R
asp
b
er
r
y
Pi
o
r
r
eso
u
r
ce
-
co
n
s
tr
ain
e
d
s
er
v
er
s
,
wh
ile
s
till
in
h
er
itin
g
th
e
ar
ch
itectu
r
al
in
tellig
en
ce
o
f
lar
g
er
L
L
aM
A
m
o
d
els
[
1
9
]
.
T
o
f
u
r
th
er
o
p
tim
ize
p
er
f
o
r
m
an
ce
with
o
u
t
o
v
er
b
u
r
d
en
i
n
g
co
m
p
u
tatio
n
al
r
eso
u
r
ce
s
,
lo
w
-
r
an
k
ad
ap
tatio
n
(
L
o
R
A)
is
em
p
lo
y
e
d
as a
k
ey
co
m
p
o
n
en
t in
th
is
s
tu
d
y
.
Fu
ll f
in
e
-
tu
n
in
g
o
f
L
L
M
s
ty
p
ically
r
eq
u
ir
es
s
u
b
s
tan
tial
m
em
o
r
y
an
d
c
o
m
p
u
tatio
n
al
o
v
er
h
ea
d
,
wh
ile
al
s
o
in
tr
o
d
u
cin
g
th
e
r
is
k
o
f
ca
tast
r
o
p
h
ic
f
o
r
g
ettin
g
,
wh
er
e
th
e
m
o
d
el
lo
s
es
its
p
r
ev
io
u
s
ly
lear
n
ed
g
en
er
al
k
n
o
wled
g
e.
L
o
R
A
ad
d
r
ess
es
th
ese
lim
itatio
n
s
b
y
u
p
d
atin
g
o
n
ly
a
s
m
all
s
u
b
s
et
o
f
p
ar
am
ete
r
s
th
r
o
u
g
h
lo
w
-
r
an
k
m
at
r
ix
d
ec
o
m
p
o
s
itio
n
,
r
ed
u
cin
g
m
em
o
r
y
r
eq
u
ir
em
e
n
ts
b
y
u
p
to
6
0
–
7
0
% a
n
d
s
ig
n
if
ican
tly
ac
ce
ler
ati
n
g
th
e
tr
ain
i
n
g
p
r
o
ce
s
s
[
2
0
]
.
B
y
lev
er
ag
in
g
th
ese
a
d
v
an
tag
es,
th
e
co
m
b
in
atio
n
o
f
L
L
aM
A
3
.
2
–
1
B
an
d
L
o
R
A
en
a
b
les
ef
f
icien
t
an
d
co
n
tex
t
-
awa
r
e
an
aly
s
is
o
f
u
s
er
in
p
u
t,
allo
win
g
s
ec
u
r
ity
s
y
s
tem
s
to
d
etec
t
S
QL
i
attac
k
s
with
lo
w
laten
cy
wh
ile
m
ain
tain
in
g
h
i
g
h
ac
cu
r
a
cy
.
I
n
a
d
d
itio
n
,
th
e
m
o
d
el
p
r
o
v
id
es
co
n
tex
tu
al
e
x
p
lan
atio
n
s
th
at
s
u
p
p
o
r
t
d
ig
ital
f
o
r
en
s
ic
an
aly
s
is
,
m
ak
in
g
it
s
u
itab
le
f
o
r
p
r
ac
tical
d
ep
lo
y
m
en
t
in
m
o
d
er
n
web
s
ec
u
r
ity
en
v
ir
o
n
m
en
ts
.
Acc
o
r
d
in
g
ly
,
th
is
s
tu
d
y
aim
s
to
d
ev
elo
p
an
in
tellig
e
n
t,
ef
f
i
cien
t,
an
d
co
n
tex
t
-
awa
r
e
SQLi
d
etec
tio
n
s
y
s
tem
th
at
ad
d
r
ess
es
th
e
lim
i
tatio
n
s
o
f
ex
is
tin
g
ap
p
r
o
ac
h
es.
Alth
o
u
g
h
p
r
ev
io
u
s
s
tu
d
ies
h
av
e
r
ep
o
r
ted
p
r
o
m
is
in
g
r
esu
lts
u
s
in
g
co
n
v
en
tio
n
al
ML
an
d
DL
ap
p
r
o
ac
h
es
f
o
r
SQLi
d
etec
tio
n
,
m
an
y
o
f
th
ese
m
eth
o
d
s
p
r
im
ar
ily
f
o
cu
s
o
n
lear
n
in
g
s
tatis
tical
p
atter
n
s
an
d
o
f
ten
s
tr
u
g
g
le
to
ca
p
tu
r
e
th
e
co
n
tex
t
u
al
s
em
a
n
tics
o
f
o
b
f
u
s
ca
ted
attac
k
p
ay
lo
a
d
s
.
Mo
r
eo
v
er
,
e
x
is
tin
g
s
tu
d
ies
r
ar
ely
in
v
esti
g
ate
th
e
u
s
e
o
f
lig
h
tweig
h
t
L
L
M
s
co
m
b
in
ed
with
p
ar
am
eter
-
ef
f
icien
t
f
in
e
-
tu
n
in
g
tech
n
iq
u
es
f
o
r
SQLi
d
etec
tio
n
.
T
h
er
e
f
o
r
e,
th
e
n
o
v
elty
o
f
t
h
is
s
tu
d
y
lies
in
th
e
in
teg
r
atio
n
o
f
L
L
aM
A
3
.
2
–
1
B
an
d
L
o
R
A
to
d
ev
elo
p
a
co
n
tex
t
-
awa
r
e
SQLi
d
etec
tio
n
f
r
am
ewo
r
k
th
at
b
alan
ce
s
s
em
an
tic
u
n
d
er
s
tan
d
in
g
,
co
m
p
u
tatio
n
al
ef
f
icien
cy
,
an
d
p
r
ac
tical
d
ep
lo
y
a
b
ilit
y
.
Fu
r
th
er
m
o
r
e
,
th
is
wo
r
k
ex
ten
d
s
b
e
y
o
n
d
attac
k
class
if
icatio
n
b
y
in
co
r
p
o
r
at
in
g
co
n
tex
t
-
awa
r
e
d
ig
ital
f
o
r
en
s
ic
r
ep
o
r
tin
g
to
s
u
p
p
o
r
t
p
o
s
t
-
in
cid
e
n
t
in
v
esti
g
atio
n
an
d
e
x
p
lain
ab
le
cy
b
er
s
ec
u
r
ity
a
n
aly
s
is
.
T
h
e
m
ain
c
o
n
tr
ib
u
tio
n
s
o
f
th
is
s
tu
d
y
ar
e
as
f
o
llo
ws
:
(
i)
t
h
e
d
e
v
elo
p
m
en
t
o
f
a
co
n
tex
t
-
awa
r
e
SQLi
d
etec
tio
n
m
o
d
el
b
ased
o
n
L
L
aM
A
3
.
2
–
1
B
with
L
o
R
A
f
in
e
-
tu
n
in
g
;
(
ii)
t
h
e
im
p
lem
en
tatio
n
o
f
a
lig
h
t
weig
h
t
d
ep
lo
y
m
e
n
t
s
tr
ateg
y
s
u
itab
le
f
o
r
ed
g
e
a
n
d
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
e
n
v
ir
o
n
m
en
ts
th
r
o
u
g
h
p
ar
am
eter
-
ef
f
i
cien
t
ad
ap
tatio
n
;
an
d
(
iii)
t
h
e
in
teg
r
atio
n
o
f
r
ea
l
-
tim
e
SQLi
d
etec
tio
n
with
co
n
tex
t
-
awa
r
e
d
i
g
ital
f
o
r
en
s
ic
r
ep
o
r
tin
g
to
s
u
p
p
o
r
t
ex
p
lai
n
ab
le
cy
b
er
s
ec
u
r
ity
o
p
er
atio
n
s
an
d
p
o
s
t
-
in
cid
en
t a
n
aly
s
is
.
2.
M
E
T
H
O
D
Fig
u
r
e
1
illu
s
tr
ates
th
e
a
r
ch
it
ec
tu
r
e
o
f
th
e
SQLi
d
etec
tio
n
m
o
d
el
,
d
ep
ictin
g
th
e
wo
r
k
f
lo
w
f
r
o
m
th
e
u
s
er
s
u
b
m
itti
n
g
in
p
u
t
th
r
o
u
g
h
th
e
ap
p
licatio
n
p
r
o
g
r
am
m
in
g
in
ter
f
ac
e
(
API
)
g
atew
ay
to
it
s
p
r
o
ce
s
s
in
g
with
in
th
e
web
ap
p
licatio
n
.
T
h
e
ar
c
h
itectu
r
e
p
o
s
itio
n
s
th
e
d
etec
t
io
n
s
y
s
tem
as
an
ad
d
itio
n
al
s
ec
u
r
ity
lay
er
th
at
an
aly
ze
s
in
co
m
in
g
u
s
er
in
p
u
t
b
ef
o
r
e
it is
ex
ec
u
te
d
b
y
th
e
d
at
ab
ase.
Fig
u
r
e
1
.
SQLi
d
etec
tio
n
ar
ch
i
tectu
r
e
On
th
e
clien
t
s
id
e,
u
s
er
s
ac
ce
s
s
th
e
web
ap
p
licatio
n
th
r
o
u
g
h
a
b
r
o
wser
.
W
h
en
th
e
web
s
ite
is
lo
ad
ed
,
u
s
er
s
s
u
b
m
it
r
eq
u
ests
in
th
e
f
o
r
m
o
f
u
s
er
in
p
u
t
(
e.
g
.
,
f
o
r
m
i
n
p
u
ts
o
r
q
u
er
y
s
tr
in
g
s
)
v
ia
an
API
.
Su
b
s
eq
u
en
tly
,
th
e
ed
g
e
d
etec
tio
n
s
y
s
tem
p
er
f
o
r
m
s
p
r
ep
r
o
ce
s
s
in
g
an
d
to
k
e
n
izatio
n
o
n
th
e
in
co
m
in
g
u
s
er
in
p
u
ts
b
ef
o
r
e
th
ey
ar
e
p
r
o
ce
s
s
ed
b
y
th
e
L
L
aM
A
3
.
2
–
1
B
m
o
d
el,
wh
ich
h
as
b
ee
n
f
in
e
-
tu
n
e
d
u
s
in
g
th
e
L
o
R
A
m
eth
o
d
.
T
h
e
m
o
d
el
an
aly
ze
s
th
e
u
s
er
in
p
u
t
to
d
ete
r
m
in
e
wh
eth
er
it
co
n
tain
s
SQLi
p
atter
n
s
.
I
f
th
e
i
n
p
u
t
is
id
e
n
tifie
d
as
m
alicio
u
s
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
TEL
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l
C
o
n
tr
o
l
,
Vo
l.
24
,
No
.
5
,
Octo
b
e
r
20
26
:
1
5
9
7
-
1
6
1
0
1600
th
e
r
eq
u
est
is
b
lo
ck
ed
,
an
d
th
e
s
y
s
tem
g
en
er
ates
a
n
o
tific
atio
n
in
d
icatin
g
a
d
etec
ted
SQLi
attem
p
t.
C
o
n
v
er
s
ely
,
if
th
e
in
p
u
t
is
class
if
ied
as
b
en
ig
n
,
ac
ce
s
s
to
th
e
d
atab
ase
is
g
r
an
ted
,
a
n
d
t
h
e
web
ap
p
licatio
n
o
p
er
ates
n
o
r
m
ally
.
