Jou nal o Compu e Technology and So wa e
ISSN: 2998-2383
Vol. 3, No. 4, 2024
Enhancing In elligen Anomaly De ec ion in Cloud Backend
Sys ems h ough Con as i e Lea ning and Sensi i i y Analysis
Ziyu Cheng
Uni e si y o Sou he n Cali o nia, Los Angeles, USA
[email protected]
Abs ac : This s udy in es iga es anomaly de ec ion in cloud backend sys ems and add esses he limi a ions o adi ional
me hods unde high-dimensional complex da a and sca ce anomaly samples. A con as i e lea ning-based algo i hm is p oposed,
which cons uc s mo e disc imina i e la en space ep esen a ions h ough ea u e mapping and ep esen a ion lea ning and
achie es e ec i e sepa a ion o no mal and abno mal pa e ns by join ly op imizing con as i e loss and classi ica ion loss. To
alida e he e ec i eness o he me hod, compa a i e expe imen s we e conduc ed on a public da ase , and he esul s show ha
he p oposed model ou pe o ms se e al mains eam app oaches in e ms o AUC, ACC, F1-Sco e, and P ecision. Sensi i i y
expe imen s we e also pe o med o analyze he e ec s o empe a u e pa ame e , lea ning a e, nega i e sample a io, and
en i onmen al dis u bance on model pe o mance. The esul s demons a e ha p ope hype pa ame e selec ion and
en i onmen al modeling no only imp o e o e all de ec ion pe o mance bu also enhance obus ness and s abili y. By combining
compa a i e expe imen s wi h sensi i i y analysis, his s udy comp ehensi ely e i ies he e ec i eness o he con as i e
lea ning-based anomaly de ec ion me hod in cloud backend scena ios and con i ms i s po en ial applica ion alue in complex
sys em ope a ions.
Keywo ds: Cloud backend; con as i e lea ning; anomaly de ec ion; sensi i i y analysis
1. In oduc ion
In he e a o apid digi aliza ion and in elligen
de elopmen , cloud compu ing has become a c i ical
ounda ion o he in o ma ion in as uc u e o en e p ises and
o ganiza ions. As he co e ha suppo s di e se applica ions,
he cloud backend plays an i eplaceable ole in p ocessing
massi e da a, handling high-concu ency eques s, and
ensu ing se ice con inui y. Howe e , wi h he inc easing
complexi y o business scena ios and he con inuous g ow h o
da a scale, anomalies in cloud backend sys ems ha e become
mo e p ominen . These anomalies may cause pe o mance
deg ada ion, se ice in e up ions, and e en lead o se e e
secu i y isks and economic losses. The e o e, how o
e icien ly and accu a ely iden i y anomalies in cloud backends
has become a key p oblem in he ield o in o ma ion
echnology[1].
T adi ional anomaly de ec ion me hods mainly ely on
s a is ical analysis o ule ma ching. These app oaches a e o en
inadequa e when aced wi h complex, di e se, and high-
dimensional cloud backend da a. S a ic h esholds and ixed
ules canno adap o dynamic en i onmen s. Wi h he wide
adop ion o mic ose ice a chi ec u es, con aine iza ion, and
dis ibu ed deploymen , sys em s uc u es ha e become mo e
lexible and complex, while anomaly pa e ns ha e become
inc easingly di e se[2]. This makes i di icul o adi ional
me hods o cap u e hidden pa e ns in high-dimensional
ea u es, esul ing in poo de ec ion pe o mance. In his
con ex , explo ing new in elligen me hods o enhance he
accu acy and obus ness o anomaly de ec ion in cloud
backends is o g ea impo ance.
In ecen yea s, con as i e lea ning has eme ged as a
powe ul sel -supe ised app oach in ep esen a ion lea ning
and classi ica ion asks. Compa ing ea u es be ween no mal
and abno mal da a enables models o lea n disc imina i e
ep esen a ions wi hou equi ing la ge amoun s o labeled da a.
Compa ed wi h adi ional supe ised me hods, con as i e
lea ning can be e cap u e in insic s uc u al ela ionships
wi hin da a, imp o ing he model's abili y o ecognize
anomaly pa e ns in complex en i onmen s. In cloud backend
scena ios, con as i e lea ning no only enhances ea u e
disc iminabili y bu also e ec i ely add esses da a imbalance
and he sca ci y o abno mal samples. This p o ides new
pe spec i es o anomaly de ec ion[3,4].
