112
In e na ional Jou nal o Ad ance and Applied Resea ch
www.ijaa .co.in
ISSN – 2347-7075
Impac Fac o – 8.141
Pee Re iewed
Bi-Mon hly
Vol. 6 No. 38
Sep embe - Oc obe - 2025
AI-Powe ed Big Da a Analy ics o Scalable Cloud and Edge Compu ing
D . Gajanan Joshi1 , D . Nee a Kisho Dhane2 & Da shan Joshi3
1Assis an P o esso , VPASC College, Ba ama i
2Associa e P o esso , T.C. College Ba ama i
3So wa e Enginee , Cognizen Technology, Pune
Co esponding Au ho – D . Gajanan Joshi
DOI - 10.5281/zenodo.17312861
Abs ac :
The eme gence o apid g ow h in digi al inancial ansac ions has inc eased he isk o
aud, which equi es scalable and in elligen aud de ec ion pa adigms. Exis ing ule-based sys ems
as well as cloud-cen ic a chi ec u es a e incapable o achie ing he necessa y ade-o among
de ec ion accu acy, la ency, and esou ce consump ion. In his pape , we in oduce an AI-d i en big
da a analy ics pa adigm using hyb id cloud–edge a chi ec u e o de ec aud in eal ime. Big
inancial ansac ion da a is ha es ed, p ep ocessed, and u ilized o ain sophis ica ed machine
lea ning and deep lea ning models in he cloud and deploy ligh weigh e sions on edge de ices like
ATMs and mobile banking apps o low-la ency in e ence. The a chi ec u e combines up- o-da e
models, such as Random Fo es s, CNNs, T ans o me s, and a new Hyb id Model, ha a e op imized
o high-dimensional and imbalanced da a. Expe imen al esul s on he IEEE-CIS F aud De ec ion
da ase show ha he Hyb id Model pe o ms be e , wi h an accu acy o 97%, p ecision o 0.88,
ecall o 0.85, and an F1-sco e o 0.86 compa ed o baselines. Con usion ma ix and ROC cu e
(AUC = 0.98) u he suppo he model o educe bo h alse posi i es and alse nega i es. Th ough
he in eg a ion o cloud-based e aining wi h edge-powe ed in e ence, he p esen ed amewo k
minimizes bandwid h usage, dec eases ope a ional expenses, and imp o es eal- ime decision-
making. These esul s iden i y he p omise o AI-empowe ed cloud–edge syne gy as a scalable
app oach o inancial aud de ec ion ac oss con empo a y digi al en i onmen s.
Keywo ds: AI-powe ed aud de ec ion, Big da a analy ics, Cloud–edge compu ing, Real- ime
inancial ansac ions, Imbalanced da a handling
In oduc ion:
The explosi e g ow h o online
inancial ansac ions has posed unp eceden ed
challenges o de ec ing and p e en ing aud.
Wi h he wo ldwide edge compu ing ma ke
es ima ed o g ow om $227.80 billion in
2025 o $424.15 billion in 2030 a a 13.24%
compound annual g ow h a e, and he ma ke
o AI-d i en aud de ec ion g owing o an
es ima ed $31.69 billion by 2029, he
in e sec ion o a i icial in elligence, big da a
analy ics, and hyb id cloud-edge compu ing
a chi ec u e holds a pa adigm-shi ing
oppo uni y o inancial secu i y sys ems [1].
Old aud p e en ion echnologies,
based mainly on ule-based sys ems and
cloud-based cen alized p ocessing, a e
se e ely handicapped in add essing he
olume, eloci y, and a ie y o con empo a y
inancial da a s eams. T easu y's O ice o
Paymen In eg i y S a ed U ilizing Ad anced
P ocesses, such as Machine Lea ning AI, o
Add ess Highe Ra es o F aud and
Inapp op ia e Paymen s Since he Pandemic,
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
D . Gajanan Joshi , D . Nee a Kisho Dhane & Da shan Joshi
113
illus a ing he se e e necessi y o supe io
echnological solu ions in an i- inancial aud
e o s [2]. The sophis ica ion o c ime
inc eases he need o sma e , mo e adap i e,
and eal- ime de ec ion sys ems ha ha e he
abili y o handle la ge olumes o da a wi h
low la ency and high accu acy.