Fig
u
r
e
2
illu
s
tr
ates
th
e
r
esear
c
h
m
et
h
o
d
o
lo
g
y
em
p
lo
y
ed
in
th
is
s
tu
d
y
.
T
h
e
r
esear
ch
p
r
o
ce
s
s
b
eg
in
s
with
d
ata
co
lle
ctio
n
f
r
o
m
Ka
g
g
le
a
n
d
p
r
o
ce
e
d
s
th
r
o
u
g
h
s
ev
er
al
s
tag
es,
cu
l
m
in
atin
g
in
th
e
f
in
al
p
h
ase
o
f
m
o
d
el
ev
al
u
atio
n
.
Fig
u
r
e
2
.
R
esear
ch
m
eth
o
d
Ph
ase
I
(
d
ata
p
r
ep
ar
atio
n
):
th
e
d
ata
p
r
ep
ar
ati
o
n
p
h
ase
aim
s
to
estab
lis
h
a
d
ataset
f
o
u
n
d
atio
n
th
at
is
r
elev
an
t
to
th
e
r
esear
ch
p
r
o
b
lem
.
T
h
is
p
h
ase
b
e
g
in
s
with
d
ata
c
o
llectio
n
,
f
o
cu
s
in
g
o
n
th
e
c
y
b
er
s
ec
u
r
ity
d
o
m
ain
,
p
ar
ticu
lar
ly
SQLi
at
tack
s
.
T
h
e
co
llected
d
ataset
in
clu
d
es
a
wid
e
r
an
g
e
o
f
u
s
er
in
p
u
t
p
atter
n
s
,
en
co
m
p
ass
in
g
b
o
th
b
en
ig
n
a
n
d
m
alicio
u
s
in
p
u
ts
,
to
r
e
p
r
e
s
en
t
r
ea
l
-
wo
r
ld
attac
k
s
ce
n
a
r
i
o
s
en
co
u
n
te
r
ed
b
y
m
o
d
er
n
web
ap
p
licatio
n
s
.
B
y
ex
p
licitly
ce
n
ter
in
g
SQLi
as
th
e
p
r
im
ar
y
th
r
ea
t,
t
h
is
p
h
ase
clea
r
ly
d
ef
in
es
th
e
p
r
o
b
lem
s
co
p
e
an
d
th
r
ea
t m
o
d
el
th
at
will b
e
lear
n
ed
b
y
th
e
l
an
g
u
ag
e
m
o
d
el
i
n
s
u
b
s
eq
u
e
n
t
s
tag
es.
Ph
ase
I
I
(
d
ata
p
r
e
-
p
r
o
ce
s
s
in
g
):
th
e
d
ata
p
r
e
-
p
r
o
ce
s
s
in
g
p
h
ase
is
r
esp
o
n
s
ib
le
f
o
r
tr
an
s
f
o
r
m
in
g
r
aw
d
ata
in
to
a
f
o
r
m
at
s
u
itab
le
f
o
r
m
o
d
el
tr
ain
in
g
.
T
h
e
co
llected
d
ataset
is
lo
ad
ed
in
to
th
e
p
r
o
c
ess
in
g
en
v
ir
o
n
m
en
t
an
d
co
n
v
er
ted
f
r
o
m
a
d
ataf
r
a
m
e
f
o
r
m
at
in
to
a
Hu
g
g
in
g
F
ac
e
d
ataset
to
en
s
u
r
e
co
m
p
atib
ilit
y
with
m
o
d
er
n
tr
ain
in
g
p
ip
elin
es.
T
h
e
d
ataset
is
th
en
s
p
lit
in
to
tr
ain
in
g
an
d
v
alid
atio
n
s
ets
to
en
ab
le
f
air
ev
alu
atio
n
an
d
p
r
ev
en
t
o
v
e
r
f
itti
n
g
.
T
h
is
s
ep
a
r
atio
n
also
s
u
p
p
o
r
ts
p
er
f
o
r
m
a
n
ce
m
o
n
ito
r
in
g
d
u
r
in
g
tr
ain
i
n
g
,
en
s
u
r
in
g
th
at
th
e
m
o
d
el
m
ain
tain
s
s
tr
o
n
g
g
en
e
r
aliza
tio
n
ca
p
ab
ilit
ies o
n
p
r
e
v
io
u
s
ly
u
n
s
ee
n
d
ata.
Ph
ase
I
I
I
(
m
o
d
elin
g
a
n
d
o
p
tim
izatio
n
):
t
h
e
m
o
d
elin
g
an
d
o
p
tim
izatio
n
p
h
ase
f
o
c
u
s
es
o
n
ef
f
icien
tly
b
u
ild
in
g
a
n
d
a
d
ap
tin
g
th
e
lan
g
u
ag
e
m
o
d
el.
T
h
e
b
ase
m
o
d
e
l
u
s
ed
in
th
is
s
tu
d
y
is
L
L
aM
A
3
.
2
with
1
b
illi
o
n
p
ar
am
eter
s
,
wh
ich
is
o
p
tim
iz
ed
u
s
in
g
4
-
b
it
q
u
an
tizatio
n
(
NF4
)
to
r
ed
u
ce
m
em
o
r
y
u
s
ag
e
an
d
co
m
p
u
tatio
n
al
o
v
er
h
ea
d
.
Fin
e
-
tu
n
i
n
g
is
p
er
f
o
r
m
ed
u
s
in
g
th
e
L
o
R
A
m
eth
o
d
,
allo
win
g
th
e
m
o
d
el
to
b
e
a
d
ap
ted
to
th
e
SQLi
d
etec
tio
n
d
o
m
ai
n
with
o
u
t
u
p
d
atin
g
all
m
o
d
el
p
ar
am
eter
s
.
T
h
is
ap
p
r
o
ac
h
r
esu
lts
in
a
r
el
ativ
ely
lig
h
tweig
h
t
lan
g
u
ag
e
m
o
d
el
with
s
tr
o
n
g
d
o
m
ain
-
s
p
ec
if
ic
ca
p
a
b
ilit
ies wh
ile
m
ain
tain
in
g
r
eso
u
r
ce
ef
f
ici
en
cy
.
Ph
ase
I
V
(
m
o
d
el
ev
alu
atio
n
):
m
o
d
el
e
v
alu
atio
n
aim
s
to
ass
ess
th
e
p
er
f
o
r
m
a
n
ce
a
n
d
ef
f
ec
ti
v
en
ess
o
f
th
e
tr
ain
ed
m
o
d
el.
I
n
th
is
s
ta
g
e,
th
e
m
o
d
el
is
ev
alu
ated
u
s
in
g
a
d
ed
icate
d
ev
alu
atio
n
d
a
taset
to
m
ea
s
u
r
e
its
ab
ilit
y
to
ac
cu
r
ately
d
etec
t a
n
d
h
an
d
le
SQLi
ca
s
es.
Per
f
o
r
m
an
ce
is
as
s
ess
ed
u
s
in
g
r
elev
an
t e
v
alu
atio
n
m
etr
ics,
in
clu
d
in
g
ac
cu
r
ac
y
,
p
r
e
cisi
o
n
,
r
ec
all,
an
d
F1
-
s
co
r
e,
to
e
n
s
u
r
e
th
at
th
e
m
o
d
el
n
o
t
o
n
ly
p
er
f
o
r
m
s
well
o
n
tr
ain
in
g
d
ata
b
u
t
also
g
en
er
a
lizes
ef
f
ec
tiv
ely
to
u
n
s
ee
n
d
ata.
T
h
e
ev
alu
atio
n
r
esu
lts
p
r
o
v
id
e
a
b
asis
f
o
r
d
eter
m
in
in
g
t
h
e
m
o
d
el
’
s
s
u
itab
ilit
y
f
o
r
d
ep
lo
y
m
en
t in
r
e
al
-
wo
r
ld
ap
p
licatio
n
s
ce
n
ar
io
s
.
2
.
1
.
Da
t
a
p
re
pa
ra
t
i
o
n
T
h
e
d
ataset
u
tili
ze
d
in
th
is
s
tu
d
y
was
o
b
tain
ed
f
r
o
m
t
h
e
SQL
-
I
n
jectio
n
-
E
x
ten
d
r
ep
o
s
ito
r
y
o
n
Kag
g
le
[
2
1
]
.
T
o
p
r
o
v
id
e
a
n
o
v
er
v
ie
w
o
f
th
e
d
ata
u
s
ed
f
o
r
m
o
d
el
tr
ain
in
g
an
d
ev
alu
atio
n
,
T
ab
le
1
p
r
esen
ts
r
ep
r
esen
tativ
e
ex
a
m
p
les
o
f
m
alicio
u
s
SQLi
p
ay
lo
ad
s
a
n
d
b
en
ig
n
u
s
er
in
p
u
ts
.
T
h
e
m
alici
o
u
s
s
am
p
les
in
clu
d
e
v
ar
io
u
s
SQLi
tech
n
iq
u
es
s
u
ch
as
tim
e
-
b
ased
attac
k
s
,
UNI
ON
-
b
ased
attac
k
s
,
an
d
d
atab
ase
f
u
n
ctio
n
ex
p
lo
itatio
n
,
wh
er
ea
s
th
e
b
en
i
g
n
s
am
p
les
co
n
s
is
t
o
f
n
o
r
m
al
u
s
er
-
g
en
er
ate
d
tex
t
co
m
m
o
n
ly
s
u
b
m
itted
th
r
o
u
g
h
web
ap
p
licatio
n
s
.
T
h
is
d
iv
er
s
ity
en
ab
les
th
e
m
o
d
el
to
lear
n
co
n
tex
tu
al
d
if
f
er
en
ce
s
b
etw
ee
n
m
alicio
u
s
an
d
non
-
m
alicio
u
s
in
p
u
ts
.
Af
ter
th
e
d
ata
p
r
ep
ar
atio
n
s
tag
e,
th
e
d
ataset
was
tr
an
s
f
o
r
m
ed
in
to
an
in
s
tr
u
ctio
n
-
b
ased
f
o
r
m
at
s
u
itab
le
f
o
r
s
u
p
er
v
is
ed
f
in
e
-
tu
n
in
g
o
f
LLM
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
E
n
h
a
n
cin
g
S
QL in
jectio
n
d
ete
ctio
n
a
cc
u
r
a
cy
w
ith
fin
e
-
tu
n
in
g
LL
a
Ma
3
.
2
1
B
u
s
in
g
…
(
Dia
s
h
F
ir
d
a
u
s
)
1601
T
ab
le
1
.
Sam
p
le
o
f
SQLi
d
ata
s
et
No
S
e
n
t
e
n
c
e
La
b
e
l
1
“
o
r
p
g
_
s
l
e
e
p
(
_
TI
M
E_
)
--
1
2
sel
e
c
t
*
f
r
o
m
u
sers
w
h
e
r
e
i
d
=
‘
1
’
o
r
@
@
1
=
1
u
n
i
o
n
s
e
l
e
c
t
1
,
v
e
r
s
i
o
n
(
)
--
1
’
1
3
O
n
e
o
f
t
h
e
o
t
h
e
r
r
e
v
i
e
w
e
r
s
…
a
f
t
e
r
w
a
t
c
h
i
n
g
j
u
s
t
1
O
z
e
p
i
s
o
d
e
y
o
u
’
l
l
b
e
h
o
o
k
e
d
.
.
.
0
4
P
e
t
t
e
r
M
a
t
t
e
i
’
s …
st
u
n
n
i
n
g
f
i
l
m
t
o
w
a
t
c
h
.
.
.