A he socie al le el, cloud se ices ha e pene a ed c i ical
indus ies such as inance, heal hca e, anspo a ion, and
ene gy. The s abili y and secu i y o hese sys ems di ec ly
a ec economic de elopmen and social ope a ions. I backend
anomalies canno be de ec ed and esol ed in ime, la ge-scale
se ice in e up ions may occu , po en ially igge ing sys emic
isks. Resea ch on con as i e lea ning-based anomaly
de ec ion in cloud backends can he e o e imp o e he le el o
in elligen ope a ions a he echnical laye , while also
enhancing esilience and eliabili y a he socie al laye . Such
wo k aligns wi h na ional s a egies o secu e and in elligen
managemen o nex -gene a ion in o ma ion in as uc u e,
ca ying bo h s a egic alue and p ac ical signi icance.
In summa y, esea ch on anomaly de ec ion in cloud
backends using con as i e lea ning no only con ibu es o he
de elopmen o in elligen anomaly de ec ion heo y a he
academic le el bu also add esses p ac ical demands o
ensu ing cloud se ice quali y in indus y[5]. I will help build
mo e in elligen and eliable backend ope a ion sys ems, educe
isks, and imp o e se ice quali y and use expe ience. This
esea ch is o g ea impo ance o ad ancing he high-quali y
de elopmen o cloud compu ing and sa egua ding he s able
ope a ion o he digi al economy.
2. Rela ed wo k
In ecen yea s, anomaly de ec ion in cloud backends has
g adually become an impo an esea ch ocus in bo h
academia and indus y. Exis ing wo k mainly ollows h ee
pa hs, including s a is ical modeling, machine lea ning, and
deep lea ning[6]. Ea ly s udies o en elied on h eshold
se ings and s a is ical dis ibu ion analysis, such as de ec ing
anomalies h ough luc ua ions in sys em logs o pe o mance
indica o s. These me hods a e simple o implemen and
compu a ionally e icien . Howe e , hey a e es ic ed o s a ic
ules. When acing high-dimensional, mul i-sou ce
he e ogeneous da a o dynamic en i onmen s, hei de ec ion
pe o mance declines signi ican ly, wi h equen alse ala ms
and missed de ec ions. This makes hem unsui able o mode n
cloud backend sys ems.
Wi h he de elopmen o machine lea ning, esea che s
in oduced classi ica ion and clus e ing me hods in o anomaly
de ec ion. These app oaches can lea n anomaly pa e ns om
his o ical da a and gene alize be e han adi ional me hods.
Common ideas include using suppo ec o machines, andom
o es s, and clus e ing analysis o iden i y anomalies in he
ea u e space. Howe e , hei pe o mance is limi ed by model
capaci y and hea y eliance on ea u e enginee ing. They
s uggle wi h complex nonlinea ela ionships. When da a
dimensions inc ease o abno mal samples a e sca ce, esul s a e
o en unsa is ac o y. Mo eo e , hese me hods usually equi e
la ge amoun s o labeled da a, which is di icul o ob ain in
p ac ical cloud backend en i onmen s[7].
The ise o deep lea ning has b ough new pe spec i es o
anomaly de ec ion in cloud backends. Le e aging he s ong
ea u e ex ac ion abili y o neu al ne wo ks, esea che s ha e
p oposed anomaly de ec ion me hods based on ime se ies
modeling, g aph neu al ne wo ks, and au oencode
econs uc ion. These app oaches can au oma ically cap u e
la en pa e ns in high-dimensional da a and imp o e de ec ion
accu acy and obus ness. Fo example, au oencode -based
me hods use econs uc ion e o s o measu e anomaly le els.
Recu en neu al ne wo ks and a en ion mechanisms a e
applied o cap u e empo al dependencies and
mul idimensional ea u e in e ac ions. Al hough hese
app oaches add ess some limi a ions o adi ional me hods,
challenges emain. They o en ely hea ily on la ge labeled
da ase s, a e sensi i e o changes in sys em s uc u es, and lack
in e p e abili y[8].