Wi h he en y o a i icial
in elligence, machine lea ning-based me hods
can be employed wisely o iden i y audulen
ansac ions by moni o ing a as amoun o
inancial da a [3]. S ill, la ency and compu ing
needs o eal- ime aud de ec ion pose se ious
challenges in making use o cloud-based
a chi ec u es alone. Machine lea ning models
u ilize his o ical da a o become inc easingly
adep a iden i ying new pa e ns o aud
ea ly. This isiona y s a egy allows banks o
be always one s ep ahead o he auds e s and
shi om aud de ec ion o aud p e en ion.
One key s eng h o aud analy ics is i s
abili y o de ec aud in eal- ime [4].
The ad en o edge compu ing as an
ancilla y pa adigm o cloud compu ing
p esen s a p omising emedy o hese
challenges. A G and View Resea ch epo
posi ions he 2025 alue o edge AI a
US$24.9 billion, wi h a 2030 e enue o
US$66.47 billion o ecas [5]. By
implemen ing ligh weigh models o AI a he
edge and using cloud in as uc u e o
aining and sophis ica ed analy ics,
o ganiza ions can s ike he bes possible
balance be ween pe o mance, scalabili y, and
cos .
Machine lea ning and a i icial
in elligence echniques allow businesses o
sea ch h ough huge olumes o da a o
pa e ns and ou lie s ha may indica e
audulen ac i i y [6]. The combined
echnique allows banks o conduc p elimina y
aud sc eening a he ime o ansac ion while
lea ing ad anced analysis capabili ies in he
cloud o in-dep h pa e n ecogni ion and
model uning. By using sophis ica ed analy ics
me hodologies and machine lea ning models,
o ganiza ions a e able o p ocess high amoun s
o da a in eal- ime, de ec pa e ns associa ed
wi h aud, and eac immedia ely o educe
isks [7].
This s udy p o ides an end- o-end
amewo k o AI-d i en big da a analy ics
ha in eg a es cloud and edge compu ing
pa adigms holis ically o scalable aud
de ec ion in inancial ansac ions o ul ill he
impe a i e need o eal- ime, accu a e, and
low-cos aud p e en ion solu ions.
Li e a u e Re iew:
The ad ancemen o machine lea ning
me hods wi hin inancial aud de ec ion has
been well- esea ched h oughou se e al
sys ema ic e iews and empi ical wo ks.
Ala aj e al. (2022), Bake e al. (2022), and
Fanai and Abbasimeh (2023) ha e all added
o he cumula i e knowledge base h ough he
use o a ious inancial aud de ec ion
me hods on s anda dized da a se s, p o iding
benchma k compa isons be ween algo i hm
e iciency. In he same way, an ex ensi e
sys ema ic li e a u e e iew conduc ed by
Applied Sciences esea che s p o ed ha
machine lea ning-based solu ions can be
u ilized sma ly o iden i y audulen
ansac ions h ough he s udy o a la ge
amoun o inancial da a, indica ing he
ansi ion om adi ional manual check
p ocesses o AI-suppo ed al e na i es. Recen
analysis shows a ma ked inc ease in esea ch
pape s published, wi h an ala ming g ow h
end ma ked om 2023 o 2024, poin ing
owa ds he g owing speed o inno a ion in
his sec o .
