1
T us Sco e P edic ion o IoT De ice Onboa ding Using T ans e
and Few-Sho Lea ning in Consume Elec onics
Ilias Poli is1, Michail Bampa sikos2, Apos olis Za as2and Ch is os Xenakis2
1Indus ial Sys ems Ins i u e, ATHENA Resea ch Cen e , Pa as, G eece
2Sys ems Secu i y Labo a o y, Uni e si y o Pi aeus, G eece
Abs ac —The apid p oli e a ion o In e ne o Things (IoT)
de ices in consume elec onics has made e icien us sco e
p edic ion essen ial o secu e de ice onboa ding. This pape
p esen s a hyb id T us Managemen amewo k ha in eg a es
ew-sho and ans e lea ning wi h a s a is ical Ma ko chain
ounda ion o add ess da a sca ci y and adap abili y challenges
in dynamic IoT en i onmen s. The ew-sho lea ning phase
enables apid adap a ion om minimal da a (as ew as 5–20
onboa ding samples), while ans e lea ning ensu es obus
c oss-domain gene alizabili y (e.g., om consume o indus ial
IoT). Comp ehensi e e alua ion demons a es ha , wi h 20
onboa ding samples, he p oposed app oach achie es ine- uned
Mean Squa ed E o (MSE) as low as 4.45 (XGBoos ) and
5.07 (Random Fo es ), R2sco es exceeding 0.96, and a e age
p edic ion e o (MAE) below 1.55. Ba ched dis ibu ed ledge
ope a ions educe o al onboa ding la ency o unde 750 ms o
i e de ices, wi h sys em h oughpu a e aging 137.7 de ices pe
second ac oss models. These esul s show he amewo k deli e s
accu a e, low-la ency, and scalable us sco e p edic ions sui able
o eal- ime onboa ding in e ol ing IoT en i onmen s.
Index Te ms—IoT, T us Sco e P edic ion, T ans e Lea ning,
Few-Sho Lea ning, De ice Onboa ding, Real-Time P edic ion
I. INTRODUCTION
The apid p oli e a ion o In e ne o Things (IoT) de ices
in consume elec onics has in oduced signi ican challenges
in ensu ing secu e and eliable de ice in e ac ions, pa icula ly
du ing onboa ding in dynamic, da a-cons ained en i onmen s.
A obus T us Managemen (TM) laye is indispensable o
apidly gauging a newcome ’s eliabili y, shielding he ne -
wo k om da a-poisoning and ze o-day h ea s, and achie ing
hese objec i es wi hin he s ic da a and esou ce cons ain s
ypical o consume -g ade IoT. Howe e , exis ing TM ap-
p oaches p esen no able limi a ions. S a is ical models, such
as ou p io Ma ko chain-based Mul i A ibu e Decision
Making (MADM) amewo k [1], a e obus and esou ce-
e icien bu exhibi limi ed adap abili y o apidly e ol ing
h ea s and new a ack ec o s. Con e sely, pu ely Machine
Lea ning (ML)-based me hods [2] o en equi e la ge aining
da ase s, limi ing hei p ac icali y in eal- ime onboa ding
scena ios—such as in sma homes, heal hca e wea ables, and
indus ial IoT—whe e only minimal da a is a ailable o new
de ices.
This mo i a es he need o a hyb id app oach ha p ese es
he e iciency and in e p e abili y o s a is ical models while
inco po a ing he adap abili y o ML echniques unde ex eme
da a sca ci y. The cen al p oblem add essed in his pape is
he e o e: how o eliably p edic he us wo hiness o a new
IoT de ice du ing i s onboa ding p ocess when only a hand ul
o labelled samples (5–20) a e a ailable, while mee ing s ic
la ency budge s (sub-500 ms) and ensu ing esilience agains
ad e sa ial manipula ion. Sol ing his p oblem equi es no
only accu a e us sco e p edic ions unde limi ed da a bu
also sys em-le el scalabili y o handle la ge de ice popula ions
in eal ime.
