Co esponding au ho : Nishan Nisan Jha
Copy igh © 2025 Au ho (s) e ain he copy igh o his a icle. This a icle is published unde he e ms o he C ea i e Commons A ibu ion Liscense 4.0.
Fede a ed lea ning o c oss-cloud obse abili y: P i acy-p ese ing model
agg ega ion ac oss dis ibu ed cloud pla o ms
Nishan Nisan Jha *
IEEE Senio Membe , USA.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 26(02), 2883-2894
Publica ion his o y: Recei ed on 04 Ap il 2025; e ised on 14 May 2025; accep ed on 16 May 2025
A icle DOI: h ps://doi.o g/10.30574/wja .2025.26.2.1892
Abs ac
This a icle p esen s a comp ehensi e amewo k o implemen ing p i acy-p ese ing c oss-cloud moni o ing using
ede a ed lea ning echniques. As o ganiza ions inc easingly adop mul i-cloud s a egies, main aining uni ied
obse abili y wi hou iola ing da a so e eign y o egula o y equi emen s becomes challenging. The inno a i e
sys em employs ede a ed lea ning a chi ec u e o de elop de ec ion models ac oss decen alized, enc yp ed
ansac ion eco ds, exchanging only model pa ame e upda es be ween seg ega ed cloud en i onmen s while
p ese ing da a locali y and p i acy. The a chi ec u e inco po a es ede a ed g aph neu al ne wo ks o disco e hidden
dependencies ac oss cloud bounda ies, secu e agg ega ion h ough homomo phic enc yp ion and secu e mul i-pa y
compu a ion, and di e en ial p i acy sa egua ds. Th ough case s udies spanning de ense, inancial se ices, and
heal hca e sec o s, A icle demons a es signi ican imp o emen s in inciden de ec ion capabili y, educ ion in alse
posi i es, and accele a ed mean ime o esolu ion while main aining s ic compliance wi h da a p o ec ion egula ions.
The esul s es ablish ede a ed lea ning as a iable solu ion o achie ing c oss-cloud obse abili y wi hou
comp omising sensi i e ope a ional da a.
Keywo ds: Fede a ed Lea ning; Mul i-Cloud Obse abili y; P i acy-P ese ing Moni o ing; C oss-Cloud
Dependencies; Da a So e eign y
1. In oduc ion
Mode n en e p ises inc easingly adop mul i-cloud s a egies o op imize cos , enhance eliabili y, and a oid endo
lock-in. By 2023, indus y analys s epo ed ha 81% o public cloud use s we e wo king wi h wo o mo e p o ide s
[1]. Howe e , his di e si ica ion has c ea ed signi ican challenges in main aining comp ehensi e obse abili y ac oss
dis ibu ed cloud en i onmen s. The isibili y in o sys em pe o mance, e o pa e ns, and secu i y inciden s becomes
agmen ed when ope a ional da a emains siloed wi hin indi idual cloud pla o ms.
This agmen a ion is pa icula ly p oblema ic o inciden de ec ion and esponse, whe e c oss-pla o m dependencies
o en c ea e cascading ailu es ha emain in isible when moni o ing sys ems ope a e in isola ion. Acco ding o a 2023
indus y su ey, o ganiza ions using mul iple cloud p o ide s expe ience 37% longe mean ime o esolu ion (MTTR)
o inciden s in ol ing c oss-cloud dependencies compa ed o single-cloud issues [1]. This ex ended esolu ion ime
di ec ly impac s business con inui y and se ice le el ag eemen s.
Regula o y amewo ks like he Gene al Da a P o ec ion Regula ion (GDPR) in Eu ope and he Cali o nia Consume
P i acy Ac (CCPA) in he Uni ed S a es ha e u he complica ed he obse abili y landscape. These egula ions impose
s ic equi emen s on da a so e eign y, limi ing how and whe e moni o ing da a can be s o ed, p ocessed, and
ans e ed. Unde GDPR A icle 44, ans e ing pe sonal da a ou side he Eu opean Economic A ea equi es speci ic
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 26(02), 2883-2894
2884
sa egua ds, e ec i ely p e en ing he cen aliza ion o log da a ha migh con ain iden i iable in o ma ion [2]. Simila ly,
CCPA g an s Cali o nia esiden s he igh o know wha pe sonal in o ma ion is collec ed and how i is used, imposing
addi ional compliance equi emen s on moni o ing sys ems ha migh p ocess use - ela ed ope a ional da a [2].
Fede a ed Lea ning (FL) has eme ged as a p omising solu ion o hese challenges. Fi s in oduced in 2016, FL enables
collabo a i e model aining wi hou cen alizing he unde lying da a [1]. In he con ex o cloud obse abili y, his
app oach allows each cloud p o ide o main ain con ol o e i s ope a ional logs while s ill con ibu ing o a sha ed
model ha can iden i y c oss-pla o m pa e ns and dependencies.
The undamen al p inciple o FL in mul i-cloud obse abili y is ha machine lea ning models a e ained locally on each
cloud pla o m's enc yp ed logs - such as in as uc u e me ics, applica ion pe o mance da a, o ope a ions logs.
Ins ead o sha ing he aw logs, only he model upda es (g adien s) a e exchanged h ough secu e channels, o en
p o ec ed by echniques like homomo phic enc yp ion o secu e mul i-pa y compu a ion (SMPC). This app oach
p ese es da a p i acy while enabling he de ec ion o complex pa e ns ha span mul iple cloud en i onmen s.