0
I
n
ter
m
s
o
f
d
ata
d
is
tr
ib
u
tio
n
,
t
h
e
d
ataset
co
n
tain
s
5
7
,
3
1
0
m
a
licio
u
s
s
en
ten
ce
s
(
L
ab
el
=
1
)
a
n
d
5
2
,
2
0
8
b
en
ig
n
s
en
ten
ce
s
(
L
ab
el
=
0
)
.
T
h
is
r
elativ
ely
b
alan
ce
d
d
is
tr
i
b
u
tio
n
s
u
p
p
o
r
ts
r
o
b
u
s
t
tr
ain
in
g
an
d
ev
alu
atio
n
o
f
tex
t
class
if
icatio
n
m
o
d
els
f
o
r
au
to
m
atic
SQLi
d
etec
tio
n
in
d
atab
ase
s
ec
u
r
ity
r
esear
c
h
.
Alt
h
o
u
g
h
th
e
d
ataset
is
d
er
iv
ed
f
r
o
m
Kag
g
le,
it
is
f
o
r
m
u
lated
as
a
u
s
er
in
p
u
t
class
i
f
icatio
n
task
,
wh
er
e
m
alicio
u
s
s
am
p
les
r
ep
r
esen
t
SQLi
p
ay
lo
ad
s
an
d
b
en
ig
n
s
am
p
les
r
ep
r
esen
t
g
en
er
al
u
s
er
in
p
u
t
tex
t
s
u
b
m
itted
th
r
o
u
g
h
web
f
o
r
m
s
.
T
h
is
ap
p
r
o
ac
h
r
e
f
lects
r
ea
l
-
wo
r
ld
s
ce
n
ar
io
s
in
wh
ic
h
SQLi
att
ac
k
s
ar
e
em
b
ed
d
e
d
with
in
u
s
er
-
p
r
o
v
id
e
d
i
n
p
u
t.
Ho
wev
er
,
alth
o
u
g
h
t
h
e
b
en
i
g
n
s
am
p
les
r
ep
r
esen
t
g
e
n
er
al
u
s
er
-
g
en
er
ated
te
x
t,
th
ey
m
ay
n
o
t
f
u
lly
ca
p
tu
r
e
th
e
d
iv
er
s
ity
an
d
co
m
p
lex
ity
o
f
r
ea
lis
tic
web
ap
p
licatio
n
in
p
u
ts
,
s
u
ch
as
lo
g
in
cr
ed
en
tials
,
u
n
if
o
r
m
r
eso
u
r
c
e
lo
ca
to
r
(
UR
L
)
p
ar
am
eter
s
,
A
PI
r
eq
u
ests
,
an
d
f
o
r
m
s
u
b
m
i
s
s
io
n
s
.
Pre
p
r
o
ce
s
s
in
g
was
ap
p
lied
to
in
tr
o
d
u
ce
ad
v
er
s
ar
ial
v
ar
iatio
n
s
s
u
ch
as
o
b
f
u
s
ca
tio
n
an
d
en
co
d
in
g
to
s
im
u
late
r
ea
lis
tic
attac
k
p
atter
n
s
.
T
h
er
ef
o
r
e,
f
u
tu
r
e
wo
r
k
will
f
o
cu
s
o
n
v
alid
atin
g
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
u
s
in
g
r
ea
l
-
wo
r
ld
d
atasets
,
in
clu
d
i
n
g
W
AF
lo
g
s
,
we
b
s
er
v
er
lo
g
s
,
an
d
s
ec
u
r
ity
in
f
o
r
m
atio
n
an
d
ev
e
n
t
m
an
ag
em
e
n
t
(
SIE
M
)
d
ata
s
o
u
r
ce
s
,
to
f
u
r
t
h
er
im
p
r
o
v
e
m
o
d
el
r
o
b
u
s
tn
ess
an
d
g
e
n
er
aliza
b
ilit
y
.
2
.
2
.
P
re
pro
ce
s
s
ing
T
h
e
p
r
ep
r
o
ce
s
s
in
g
p
ip
elin
e
is
d
esig
n
ed
to
s
y
s
tem
atica
lly
tr
an
s
f
o
r
m
th
e
r
aw
d
ataset
in
to
a
s
tr
u
ctu
r
ed
f
o
r
m
at
s
u
itab
le
f
o
r
in
s
tr
u
ctio
n
-
b
ased
lear
n
in
g
.
First,
(
i
)
d
a
ta
lo
ad
in
g
is
p
er
f
o
r
m
e
d
b
y
im
p
o
r
tin
g
th
e
d
ataset
f
r
o
m
a
co
m
m
a
-
s
ep
ar
ated
v
alu
es
(
C
SV
)
f
ile
an
d
co
n
v
er
tin
g
it
in
to
a
Pan
d
as
d
ata
f
r
am
e
f
o
r
e
f
f
icien
t
p
r
o
ce
s
s
in
g
.
Nex
t,
(
ii
)
d
ata
ty
p
e
s
tan
d
ar
d
iz
atio
n
is
ap
p
lied
b
y
co
n
v
er
tin
g
th
e
s
en
ten
ce
an
d
lab
el
co
lu
m
n
s
in
to
s
tr
in
g
d
ata
ty
p
es
to
en
s
u
r
e
c
o
n
s
is
ten
cy
an
d
p
r
ev
en
t
ty
p
e
-
r
elate
d
e
r
r
o
r
s
d
u
r
in
g
tr
ain
in
g
.
Su
b
s
eq
u
en
tly
,
(
iii
)
lab
el
id
en
tific
atio
n
is
co
n
d
u
cte
d
b
y
ex
tr
ac
tin
g
u
n
i
q
u
e
class
lab
el
s
d
ir
ec
tly
f
r
o
m
th
e
d
ataset
to
m
ain
tain
f
lex
i
b
ilit
y
f
o
r
p
o
ten
tial
e
x
ten
s
io
n
s
to
m
u
lti
-
class
clas
s
if
icatio
n
task
s
.
T
h
e
d
ataset
is
th
e
n
d
i
v
id
ed
in
to
(
iv
)
d
ata
s
p
lits
in
to
tr
ain
in
g
a
n
d
ev
alu
atio
n
s
u
b
s
et
s
u
s
in
g
an
8
0
:2
0
r
atio
,
with
a
f
ix
ed
r
an
d
o
m
_
s
tate
=
4
2
to
en
s
u
r
e
r
ep
r
o
d
u
cib
ilit
y
wh
ile
p
r
eser
v
in
g
class
d
is
tr
ib
u
tio
n
.
T
o
s
u
p
p
o
r
t
in
s
tr
u
ctio
n
-
b
ased
lear
n
in
g
,
(
v
)
p
r
o
m
p
t
co
n
s
tr
u
ctio
n
is
ca
r
r
ie
d
o
u
t
b
y
tr
a
n
s
f
o
r
m
in
g
ea
ch
d
ata
in
s
tan
ce
in
to
a
s
tr
u
ctu
r
ed
f
o
r
m
at
co
n
s
is
tin
g
o
f
task
in
s
tr
u
ctio
n
s
,
lab
el
o
p
tio
n
s
,
in
p
u
t
tex
t,
an
d
th
e
co
r
r
esp
o
n
d
in
g
o
u
tp
u
t
lab
el.
Fu
r
th
er
m
o
r
e,
(
vi
)
t
o
k
en
izatio
n
is
ap
p
lied
u
s
in
g
th
e
to
k
e
n
izer
ass
o
ciate
d
with
th
e
L
L
aM
A
3
.
2
–
1
B
m
o
d
el,
wh
e
r
e
tex
t
in
p
u
ts
ar
e
co
n
v
er
ted
i
n
to
to
k
en
I
Ds
with
ap
p
r
o
p
r
iat
e
p
ad
d
in
g
an
d
tr
u
n
ca
tio
n
to
en
s
u
r
e
u
n
if
o
r
m
s
eq
u
en
ce
len
g
t
h
.
T
h
is
is
f
o
llo
wed
b
y
(
v
ii
)
s
p
ec
ial
to
k
en
h
an
d
lin
g
,
wh
er
e
p
ad
d
i
n
g
an
d
en
d
-
of
-
s
e
q
u
en
ce
(
E
OS)
to
k
en
s
ar
e
in
co
r
p
o
r
ated
t
o
m
ain
tain
p
r
o
p
e
r
s
eq
u
en
ce
b
o
u
n
d
ar
ies.
F
in
ally
,
(
v
iii
)
ev
alu
atio
n
d
ata
p
r
ep
ar
atio
n
is
p
er
f
o
r
m
e
d
b
y
g
en
er
atin
g
u
n
la
b
eled
p
r
o
m
p
ts
f
r
o
m
th
e
ev
alu
atio
n
d
ataset
to
s
u
p
p
o
r
t
i
n
f
er
en
ce
a
n
d
p
er
f
o
r
m
a
n
ce
test
in
g
.
T
h
r
o
u
g
h
th
is
p
ip
eli
n
e,
th
e
o
r
ig
in
al
tab
u
lar
d
ataset
is
ef
f
ec
tiv
ely
tr
an
s
f
o
r
m
ed
in
to
a
m
o
d
el
-
r
ea
d
y
f
o
r
m
at
s
u
itab
le
f
o
r
s
u
p
er
v
is
ed
f
in
e
-
tu
n
in
g
o
f
LLM
s
.
I
n
th
is
tr
an
s
f
o
r
m
atio
n
,
ea
ch
s
am
p
le
is
r
ef
o
r
m
u
lated
as
a
n
in
s
tr
u
ct
io
n
-
f
o
llo
win
g
class
if
icatio
n
t
ask
co
n
s
is
tin
g
o
f
a
p
r
o
m
p
t,
in
p
u
t
tex
t,
a
n
d
t
h
e
c
o
r
r
esp
o
n
d
in
g
lab
el.
T
a
b
le
2
p
r
esen
ts
r
ep
r
esen
tativ
e
ex
am
p
l
es
o
f
th
e
c
o
n
v
e
r
ted
d
ataset
f
o
r
m
at
u
s
ed
d
u
r
in
g
m
o
d
el
tr
ain
in
g
.
T
ab
le
2
.
C
o
n
v
er
s
io
n
o
f
d
ataf
r
a
m
e
in
to
h
u
g
g
in
g
f
ace
d
ataset
No
Te
x
t
1
C
l
a
s
si
f
y
t
h
e
f
o
l
l
o
w
i
n
g
i
n
p
u
t
i
n
t
o
o
n
e
o
f
t
h
e
s
e
l
a
b
e
l
s:
[
‘
0
’
,
‘
1
’
]
I
n
p
u
t
:
“
o
r
p
g
_
s
l
e
e
p
(
TI
M
E)
--
A
n
sw
e
r
w
i
t
h
o
n
l
y
o
n
e
o
f
t
h
e
l
a
b
e
l
s
a
b
o
v
e
.
L
a
b
e
l
:
1
5
C
l
a
s
si
f
y
t
h
e
f
o
l
l
o
w
i
n
g
i
n
p
u
t
i
n
t
o
o
n
e
o
f
t
h
e
s
e
l
a
b
e
l
s:
[
‘
0
’
,
‘
1
’
]
I
n
p
u
t
:
s
e
l
e
c
t
*
f
r
o
m
u
sers
w
h
e
r
e
i
d
=
‘
1
’
o
r
@
@
1
=
1
u
n
i
o
n
se
l
e
c
t
1
,
v
e
r
si
o
n
(
)
--
‘
A
n
sw
e
r
w
i
t
h
o
n
l
y
o
n
e
o
f
t
h
e
l
a
b
e
l
s
a
b
o
v
e
.