In ecen yea s, con as i e lea ning has been in oduced
in o anomaly de ec ion as a sel -supe ised pa adigm, b inging
b eak h oughs o cloud backend scena ios. By cons uc ing
posi i e and nega i e sample pai s and emphasizing he
disc iminabili y o ea u e ep esen a ions, i can achie e s ong
obus ness e en wi h limi ed labels. In cloud backend anomaly
de ec ion, con as i e lea ning alle ia es he p oblems o
insu icien labels and da a imbalance. I also s eng hens he
model's abili y o cap u e hidden anomaly pa e ns. Al hough
ini ial s udies ha e explo ed his di ec ion, challenges emain in
e ec i e ea u e cons uc ion, adap a ion o di e se anomaly
ca ego ies, and eal- ime pe o mance in la ge-scale dis ibu ed
en i onmen s. Cu en p og ess p o ides a solid ounda ion o
u u e esea ch, while also e ealing u gen issues ha need o
be sol ed.
3. P oposed App oach
The co e idea o his esea ch is o cons uc ea u e
ep esen a ions ha can dis inguish no mal and abno mal
pa e ns h ough con as i e lea ning, he eby achie ing cloud
backend anomaly iden i ica ion. Speci ically, i is necessa y o
i s pe o m ea u e modeling on he inpu high-dimensional
sys em moni o ing da a. This pape also gi es he o e all
model a chi ec u e, and i s expe imen al esul s a e shown in
Figu e 1.
Figu e 1. O e all model a chi ec u e
Le he o iginal inpu sequence be
},...,,{ 21 T
xxxX
,
whe e each
d
Rx
ep esen s he mul idimensional
ea u es collec ed a ime . To ob ain a s able and exp essi e
po en ial ep esen a ion, we use a nonlinea mapping unc ion
)(
o map i o he la en space and ob ain a ep esen a ion
ec o :
k
Rhx h ),(
This ep esen a ion can e ec i ely cap u e he complex
ela ionships be ween di e en ea u es and lay he ounda ion
o subsequen disc imina i e lea ning.
A e ea u e mapping is comple ed, he objec i e unc ion
o con as i e lea ning needs o be cons uc ed. The co e o
con as i e lea ning is o maximize he simila i y be ween
samples o he same class while minimizing he simila i y
be ween samples o di e en classes. The speci ic o m can be
exp essed by he con as i e loss unc ion wi h no malized
empe a u e scaling:
N
kki
ji
con hhsim
hhsim
L
1)/),(exp(
)/),(exp(
log
He e,
)(sim
ep esen s cosine simila i y,
is he
empe a u e pa ame e ,
i
h
and
j
h
ep esen pai s o
posi i e samples, and he denomina o includes all nega i e
samples. This loss unc ion enables he model o au oma ically
close he ep esen a ion o simila pa e ns and dis inguish
ea u es om di e en ca ego ies.
In he speci ic modeling o abno mali y disc imina ion,
he ep esen a ion ob ained by con as i e lea ning can be
combined wi h he disc iminan unc ion. A linea classi ica ion
head
)(
g
is de ined, whose unc ion is o map he la en
ep esen a ion o he bina y classi ica ion space o abno mali y
and no mali y, and he ou pu is:
)()( bWhhgy
He e, W and b a e pa ame e s,
)(
ep esen s he
sigmoid unc ion, and he inal ou pu
]1,0[y
ep esen s
he p obabili y ha he sample is an anomaly. This s uc u e
can u he achie e explici anomaly iden i ica ion based on he
ep esen a ion o con as i e lea ning.
To enhance he gene aliza ion abili y o he model in
complex cloud backend scena ios, his s udy also in oduced a
join op imiza ion s a egy ha combines con as loss and
disc iminan loss. The disc iminan loss is usually in he o m
o bina y c oss-en opy:
)]1log()1(log[ yzyzLcls
Among hem,
]1,0[z
is he ue label and
y
is he
p edic ed p obabili y. The inal op imiza ion goal is o
comp ehensi ely conside he wo:
clscon LLL )1(
He e,
]1,0[
is a balancing ac o ha adjus s he
con ibu ion a io be ween con as i e lea ning and
disc imina ion asks. This allows he model o no only lea n
highly disc imina i e ep esen a ions bu also di ec ly op imize
o he disc imina ion ask, hus be e adap ing o he anomaly
de ec ion equi emen s o he cloud backend.
4. Expe imen esul
4.1 da ase
The da ase used in his s udy is he Yahoo Webscope S5
anomaly de ec ion da ase . I consis s o eal se e
pe o mance and business eques sequences and is widely used
in anomaly de ec ion and ime se ies modeling asks. The
da ase includes a ious ypes o indica o s, such as CPU
u iliza ion, memo y usage, and ne wo k h oughpu .