Recen implemen a ions ha e demons a ed signi ican imp o emen s in inciden p edic ion and anomaly de ec ion. A
2022 s udy showed ha ede a ed models ained ac oss h ee majo cloud p o ide s achie ed a 42% imp o emen in
ea ly wa ning capabili y o c oss-pla o m inciden s compa ed o isola ed moni o ing sys ems [1]. This imp o emen
was achie ed wi hou iola ing da a so e eign y equi emen s o exposing sensi i e ope a ional da a.
This esea ch aims o ad ance he s a e o he a in p i acy-p ese ing c oss-cloud obse abili y h ough ede a ed
lea ning. Speci ically, we seek o: (1) de elop no el a chi ec u al amewo ks o secu e model aining ac oss cloud
bounda ies; (2) implemen and e alua e ede a ed g aph neu al ne wo ks o dependency disco e y; (3) quan i y he
pe o mance imp o emen s in eal-wo ld de ense, inancial, and heal hca e applica ions; and (4) es ablish bes
p ac ices o egula o y compliance in mul i-cloud moni o ing sys ems.
The signi icance o his wo k ex ends beyond echnical inno a ion o add ess c i ical business and egula o y
equi emen s. As o ganiza ions con inue o dis ibu e wo kloads ac oss mul iple cloud en i onmen s, he abili y o
main ain comp ehensi e obse abili y wi hou comp omising da a so e eign y becomes essen ial o ope a ional
esilience, secu i y, and compliance. Fede a ed lea ning o e s a pa h owa d his goal by enabling collabo a i e
in elligence wi hou cen alized da a pools.
2. A chi ec u al F amewo k o P i acy-P ese ing Cloud Moni o ing
The implemen a ion o ede a ed lea ning (FL) o c oss-cloud obse abili y equi es a ca e ully designed a chi ec u al
amewo k ha balances p i acy p ese a ion wi h e ec i e model aining. This sec ion p esen s a comp ehensi e
app oach o deploying FL ac oss dis ibu ed cloud p o ide s while main aining da a so e eign y and egula o y
compliance.
A ypical mul i-cloud deploymen consis s o wo kloads dis ibu ed ac oss 3-5 majo cloud se ice p o ide s, each
gene a ing be ween 10-50 GB o log da a daily [3]. T adi ional app oaches o uni ied moni o ing would equi e
cen alizing his da a, po en ially iola ing da a so e eign y equi emen s. Ins ead, he p oposed a chi ec u e main ains
da a locali y while enabling collabo a i e model aining h ough a ou -laye app oach: da a p epa a ion, local model
aining, secu e agg ega ion, and global model dis ibu ion.
A he da a p epa a ion laye , each cloud en i onmen implemen s s anda dized log p ep ocessing o no malize a ied
da a o ma s. Recen benchma ks show ha implemen ing consis en ea u e ex ac ion ac oss cloud p o ide s can
educe model con e gence ime by up o 43% [3]. This s anda diza ion includes no malizing imes amps o UTC,
ca ego izing log se e i y le els on a uni ied scale, and ex ac ing common ea u es such as se ice iden i ie s, e o
codes, and pe o mance me ics. Impo an ly, pe sonally iden i iable in o ma ion (PII) and o he sensi i e da a a e
emo ed o anonymized a his s age, applying di e en ial p i acy echniques wi h a ypical p i acy budge (ε) o 1-5
o balance u ili y and p i acy p o ec ion [3].
The local model aining me hodology employs specialized neu al ne wo k a chi ec u es adap ed o ime-se ies da a
analysis. Recu en neu al ne wo ks (RNNs) and Long Sho -Te m Memo y (LSTM) ne wo ks ha e demons a ed
supe io pe o mance o anomaly de ec ion in cloud in as uc u e logs, wi h LSTM models showing a 17%
imp o emen in p ecision o ou age p edic ion compa ed o adi ional s a is ical me hods [3]. Each cloud p o ide
ains hese models locally on hei p op ie a y logs, wi h aining ypically pe o med on dedica ed GPU ins ances o
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 26(02), 2883-2894
2885
minimize impac on p oduc ion wo kloads. Benchma ks indica e ha local aining on high-pe o mance compu e
ins ances wi h 16 CPUs can p ocess app oxima ely 24 hou s o log da a in 30-45 minu es [3].
Secu e agg ega ion ep esen s he c i ical p i acy-p ese ing componen o he a chi ec u e. Two p ima y app oaches
ha e shown p ac ical iabili y: homomo phic enc yp ion (HE) and secu e mul i-pa y compu a ion (SMPC).
Homomo phic enc yp ion allows compu a ions o be pe o med on enc yp ed da a wi hou dec yp ion, he eby
enabling he agg ega ion o model upda es wi hou exposing he unde lying da a. Cu en implemen a ions using pa ial
homomo phic enc yp ion add app oxima ely 200-300ms o la ency pe agg ega ion ound bu p o ide s ong
ma hema ical gua an ees agains da a exposu e [4]. SMPC, al e na i ely, dis ibu es he compu a ion ac oss mul iple
pa ies such ha no single pa y can access he comple e in o ma ion. In a h ee-cloud implemen a ion, SMPC p o ocols
ha e demons a ed 99.98% da a p o ec ion wi h 15-20% compu a ional o e head compa ed o non-secu e agg ega ion
[4].