L
a
b
e
l
:
1
6
C
l
a
s
si
f
y
t
h
e
f
o
l
l
o
w
i
n
g
i
n
p
u
t
i
n
t
o
o
n
e
o
f
t
h
e
s
e
l
a
b
e
l
s:
[
‘
0
’
,
‘
1
’
]
I
n
p
u
t
:
O
n
e
o
f
t
h
e
o
t
h
e
r
r
e
v
i
e
w
e
r
s
…
w
a
t
c
h
i
n
g
j
u
s
t
1
O
z
e
p
i
s
o
d
e
y
o
u
’
l
l
b
e
h
o
o
k
e
d
.
.
.
A
n
sw
e
r
w
i
t
h
o
n
l
y
o
n
e
o
f
t
h
e
l
a
b
e
l
s
a
b
o
v
e
.
L
a
b
e
l
:
0
10
C
l
a
s
si
f
y
t
h
e
f
o
l
l
o
w
i
n
g
i
n
p
u
t
i
n
t
o
o
n
e
o
f
t
h
e
s
e
l
a
b
e
l
s:
[
‘
0
’
,
‘
1
’
]
I
n
p
u
t
:
P
e
t
e
r
M
a
t
t
e
i
’
s
…
i
s a
v
i
s
u
a
l
l
y
s
t
u
n
n
i
n
g
f
i
l
m
t
o
w
a
t
c
h
...
A
n
sw
e
r
w
i
t
h
o
n
l
y
o
n
e
o
f
t
h
e
l
a
b
e
l
s
a
b
o
v
e
.
L
a
b
e
l
:
0
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
TEL
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l
C
o
n
tr
o
l
,
Vo
l.
24
,
No
.
5
,
Octo
b
e
r
20
26
:
1
5
9
7
-
1
6
1
0
1602
As
s
h
o
wn
in
T
ab
le
2
,
th
e
d
ataset
co
n
s
is
t
s
o
f
tex
t
s
am
p
les
th
at
h
av
e
b
ee
n
tr
an
s
f
o
r
m
ed
in
to
an
in
s
tr
u
ctio
n
-
b
ased
f
o
r
m
at
f
o
r
S
QL
i
d
etec
tio
n
.
E
ac
h
d
ata
in
s
tan
ce
in
clu
d
es
a
clea
r
class
if
ic
atio
n
in
s
tr
u
ctio
n
,
a
p
r
ed
ef
in
e
d
s
et
o
f
la
b
els
[
0
,
1
]
,
a
n
d
an
in
p
u
t
th
at
r
ep
r
es
en
ts
u
s
er
-
p
r
o
v
id
e
d
tex
t
s
u
b
m
itted
th
r
o
u
g
h
web
ap
p
licatio
n
in
ter
f
ac
es.
T
h
ese
in
p
u
ts
v
a
r
y
b
etwe
en
p
o
ten
ti
ally
m
alicio
u
s
SQLi
p
ay
lo
a
d
s
an
d
b
e
n
ig
n
u
s
er
in
p
u
ts
.
T
h
e
m
alicio
u
s
s
am
p
l
es
ex
h
ib
it
co
m
m
o
n
SQLi
p
a
tter
n
s
s
u
ch
as
tim
e
-
b
ased
att
ac
k
s
,
u
n
io
n
-
b
ased
q
u
er
ies,
an
d
s
y
s
tem
f
u
n
ctio
n
ex
p
lo
itatio
n
.
I
n
co
n
tr
ast,
th
e
b
en
ig
n
s
am
p
les
r
ep
r
esen
t
g
en
er
al
u
s
er
in
p
u
t,
in
clu
d
in
g
n
at
u
r
al
lan
g
u
ag
e
t
ex
t
co
m
m
o
n
l
y
f
o
u
n
d
in
web
f
o
r
m
s
s
u
ch
as
s
ea
r
ch
q
u
er
i
es,
co
m
m
en
ts
,
an
d
d
escr
ip
tiv
e
in
p
u
ts
.
T
h
is
f
o
r
m
u
latio
n
r
ef
lects
r
ea
l
-
wo
r
ld
s
ce
n
ar
io
s
wh
er
e
SQLi
attac
k
s
ar
e
em
b
ed
d
ed
with
in
u
s
er
-
p
r
o
v
id
ed
in
p
u
t
r
ath
er
t
h
a
n
is
o
lated
SQL
q
u
er
ies.
B
y
co
m
b
in
in
g
s
tr
u
ctu
r
ed
SQL
p
ay
lo
ad
s
with
g
en
er
al
u
s
er
in
p
u
t,
t
h
e
m
o
d
el
is
tr
ain
ed
to
d
is
tin
g
u
is
h
m
alicio
u
s
p
atter
n
s
f
r
o
m
n
o
r
m
al
in
p
u
t
b
eh
av
io
r
b
ased
o
n
co
n
tex
tu
al
u
n
d
er
s
tan
d
in
g
.
B
y
f
r
am
in
g
th
e
task
as
an
in
s
tr
u
ctio
n
-
f
o
llo
win
g
p
r
o
b
lem
,
th
e
d
ataset
b
ec
o
m
es
s
u
itab
le
f
o
r
s
u
p
er
v
is
ed
f
i
n
e
-
t
u
n
in
g
o
f
LLM
s
,
en
ab
lin
g
r
o
b
u
s
t
ev
alu
atio
n
o
f
SQLi
d
etec
tio
n
ac
r
o
s
s
d
iv
er
s
e
in
p
u
t p
atter
n
s
.
2
.
3
.
M
o
delin
g
a
nd
o
ptim
iza
t
io
n
I
n
th
is
s
tu
d
y
,
th
e
L
L
aM
A
3
.
2
–
1
B
m
o
d
el
was
em
p
lo
y
e
d
as
th
e
b
ase
lan
g
u
ag
e
m
o
d
el
f
o
r
s
u
p
er
v
is
ed
f
in
e
-
tu
n
in
g
.
T
h
e
m
o
d
el
an
d
to
k
en
izer
wer
e
lo
ad
ed
u
s
in
g
th
e
Hu
g
g
in
g
Face
T
r
an
s
f
o
r
m
er
s
lib
r
ar
y
.
T
o
r
e
d
u
ce
m
em
o
r
y
c
o
n
s
u
m
p
tio
n
an
d
e
n
ab
le
ef
f
icien
t
tr
ain
in
g
o
n
lim
ited
h
ar
d
war
e,
NF4
was
ap
p
lied
u
s
in
g
th
e
B
its
An
d
B
y
tes
co
n
f
ig
u
r
atio
n
with
NF4
q
u
an
tizatio
n
an
d
f
lo
at1
6
co
m
p
u
tatio
n
.
T
h
e
m
o
d
el
was
au
to
m
atica
lly
d
is
tr
ib
u
ted
ac
r
o
s
s
av
ailab
le
d
ev
ices
u
s
in
g
_
=
“
”
.
Ad
d
itio
n
ally
,
th
e
ca
ch
e
m
ec
h
an
is
m
was
d
is
ab
led
to
s
u
p
p
o
r
t
tr
ain
i
n
g
s
tab
ilit
y
,
an
d
th
e
to
k
e
n
izer
p
a
d
d
in
g
to
k
e
n
was
alig
n
ed
wit
h
th
e
E
OS
to
k
e
n
to
en
s
u
r
e
p
r
o
p
er
s
eq
u
en
ce
h
an
d
l
in
g
d
u
r
in
g
tr
ain
in
g
.
T
h
is
co
n
f
ig
u
r
atio
n
allo
ws
ef
f
icien
t
f
i
n
e
-
tu
n
in
g
o
f
an
LLM
wh
ile
m
ain
tain
in
g
r
ea
s
o
n
ab
le
co
m
p
u
tatio
n
al
r
eq
u
ir
e
m
en
ts
T
h
e
co
r
e
o
f
th
is
s
y
s
tem
is
th
e
L
L
aM
A
3
.
2
–
1
B
m
o
d
el,
o
p
tim
ized
u
s
in
g
NF4
to
en
s
u
r
e
co
m
p
u
tatio
n
al
ef
f
icien
cy
o
n
lim
ited
h
ar
d
wa
r
e
r
eso
u
r
ce
s
.
T
h
e
d
etailed
en
v
ir
o
n
m
e
n
t
an
d
m
o
d
el
s
etu
p
ar
e
s
u
m
m
ar
ized
in
T
ab
le
3
.
T
h
is
co
n
f
ig
u
r
atio
n
allo
ws
th
e
d
ep
lo
y
m
en
t
o
f
a
SL
M
ar
ch
itectu
r
e
with
a
s
ig
n
if
ican
tly
r
ed
u
ce
d
m
em
o
r
y
f
o
o
t
p
r
in
t.
T
ab
le
3
.
E
n
v
ir
o
n
m
en
t a
n
d
m
o
d
el
co
n
f
ig
u
r
atio
n
No
C
o
m
p
o
n
e
n
t
C
o
n
f
i
g
u
r
a
t
i
o
n
1
B
a
se
m
o
d
e
l
LLa
M
A
3
.
2
–
1
B
(
M
e
t
a
-
LLa
M
A
/
L
La
M
A
-
3
.
2
-
1
B
)
2
M
o
d
e
l
t
y
p
e
C
a
u
sa
l
l
a
n
g
u
a
g
e
mo
d
e
l
3
Li
b
r
a
r
y
H
u
g
g
i
n
g
f
a
c
e
t
r
a
n
sf
o
r
m
e
r
s
4
Q
u
a
n
t
i
z
a
t
i
o
n
m
e
t
h
o
d
N
F
4
(
B
i
t
sA
n
d
B
y
t
e
s)
5
Q
u
a
n
t
i
z
a
t
i
o
n
t
y
p
e
N
F
4
6
D
o
u
b
l
e
q
u
a
n
t
i
z
a
t
i
o
n
D
i
sab
l
e
d
7
C
o
m
p
u
t
e
d
a
t
a
t
y
p
e
F
l
o
a
t
1
6
8
D
e
v
i
c
e
ma
p
p
i
n
g
A
u
t
o
ma
t
i
c
(
_
=
“
”
)
9
C
a
c
h
e
u
s
a
g
e
D
i
sab
l
e
d
(
_
ℎ
=
)
10
P
r
e
t
r
a
i
n
i
n
g
t
e
n
so
r
p
a
r
a
l
l
e
l
i
sm
1
11
To
k
e
n
i
z
e
r
A
u
t
o
T
o
k
e
n
i
z
e
r
12
P
a
d
d
i
n
g
t
o
k
e
n
EO
S
t
o
k
e
n
13
F
r
a
mew
o
r
k
P
y
T
o
r
c
h
T
ab
le
3
s
u
m
m
ar
izes
th
e
en
v
ir
o
n
m
en
t
an
d
m
o
d
el
c
o
n
f
ig
u
r
at
io
n
u
s
ed
in
th
is
s
tu
d
y
.
T
h
e
b
ase
m
o
d
el
em
p
lo
y
ed
is
L
L
aM
A
3
.
2
–
1
B
,
wh
ich
is
a
ca
u
s
al
lan
g
u
ag
e
m
o
d
el
ac
ce
s
s
ed
th
r
o
u
g
h
th
e
Hu
g
g
in
g
Face
T
r
an
s
f
o
r
m
e
r
s
lib
r
ar
y
.