Anomalous in e als a e labeled, which allows o di ec
e alua ion o model pe o mance in cloud backend scena ios.
The da ase is mode a e in size. I has a ce ain le el o
complexi y while main aining expe imen al easibili y, making
i an impo an benchma k in anomaly de ec ion esea ch.
The da ase is cha ac e ized by di e se and complex
anomaly pa e ns. I con ains bo h sho - e m bu s anomalies
and long- e m pe sis en anomalies. These cha ac e is ics
closely esemble he ac ual beha io o cloud backend sys ems,
which o en exhibi dynamic and nonlinea anomalies unde
di e en se ice loads and esou ce scheduling. The e o e, he
use o his da ase no only e alua es he ep esen a ion abili y
o models bu also e ec i ely simula es he eal challenges o
anomaly de ec ion in cloud backends. I holds s ong p ac ical
alue.
In addi ion, he Yahoo Webscope S5 da ase has been
widely adop ed in he esea ch communi y and p o ides a
common benchma k o me hod compa ison. S udies based on
his da ase can clea ly e eal he s eng hs and limi a ions o
con as i e lea ning in anomaly de ec ion asks. They also lay
he ounda ion o u u e ex ensions o la ge -scale and highe -
dimensional eal-wo ld cloud backend da a. In summa y, he
choice o his da ase is bo h ep esen a i e and scien i ically
sound.
4.2 Expe imen al Resul s
This pape i s conduc s a compa a i e expe imen , and he
expe imen al esul s a e shown in Table 1.
Table 1: Compa a i e expe imen al esul s
Me hod
AUC
ACC
F1-Sco e
P ecision
MLP[9]
0.872
0.841
0.835
0.828
1DCNN[10]
0.896
0.859
0.849
0.843
LSTM[11]
0.911
0.872
0.861
0.855
T ans o me [12]
0.928
0.884
0.874
0.869
Ou s
0.957
0.913
0.902
0.896
F om he esul s in Table 1, i can be obse ed ha he
adi ional MLP model pe o ms he wo s ac oss all me ics.
The AUC, ACC, F1-Sco e, and P ecision a e 0.872, 0.841,
0.835, and 0.828, espec i ely. This indica es ha elying only
on shallow ully connec ed laye s canno e ec i ely cap u e he
complex nonlinea ela ionships and empo al dependencies in
cloud backend da a. As a esul , i s pe o mance in anomaly
de ec ion asks is clea ly limi ed. The pa icula ly low F1-Sco e
sugges s ha he model s uggles o balance ecall and
p ecision, making i di icul o main ain s able pe o mance
unde high-dimensional and di e se anomaly pa e ns.
Wi h he inc ease in model complexi y, bo h 1DCNN and
LSTM demons a e s onge de ec ion abili y han MLP. The
1DCNN le e ages con olu ion ope a ions o ex ac local
pa e ns, leading o imp o emen s in AUC and F1-Sco e. The
LSTM cap u es empo al dependencies h ough memo y uni s,
which u he imp o es ACC and P ecision, eaching 0.872 and
0.855, espec i ely. These esul s highligh he impo ance o
empo al ea u es o anomaly de ec ion in cloud backends.
S a ic ea u e modeling alone canno achie e op imal
pe o mance.
The T ans o me model ou pe o ms LSTM in o e all
pe o mance, wi h consis en ly high esul s ac oss all me ics.
The AUC eaches 0.928, and he F1-Sco e is 0.874. I s
ad an age comes mainly om he mul i-head sel -a en ion
mechanism, which can model long- ange dependencies and
cap u e bo h global and local ea u es. Fo complex anomalies
ha span di e en ime anges in cloud backends, he
T ans o me demons a es s ong ep esen a ional powe .
Howe e , al hough he esul s a e supe io o adi ional
me hods, he e is s ill oom o imp o emen , especially when
dealing wi h anomalies ha ha e uzzy bounda ies.
In compa ison, he p oposed me hod achie es he bes
pe o mance ac oss all me ics. The AUC inc eases o 0.957,
he ACC eaches 0.913, and he F1-Sco e and P ecision ise o
0.902 and 0.896, espec i ely. Compa ed wi h he T ans o me ,
he p oposed me hod imp o es de ec ion abili y by 2 o 3
pe cen age poin s. This ad an age is a ibu ed o he
in oduc ion o con as i e lea ning, which b ings no mal
samples close oge he in he ep esen a ion space while
e ec i ely dis inguishing anomalies. This s eng hens he
disc imina i e powe o he ea u es. The esul s con i m he
e ec i eness o he con as i e lea ning-based anomaly
de ec ion app oach in handling complex da a and imbalanced
scena ios. They also p o ide s ong suppo o building mo e
in elligen and eliable cloud backend ope a ion sys ems.