Sys em design conside a ions o model con e gence ocus on add essing he challenges unique o ede a ed
en i onmen s. Non-IID (Independen and Iden ically Dis ibu ed) da a dis ibu ions ac oss cloud p o ide s can slow
con e gence by 25-40% compa ed o cen alized aining [4]. To mi iga e his, he a chi ec u e implemen s adap i e
lea ning a e scheduling and pe iodic model synch oniza ion. Empi ical es ing has shown ha synch onizing model
upda es e e y 50-100 local aining ba ches p o ides an op imal balance be ween communica ion o e head and
con e gence speed, wi h Fede a ed A e aging algo i hms achie ing 92% o cen alized accu acy a e 10 ounds o
agg ega ion [4].
Communica ion e iciency p esen s ano he key design conside a ion, as bandwid h be ween cloud p o ide s is o en
limi ed and cos ly. The implemen a ion uses g adien comp ession echniques, educing in e -cloud communica ion
olume by 60-85% wi h minimal impac on model quali y [4]. Speci ically, adap i e h eshold-based spa si ica ion
ansmi s only g adien alues exceeding a dynamically calcula ed h eshold, ypically se a 10^-3 o he i s
agg ega ion ound and dec easing by 10% in subsequen ounds [4].
Implemen a ion challenges o p i acy-p ese ing cloud moni o ing ex end beyond echnical aspec s o o ganiza ional
and ope a ional conce ns. C oss-cloud communica ion equi es es ablishing secu e channels be ween p o ide s,
ypically implemen ed h ough dedica ed VPN connec ions o p i a e pee ing a angemen s wi h end- o-end
enc yp ion using AES-256. Access con ol mechanisms limi pa icipa ion in he ede a ed sys em o au hen ica ed
nodes, wi h mu ual TLS au hen ica ion p o iding he ounda ion o us ed communica ion. Pe o mance es ing shows
ha hese secu i y measu es add 50-75ms o la ency pe c oss-cloud message exchange [4].
Figu e 1 Fede a ed Lea ning A chi ec u e o C oss-Cloud Obse abili y [3, 4]
Faul ole ance ep esen s ano he c i ical implemen a ion conside a ion, as dis ibu ed sys ems mus main ain
ope a ion despi e indi idual node ailu es. The a chi ec u e employs a leade elec ion p o ocol o designa e an
agg ega ion coo dina o , wi h au oma ic ailo e i he p ima y coo dina o becomes una ailable. Benchma k es ing
indica es ha leade e-elec ion can comple e wi hin 3-5 seconds when a node ailu e is de ec ed, main aining sys em
a ailabili y a 99.95% e en du ing cloud p o ide ou ages [4].
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 26(02), 2883-2894
2886
The echnical equi emen s o implemen ing his a chi ec u e include: (1) homogeneous model a chi ec u es ac oss
cloud en i onmen s o ensu e compa ibili y o g adien s du ing agg ega ion; (2) dedica ed compu e esou ces o local
aining o minimize impac on p oduc ion wo kloads; (3) s anda dized APIs o model upda e exchange; and (4)
c yp og aphic in as uc u e suppo ing a leas 2048-bi RSA o equi alen o secu e communica ion. Cloud p o ide s
pa icipa ing in he ede a ion mus alloca e app oxima ely 5-8% addi ional compu a ional esou ces compa ed o
s andalone moni o ing, wi h s o age equi emen s inc easing by 12-15% o accommoda e model e sioning and
aining da a p ese a ion [3].
3. C oss-cloud dependency disco e y using ede a ed g aph neu al ne wo ks
The disco e y and modeling o dependencies ac oss dis ibu ed cloud en i onmen s p esen s a signi ican challenge o
uni ied obse abili y. Fede a ed G aph Neu al Ne wo ks (FGNNs) ha e eme ged as a powe ul app oach o iden i ying
hese complex c oss-cloud ela ionships while p ese ing da a p i acy. This sec ion explo es he s uc u e,
implemen a ion, and e ec i eness o FGNNs o dependency disco e y in mul i-cloud en i onmen s.
FGNNs ex end adi ional G aph Neu al Ne wo ks (GNNs) o ope a e in a ede a ed lea ning con ex , enabling
dis ibu ed aining ac oss o ganiza ional bounda ies. The model s uc u e consis s o a g aph ep esen a ion whe e
nodes ep esen cloud esou ces (e.g., i ual machines, s o age se ices, ne wo k componen s) and edges ep esen
in e ac ions o dependencies be ween hese esou ces. A ypical FGNN implemen a ion o c oss-cloud moni o ing
inco po a es 3-8 g aph con olu ional laye s wi h 64-256 hidden uni s pe laye , achie ing a balance be ween model
complexi y and aining e iciency [5]. The node ea u es ypically include 30-50 ime-se ies me ics such as CPU
u iliza ion, memo y consump ion, eques la ency, and e o a es, while edge ea u es cap u e in e ac ion pa e ns
such as API call equencies and da a ans e olumes. Benchma ks indica e ha his a chi ec u e can e ec i ely model
dependencies ac oss clouds wi h up o 10,000 nodes and 50,000 edges while main aining aining con e gence wi hin
24-48 hou s on s anda d cloud GPU ins ances [5].