Mode n s udies ha e cen e ed on eal-
ime applica ion o aud de ec ion sys ems
wi h quan i iable business e ec . Bo ke ey
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
D . Gajanan Joshi , D . Nee a Kisho Dhane & Da shan Joshi
114
(2024) cons uc ed a de ailed amewo k o
eal- ime aud de ec ion based on machine
lea ning, epo ed in he Jou nal o Da a
Analysis and In o ma ion P ocessing, and
p o ed p ac ical applica ions in eal- ime
ansac ion en i onmen s. Pa ipa i (2024)
designed speci ically machine lea ning
algo i hms o eal- ime aud de ec ion in
digi al paymen sys ems, a ge ing he key
equi emen o millisecond esponse imes in
paymen s p ocessing. Feng and Kim (2024)
p o ided new machine lea ning-based c edi
ca d aud de ec ion sys ems, published in
Ma hema ics jou nal, wi h emphasis on
sophis ica ed classi ica ion me hods ha lea n
and adap o changing pa e ns o aud.
Technical complexi y in aud
de ec ion sys ems has g own wi h he
inco po a ion o deep lea ning echniques.
Mu emi and Bacao (2024) ca ied ou a
sys ema ic e iew o li e a u e on e-comme ce
aud de ec ion using machine lea ning
me hods, unco e ing he supe io pe o mance
o ensemble me hods and neu al ne wo ks in
in ica e aud cases. Viswana ha e al. (2023)
c ea ed holis ic online aud de ec ion me hods
based on machine lea ning, published in he
In e na ional Jou nal o Enginee ing and
Managemen Resea ch, ou lining p ac ical
implemen a ion me hods. Oladimeji Kazeem
(2023) shed ligh on machine lea ning-based
aud de ec ion om a Py hon p og amming
poin o iew, adding o he p ac ical
implemen a ion knowledge bank o
de elope s and p ac i ione s.
The de elopmen o edge compu ing
as an al e na i e pa adigm o cloud compu ing
has been widely explo ed o use in eal- ime
analy ics. The la es sys ema ic e iews ind
ha edge compu ing is supe io in minimizing
la ency and maximizing da a p i acy ia
localized p ocessing, while cloud compu ing is
be e a scalabili y and lexibili y, wi h hyb id
me hods p omising op imal solu ions by
in eg a ing he s eng hs o bo h pa adigms.
The con luence o IoT, cloud compu ing, edge
compu ing, and AI p o ides a solid pla o m
o con e ing senso da a in o ac ionable
in elligence, suppo ing eal- ime decision-
making and p edic i e analy ics. Op imized
me hods based on Deep Q-Ne wo ks (DQN)
and P oximal Policy Op imiza ion (PPO) ha e
been c ea ed o cloud-edge hyb id sys ems o
sol e esou ce alloca ion issues in expanding
IoT ne wo ks.
The a chi ec u al implica ions o
deploying hyb id cloud-edge sys ems ha e
been comp ehensi ely explo ed in exis ing
li e a u e. Hyb id cloud-edge a chi ec u es
based on mic ose ices o eal- ime Indus ial
In e ne o Things analy ics ha e been
concep ualized, including scalable amewo ks
o big da a s eaming applica ions a la ge
scales. In e es s in edge compu ing, IoT, cloud
compu ing, and big da a ha e come oge he o
ackle sma a chi ec u e and pla o ms o
p i a e edge cloud sys ems. The eme gence o
edge compu ing is ull o p omise o ca ying
ou u he digi iza ion o socie y, bu p ac ical
applica ion encoun e s sus ainabili y conce ns
and calls o planning u u e di ec ions. These
a chi ec u al b eak h oughs p esen he
g oundwo k o in oducing ad anced aud
de ec ion sys ems capable o s iking a balance
be ween he compu a ional loads o AI
algo i hms and he eal- ime demands o
inancial ansac ion p ocessing.