Figu e 2. Compa ison o he in luence o he empe a u e pa ame e τon expe imen al esul s
F om he esul s in Figu e 2, i can be seen ha he
empe a u e pa ame e τ has a signi ican impac on model
pe o mance. In e ms o AUC, he model eaches i s bes le el
when τ is se o 0.2, which is clea ly be e han o he alues.
This indica es ha an app op ia e empe a u e pa ame e can
enhance he sepa a ion be ween posi i e and nega i e samples
in con as i e lea ning. As a esul , he model can mo e
e ec i ely cap u e hidden anomaly pa e ns in cloud backend
da a. When τ is oo small o oo la ge, he di e ences among
samples canno be ully u ilized, which leads o weake
disc imina ion.
Fo ACC, a simila end can be obse ed, wi h a peak a τ
= 0.2. This sugges s ha when he empe a u e pa ame e is se
o a mode a e le el, he model achie es he bes o e all
classi ica ion accu acy. I can be e balance he disc imina ion
abili y ac oss di e en classes. Too low a τ alue o ces
ea u es o be o e ly concen a ed, making i di icul o co e
di e se anomaly pa e ns. Too high a τ alue dilu es he
di e ences among ea u es, esul ing in blu ed decision
bounda ies. The e o e, a mode a e empe a u e se ing can
s eadily imp o e de ec ion pe o mance in he complex
en i onmen o cloud backends.
The end o he F1-Sco e u he con i ms his obse a ion,
wi h he highes alue ob ained a τ = 0.2. This means ha a
his pa ame e alue, he model achie es he bes balance
be ween ecall and p ecision. In cloud backend scena ios,
anomalies a e o en sca ce and di e se. Main aining high ecall
while p ese ing p ecision is he e o e c i ical. The
expe imen al esul s show ha a easonable empe a u e
pa ame e helps he model emain obus when acing spa se
anomaly samples. I also p e en s pe o mance deg ada ion
caused by o e i ing o unde i ing.
The a ia ion in P ecision also shows a peak a τ = 0.2. This
demons a es ha a his empe a u e, he model no only
iden i ies anomalies e ec i ely bu also educes alse ala ms.
This is especially impo an o p ac ical cloud backend
ope a ions. Bo h o e ly high and o e ly low τ alues lead o a
d op in P ecision, indica ing ha he model is p one o
misclassi ica ion in anomaly de ec ion. O e all, he
expe imen al esul s e i y he sensi i i y and dependence o
he p oposed me hod on he empe a u e pa ame e wi hin he
con as i e lea ning amewo k. They also highligh he
impo ance o pa ame e selec ion o imp o ing anomaly
de ec ion pe o mance in cloud backends.
This pape also p o ides a de ailed analysis o he impac o
he lea ning a e on he expe imen al ou comes, wi h a
pa icula ocus on how di e en alues o his pa ame e
in luence he s abili y and e ec i eness o he aining p ocess.
The in es iga ion highligh s he ole o he lea ning a e as a
c ucial hype pa ame e ha di ec ly a ec s he con e gence
speed o he model and i s abili y o cap u e complex pa e ns
wi hin he da a. To clea ly illus a e his aspec , he
co esponding esul s a e sys ema ically p esen ed in Figu e 3,
o e ing an in ui i e ep esen a ion o he ela ionship be ween
lea ning a e se ings and he o e all model beha io .
Figu e 3. The impac o he lea ning a e on expe imen al esul s
F om he esul s in Figu e 3, i can be obse ed ha he
lea ning a e has a clea impac on model pe o mance. In
e ms o AUC, he highes alue is achie ed when he lea ning
a e is 1×10⁻⁴. This indica es ha unde his se ing, he model
can be e sepa a e no mal and abno mal pa e ns. When he
lea ning a e is oo small, such as 1×10⁻⁵, he model
con e ges slowly, leading o poo pe o mance. When he
lea ning a e is oo la ge, such as 1×10⁻³, he aining p ocess
becomes uns able and pe o mance dec eases. These indings
show ha a p ope lea ning a e is c i ical o con as i e
lea ning o achie e s ong ep esen a ion abili y in cloud
backend anomaly de ec ion.