The aining app oach o FGNNs in mul i-cloud scena ios ollows a specialized ede a ed lea ning p o ocol adap ed o
g aph da a. Each cloud p o ide main ains a subg aph ep esen ing i s in e nal esou ces and di ec ly obse able
ex e nal dependencies. Local aining occu s on hese subg aphs using s ochas ic g adien descen wi h a ypical
lea ning a e o 0.001-0.005 and ba ch sizes o 32-128 samples [5]. The agg ega ion p ocess me ges model upda es
ac oss clouds while p ese ing he p i acy o local g aph s uc u es. Compa a i e analysis shows ha his ede a ed
app oach achie es 87-92% o he accu acy ob ained by a hypo he ical cen alized model (which would iola e p i acy
cons ain s) a e 15-20 ounds o aining [5].
A key inno a ion in he FGNN app oach is he abili y o disco e hidden dependencies ha span mul iple cloud
p o ide s. By analyzing pa e ns in se ice beha io wi hou di ec access o he unde lying sys ems, hese models can
iden i y co ela ions ha would emain in isible in isola ed moni o ing en i onmen s. Field implemen a ions ha e
demons a ed he abili y o de ec up o 78% o c oss-cloud dependencies wi h a alse posi i e a e below 8%,
signi ican ly ou pe o ming adi ional co ela ion-based app oaches ha ypically iden i y only 30-45% o c oss-cloud
dependencies [5].
Di e en ial p i acy mechanisms o m an essen ial componen o he FGNN amewo k, ensu ing ha sensi i e
in o ma ion abou speci ic cloud esou ces canno be e e se-enginee ed om he sha ed model upda es. The
implemen a ion applies noise calib a ed o he sensi i i y o he g adien upda es, ypically u ilizing he Gaussian
mechanism wi h a noise scale (σ) o 2.0-4.0 and a p i acy budge (ε) o 1.0-3.0 pe aining ound [6]. This con igu a ion
esul s in a cumula i e p i acy loss o ε < 10 o e a comple e aining cycle o 20 ounds, aligning wi h indus y
s anda ds o sensi i e ope a ional da a [6]. Analysis o p i acy-u ili y ade o s shows ha his le el o p o ec ion
educes model accu acy by only 3-7% compa ed o non-p i a e aining while p o iding heo e ical gua an ees agains
in o ma ion leakage [6].
Beyond simple noise addi ion, he p i acy-p ese ing mechanism inco po a es se e al ad anced echniques o
main ain u ili y. These include g adien clipping a a h eshold o 1.0-3.0 o bound sensi i i y, adap i e p i acy budge
alloca ion ha assigns mo e budge o c i ical aining phases, and secu e agg ega ion p o ocols ha ensu e no
indi idual p o ide 's upda es a e exposed in aw o m. Empi ical e alua ion demons a es ha his comp ehensi e
app oach main ains a p i acy gua an ee o (ε, δ) = (8.2, 10^-5) o a ypical aining p ocess while p ese ing 91% o
model u ili y compa ed o non-p i a e aining [6].
Me ics o measu ing c oss-pla o m co ela ion e ec i eness ex end beyond adi ional classi ica ion me ics o
add ess he unique challenges o dependency disco e y. The p ima y e alua ion me ics include dependency ecall ( he
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 26(02), 2883-2894
2887
pe cen age o ac ual dependencies disco e ed), dependency p ecision ( he pe cen age o disco e ed dependencies ha
a e genuine), and ime- o-de ec ion (how quickly he sys em iden i ies a newly o med dependency). Field deploymen s
ac oss mul iple indus y sec o s show ha FGNN-based sys ems achie e dependency ecall a es o 75-85% and
p ecision a es o 82-90%, wi h an a e age ime- o-de ec ion o 4.3 hou s o new dependencies [6]. Tempo al
consis ency, measu ed as he s abili y o iden i ied dependencies ac oss aining ounds, eaches 94% a e model
con e gence, indica ing obus and eliable de ec ion [6].
Ope a ional me ics u he demons a e he p ac ical e ec i eness o FGNN app oaches. In p oduc ion en i onmen s,
hese sys ems ha e educed alse posi i e ale s o c oss-cloud inciden s by 62-71% compa ed o adi ional h eshold-
based moni o ing [6]. The mean ime o esolu ion (MTTR) o inciden s in ol ing mul iple cloud p o ide s dec eased
by 47% in inancial se ices applica ions and 53% in e-comme ce pla o ms a e implemen ing FGNN-based
dependency disco e y, di ec ly ansla ing o imp o ed se ice eliabili y and educed ope a ional cos s [6].
Compa a i e analysis wi h adi ional moni o ing app oaches highligh s he signi ican ad an ages o he FGNN
me hodology. Con en ional co ela ion-based echniques ypically ely on Pea son o Spea man co ela ion coe icien s
applied o ime-se ies me ics, achie ing de ec ion a es o only 35-45% o c oss-cloud dependencies [5]. Rule-based
sys ems a e sligh ly be e a 40-55% bu equi e ex ensi e manual con igu a ion and domain expe ise. S a is ical
anomaly de ec ion me hods using ARIMA o exponen ial smoo hing iden i y app oxima ely 50-60% o dependencies
bu gene a e alse posi i e a es o 20-30% [5]. In con as , he FGNN app oach no only achie es supe io de ec ion
a es (75-85%) bu also educes alse posi i es o below 10% while equi ing minimal manual con igu a ion a e ini ial
deploymen [5].