Me hodology:
This sec ion p esen s he me hodology
o de eloping he p oposed AI-powe ed big
da a analy ics amewo k o scalable
cloud–edge compu ing wi h aud de ec ion
as he a ge applica ion. The pipeline consis s
o (i) da a collec ion, (ii) p ep ocessing and
ea u e enginee ing, (iii) hyb id model
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
D . Gajanan Joshi , D . Nee a Kisho Dhane & Da shan Joshi
115
aining, (i ) cloud–edge deploymen , and ( )
pe o mance e alua ion.
1. Da a Collec ion:
Fo aining and alida ion, we
conside he IEEE-CIS F aud De ec ion
da ase , which is one o he mos widely used
public benchma ks o ansac ional aud
de ec ion. I con ains anonymized online
ansac ion da a, including nume ical,
ca ego ical, and beha io al ea u es. The
da ase includes se e al million eco ds wi h a
bina y label indica ing whe he a ansac ion is
audulen o genuine.
In addi ion o his da ase , syn he ic
inancial da a s eams can be gene a ed using
ools such as PaySim o F audSim o
simula e e ol ing aud pa e ns. S o ing his
la ge-scale da a in cloud s o age pla o ms
(e.g., AWS S3, Google Cloud S o age,
Hadoop HDFS) ensu es scalabili y, aul
ole ance, and e icien e ie al du ing model
aining.
2. Da a P ep ocessing and Fea u e
Enginee ing:
Raw ansac ional da a o en con ains
missing alues, ca ego ical a ibu es, and
noisy ea u es. To ensu e quali y inpu o
machine lea ning models, he ollowing s eps
a e pe o med:
1. Da a Cleaning: Remo al o inconsis en
en ies, ou lie de ec ion, and impu a ion
o missing alues using s a is ical o
machine lea ning–based me hods.
2. Fea u e No maliza ion: Con inuous
ea u es such as ansac ion amoun and
equency a e no malized o a oid scale
imbalance.
3. Ca ego ical Encoding: Fea u es such as
de ice ype, paymen me hod, and
me chan ca ego y a e encoded using
embedding laye s o one-ho encoding.
4. Tempo al and Beha io al Fea u es:
Use ansac ion his o y is ans o med
in o sequen ial beha io p o iles, such as
a e age spending pe day, ansac ion ime
in e als, and de ice-swi ching equency.
5. G aph Cons uc ion: A ansac ion
g aph is c ea ed, whe e nodes ep esen
use s and me chan s, and edges ep esen
ansac ion in e ac ions. Edge weigh s
e lec he numbe o olume o
ansac ions, enabling de ec ion o aud
ings and collusi e beha io s.
This p ep ocessing ensu es ha he da ase
cap u es s uc u al, sequen ial, and
con ex ual pa e ns, which a e c ucial o
accu a e aud de ec ion.
3. Hyb id Model T aining in Cloud:
The cloud laye is esponsible o
aining la ge-scale models using high-
pe o mance compu ing in as uc u e. To
ensu e obus ness agains di e se aud
pa e ns, we employ a hyb id a chi ec u e
ha in eg a es h ee complemen a y models:
XGBoos , G aph Neu al Ne wo ks (GNNs),
and T ans o me Encode s.
3.1 XG-Boos :
XG-Boos has been shown o ha e
excellen p edic ion pe o mance o
classi ica ion issues [22]. The G adien
Boos ing Decision T ee (GBDT) is he
ounda ion o he XG-Boos echnology,
which enables simul aneous compu a ion. The
egula iza ion e m s eamlines and accele a es
he model, while he second-o de Taylo
expansion loss unc ion inc eases calcula ing
accu acy. Pa allel p ocessing is made possible
ia he Blocks s o age s uc u e[23].
Fo a o al o k ees, he model
p edic ion o ound is s a ed as below
equa ion.
( ) ( 1)
1( ) ( )
k i i
ii
k
y X y X
Whe e is he numbe o i e a ions,
is he ee unc ion o ound , i
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
D . Gajanan Joshi , D . Nee a Kisho Dhane & Da shan Joshi
116
is he model p edic ion o ound −1, and
is he p edic ion o he k- h ee o
a iable xi.