T
o
im
p
r
o
v
e
m
em
o
r
y
ef
f
icien
cy
an
d
en
ab
le
tr
ain
in
g
o
n
lim
ited
co
m
p
u
tatio
n
al
r
eso
u
r
ce
s
,
th
e
m
o
d
el
was
lo
ad
ed
u
s
in
g
NF4
with
th
e
B
i
ts
An
d
B
y
tes
f
r
am
ewo
r
k
.
T
h
e
NF4
q
u
a
n
tizatio
n
ty
p
e
an
d
f
lo
at1
6
co
m
p
u
tatio
n
wer
e
s
elec
ted
to
b
alan
ce
p
er
f
o
r
m
an
ce
an
d
n
u
m
er
ical
s
tab
ilit
y
,
wh
ile
d
o
u
b
l
e
q
u
an
tizatio
n
was
d
is
ab
led
to
s
im
p
lify
th
e
q
u
a
n
tizatio
n
p
r
o
ce
s
s
.
T
h
e
m
o
d
el
was
au
to
m
atica
lly
d
is
tr
ib
u
te
d
ac
r
o
s
s
av
ailab
le
h
ar
d
war
e
u
s
in
g
an
au
to
m
atic
d
ev
ice
m
ap
p
in
g
s
tr
ateg
y
.
C
ac
h
e
u
s
ag
e
was
d
is
ab
led
to
s
u
p
p
o
r
t
s
tab
le
tr
ain
in
g
b
eh
av
io
r
,
an
d
th
e
p
r
etr
ain
in
g
ten
s
o
r
p
ar
allelis
m
p
a
r
am
e
ter
was
s
et
to
o
n
e.
Fo
r
te
x
t
p
r
o
ce
s
s
in
g
,
th
e
Au
to
T
o
k
en
ize
r
was
u
s
ed
,
wi
th
th
e
p
a
d
d
in
g
to
k
en
alig
n
e
d
to
th
e
E
OS
to
k
e
n
to
en
s
u
r
e
co
n
s
is
ten
t
in
p
u
t
h
an
d
lin
g
.
Ov
er
all,
t
h
e
co
n
f
ig
u
r
atio
n
p
r
esen
ted
in
t
h
is
T
ab
le
3
en
ab
les
ef
f
icien
t
f
in
e
-
t
u
n
in
g
o
f
an
LLM
wh
ile
m
ain
tain
in
g
m
a
n
ag
ea
b
le
c
o
m
p
u
tatio
n
al
r
eq
u
i
r
em
en
ts
.
Fig
u
r
e
3
illu
s
tr
ates
th
e
ar
ch
ite
ctu
r
e
o
f
L
o
R
A
.
T
o
ad
ap
t
t
h
e
p
r
e
-
tr
ain
ed
lan
g
u
a
g
e
m
o
d
el
to
th
e
SQLi
d
etec
tio
n
task
,
th
is
s
tu
d
y
em
p
lo
y
ed
L
o
R
A
,
a
p
ar
am
eter
-
ef
f
icien
t
f
in
e
-
t
u
n
in
g
tech
n
i
q
u
e
d
esig
n
ed
to
r
e
d
u
ce
co
m
p
u
tatio
n
al
o
v
er
h
ea
d
.
I
n
s
tead
o
f
u
p
d
atin
g
th
e
en
tire
s
et
o
f
m
o
d
el
p
ar
am
eter
s
,
L
o
R
A
in
tr
o
d
u
ce
s
tr
ain
a
b
le
lo
w
-
r
an
k
m
atr
ices
in
to
s
elec
ted
lay
er
s
wh
ile
p
r
eser
v
in
g
th
e
o
r
ig
in
al
p
r
e
-
tr
ain
e
d
weig
h
t
s
.
T
h
is
m
ec
h
an
is
m
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
E
n
h
a
n
cin
g
S
QL in
jectio
n
d
ete
ctio
n
a
cc
u
r
a
cy
w
ith
fin
e
-
tu
n
in
g
LL
a
Ma
3
.
2
1
B
u
s
in
g
…
(
Dia
s
h
F
ir
d
a
u
s
)
1603
s
ig
n
if
ican
tly
d
ec
r
ea
s
es m
em
o
r
y
co
n
s
u
m
p
tio
n
an
d
tr
ai
n
in
g
co
s
ts
wh
ile
m
ain
tain
in
g
th
e
m
o
d
el
’
s
ab
ilit
y
to
lear
n
d
o
m
ain
-
s
p
ec
if
ic
k
n
o
wled
g
e.
C
o
n
s
eq
u
en
tly
,
L
o
R
A
o
f
f
er
s
an
ef
f
ec
tiv
e
s
o
lu
tio
n
f
o
r
d
e
p
lo
y
in
g
LLM
s
in
en
v
ir
o
n
m
en
ts
with
lim
ited
co
m
p
u
tatio
n
al
r
eso
u
r
ce
s
.
Fig
u
r
e
3
.
Ar
c
h
itectu
r
e
o
f
L
o
R
A
[
2
2
]
L
o
R
A
en
ab
les
p
ar
am
eter
-
e
f
f
ic
ien
t
f
in
e
-
tu
n
i
n
g
b
y
in
jectin
g
tr
ain
ab
le
lo
w
-
r
an
k
m
atr
ices
in
t
o
s
elec
ted
lay
er
s
o
f
th
e
b
ase
lan
g
u
a
g
e
m
o
d
el
wh
ile
k
ee
p
in
g
th
e
o
r
ig
i
n
al
m
o
d
el
weig
h
ts
f
r
o
ze
n
.
I
n
th
is
s
tu
d
y
,
L
o
R
A
is
co
n
f
ig
u
r
ed
with
a
r
a
n
k
=
64
an
d
a
s
ca
lin
g
f
ac
to
r
=
16
,
as
s
u
m
m
a
r
ized
in
T
ab
le
4
.
I
n
a
d
d
itio
n
,
th
e
o
v
er
all
tr
ain
in
g
c
o
n
f
ig
u
r
atio
n
,
in
clu
d
in
g
lear
n
in
g
r
ate,
b
atch
s
ize,
o
p
tim
izer
,
an
d
t
r
ain
in
g
s
tr
ateg
y
,
is
p
r
esen
ted
in
T
ab
le
5
to
e
n
s
u
r
e
clar
ity
an
d
r
ep
r
o
d
u
cib
ilit
y
.
T
h
is
ap
p
r
o
ac
h
m
in
im
izes
th
e
r
is
k
o
f
ca
tast
r
o
p
h
ic
f
o
r
g
ettin
g
a
n
d
ac
ce
ler
ates
t
h
e
tr
ain
in
g
p
r
o
ce
s
s
.
Fu
r
th
er
m
o
r
e,
co
m
p
ar
ativ
e
s
tu
d
ies
c
o
n
f
ir
m
th
at
L
o
R
A
s
ig
n
if
ican
tly
r
ed
u
ce
s
co
m
p
u
ta
tio
n
al
o
v
e
r
h
ea
d
co
m
p
ar
ed
to
f
u
ll
f
in
e
-
t
u
n
in
g
,
m
ak
i
n
g
it
an
id
ea
l
s
tr
ateg
y
f
o
r
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
cy
b
e
r
s
ec
u
r
ity
d
ep
lo
y
m
en
ts
.
T
h
is
ef
f
ici
en
cy
is
cr
u
cial
wh
en
p
r
o
ce
s
s
in
g
lar
g
e
-
s
ca
le
th
r
ea
t
d
atasets
in
r
ea
l
-
tim
e
en
v
ir
o
n
m
en
ts
[
2
3
]
.
T
ab
le
4
.
L
o
R
A
co
n
f
ig
u
r
atio
n
f
o
r
f
in
e
-
t
u
n
in
g
No
P
a
r
a
me
t
e
r
V
a
l
u
e
D
e
scri
p
t
i
o
n
1
F
i
n
e
-
t
u
n
i
n
g
m
e
t
h
o
d
Lo
R
A
P
a
r
a
me
t
e
r
-
e
f
f
i
c
i
e
n
t
f
i
n
e
-
t
u
n
i
n
g
a
p
p
r
o
a
c
h
2
Ta
sk
t
y
p
e
C
a
u
sa
l
l
a
n
g
u
a
g
e
mo
d
e
l
i
n
g
M
o
d
e
l
p
r
e
d
i
c
t
s
n
e
x
t
t
o
k
e
n
a
u
t
o
r
e
g
r
e
ss
i
v
e
l
y
3
R
a
n
k
(
)
64
D
i
me
n
si
o
n
o
f
l
o
w
-
r
a
n
k
m
a
t
r
i
c
e
s
4
Lo
R
A
a
l
p
h
a
16
S
c
a
l
i
n
g
f
a
c
t
o
r
f
o
r
L
o
R
A
u
p
d
a
t
e
s
5
Bias
N
o
n
e
N
o
b
i
a
s
p
a
r
a
m
e
t
e
r
s
w
e
r
e
t
r
a
i
n
e
d
6
Ta
r
g
e
t
m
o
d
u
l
e
s
S
e
l
e
c
t
e
d
mo
d
e
l
l
a
y
e
r
s
M
o
d
u
l
e
s a
d
a
p
t
e
d
u
si
n
g
L
o
R
A
7
Tr
a
i
n
a
b
l
e
p
a
r
a
m
e
t
e
r
s
Lo
R
A
l
a
y
e
r
s o
n
l
y
B
a
se
m
o
d
e
l
w
e
i
g
h
t
s fr
o
z
e
n
8
P
a
r
a
me
t
e
r
-
e
f
f
i
c
i
e
n
t
f
i
n
e
-
t
u
n
i
n
g
(
P
EFT
)
H
u
g
g
i
n
g
F
a
c
e
P
EFT
U
sed
f
o
r
L
o
R
A
i
m
p
l
e
me
n
t
a
t
i
o
n
T
ab
le
5
.
T
r
ai
n
in
g
h
y
p
e
r
p
ar
am
eter
s
No
P
a
r
a
me
t
e
r
V
a
l
u
e
1
Ep
o
c
h
s
1
2
Le
a
r
n
i
n
g
r
a
t
e
2
×
1
0
⁻
⁴
3
P
e
r
-
d
e
v
i
c
e
b
a
t
c
h
s
i
z
e
1
4
G
r
a
d
i
e
n
t
a
c
c
u
m
u
l
a
t
i
o
n
8
5
Ef
f
e
c
t
i
v
e
b
a
t
c
h
s
i
z
e
8
6
O
p
t
i
mi
z
e
r
P
a
g
e
d
A
d
a
m
W
(
3
2
-
b
i
t
)
7
W
e
i
g
h
t
d
e
c
a
y
0
.
0
0
1
8
LR
s
c
h
e
d
u
l
e
r
C
o
s
i
n
e
9
W
a
r
mu
p
r
a
t
i
o
0
.
0
3
10
M
a
x
g
r
a
d
i
e
n
t
n
o
r
m
0
.
3
11
G
r
a
d
i
e
n
t
c
h
e
c
k
p
o
i
n
t
i
n
g
En
a
b
l
e
d
12
P
r
e
c
i
s
i
o
n
F
P
1
6
13
Ev
a
l
u
a
t
i
o
n
s
t
r
a
t
e
g
y
S
t
e
p
s (e
v
e
r
y
5
0
st
e
p
s)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
TEL
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l
C
o
n
tr
o
l
,
Vo
l.
24
,
No
.
5
,
Octo
b
e
r
20
26
:
1
5
9
7
-
1
6
1
0
1604
T
ab
le
4
p
r
esen
ts
th
e
co
n
f
ig
u
r
atio
n
o
f
th
e
L
o
R
A
m
eth
o
d
u
s
ed
in
th
is
s
tu
d
y
.