The compu a ional e iciency o FGNNs also compa es a o ably wi h al e na i e app oaches. While cen alized machine
lea ning me hods would equi e ans e ing and p ocessing 15-20TB o daily log da a o a la ge en e p ise
deploymen , he ede a ed app oach educes c oss-cloud da a ans e by 99.5%, ansmi ing only enc yp ed model
upda es o app oxima ely 50-100MB pe aining ound [5]. This d ama ic educ ion in da a mo emen no only
add esses p i acy conce ns bu also signi ican ly educes bandwid h cos s and la ency in c oss-cloud ope a ions.
Figu e 2 FGNNs Ou pe o m T adi ional Me hods in Dependency De ec ion and E iciency [5, 6]
4. Case S udy: De ense-C i ical In as uc u e Moni o ing
De ense o ganiza ions wi h mul ina ional ope a ions ace unique challenges in implemen ing comp ehensi e cloud
moni o ing solu ions. These challenges include s ic secu i y equi emen s, da a so e eign y conce ns ac oss di e en
na ions, and he need o esilien ope a ions du ing po en ial cybe a acks. This case s udy examines he
implemen a ion o a ede a ed lea ning-based moni o ing sys em o de ense-c i ical in as uc u e, analyzing bo h i s
echnical pe o mance and ope a ional bene i s.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 26(02), 2883-2894
2888
A majo de ense alliance implemen ed a mul i-cloud in as uc u e spanning h ee majo cloud se ice p o ide s ac oss
geog aphically dis ibu ed egions o suppo i s command, con ol, communica ions, compu e s, in elligence,
su eillance, and econnaissance (C4ISR) sys ems [7]. The in as uc u e consis ed o app oxima ely 12,000 i ual
machines, 8,500 con aine ins ances, and 4,200 managed se ices dis ibu ed ac oss membe na ions. P io o
implemen ing he ede a ed lea ning solu ion, he en i onmen expe ienced an a e age o 27.3 hou s o Mean Time o
De ec ion (MTTD) o c oss-cloud inciden s, wi h 18% o se ious inciden s emaining unde ec ed un il hey impac ed
ope a ional capabili ies [7]. This agmen ed obse abili y p esen s a signi ican secu i y isk, pa icula ly o
sophis ica ed h ea ac o s who delibe a ely exploi c oss-cloud blind spo s.
The ede a ed moni o ing implemen a ion began wi h a pilo in 2022, ollowed by ull deploymen ac oss pa icipa ing
na ions in ea ly 2023. The solu ion a chi ec u e ollowed a hie a chical app oach wi h h ee ie s: (1) Local moni o ing
wi hin each cloud en i onmen using p o ide -na i e ools; (2) Regional agg ega ion nodes ha pe o med ede a ed
aining o speci ic geog aphic a eas; and (3) A global coo dina ion laye ha managed model dis ibu ion and secu e
agg ega ion [7]. This ie ed app oach balanced local so e eign y equi emen s wi h he need o comp ehensi e
isibili y. The implemen a ion u ilized homomo phic enc yp ion o g adien p o ec ion, wi h 4096-bi enc yp ion keys
and a e-enc yp ion schedule e e y 72 hou s o minimize he isk o c yp og aphic a acks [7].
Quan i a i e pe o mance analysis conduc ed o e nine mon hs o ope a ion demons a ed signi ican imp o emen s
in h ea de ec ion capabili ies. The mos no able me ic was a 31.7% educ ion in alse nega i es o c oss-cloud
secu i y inciden s compa ed o he p e ious non- ede a ed moni o ing app oach [7]. The sys em achie ed a mean F1
sco e o 0.873 o anomaly de ec ion ac oss cloud bounda ies, compa ed o 0.641 o adi ional co ela ion-based
me hods. De ec ion la ency o c oss-cloud inciden s dec eased om 27.3 hou s o 8.6 hou s on a e age, wi h 94% o
c i ical inciden s de ec ed wi hin 4 hou s o ini ial indica o s appea ing [7]. These imp o emen s ansla ed di ec ly o
enhanced ope a ional secu i y and educed ulne abili y windows.
Du ing a majo cybe de ense exe cise conduc ed in mid-2023, he ede a ed moni o ing sys em was subjec ed o a
ealis ic a ack scena io in ol ing sophis ica ed ac ics a ge ed speci ically a exploi ing cloud bounda y ulne abili ies
[8]. The exe cise, which in ol ed o e 200 cybe secu i y p o essionals ac oss mul iple na ions, simula ed a coo dina ed
a ack campaign agains c i ical de ense in as uc u e. The pe o mance analysis e ealed ha he ede a ed sys em
de ec ed 87% o he simula ed a ack echniques, compa ed o 52% o adi ional secu i y in o ma ion and e en
managemen (SIEM) sys ems ope a ing wi hou c oss-cloud isibili y [8]. Pa icula ly no ewo hy was he sys em's
abili y o iden i y da a ex il a ion a emp s ha le e aged mul iple cloud p o ide s as elay poin s, de ec ing 92% o
hese a emp s wi h an a e age ime- o-de ec ion o 17 minu es [8].