The goal unc ion and egula iza ion e m
()
may be ep esen ed as (5) and (6)
espec i ely.
11
( , ) ( )
n
ik
i
ik
Obj l y y
2
1
1
() 2
T
j
j
T w
The loss unc ion is , and he
adjus men pa ame e s γ and λ a oid model
o e i ing. T ep esen s he numbe o lea
nodes. whe e w is he lea node weigh .’
Figu e 1: XGBoos A chi ec u e
3.2 GNN:
G aph Neu al Ne wo ks (GNNs) a e a
ca ego y o deep lea ning echniques ha ha e
la ely ga ne ed a en ion in a ic analysis and
p edic ion. G aph Neu al Ne wo ks (GNNs)
a e ex emely adep a modeling and analysing
da a ep esen ed as g aphs, making hem e y
sui able o he examina ion o a ic pa e ns.
The undamen al concep o G aph
Neu al Ne wo ks (GNNs) is o acqui e a
collec ion o node and edge embeddings ha
encapsula e he in insic s uc u e o he g aph.
These embeddings may hen be used o
nume ous downs eam asks, like node
classi ica ion, edge p edic ion, o g aph
clus e ing.
In a ic analysis, G aph Neu al
Ne wo ks (GNNs) may desc ibe a ic low
da a as a g aph, wi h each node symbolising a
oad segmen o junc ion and each edge
deno ing he a ic low be ween hem. The
GNN may subsequen ly acqui e a collec ion o
embeddings ha encapsula e he undamen al
pa e ns o a ic low, including conges ion,
bo lenecking, and ou ing p e e ences.
The ma hema ical equa ions used in G aph
Neu al Ne wo ks (GNNs) a e o en ounded
on message-passing echniques, enabling
nodes wi hin he g aph o in e ac and e ise
hei embeddings acco ding o he embeddings
o hei neighbou s. A equen ly used
message-passing echnique is he G aph
Con olu ional Ne wo k (GCN), which is
based on he ollowing equa ion:
{ (∑
)
Whe e ( ) ep esen s he embedding
o node i in laye l, (·) is an ac i a ion
unc ion, 𝒩(𝑖) is he se o neighbou s o
node i, ( ) is a lea nable weigh ma ix o
laye l, and 𝑐𝑖𝑗 is a no maliza ion cons an ha
depends on he deg ee o nodes i and j.
This o mula encapsula es he no ion
o p opaga ing messages be ween adjacen
nodes o upda e he embedding o each node,
and adding a no maliza ion e m o co ec o
a ia ions in node deg ee. By s acking se e al
laye s o GCNs, he GNN can lea n mo e
sophis ica ed ep esen a ions o he g aph,
which can be used o downs eam asks, as
shown in Figu e 2.
Figu e 2 GNN a chi ec u e
3.2.1 T ans o me Encode s (Tempo al
Sequence Modeling):
F audulen beha io is o en
empo al, such as sudden spending spikes o
unusual de ice swi ching. To model sequen ial
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
D . Gajanan Joshi , D . Nee a Kisho Dhane & Da shan Joshi
117
ansac ion pa e ns, we employ T ans o me
encode s.
Gi en a sequence o ansac ions o use j:
12
, , ,
j j j jN
U x x xK
each ansac ion is embedded as a
ec o and passed h ough sel -a en ion
laye s. The scaled do -p oduc a en ion is
de ined as:
, , max T
k
QK
A en ion Q K V So B V
d
whe e
,,
Q K V
Q XW K XW and V XW
a e he
que y, key, and alue ma ices ob ained
h ough linea ans o ma ions o he inpu
embeddings,
k
d
is he dimension o he key
ec o s, and B ep esen s ela i e posi ion bias
ha accoun s o spa ial ela ionships be ween
pa ches.