T
h
e
T
ab
le
4
d
etails
th
e
k
ey
h
y
p
er
p
a
r
am
eter
s
th
at
co
n
tr
o
l
h
o
w
L
o
R
A
ad
ap
ts
th
e
b
ase
lan
g
u
ag
e
m
o
d
el,
in
cl
u
d
in
g
th
e
r
a
n
k
,
s
ca
lin
g
f
ac
to
r
,
a
n
d
d
r
o
p
o
u
t
r
ate.
T
h
ese
s
ettin
g
s
d
eter
m
in
e
t
h
e
ca
p
ac
ity
an
d
s
tab
ilit
y
o
f
th
e
ad
a
p
tatio
n
wh
ile
k
ee
p
in
g
th
e
n
u
m
b
er
o
f
tr
ain
ab
le
p
ar
a
m
eter
s
lo
w.
B
y
ap
p
ly
in
g
L
o
R
A
o
n
ly
to
s
elec
ted
tar
g
et
m
o
d
u
le
s
an
d
f
r
ee
zin
g
th
e
o
r
ig
in
al
m
o
d
el
weig
h
ts
,
th
e
f
in
e
-
tu
n
in
g
p
r
o
ce
s
s
b
ec
o
m
es
m
o
r
e
ef
f
icie
n
t
in
ter
m
s
o
f
m
em
o
r
y
u
s
ag
e
an
d
tr
ain
in
g
tim
e,
m
a
k
in
g
it su
itab
le
f
o
r
L
L
M
tr
ai
n
in
g
o
n
lim
ited
co
m
p
u
tatio
n
al
r
eso
u
r
ce
s
.
T
ab
le
5
p
r
esen
ts
th
e
m
ain
tr
ain
in
g
h
y
p
e
r
p
ar
am
eter
s
u
s
ed
in
th
e
f
in
e
-
tu
n
in
g
p
r
o
c
ess
.
T
h
ese
p
ar
am
eter
s
d
ef
in
e
th
e
o
p
tim
i
za
tio
n
s
tr
ateg
y
,
tr
ain
in
g
d
u
r
a
tio
n
,
an
d
m
em
o
r
y
m
an
a
g
em
e
n
t
s
ettin
g
s
ap
p
lied
d
u
r
in
g
m
o
d
el
tr
ain
in
g
.
T
h
e
T
ab
le
5
s
h
o
ws
th
at
th
e
m
o
d
el
w
as
tr
ain
ed
f
o
r
a
s
in
g
le
ep
o
ch
u
s
in
g
a
lear
n
in
g
r
ate
o
f
2
×1
0
⁻⁴.
A
s
in
g
le
ep
o
ch
wa
s
s
elec
ted
b
ec
au
s
e
p
r
elim
in
ar
y
ex
p
er
im
e
n
ts
s
h
o
wed
co
n
v
e
r
g
en
ce
o
f
e
v
alu
atio
n
lo
s
s
with
o
u
t
ad
d
itio
n
al
p
e
r
f
o
r
m
an
ce
g
ai
n
s
f
r
o
m
lo
n
g
e
r
tr
a
in
in
g
,
wh
ile
also
r
e
d
u
cin
g
co
m
p
u
tatio
n
al
co
s
ts
.
Gr
ad
ien
t
ac
cu
m
u
latio
n
an
d
ch
ec
k
p
o
i
n
tin
g
wer
e
em
p
lo
y
e
d
to
im
p
r
o
v
e
tr
ain
in
g
ef
f
icien
cy
,
wh
ile
a
co
s
in
e
lear
n
in
g
r
ate
s
ch
e
d
u
ler
with
w
ar
m
u
p
was u
s
ed
to
s
tab
ilize
th
e
o
p
tim
izatio
n
p
r
o
ce
s
s
.
2
.
4
.
M
o
del
ev
a
lua
t
i
o
n m
et
ri
cs
T
o
a
s
s
e
s
s
t
h
e
p
e
r
f
o
r
m
a
n
c
e
o
f
t
h
e
p
r
o
p
o
s
e
d
m
o
d
e
l
,
w
e
e
m
p
l
o
y
e
d
a
c
o
n
f
u
s
i
o
n
m
a
t
r
i
x
a
n
d
d
e
r
i
v
e
d
s
t
a
n
d
a
r
d
c
l
ass
i
f
i
c
at
i
o
n
m
e
t
r
ic
s
:
a
cc
u
r
a
c
y
,
p
r
e
c
is
i
o
n
,
r
e
ca
l
l
,
a
n
d
F
1
-
s
c
o
r
e
.
T
h
es
e
m
e
t
r
i
c
s
p
r
o
v
i
d
e
a
c
o
m
p
r
e
h
e
n
s
i
v
e
v
i
ew
o
f
t
h
e
m
o
d
e
l
’
s
a
b
i
l
it
y
t
o
m
i
n
i
m
i
ze
b
o
t
h
f
a
l
s
e
p
o
s
it
i
v
e
s
(
b
e
n
i
g
n
q
u
e
r
i
es
f
l
a
g
g
e
d
a
s
a
tt
a
c
k
s
)
a
n
d
f
a
l
s
e
n
e
g
a
ti
v
e
s
(
at
t
a
c
k
s
b
y
p
a
s
s
i
n
g
d
e
t
e
c
t
i
o
n
)
.
T
h
e
e
v
al
u
a
ti
o
n
m
e
t
r
i
c
s
a
r
e
c
a
l
c
u
l
a
t
e
d
u
s
i
n
g
(
1
)
t
o
(
4
)
[
2
4
]
:
A
cc
u
r
a
cy
=
+
+
+
+
(
1
)
P
r
ec
is
io
n
=
+
(
2
)
R
ec
a
ll
=
+
(
3
)
1
-
S
co
r
e
=
2
×
(
P
r
e
c
is
ion
×
R
e
c
all
)
P
r
e
c
is
ion
+
R
e
c
all
(
4
)
T
h
e
p
r
im
ar
y
p
er
f
o
r
m
a
n
ce
in
d
icato
r
s
—
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e
—
ar
e
c
o
m
p
u
ted
to
m
ea
s
u
r
e
th
e
m
o
d
el
’
s
o
v
e
r
all
p
er
f
o
r
m
an
ce
.
W
ith
in
th
e
co
n
f
u
s
io
n
m
atr
ix
,
T
P
an
d
T
N
d
en
o
te
p
o
s
itiv
e
an
d
n
eg
ativ
e
s
am
p
les
th
at
ar
e
class
if
ied
co
r
r
ec
tly
,
wh
er
ea
s
FP
an
d
FN
r
ef
er
to
in
s
tan
ce
s
in
wh
ich
th
e
s
y
s
tem
p
r
o
d
u
ce
s
in
co
r
r
ec
t
p
r
ed
ictio
n
s
.
T
h
e
m
ath
em
atica
l
d
ef
in
itio
n
s
o
f
ea
ch
m
etr
ic
ar
e
p
r
o
v
i
d
ed
in
(
1
)
–
(
4
)
.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
I
n
t
h
i
s
s
t
u
d
y
,
t
h
e
e
x
p
e
r
i
m
e
n
ts
w
e
r
e
c
o
n
d
u
c
t
e
d
u
s
i
n
g
a
b
en
c
h
m
a
r
k
d
a
t
a
s
e
t
d
e
r
i
v
e
d
f
r
o
m
K
a
g
g
l
e
,
d
e
s
i
g
n
e
d
f
o
r
SQ
L
i
d
et
e
c
ti
o
n
.
T
h
e
d
a
t
a
s
et
c
o
n
s
i
s
t
s
o
f
b
o
t
h
m
a
l
ic
i
o
u
s
S
Q
L
i
i
n
p
u
t
s
a
n
d
b
e
n
i
g
n
te
x
t
s
a
m
p
le
s
,
w
h
i
c
h
w
e
r
e
u
s
e
d
t
o
t
r
a
i
n
a
n
d
e
v
a
l
u
a
te
t
h
e
p
r
o
p
o
s
e
d
m
o
d
e
l
.
M
o
d
e
l
t
r
a
i
n
i
n
g
w
as
p
e
r
f
o
r
m
e
d
u
s
i
n
g
th
e
L
L
a
M
A
3
.
2
–
1
B
a
r
c
h
i
t
e
c
t
u
r
e
wi
t
h
t
h
e
t
r
a
i
n
i
n
g
c
o
n
f
i
g
u
r
a
t
i
o
n
d
e
s
c
r
i
b
e
d
i
n
T
a
b
l
e
4
.
T
h
e
t
r
a
i
n
i
n
g
a
n
d
e
v
a
l
u
a
t
i
o
n
p
r
o
c
e
s
s
es
w
e
r
e
m
o
n
i
t
o
r
e
d
u
s
i
n
g
l
o
s
s
v
a
l
u
e
s
,
wh
e
r
e
t
h
e
t
r
ai
n
i
n
g
l
o
s
s
i
s
p
r
es
e
n
te
d
i
n
F
i
g
u
r
e
4
,
a
n
d
t
h
e
e
v
al
u
a
t
io
n
l
o
s
s
i
s
s
h
o
w
n
i
n
F
i
g
u
r
e
5
.
T
h
e
r
es
u
l
ts
d
is
c
u
s
s
e
d
i
n
t
h
is
s
ec
t
i
o
n
h
i
g
h
li
g
h
t
t
h
e
l
ea
r
n
i
n
g
b
e
h
a
v
i
o
r
o
f
t
h
e
m
o
d
e
l
an
d
i
t
s
e
f
f
ec
t
i
v
e
n
es
s
i
n
d
i
s
ti
n
g
u
i
s
h
i
n
g
b
e
t
we
e
n
m
a
l
ic
i
o
u
s
a
n
d
n
o
n
-
m
a
l
i
c
i
o
u
s
i
n
p
u
t
s
.
Fig
u
r
e
4
.
T
r
ain
in
g
l
o
s
s
p
r
o
g
r
e
s
s
io
n
o
v
er
g
lo
b
al
s
tep
s
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
E
n
h
a
n
cin
g
S
QL in
jectio
n
d
ete
ctio
n
a
cc
u
r
a
cy
w
ith
fin
e
-
tu
n
in
g
LL
a
Ma
3
.
2
1
B
u
s
in
g
…
(
Dia
s
h
F
ir
d
a
u
s
)
1605
Fig
u
r
e
5
.
E
v
alu
atio
n
l
o
s
s
p
r
o
g
r
ess
io
n
o
v
er
g
l
o
b
al
s
tep
s
Fig
u
r
e
4
illu
s
tr
ates
th
e
tr
en
d
o
f
tr
ain
i
n
g
lo
s
s
th
r
o
u
g
h
o
u
t
th
e
f
in
e
-
t
u
n
in
g
p
r
o
ce
s
s
ac
r
o
s
s
g
lo
b
al
tr
ain
in
g
s
tep
s
.
Var
iatio
n
s
in
t
h
e
lo
s
s
v
alu
es
ar
e
o
b
s
er
v
ed
d
u
r
in
g
tr
ain
in
g
,
wh
ich
is
ex
p
ec
ted
wh
en
s
to
ch
asti
c
o
p
tim
izatio
n
,
s
m
all
b
atch
s
izes,
an
d
g
r
ad
ien
t
ac
c
u
m
u
latio
n
ar
e
em
p
lo
y
e
d
.