The exe cise also assessed he sys em's esilience agains ad e sa ial machine lea ning echniques. When subjec ed o
g adien poisoning a acks designed o deg ade model pe o mance, he ede a ed a chi ec u e demons a ed obus
de ense capabili ies, main aining 89% o baseline de ec ion accu acy despi e 15% o pa icipa ing nodes being
comp omised [8]. This esilience was a ibu ed o he secu e agg ega ion p o ocols and he implemen a ion o
Byzan ine- esis an ede a ed a e aging algo i hms ha could iden i y and mi iga e he impac o comp omised model
upda es.
Beyond echnical pe o mance, he exe cise e ealed signi ican ope a ional bene i s. Command s a epo ed a 43%
educ ion in ime equi ed o achie e si ua ional awa eness du ing simula ed inciden s, and a 37% imp o emen in he
accu acy o a ibu ion o a ack sou ces [8]. The abili y o isualize c oss-cloud a ack pa e ns enabled mo e e ec i e
coo dina ion o de ensi e esponses, wi h con ainmen ac ions execu ed on a e age 68 minu es as e han in p e ious
exe cises [8].
A comp ehensi e cos -bene i analysis o he implemen a ion quan i ied bo h he di ec and indi ec bene i s o he
ede a ed moni o ing app oach. The ini ial implemen a ion equi ed an in es men o app oxima ely $4.7 million,
including in as uc u e enhancemen s, specialized aining o cybe secu i y pe sonnel, and in eg a ion se ices [7].
Annual ope a ional cos s we e es ima ed a $1.2 million, p ima ily o dedica ed compu e esou ces and ongoing
c yp og aphic key managemen . Agains hese cos s, he analysis iden i ied annual bene i s o $12.8 million, de i ed
om se e al sou ces [7]:
• Reduced inciden esponse cos s: $3.2 million annually, based on he 31.7% educ ion in unde ec ed inciden s
and he a e age cos o $175,000 pe majo secu i y inciden .
• A oided ope a ional dis up ion: $5.7 million, calcula ed om he educ ion in se ice down ime mul iplied by
he es ima ed cos o dis up ion o c i ical ope a ions.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 26(02), 2883-2894
2889
• Dec eased eco e y e o s: $2.3 million in educed pe sonnel ime and esou ces dedica ed o eco e y om
secu i y b eaches.
• In as uc u e op imiza ion: $1.6 million om iden i ied ine iciencies in c oss-cloud esou ce u iliza ion ha
we e disco e ed h ough he enhanced moni o ing isibili y.
This cos -bene i analysis yielded a Re u n on In es men (ROI) o 172% o e a h ee-yea ho izon and a payback pe iod
o 14 mon hs [7]. These igu es we e conside ed conse a i e as hey excluded di icul - o-quan i y bene i s such as
enhanced p o ec ion o classi ied in o ma ion and imp o ed ope a ional secu i y pos u e.
A signi ican componen o he implemen a ion's success was a ibu ed o i s abili y o suppo p eemp i e h ea
de ec ion ac oss cloud bounda ies. By iden i ying abno mal pa e ns in seemingly benign ac i i ies ac oss mul iple
en i onmen s, he sys em could de ec he ea ly s ages o ad anced pe sis en h ea s (APTs) be o e hey p og essed
o ac ual da a ex il a ion o se ice dis up ion [8]. In one documen ed ins ance du ing he ope a ional deploymen , he
sys em iden i ied unusual au hen ica ion pa e ns occu ing simul aneously ac oss wo cloud en i onmen s ha , when
analyzed indi idually, appea ed wi hin no mal pa ame e s. This ea ly de ec ion enabled secu i y eams o iden i y and
mi iga e a sophis ica ed c eden ial-ha es ing campaign 37 days be o e he a acke s a emp ed o access sensi i e
sys ems [8].
The de ense o ganiza ion has since es ablished a oadmap o u he enhancing he sys em, wi h planned
imp o emen s including he in eg a ion o ede a ed ein o cemen lea ning o au oma ed esponse
ecommenda ions, expansion o include edge compu ing nodes in ac ical en i onmen s, and enhanced p i acy-
p ese ing echniques o suppo inclusion o in elligence da a sou ces wi h s ic e so e eign y equi emen s [7]. The
documen ed success o his implemen a ion has led o adop ion o simila app oaches by o he de ense and in elligence
o ganiza ions, wi h an es ima ed 35% o simila o ganiza ions planning implemen a ions by 2025 [7].
Figu e 3 Enhancing De ense In as uc u e Secu i y [7, 8]
5. Indus ial Applica ions and Pe o mance Me ics
The adop ion o ede a ed lea ning (FL) o mul i-cloud obse abili y has expanded beyond de ense applica ions o
a ious indus y e icals wi h c i ical equi emen s o bo h comp ehensi e moni o ing and da a p i acy. This sec ion
examines implemen a ions ac oss inancial se ices, heal hca e, and o he sec o s, analyzing pe o mance me ics and
key success ac o s o each domain.
The inancial se ices sec o , wi h i s s ingen equi emen s o ansac ion in eg i y and egula o y compliance,
ep esen s a p ima y adop e o ede a ed lea ning o c oss-cloud moni o ing. A no able implemen a ion in ol es a
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 26(02), 2883-2894
2890
global paymen ne wo k p ocessing o e 15.5 million daily ansac ions ac oss 11,000+ inancial ins i u ions in 200+
coun ies [9]. This ne wo k ope a es a dis ibu ed in as uc u e spanning i e majo cloud p o ide s and 17 geog aphic
egions, wi h each componen subjec o di e en egula o y amewo ks. P io o implemen ing ede a ed lea ning,
he o ganiza ion expe ienced an a e age o 7.3 minu es o ansac ion p ocessing delays pe mon h due o unde ec ed
c oss-cloud dependencies, a ec ing app oxima ely 175,000 high- alue ansac ions annually and esul ing in $12M in
ope a ional penal ies [9].