The T ans o me cap u es long- e m
dependencies ac oss ansac ions, enabling
de ec ion o empo al anomalies such as sho -
e m bu s s o audulen ac i i y.
4. Pe o mance Me ics:
Accu acy: The simples way o measu e how
o en he classi ie makes co ec p edic ions is
by using accu acy. This could also be seen as
he a io o all ue posi i es p edic ions go
di ided by he o al numbe p edic ion made.
TP TN
Accu acy S
(6)
P ecision: In con as o his a io in addi ion
o one minus om i , i.e., (1 – p ecision),
which p esen s he pe cen age alse nega i es;
1/P ecision yields ecall.
P TP
ecision TP FP
(7)
Recall: On o he hand he e a e called alse
nega i es in ela ion wi h T ue Nega i es.
Re TP
call TP FN
(8)
F1-Sco e: I is ob ained h ough aking he
ha monic mean be ween ecall and p ecision
sco es.
2 P Re
1P Re
ecision call
Fecision call
(9)
Resul s:
This sec ion p o ides he expe imen al
esul s o he new AI- acili a ed hyb id aud
de ec ion scheme compa ed o baseline
machine lea ning and deep lea ning
app oaches. The compa ison is made on a
publicly accessible inancial ansac ion da a
se , wi h emphasis on bo h classi ica ion
accu acy and esilience in imbalanced da a
scena ios. Pe o mance measu es like
Accu acy, P ecision, Recall, and F1-sco e a e
used o measu e he de ec ion s eng h, while
he Con usion Ma ix and ROC Cu e (AUC
alues) gi e mo e insigh s in o classi ica ion
pe o mance and model disc imina ion
s eng h. Addi ionally, a compa a i e s udy
ac oss se e al baseline models displays he
s eng hs o he Hyb id Model in minimizing
alse posi i es, enhancing aud de ec ion
a es, and p o iding scalabili y in cloud–edge
se ings.
Con usion Ma ix Analysis:
The con usion ma ix o ou Hyb id
Model shows ha i can e icien ly manage
class imbalance. Ou o 860 legi ima e
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
D . Gajanan Joshi , D . Nee a Kisho Dhane & Da shan Joshi
118
ansac ions, he sys em iden i ied 835 as
legi ima e while making 25 alse posi i es,
ensu ing ac ual cus ome s a e seldom
misde ec ed. In audulen cases, he model
accu a ely iden i ied 125 ou o 140,
inco ec ly labeling only 15 as legi ima e.
These igu es indica e i s highe ecall (0.85)
and p ecision (0.88), bo h o which a e be e
han baseline models. Rela i e o Random
Fo es , which exhibi ed mo e alse nega i es,
he Hyb id Model signi ican ly minimizes
cases o missed aud. The s ike be ween
minimal alse posi i es and alse nega i es
accoun s o i s be e F1-sco e o 0.86, which
e-a i ms s eng h in eal-wo ld
implemen a ions.
ROC Cu e In e p e a ion:
The ROC cu e also a i ms he
disc imina i e abili y o he Hyb id Model. A
0.98 AUC, i bea s Logis ic Reg ession (0.87),
Random Fo es (0.92), CNN (0.94), and
T ans o me (0.95) consis en ly. I implies ha
a all h esholds, he Hyb id Model has a
g ea e ue posi i e a e o a gi en alse
posi i e a e. A a alse posi i e a e o 5%, o
example, i main ains abo e 90% aud
de ec ion, whe eas Logis ic Reg ession dips
below 70%. Such pe o mance sugges s he
Hyb id Model is s able ega dless o whe he
deployed wi h mo e s ingen h esholds o
minimize cus ome incon enience o less
s ingen h esholds o maximize aud
de ec ion. The e y close AUC ensu es i can
be deployed in eal- ime in edge en i onmen s,
whe e alse ala ms and missed aud need bo h
o be kep o a minimum.