Fo ACC, a simila end is obse ed, wi h he bes esul s
a ound 1×10⁻⁴. This u he con i ms he impo ance o a
mode a e lea ning a e. Too low a lea ning a e esul s in e y
small upda es, making i di icul o he model o cap u e
complex anomaly pa e ns. Too high a lea ning a e causes
oscilla ions and educes classi ica ion accu acy. In cloud
backend scena ios, highe ACC means ha anomaly de ec ion
sys ems can main ain s ong disc imina i e abili y mo e
consis en ly, wi h ewe alse ala ms and missed de ec ions.
The end o he F1-Sco e also shows ha lea ning a e
selec ion a ec s he balance be ween p ecision and ecall. A
1×10⁻⁴, he model eaches i s peak F1-Sco e. This sugges s
ha his lea ning a e main ains high ecall while a oiding a
d op in p ecision. This is c i ical in anomaly de ec ion, since
anomalies in cloud backends a e o en a e. Low ecall would
cause se ious missed de ec ions, while low p ecision would
esul in many alse ala ms. A easonable lea ning a e ensu es
he bes balance be ween he wo.
The esul s o P ecision u he suppo his poin . The
model achie es i s highes P ecision a 1×10⁻⁴, showing ha
his lea ning a e educes misclassi ica ion and imp o es he
eliabili y o de ec ion. When he lea ning a e is oo high o
oo low, he P ecision dec eases signi ican ly. This indica es
ha pa ame e uning is essen ial o he deploymen o
anomaly de ec ion models in cloud backends. A easonable
lea ning a e no only imp o es pe o mance bu also enhances
obus ness and usabili y in eal complex en i onmen s.
This pape also p esen s a sensi i i y expe imen on he
nega i e sample a io o F1-Sco e, and he expe imen al esul s
a e shown in Figu e 4.
Figu e 4. Sensi i i y expe imen o he nega i e sample
a io o F1-Sco e
F om he esul s in Figu e 4, i can be obse ed ha he
a io o nega i e samples has a clea e ec on he F1-Sco e.
When he a io o nega i e samples is 0.5, he F1-Sco e is
ela i ely low, a ound 0.889. This indica es ha an insu icien
numbe o nega i e samples makes i di icul o he model o
lea n he bounda y be ween no mal and abno mal da a. In he
con as i e lea ning amewo k, oo ew nega i e samples
weaken he con as i e signal and educe he o e all anomaly
de ec ion abili y.
When he a io o nega i e samples inc eases o 1.0, he F1-
Sco e ises signi ican ly o 0.896, and he model's
disc imina i e abili y imp o es. This change shows ha an
app op ia e inc ease in nega i e samples helps he model
cap u e he di e ences be ween no mal and abno mal da a
mo e e ec i ely. I also imp o es he balance be ween ecall
and p ecision. In complex cloud backend scena ios, di e si y in
nega i e samples suppo s he de ec ion o di e en ypes o
anomalies.
When he a io o nega i e samples u he inc eases o 2.0,
he F1-Sco e eaches i s highes alue o 0.902. A his poin ,
he model ecei es he s onges con as i e signal. The model
achie es he bes ade-o be ween p ecision and ecall, which
enhances o e all de ec ion pe o mance. This sugges s ha
inc easing he numbe o nega i e samples is bene icial up o a
ce ain poin , bu imp o emen s do no con inue beyond he
op imal a io.
When he a io con inues o inc ease o 3.0 and 4.0, he F1-
Sco e begins o decline, d opping o 0.897 and 0.891,
espec i ely. This indica es ha oo many nega i e samples
in oduce noise and weaken he e ec i eness o con as i e
lea ning. As a esul , he model's decision bounda y becomes
blu ed. Fo anomaly de ec ion in cloud backends, his inding
highligh s he impo ance o choosing an app op ia e nega i e
sample a io. A easonable a io can signi ican ly imp o e he
e ec i eness o con as i e lea ning models, while a ios ha
a e oo high o oo low will ha m pe o mance.
This pape also p esen s an expe imen on he
en i onmen al sensi i i y o sampling equency ji e o AUC,
and he expe imen al esul s a e shown in Figu e 5.