The paymen ne wo k's ede a ed moni o ing solu ion ocuses speci ically on API ansac ion lows ha c oss cloud
bounda ies. The implemen a ion uses a specialized FL a chi ec u e wi h ime-se ies con olu ional neu al ne wo ks
(CNNs) ained on 67 dis inc pe o mance me ics pe se ice endpoin . Da a emains wi hin each cloud en i onmen ,
wi h only model upda es exchanged ia secu e channels using 3072-bi RSA enc yp ion [9]. Pe o mance analysis
conduc ed o e 12 mon hs demons a es a 26.3% educ ion in c oss-cloud API ailu es compa ed o p e ious
moni o ing app oaches. Du ing peak p ocessing pe iods ( ypically expe iencing 2,300 ansac ions pe second), he
sys em iden i ied 94.7% o eme ging pe o mance deg ada ions be o e hey impac ed end-use s, compa ed o 61.2%
wi h adi ional h eshold-based moni o ing [9].
Financial ansac ion moni o ing p esen s unique challenges due o he c i ical na u e o paymen p ocessing and
se lemen sys ems. The FL implemen a ion demons a ed pa icula alue in acing dependencies be ween message
queuing sys ems hos ed in one cloud and ansac ion p ocessing componen s in ano he . Du ing a signi ican egional
ou age a ec ing a majo cloud p o ide in Q2 2023, he sys em iden i ied abno mal beha io pa e ns in dependen
sys ems 7.5 minu es be o e adi ional moni o ing de ec ed issues, enabling p eemp i e e ou ing o app oxima ely
42,000 ansac ions alued a $1.7B o al e na e p ocessing pa hs [9]. The o ganiza ion's inciden pos -mo em analysis
c edi ed he ea ly de ec ion wi h a oiding an es ima ed $3.2M in ope a ional penal ies and epu a ional damage [9].
Cos analysis o he inancial sec o implemen a ion shows an ini ial in es men o $5.3M wi h annual ope a ional cos s
o $1.8M, o se by $7.2M in annual sa ings om educed ou ages and imp o ed ope a ional e iciency. The calcula ed
ROI eached 189% o e h ee yea s, wi h a payback pe iod o 19 mon hs [9]. Beyond inancial me ics, he o ganiza ion
epo ed a 41% educ ion in ime spen in es iga ing c oss-cloud inciden s and a 37% dec ease in alse posi i e ale s,
allowing secu i y and ope a ions eams o ocus on genuine se ice imp o emen s a he han noise educ ion [9].
In he heal hca e sec o , ede a ed lea ning implemen a ions ace he dual challenges o s ic da a so e eign y
equi emen s and li e-c i ical eliabili y needs. A p ominen case s udy in ol es a Eu opean heal hca e ne wo k
managing pa ien da a ac oss 23 hospi als in h ee coun ies, wi h s ic GDPR compliance equi emen s [10]. Each
hospi al main ained pa ien eco ds in local cloud en i onmen s wi hin na ional bo de s o sa is y legal equi emen s,
bu s ill equi ed uni ied moni o ing o sha ed se ices such as diagnos ic imaging sys ems, elec onic heal h eco d
(EHR) pla o ms, and pha macy managemen [10].
The heal hca e implemen a ion ocused hea ily on p i acy-p ese ing mechanisms, employing bo h di e en ial p i acy
and secu e mul i-pa y compu a ion. The di e en ial p i acy implemen a ion used he Laplace mechanism wi h a
p i acy budge (ε) o 0.8 pe aining ound and a δ alue o 10^-6, exceeding GDPR equi emen s o sensi i e heal h
da a p o ec ion [10]. Secu e mul i-pa y compu a ion p o ec ed model agg ega ion, ensu ing ha no pa icipa ing node
could access he comple e model o i s upda es. These p o ec i e measu es educed model accu acy by only 3.2%
compa ed o a hypo he ical non-p i a e implemen a ion, ep esen ing an excellen p i acy-u ili y adeo o
heal hca e applica ions [10].
Pe o mance me ics o he heal hca e implemen a ion demons a e signi ican ope a ional bene i s. The sys em
de ec ed a egion-speci ic s o age ailu e a ec ing pa ien imaging da a 13 minu es be o e adi ional moni o ing
sys ems igge ed ale s, allowing au oma ed ailo e p ocesses o engage be o e any diagnos ic p ocedu es we e
impac ed [10]. O e six mon hs o ope a ion, he implemen a ion educed unplanned down ime o c oss-cloud se ices
by 42.7%, om 27.3 minu es mon hly o 15.6 minu es. Gi en ha each minu e o down ime in c i ical heal hca e
sys ems a ec s app oxima ely 17 pa ien in e ac ions, his imp o emen di ec ly enhanced ca e deli e y o an
es ima ed 11,900 pa ien s annually [10].