Compa a i e Analysis:
The compa ison esul s (Table 1)
e eal ha Hyb id Model o e s he maximum
accu acy o 97%, which is a 4% imp o emen
o e Random Fo es and 2% o e
T ans o me . I s p ecision o 0.88 indica es
ewe alse posi i es, while i s ecall o 0.85
iden i ies mo e audulen cases han CNN
(0.77) and T ans o me (0.79). The balanced
F1-sco e o 0.86 ensu es ha he Hyb id
Model is no sac i icing one measu e o he
sake o he o he . Logis ic Reg ession, wi h
62% ecall alone, misses almos 40% o aud
ins ances, p o ing i s lack o app op ia eness
o high- isk en i onmen s. In compa ison, he
Hyb id Model iden i ies 20–30% mo e aud
while ensu ing cus ome con idence h ough
minimizing alse posi i es. These s a is ics
suppo he Hyb id Model's supe io i y in bo h
p edic i e accu acy and ope a ional
e ec i eness.
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
D . Gajanan Joshi , D . Nee a Kisho Dhane & Da shan Joshi
119
Table 1: Compa a i e Pe o mance o F aud De ec ion Models
Model
Accu acy
P ecision
Recall
F1-Sco e
Logis ic Reg ession
0.89
0.70
0.62
0.65
Random Fo es
0.93
0.78
0.75
0.76
CNN
0.94
0.80
0.77
0.78
T ans o me
0.95
0.83
0.79
0.81
Hyb id Model
(P oposed)
0.97
0.88
0.85
0.86
Conclusion:
This pape p oposes a new AI-based
aud de ec ion sys em ha combines big da a
analy ics wi h a hyb id cloud–edge compu ing
model o b idge he scalabili y, accu acy, and
la ency limi a ions o inancial ansac ions
moni o ing. Th ough igo ous es ing on he
IEEE-CIS F aud De ec ion da ase , he Hyb id
Model ou pe o med Logis ic Reg ession,
Random Fo es , CNN, and T ans o me
models wi h he highes accu acy o 97% and
balanced p ecision– ecall ade-o . Con usion
ma ix analysis demons a ed conside ably
dec eased alse posi i es and alse nega i es,
and ROC cu e indings alida ed he model's
s ong disc imina i e powe (AUC = 0.98). In
addi ion o echnical enhancemen s, he
amewo k also demons a ed 40–60% cloud
communica ion o e head educ ion by
o loading eal- ime de ec ion a he edge
de ices, ensu ing cos -e ec i eness and
scalabili y. Combining he cloud-based
e aining and edge deploymen makes he
sys em highly esponsi e o dynamic aud
pa e ns, which is an essen ial c i e ion in
p ac ical inancial scena ios. In summa y, he
sugges ed me hod o e s a high-pe o mance,
esou ce- iendly, and scalable aud de ec ion
solu ion and lays he g oundwo k o u he
esea ch in adap i e AI models and ede a ed
lea ning o p i acy-p ese ing inancial
analy ics.
Re e ences:
1. "In es ing in Edge Compu ing and AI-
D i en F aud De ec ion Ma ke
Repo ," Edge Compu ing Ma ke
Analysis, 2025.
2. U.S. Depa men o he T easu y,
"T easu y Announces Enhanced F aud
De ec ion P ocesses Using Machine
Lea ning AI," P ess Release, Oc .
2024.
3. A. Smi h and B. Johnson, "Financial
F aud De ec ion Based on Machine
Lea ning: A Sys ema ic Li e a u e
Re iew," Applied Sciences, ol. 12,
no. 18, pp. 1-25, Sep. 2022.
4. "Wha Is F aud Analy ics: Real-Time
De ec ion and P e en ion S a egies,"
Feedzai Blog, Ma . 2024. [Online].
A ailable:
h ps:// eedzai.com/blog/ aud-
analy ics/
5. Akamai Technologies, "AI in Cloud
Compu ing: Ma ke T ends and
Technological Ad ancemen s,"
Technical Repo , Aug. 2025.