Figu e 5. Expe imen on he en i onmen al sensi i i y o
sampling equency ji e o AUC
F om he esul s in Figu e 5, i can be seen ha sampling
equency ji e has a clea e ec on he AUC pe o mance o
he model. When he ji e is 0 pe cen , he AUC is 0.954,
which shows ha he model can main ain high anomaly
de ec ion abili y in an in e e ence- ee en i onmen . When he
ji e inc eases o 5 pe cen , he AUC ises o 0.956, showing a
sligh imp o emen . This indica es ha mode a e sampling
dis u bance can enhance he model's abili y o adap o
unce ain y and imp o e de ec ion pe o mance o some ex en .
When he ji e inc eases u he o 10 pe cen , he AUC
eaches a peak o 0.957, which is he bes esul in his
expe imen . This shows ha mode a e en i onmen al
dis u bance does no weaken s abili y. Ins ead, i encou ages
he model o lea n mo e obus ea u e ep esen a ions. This is
impo an o cloud backend scena ios, as sys em ope a ion
o en in ol es signal luc ua ions and noise. A model ha
pe o ms be e unde mode a e dis u bance demons a es
s onge adap abili y in eal applica ions.
Howe e , when he ji e inc eases o 20 pe cen , he AUC
d ops signi ican ly o 0.953. This indica es ha excessi e
sampling ji e begins o dis o he o iginal da a dis ibu ion,
making i ha de o he model o dis inguish be ween no mal
and abno mal pa e ns. In complex cloud backend
en i onmen s, excessi e luc ua ions weaken he cla i y o he
decision bounda y, which leads o pe o mance deg ada ion.
When he ji e eaches 30 pe cen , he AUC dec eases
u he o 0.949. This shows ha unde high-in ensi y
dis u bances, he model's de ec ion abili y is s ongly a ec ed.
The esul e i ies he ulne abili y o he model unde ex eme
en i onmen al dis u bance and emphasizes he impo ance o
easonable en i onmen al modeling. O e all, he indings show
ha mode a e ji e can imp o e he obus ness o con as i e
lea ning models, bu excessi e dis u bance signi ican ly
educes pe o mance. This e eals he key issue o balancing
obus ness and sensi i i y o noise in cloud backend anomaly
de ec ion.
5. Conclusion
This s udy ocuses on anomaly de ec ion in cloud backends
and p oposes a con as i e lea ning-based amewo k. The
e ec i eness and obus ness o he amewo k a e e i ied
om di e en pe spec i es. The s udy i s analyzes he
complexi y and di e si y o cloud backend sys ems and
highligh s he limi a ions o adi ional me hods unde high-
dimensional da a, dynamic en i onmen s, and sca ci y o
anomaly samples. By in oducing con as i e lea ning, he
model can au oma ically lea n disc imina i e ea u e
ep esen a ions wi hou elying on la ge labeled da ase s, which
signi ican ly imp o es anomaly de ec ion pe o mance. The
esul s show ha he p oposed me hod no only ou pe o ms
adi ional and mains eam app oaches on o e all me ics bu
also demons a es s ong adap abili y and s abili y unde
di e en pa ame e s and en i onmen al dis u bances.
The signi icance o his wo k lies no only a he
algo i hmic le el bu also in i s p ac ical alue. In c i ical
indus ies such as inance, heal hca e, anspo a ion, and
ene gy, he accu acy o backend anomaly de ec ion di ec ly
a ec s se ice quali y and sys em secu i y. The p oposed
con as i e lea ning amewo k educes alse ala ms and
missed de ec ions and enhances sys em obus ness in complex
en i onmen s. This means ha in u u e in elligen ope a ion
scena ios, he me hod has he po en ial o p o ide en e p ises
wi h mo e eliable echnical suppo , educe isks, imp o e use
expe ience, and p omo e he heal hy de elopmen o he cloud
compu ing indus y.
A he same ime, his s udy also e eals bo h he po en ial
and limi a ions o con as i e lea ning in anomaly de ec ion
asks. On one hand, he esul s show ha p ope pa ame e
se ings and mode a e en i onmen al dis u bances can u he
imp o e obus ness and disc imina i e abili y. On he o he
hand, he indings also indica e ha excessi e dis u bances o
se e e imbalance in sample a ios may s ill deg ade model
pe o mance. The e o e, how o adap i ely adjus pa ame e s
acco ding o business equi emen s and how o cons uc mo e
di e se and high-quali y da ase s emain impo an challenges
o u u e wo k. These indings p o ide new di ec ions o
u he esea ch and also lay he ounda ion o c oss-domain
applica ions.