Desp
ite
th
e
s
e
f
lu
ctu
atio
n
s
,
th
e
o
v
er
all
lo
s
s
r
em
ain
s
with
in
a
s
tab
le
r
an
g
e,
in
d
icatin
g
t
h
at
th
e
o
p
tim
izatio
n
p
r
o
ce
s
s
p
r
o
ce
ed
s
in
a
co
n
tr
o
lle
d
m
an
n
er
.
T
h
e
h
ig
h
est
tr
ain
in
g
lo
s
s
v
alu
e
o
f
5
.
1
3
9
1
was
r
ec
o
r
d
ed
at
g
lo
b
al
s
tep
2
d
u
r
in
g
th
e
in
itial
ad
ap
tatio
n
s
tag
e
o
f
th
e
m
o
d
el.
As
tr
ain
in
g
p
r
o
g
r
ess
ed
,
th
e
lo
s
s
g
en
er
al
ly
d
ec
lin
ed
an
d
ex
h
ib
ite
d
a
m
o
r
e
s
tab
le
p
atter
n
.
T
h
e
m
in
im
u
m
lo
s
s
v
alu
e
o
f
0
.
4
was
ac
h
iev
ed
at
g
lo
b
al
s
tep
7
6
4
3
,
s
u
g
g
esti
n
g
ef
f
ec
tiv
e
o
p
tim
izatio
n
d
u
r
in
g
f
in
e
-
tu
n
in
g
.
Ov
er
all,
th
e
d
ec
r
ea
s
in
g
tr
en
d
in
tr
ain
i
n
g
lo
s
s
d
em
o
n
s
tr
ates
th
e
m
o
d
el
’
s
ca
p
ab
ilit
y
to
lear
n
r
elev
an
t
r
ep
r
esen
tatio
n
s
f
r
o
m
th
e
d
ataset
with
o
u
t
ex
h
ib
itin
g
s
ig
n
s
o
f
d
iv
er
g
e
n
ce
o
r
u
n
s
t
ab
le
b
eh
av
i
o
r
.
T
h
is
o
b
s
er
v
atio
n
is
co
n
s
is
ten
t
with
th
e
ch
ar
ac
ter
is
tics
o
f
L
o
R
A
-
b
ased
f
in
e
-
tu
n
in
g
p
e
r
f
o
r
m
ed
u
s
in
g
a
lim
ited
n
u
m
b
er
o
f
e
p
o
ch
s
an
d
a
r
elati
v
ely
s
m
all
ef
f
ec
tiv
e
b
atc
h
s
ize.
C
o
m
p
lem
en
tin
g
th
e
tr
ai
n
in
g
lo
s
s
,
Fig
u
r
e
5
p
r
esen
ts
th
e
ev
alu
atio
n
lo
s
s
o
n
th
e
v
ali
d
atio
n
s
et.
Fig
u
r
e
5
p
r
esen
ts
t
h
e
ev
al
u
atio
n
lo
s
s
tr
en
d
o
v
er
g
lo
b
al
tr
ain
in
g
s
tep
s
.
At
th
e
ea
r
ly
s
ta
g
e
o
f
tr
ain
i
n
g
,
th
e
ev
alu
atio
n
lo
s
s
s
h
o
ws
a
r
elativ
ely
h
ig
h
v
alu
e,
r
ef
lectin
g
th
e
m
o
d
el
’
s
lim
ited
g
en
e
r
aliza
tio
n
ca
p
ab
ilit
y
b
e
f
o
r
e
s
u
f
f
icien
t
lear
n
in
g
h
as
tak
en
p
lace
.
As
tr
ain
in
g
p
r
o
g
r
ess
es,
th
e
ev
alu
atio
n
lo
s
s
d
ec
r
ea
s
es
s
tead
ily
,
in
d
icatin
g
im
p
r
o
v
e
d
m
o
d
el
p
e
r
f
o
r
m
an
c
e
o
n
th
e
ev
alu
atio
n
d
ataset.
B
ased
o
n
th
e
o
b
s
er
v
ed
r
esu
lts
,
th
e
m
ax
im
u
m
ev
alu
atio
n
lo
s
s
o
f
3
.
3
1
2
4
o
cc
u
r
s
at
g
lo
b
al
s
tep
5
0
,
wh
ich
co
r
r
esp
o
n
d
s
to
th
e
in
itial
p
h
a
s
e
o
f
tr
ain
in
g
.
T
h
e
ev
alu
atio
n
lo
s
s
co
n
tin
u
es
to
d
ec
lin
e
an
d
r
ea
c
h
es
its
m
in
im
u
m
v
al
u
e
o
f
2
.
6
6
4
0
at
g
lo
b
al
s
tep
1
0
,
8
0
0
,
s
u
g
g
esti
n
g
th
at
th
e
m
o
d
el
ac
h
i
ev
es
its
b
est
g
en
er
aliza
tio
n
p
e
r
f
o
r
m
a
n
ce
to
war
d
th
e
en
d
o
f
t
h
e
tr
ain
in
g
p
r
o
ce
s
s
.
Ov
er
all,
th
e
co
n
s
is
ten
t
r
ed
u
cti
o
n
an
d
s
tab
ilizatio
n
o
f
th
e
e
v
alu
atio
n
lo
s
s
in
d
icate
th
at
t
h
e
m
o
d
el
s
u
cc
ess
f
u
lly
lear
n
s
r
elev
an
t
p
atter
n
s
f
r
o
m
th
e
tr
ain
in
g
d
ata
with
o
u
t
ex
h
ib
itin
g
s
ig
n
s
o
f
o
v
er
f
itti
n
g
.
T
h
is
b
eh
a
v
io
r
d
em
o
n
s
tr
ates
th
at
th
e
ap
p
lied
tr
ain
in
g
co
n
f
i
g
u
r
atio
n
an
d
L
o
R
A
-
b
ased
f
in
e
-
tu
n
in
g
s
tr
ateg
y
ef
f
ec
tiv
ely
s
u
p
p
o
r
t
s
tab
le
g
en
er
aliza
tio
n
o
n
u
n
s
ee
n
d
ata.
B
ased
o
n
th
e
ev
alu
atio
n
lo
s
s
an
aly
s
is
,
th
e
m
o
d
el
d
em
o
n
s
tr
ates
s
tab
le
an
d
co
n
s
is
ten
t
g
en
er
aliza
tio
n
p
er
f
o
r
m
an
ce
o
n
th
e
ev
alu
atio
n
d
ataset.
T
o
f
u
r
th
er
ass
ess
th
e
class
if
icatio
n
p
er
f
o
r
m
an
ce
i
n
t
er
m
s
o
f
p
r
e
d
ictio
n
o
u
tco
m
es,
th
e
ev
al
u
atio
n
is
co
n
tin
u
ed
u
s
in
g
a
co
n
f
u
s
io
n
m
atr
ix
,
wh
ich
p
r
o
v
id
es
a
d
etail
ed
v
iew
o
f
tr
u
e
an
d
f
alse p
r
ed
ictio
n
s
ac
r
o
s
s
class
es.
T
h
e
r
esu
lts
o
f
th
e
c
o
n
f
u
s
io
n
m
atr
ix
ar
e
p
r
esen
ted
in
Fig
u
r
e
6
.
Fig
u
r
e
6
.
C
o
n
f
u
s
io
n
m
atr
i
x
o
f
th
e
f
in
e
-
tu
n
ed
L
L
aM
A
3
.
2
m
o
d
el
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
TEL
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l
C
o
n
tr
o
l
,
Vo
l.
24
,
No
.
5
,
Octo
b
e
r
20
26
:
1
5
9
7
-
1
6
1
0
1606
Fig
u
r
e
6
p
r
o
v
id
es
a
d
etailed
o
v
e
r
v
iew
o
f
t
h
e
class
if
icatio
n
o
u
tco
m
es
g
en
er
ated
b
y
th
e
p
r
o
p
o
s
e
d
m
o
d
el
o
n
th
e
ev
al
u
atio
n
d
atas
et.
T
h
e
co
n
f
u
s
io
n
m
at
r
ix
in
d
i
ca
tes
th
at
1
0
,
3
2
6
b
en
i
g
n
s
am
p
les
wer
e
co
r
r
ec
tly
id
en
tifie
d
as
n
o
n
-
m
alicio
u
s
(
t
r
u
e
n
e
g
ativ
es
/TN
)
,
wh
ile
1
1
,
5
5
8
SQLi
s
am
p
les
wer
e
ac
cu
r
ately
class
if
ied
as
m
alicio
u
s
(
tr
u
e
p
o
s
itiv
es
/TP
)
.
Misclas
s
if
icatio
n
ca
s
es
wer
e
lim
ited
,
co
n
s
is
tin
g
o
f
f
o
u
r
f
alse
p
o
s
itiv
es
an
d
s
ix
teen
f
alse
n
eg
ativ
es.
T
h
ese
f
in
d
in
g
s
d
em
o
n
s
tr
ate
th
e
m
o
d
el
’
s
s
tr
o
n
g
ca
p
a
b
ilit
y
to
d
i
f
f
er
en
tiate
b
etwe
en
leg
itima
te
u
s
er
in
p
u
ts
an
d
SQLi
attem
p
ts
.
Fu
r
th
er
m
o
r
e,
th
e
r
elativ
ely
lo
w
n
u
m
b
er
o
f
f
alse
alar
m
s
s
u
g
g
ests
th
at
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
ca
n
ef
f
ec
tiv
ely
m
in
im
ize
u
n
n
ec
ess
ar
y
b
lo
c
k
in
g
o
f
b
en
i
g
n
r
e
q
u
ests
wh
ile
m
ain
tain
in
g
h
ig
h
s
en
s
itiv
ity
t
o
war
d
m
alicio
u
s
p
atter
n
s
.
Ov
er
all,
th
e
co
n
f
u
s
io
n
m
atr
ix
co
n
f
ir
m
s
th
e
r
eliab
ilit
y
an
d
r
o
b
u
s
tn
ess
o
f
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
f
o
r
SQLi
d
etec
ti
o
n
[
2
5
]
.
T
ab
le
6
p
r
esen
ts
a
b
en
ch
m
ar
k
co
m
p
ar
is
o
n
b
etwe
en
th
e
p
r
o
p
o
s
ed
L
L
aM
A
3
.
2
–
L
o
R
A
m
o
d
el
an
d
v
ar
io
u
s
tr
a
d
itio
n
al
ML
an
d
D
L
b
aselin
es.
T
h
e
R
F
m
o
d
el
ac
h
iev
es
an
ac
c
u
r
ac
y
o
f
9
7
.
4
8
%
,
with
a
p
r
ec
is
io
n
o
f
9
8
.
7
5
%,
a
r
ec
all
o
f
9
6
.
4
9
%,
an
d
an
F1
-
s
co
r
e
o
f
9
7
.
6
1
%,
i
n
d
icatin
g
s
tr
o
n
g
an
d
b
alan
ce
d
p
er
f
o
r
m
an
ce
.
T
h
e
XGBo
o
s
t
m
o
d
el
r
ec
o
r
d
s
an
a
cc
u
r
ac
y
o
f
9
6
.
9
6
%,
p
r
ec
is
io
n
o
f
9
9
.
3
7
%,
r
ec
all
o
f
9
4
.
8
9
%,
an
d
an
F1
-
s
co
r
e
o
f
9
7
.
0
8
%,
s
h
o
win
g
h
ig
h
p
r
ec
i
s
io
n
b
u
t
s
lig
h
tly
lo
wer
r
ec
all.