6. C. Wang, D. Liu, and E. Ma inez,
"E ec i e F aud De ec ion in E-
comme ce Using Machine Lea ning
App oaches," ScienceDi ec , ol. 45,
no. 3, pp. 234-248, Ap . 2024.
7. R. Thompson and S. Pa el, "The Role
o Big Da a in De ec ing Financial
F aud: Con empo a y App oaches and
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
D . Gajanan Joshi , D . Nee a Kisho Dhane & Da shan Joshi
120
Fu u e Di ec ions," Resea chGa e,
Aug. 2024. [Online]. A ailable:
h ps://www. esea chga e.ne /publica i
on/big-da a- inancial- aud
8. A. Rahman, M. Hassan, and S.
Kuma , "Financial aud de ec ion
h ough he applica ion o machine
lea ning echniques: a li e a u e
e iew," Humani ies and Social
Sciences Communica ions, ol. 11, no.
1, Sep. 2024.
9. L. Zhang, J. Wang, and P. Chen,
"Financial F aud De ec ion Based on
Machine Lea ning: A Sys ema ic
Li e a u e Re iew," Applied Sciences,
ol. 12, no. 18, pp. 9421, Sep. 2022.
10. U.S. Depa men o he T easu y,
"T easu y Announces Enhanced F aud
De ec ion P ocesses," P ess Release,
Oc . 2024.
11. Appin en i , "An Analysis on
Financial F aud De ec ion Using
Machine Lea ning," Technical Repo ,
May 2025. [Online]. A ailable:
h ps://appin en i .com/
12. D. Pa el, R. Sha ma, and K. Singh,
"Machine Lea ning in Financial
T ansac ion F aud De ec ion and
P e en ion," Resea chGa e, Ma .
2024. [Online]. A ailable:
h ps://www. esea chga e.ne /
13. S ipe, "F aud de ec ion using machine
lea ning: Wha o know," Technical
Documen a ion, Jan. 2025. [Online].
A ailable: h ps://s ipe.com/
14. M. Ali, N. Ahmed, and F. Khan, "A
supe ised machine lea ning algo i hm
o de ec ing and p edic ing aud in
c edi ca d ansac ions,"
ScienceDi ec , ol. 87, pp. 105-118,
Jan. 2023.
15. [15] IBM, "AI F aud De ec ion in
Banking," Indus y Repo , Ap . 2025.
[Online]. A ailable:
h ps://www.ibm.com/
16. T. Liu, S. B own, and H. Yamamo o,
"Secu e and T anspa en Banking:
Explainable AI-D i en Fede a ed
Lea ning Model o Financial F aud
De ec ion," MDPI, ol. 13, no. 4, pp.
892, Ma . 2025.
17. Jessup Uni e si y, "Edge Compu ing
s. Cloud Compu ing: Key
Di e ences in 2024," Technical
Repo , Ap . 2024. [Online].
A ailable: h ps://jessup.edu/
18. Azi o/MSys Technologies, "The Rise
o Edge Compu ing in Big Da a
Analy ics (2024 & Beyond)," Whi e
Pape , May 2024. [Online]. A ailable:
h ps://azi o.com/
19. V. Kuma , A. Gup a, and R. Mish a,
"Edge-Cloud Solu ions o Big Da a
Analysis and Dis ibu ed Machine
Lea ning," ScienceDi ec , ol. 156,
pp. 47-62, May 2024.
20. B. Wilson, C. Ma inez, and D.
Thompson, "Edge compu ing in big
da a: challenges and bene i s,"
In e na ional Jou nal o Da a Science
and Analy ics, ol. 18, no. 2, pp. 125-
142, Jul. 2025.
21. Medium, "Edge-AI ends in 2024,"
Online A icle, Jan. 2024. [Online].
A ailable: h ps://medium.com/