Looking ahead, he con ibu ions o his s udy can be
ex ended o b oade in elligen sys em ope a ion scena ios. Fo
example, combining con as i e lea ning wi h ede a ed
lea ning amewo ks in la ge-scale dis ibu ed sys ems can
enable anomaly de ec ion ac oss da a cen e s. Inco po a ing
g aph neu al ne wo ks and ime se ies modeling echniques can
u he enhance he model's abili y o unde s and mul i-sou ce
he e ogeneous da a. In addi ion, in eg a ing anomaly de ec ion
wi h au oma ed decision-making sys ems would allow no only
ecogni ion o anomalies bu also he igge ing o in elligen
esponses and op imiza ion s a egies. In summa y, his s udy
ad ances he heo e ical de elopmen o cloud backend
anomaly de ec ion and p o ides p ac ical suppo o eal-wo ld
applica ions, o e ing signi ican implica ions o u u e
in elligen ope a ions and cloud compu ing secu i y.
Re e ences
[1] M. Jin, Y. Liu, Y. Zheng, L. Chi, Y. F. Li, and S. Pan, "Anemone: G aph
anomaly de ec ion wi h mul i-scale con as i e lea ning," P oceedings o
he 2021 30 h ACM In e na ional Con e ence on In o ma ion &
Knowledge Managemen , pp. 3122-3126, 2021.
[2] Yu H, Yang W, Cui B, e al. Enhanced anomaly a ic de ec ion
amewo k using BiGAN and con as i e lea ning[J]. Cybe secu i y,
2024, 7(1): 71.
[3] Kong S, Ai J, Lu M. CL-MMAD: A con as i e lea ning based
mul imodal so wa e un ime anomaly de ec ion me hod[J]. Applied
Sciences, 2023, 13(6): 3596.
[4] H. Sun, Y. Huang, L. Han, C. Fu, and C. Zhou, "HCL-MTSAD:
Hie a chical con as i e consis ency lea ning o accu a e de ec ion o
indus ial mul i a ia e ime se ies anomalies," a Xi p ep in
a Xi :2404.08224, 2024.
[5] H. C. V. Ngu and K. M. Lee, "CL-TAD: A con as i e-lea ning-based
me hod o ime se ies anomaly de ec ion," Applied Sciences, ol. 13, no.
21, p. 11938, 2023.
[6] Y. Tian, G. Pang, Y. Chen, R. Singh, J. W. Ve jans, and G. Ca nei o,
“Weakly ‑supe ised ideo anomaly de ec ion wi h con as i e
lea ning o long and sho ‑ ange empo al ea u es,”a Xi p ep in
a Xi :2101.10030, 2021.
[7] T. Wu, Q. Chen, D. Zhao, J. Wang, and L. Jiang, "Domain adap a ion o
ime se ies ia con as i e lea ning wi h ask-speci ic consis ency,"
Applied In elligence, ol. 54, no. 23, pp. 12576-12588, 2024.
[8] Kong X, Zhang W, Wang H, e al. Fede a ed g aph anomaly de ec ion ia
con as i e sel -supe ised lea ning[J]. IEEE T ansac ions on Neu al
Ne wo ks and Lea ning Sys ems, 2024, 36(5): 7931-7944.
[9] Fa zad A, Gulli e T A. Log message anomaly de ec ion wi h uzzy C-
means and MLP[J]. Applied In elligence, 2022, 52(15): 17708-17717.
[10] A ha A, Mozumde M A I, Ali S, e al. Deep lea ning-based anomaly
de ec ion using one-dimensional con olu ional neu al ne wo ks (1D
CNN) in machine cen e s (MCT) and compu e nume ical con ol (CNC)
machines[J]. Pee J Compu e Science, 2024, 10: e2389.
[11] A. Iqbal, R. Amin, F. S. Alsubaei, and A. Alzah ani, "Anomaly de ec ion
in mul i a ia e ime se ies da a using deep ensemble models," PLOS
ONE, ol. 19, no. 6, p. e0303890, 2024.
[12] Y. Zhang, "A mul i-scale empo al ea u e ex ac ion app oach o
ne wo k a ic anomaly de ec ion," In e na ional Jou nal o In o ma ion
Secu i y and P i acy (IJISP), ol. 18, no. 1, pp. 1-20, 2024.