Similar
ly
,
th
e
DT
m
o
d
el
ac
h
iev
es
9
6
.
6
1
%
ac
cu
r
ac
y
,
9
9
.
0
6
%
p
r
ec
is
io
n
,
9
4
.
5
4
%
r
ec
all,
a
n
d
a
n
F1
-
s
co
r
e
o
f
9
6
.
7
5
%.
Am
o
n
g
DL
ap
p
r
o
ac
h
es,
th
e
C
NN
m
o
d
el
attain
s
an
ac
c
u
r
ac
y
o
f
9
7
.
8
4
%,
p
r
ec
is
io
n
o
f
9
9
.
5
2
%,
r
ec
all
o
f
9
6
.
4
2
%,
an
d
an
F1
-
s
co
r
e
o
f
9
7
.
9
4
%,
d
em
o
n
s
tr
atin
g
co
m
p
etitiv
e
p
e
r
f
o
r
m
a
n
ce
co
m
p
a
r
ed
to
tr
ad
iti
o
n
al
ML
m
o
d
els.
T
h
e
L
STM
m
o
d
el
s
h
o
ws
lo
wer
b
u
t
co
n
s
is
ten
t
r
esu
lts
,
with
an
ac
cu
r
ac
y
o
f
9
2
.
2
5
%,
p
r
ec
is
io
n
o
f
9
2
.
8
4
%,
r
ec
all
o
f
9
2
.
6
0
%,
an
d
an
F1
-
s
co
r
e
o
f
9
2
.
7
2
%.
T
h
e
b
aselin
e
L
L
aM
A
3
.
2
–
1
B
m
o
d
el
with
o
u
t
f
in
e
-
tu
n
in
g
ac
h
iev
es
a
n
ac
cu
r
ac
y
o
f
o
n
l
y
5
0
.
2
1
%,
d
esp
ite
a
h
ig
h
p
r
ec
is
io
n
o
f
9
9
.
2
6
%.
Ho
wev
er
,
its
r
ec
all
d
r
o
p
s
s
ig
n
if
ican
tly
to
5
.
7
8
%,
r
es
u
ltin
g
in
a
lo
w
F1
-
s
co
r
e
o
f
1
0
.
9
3
%,
in
d
icatin
g
p
o
o
r
d
etec
tio
n
ca
p
ab
ilit
y
with
o
u
t
task
-
s
p
ec
if
ic
ad
ap
tatio
n
.
I
n
co
n
tr
ast,
th
e
f
in
e
-
tu
n
ed
L
L
aM
A
3
.
2
–
1
B
m
o
d
e
l
u
s
in
g
L
o
R
A
ac
h
iev
es
th
e
b
est
o
v
er
all
p
er
f
o
r
m
an
ce
,
wi
th
an
ac
cu
r
ac
y
o
f
9
9
.
9
1
%,
p
r
ec
is
io
n
o
f
9
9
.
9
6
%,
r
ec
all
o
f
9
9
.
8
6
%,
an
d
a
n
F1
-
s
co
r
e
o
f
9
9
.
9
1
%.
T
h
ese
r
e
s
u
lts
co
n
f
ir
m
th
at
L
o
R
A
-
b
ased
f
in
e
-
tu
n
i
n
g
s
u
b
s
t
an
tially
im
p
r
o
v
es th
e
ef
f
ec
tiv
e
n
ess
o
f
LLM
s
f
o
r
SQLi
d
etec
t
io
n
.
T
ab
le
6
.
C
o
m
p
a
r
is
o
n
o
f
m
o
d
el
p
er
f
o
r
m
an
ce
a
g
ain
s
t b
aselin
es
No
M
o
d
e
l
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
s
c
o
r
e
1
RF
9
7
.
4
8
%
9
8
.
7
5
%
9
6
.
4
9
%
9
7
.
6
1
%
2
X
G
B
o
o
st
9
6
.
9
6
%
9
9
.
3
7
%
9
4
.
8
9
%
9
7
.
0
8
%
3
DT
9
6
.
6
1
%
9
9
.
0
6
%
9
4
.
5
4
%
9
6
.
7
5
%
4
C
N
N
9
7
.
8
4
%
9
9
.
5
2
%
9
6
.
4
2
%
9
7
.
9
4
%
5
LSTM
9
2
.
2
5
%
9
2
.
8
4
%
9
2
.
6
0
%
9
2
.
7
2
%
6
LLa
M
a
3
.
2
1B
5
0
.
2
1
%
9
9
.
2
6
%
5
.
7
8
%
1
0
.
9
3
%
7
F
i
n
e
t
u
n
i
n
g
LL
a
M
a
3
.
2
1B
(
Lo
R
a
)
9
9
.
9
1
%
9
9
.
9
6
%
9
9
.
8
6
%
9
9
.
9
1
%
I
n
ad
d
itio
n
t
o
th
e
ex
p
e
r
im
en
t
al
ev
alu
atio
n
,
th
e
p
r
o
p
o
s
ed
m
o
d
el
h
as
b
ee
n
im
p
lem
en
ted
in
a
web
-
b
ased
ap
p
licatio
n
,
as
illu
s
tr
at
ed
in
Fig
u
r
e
1
.
T
h
e
web
-
b
ase
d
s
y
s
tem
allo
ws
u
s
er
s
to
s
u
b
m
it
SQL
q
u
er
ies
o
r
tex
tu
al
in
p
u
ts
an
d
r
ec
ei
v
e
r
ea
l
-
tim
e
class
if
icatio
n
r
esu
lts
in
d
icatin
g
wh
eth
er
th
e
in
p
u
t
is
m
alicio
u
s
o
r
b
en
ig
n
.
T
h
is
im
p
lem
en
tatio
n
d
em
o
n
s
tr
ates
th
at
th
e
f
in
e
-
tu
n
ed
m
o
d
el
ca
n
b
e
ef
f
ec
tiv
ely
in
teg
r
ated
in
to
p
r
ac
tical
s
ec
u
r
ity
s
y
s
tem
s
,
s
u
ch
as
W
AF
s
o
r
d
atab
ase
s
ec
u
r
ity
m
o
n
ito
r
in
g
p
latf
o
r
m
s
.
B
ased
o
n
Fig
u
r
e
1
,
th
e
s
y
s
tem
ar
ch
itectu
r
e
co
n
f
ir
m
s
th
e
f
ea
s
ib
ilit
y
o
f
d
ep
lo
y
in
g
t
h
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
in
r
ea
l
-
wo
r
ld
en
v
ir
o
n
m
e
n
ts
b
ey
o
n
d
o
f
f
lin
e
e
x
p
er
im
e
n
tatio
n
.
Fig
u
r
e
7
p
r
esen
ts
ex
a
m
p
les
o
f
th
e
s
y
s
tem
’
s
r
ea
l
-
tim
e
d
etec
tio
n
ca
p
ab
ilit
y
u
s
in
g
b
o
th
m
alicio
u
s
an
d
b
en
ig
n
in
p
u
ts
.
T
h
e
m
alicio
u
s
ex
am
p
le
co
n
s
is
ts
o
f
a
lo
g
in
r
eq
u
est
co
n
tain
in
g
th
e
SQLi
p
ay
lo
ad
`
’
OR
1
=1
--
`
,
wh
er
ea
s
th
e
b
en
ig
n
ex
am
p
le
co
n
s
is
ts
o
f
a
f
ilm
r
ev
iew
s
u
b
m
itted
as
o
r
d
in
ar
y
u
s
er
-
g
en
er
ated
tex
t.
T
h
e
r
esu
lts
d
em
o
n
s
tr
ate
t
h
at
th
e
p
r
o
p
o
s
ed
s
y
s
tem
is
ca
p
ab
le
o
f
ac
cu
r
ately
d
is
tin
g
u
is
h
in
g
m
alicio
u
s
SQLi
attem
p
ts
f
r
o
m
leg
itima
te
u
s
er
i
n
p
u
ts
in
p
r
a
ctica
l d
ep
lo
y
m
en
t
s
ce
n
ar
io
s
.
As
a
r
esu
lt,
th
e
s
y
s
tem
d
is
p
la
y
s
a
n
o
tific
atio
n
in
d
icatin
g
th
at
th
e
u
s
er
is
attem
p
tin
g
a
B
o
o
lean
-
b
ased
SQLi
attac
k
,
an
d
ac
ce
s
s
to
t
h
e
s
y
s
tem
is
d
en
ied
.
Fu
r
th
er
m
o
r
e,
f
o
r
d
ig
ital
f
o
r
en
s
ic
p
u
r
p
o
s
es,
th
e
m
o
d
el
p
r
o
v
id
es a
co
n
te
x
tu
al
ex
p
lan
at
io
n
id
en
tify
in
g
th
e
d
etec
ted
att
ac
k
alo
n
g
with
r
ec
o
m
m
en
d
e
d
m
itig
atio
n
ac
tio
n
s
.
T
h
e
o
u
tp
u
t
o
f
th
is
d
ig
ital
f
o
r
en
s
ic
an
aly
s
is
ca
n
b
e
o
b
s
er
v
e
d
in
Fig
u
r
e
8
.
T
h
is
ca
p
ab
ilit
y
s
u
p
p
o
r
ts
n
o
t
o
n
ly
r
ea
l
-
tim
e
attac
k
p
r
ev
en
tio
n
b
u
t
also
p
o
s
t
-
in
cid
e
n
t
in
v
esti
g
a
tio
n
b
y
ass
is
tin
g
ad
m
in
is
tr
ato
r
s
in
u
n
d
er
s
tan
d
in
g
th
e
attac
k
ch
ar
ac
ter
is
tics
an
d
a
p
p
r
o
p
r
iate
r
esp
o
n
s
e
m
ea
s
u
r
es.
Fig
u
r
e
8
p
r
esen
ts
th
e
d
ig
ital
f
o
r
en
s
ic
r
ep
o
r
t
g
e
n
er
ated
b
y
th
e
p
r
o
p
o
s
ed
s
y
s
tem
f
o
r
d
ete
cted
SQLi
ev
en
ts
.
T
h
e
r
ep
o
r
t
r
ec
o
r
d
s
d
etailed
in
f
o
r
m
atio
n
f
o
r
ea
ch
d
etec
ted
attac
k
,
in
clu
d
i
n
g
t
h
e
tim
estam
p
,
s
o
u
r
ce
in
ter
n
et
p
r
o
to
c
o
l
(
IP
)
ad
d
r
e
s
s
,
u
s
er
ag
en
t,
tar
g
et
f
ield
,
d
etec
ted
p
ay
l
o
ad
,
attac
k
t
y
p
e,
r
ec
o
m
m
e
n
d
ed
m
itig
atio
n
,
an
d
s
y
s
tem
ac
tio
n
.
T
h
e
r
esu
lts
s
h
o
w
th
at
th
e
s
y
s
tem
is
ab
le
to
ac
cu
r
ately
id
e
n
tify
v
ar
io
u
s
ty
p
es
o
f
SQLi
attac
k
s
,
s
u
ch
as
B
o
o
lean
-
b
ased
,
T
im
e
-
b
ased
,
UNI
ON
-
b
ased
,
an
d
Gen
e
r
ic
SQ
L
i
p
atter
n
s
.
E
ac
h
m
alicio
u
s
p
ay
lo
ad
is
clea
r
ly
h
ig
h
lig
h
ted
,
allo
win
g
s
ec
u
r
ity
a
n
aly
s
ts
to
q
u
ick
ly
u
n
d
er
s
tan
d
th
e
s
tr
u
ctu
r
e
o
f
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.