The heal hca e implemen a ion paid pa icula a en ion o pe o mance a ound medical da a access pa e ns. By
aining on anonymized ac i i y logs a he han pa ien da a i sel , he sys em iden i ied abno mal access pa e ns ha
could indica e secu i y issues o sys em mal unc ions. In one documen ed case, he ede a ed model de ec ed unusual
c oss-cloud au hen ica ion pa e ns ha led o he disco e y o a miscon igu ed iden i y ede a ion se ice ha had
c ea ed a secu i y ulne abili y a ec ing app oxima ely 10,800 pa ien eco ds [10]. This de ec ion occu ed 36 hou s
be o e he ulne abili y could be exploi ed, based on subsequen h ea in elligence analysis [10].
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 26(02), 2883-2894
2891
Compa a i e analysis ac oss indus y e icals e eals bo h common bene i s and domain-speci ic pe o mance
a ia ions. Telecommunica ions implemen a ions (5 documen ed cases) achie ed an a e age educ ion in MTTR o
c oss-cloud inciden s o 31.7%, simila o inancial se ices a 29.5% [9]. Howe e , manu ac u ing sec o
implemen a ions (7 cases) demons a ed a highe ROI o 213% o e h ee yea s, a ibu ed o he di ec co ela ion
be ween sys em down ime and p oduc ion losses [9]. Re ail implemen a ions showed he mos signi ican educ ion in
alse posi i es a 47.3%, likely due o he mo e p edic able wo kload pa e ns in consume - acing applica ions [9].
Scalabili y measu emen s indica e ha ede a ed lea ning implemen a ions main ain e ec i eness as he numbe o
cloud en i onmen s inc eases. Pe o mance analysis ac oss 23 di e en mul i-cloud deploymen s shows ha de ec ion
accu acy dec eases by only 2.3% on a e age when expanding om h ee o se en cloud en i onmen s, while aining
ime inc eases nea -linea ly a app oxima ely 14% pe addi ional cloud en i onmen [10]. This a o able scaling
cha ac e is ic makes ede a ed lea ning sui able o e en he mos complex mul i-cloud a chi ec u es ypical in global
en e p ises.
Reliabili y measu emen s ocus on bo h he moni o ing sys em i sel and i s impac on ope a ional esilience. Ac oss
documen ed implemen a ions, he ede a ed moni o ing sys ems main ained 99.97% a ailabili y, exceeding he 99.92%
achie ed by adi ional cen alized moni o ing app oaches [10]. This high a ailabili y is a ibu ed o he inhe en ly
dis ibu ed na u e o ede a ed lea ning, whe e he ailu e o indi idual nodes has minimal impac on o e all sys em
pe o mance. Mo e impo an ly, he sys ems demons a ed a 41.3% a e age imp o emen in co ec ly p edic ing
po en ial se ice dis up ions 10+ minu es be o e use impac , p o iding ope a ions eams wi h c i ical ime o
mi iga ion ac ions [10].
Common implemen a ion challenges iden i ied ac oss sec o s include: (1) ini ial model con e gence di icul ies when
cloud en i onmen s ha e signi ican ly di e en wo kload cha ac e is ics, equi ing 27-31% mo e aining ounds o
achie e s able pe o mance; (2) in eg a ion complexi y wi h exis ing secu i y in as uc u e, necessi a ing an a e age
o 47 pe son-days o specialized in eg a ion wo k pe cloud en i onmen ; and (3) he need o s anda dized ea u e
ex ac ion ac oss he e ogeneous moni o ing sys ems, which ypically accoun s o 35% o implemen a ion e o [10].
Despi e hese challenges, all documen ed implemen a ions achie ed posi i e ROI wi hin 24 mon hs, wi h an a e age
payback pe iod o 17.3 mon hs ac oss indus ies [10].
Table 1 Compa a i e Analysis o Key Implemen a ion Ou comes [9, 10]
Indus y Sec o
Key Pe o mance Imp o emen s
Re u n on In es men (ROI)
Financial Se ices
26.3% educ ion in c oss-cloud API ailu es
94.7% ea ly de ec ion o pe o mance deg ada ions
37% dec ease in alse posi i e ale s
189% o e h ee yea s wi h
19-mon h payback pe iod
Heal hca e
42.7% educ ion in unplanned down ime
Enhanced p o ec ion o 10,800+ pa ien eco ds
13-minu e ea ly de ec ion o s o age ailu es
No explici ly s a ed, bu
posi i e ROI wi hin 24 mon hs
Telecommunica ions
31.7% educ ion in Mean Time o Resolu ion (MTTR)
Simila pe o mance cha ac e is ics o inancial se ices
Wi hin indus y a e age o
posi i e ROI in 17.3 mon hs
Manu ac u ing
Di ec co ela ion be ween sys em up ime and p oduc ion
e iciency
Highes c oss-sec o ROI
213% o e h ee yea s
Re ail
47.3% educ ion in alse posi i es
Bene i s om p edic able wo kload pa e ns
Wi hin indus y a e age o
posi i e ROI in 17.3 mon hs
6. Fu u e Resea ch Di ec ions
This pape has p esen ed a comp ehensi e examina ion o ede a ed lea ning (FL) applica ions in mul i-cloud
obse abili y, demons a ing signi ican imp o emen s in c oss-pla o m moni o ing while p ese ing da a p i acy and
so e eign y. As his ield con inues o e ol e, se e al key esea ch di ec ions eme ge ha can u he enhance he
capabili ies, pe o mance, and secu i y o hese sys ems.