To add ess hese challenges, his pape p oposes a no el
hyb id TM amewo k ha in eg a es: (i) Few Sho Lea ning
(FSL) o apid adap a ion om minimal onboa ding da a;
(ii) ans e lea ning o c oss-domain gene alisa ion (e.g.,
consume - o-indus ial IoT); and (iii) a ounda ional Ma ko
chain-based s a is ical model o in e p e able and de e minis-
ic us es ima ion. The no el y o ou app oach lies in his
explici in eg a ion: unlike p io TM solu ions, we combine
he s a is ical obus ness o a Ma ko /MADM us unc ion
wi h he adap abili y o ML p edic o s and enhance i u he
wi h a ma hema ically jus i ied blending s a egy be ween base
and ine- uned models. This makes he amewo k esilien
o poisoning o sca ce onboa ding samples while ensu ing
adap abili y o de ice-speci ic beha iou s.
Ou expe imen al e alua ion, using Random Fo es , k-NN,
and XGBoos as us p edic o s, demons a es ha he p o-
posed amewo k achie es ine- uned Mean Squa ed E o
(MSE) as low as 4.45 (XGBoos ) and 5.07 (Random Fo -
es ), R2sco es exceeding 0.96, and Mean Absolu e E o
(MAE) below 1.55 wi h 20 onboa ding samples. The app oach
achie es sys em h oughpu up o 175 de ices pe second (k-
NN) and consis en ly a e ages abo e 137 de ices/s ac oss
models. Onboa ding la ency is educed o unde 750 ms o
i e de ices using ba ched dis ibu ed ledge (DLT) ope a ions.
To u he imp o e p edic ion accu acy and adap abili y unde
ex eme da a sca ci y, we employ a ma hema ically jus i ied
blending o he base and ine- uned models, wi h op imal
weigh s (α) empi ically de e mined ia g id sea ch o min-
imise alida ion e o . This blended p edic ion s a egy ensu es
bo h obus ness and apid domain adap a ion, as alida ed by
de ailed abla ion and sensi i i y analyses.
To ou knowledge, his is he i s TM amewo k o
explici ly in eg a e ew-sho lea ning, ans e lea ning, and
a Ma ko chain-based us es ima ion model o secu e IoT
de ice onboa ding. The con ibu ions o his pape can be
summa ised as ollows:
•P oblem o mula ion: we o mally de ine he challenge o
us sco e p edic ion unde sca ce onboa ding da a and
2
s ic la ency cons ain s, a gap no di ec ly add essed in
exis ing TM esea ch.
•No el hyb id amewo k: we in oduce he i s TM
amewo k ha uses s a is ical Ma ko -based modelling,
ans e lea ning, and ew-sho adap a ion, augmen ed
wi h a p incipled blending mechanism.
•Scalable, low-la ency e alua ion: we demons a e expe i-
men ally ha he amewo k achie es high accu acy, sub-
second onboa ding, and h oughpu exceeding 137 de-
ices/s, alida ing i s applicabili y o eal-wo ld consume
elec onics.
The emainde o he pape is o ganised as ollows. Sec-
ion II e iews ela ed wo k on TM in IoT. Sec ion III in o-
duces he algo i hmic design and logic. Sec ion IV p esen s
pe o mance e alua ion using eal-wo ld use cases. Sec ion V
discusses compa a i e ad an ages and secu i y bene i s. Sec-
ion VI concludes he pape and ou lines u u e esea ch
di ec ions.
II. LITERATURE REVIEW
This li e a u e e iew su eys key TM app oaches in IoT
ne wo ks, ocusing on me hodologies used o compu e de ice
us sco es based on ac o s such as Quali y o Se ice
(QoS), delays, secu i y, and eedback. By examining exis ing
solu ions, we unde sco e how he p oposed hyb id amewo k
ad ances he s a e o he a . Special emphasis is placed on
ou p io wo k le e aging a Ma ko chain-based me hod [1],
a c i ical ounda ional elemen in his esea ch.
A. Machine Lea ning App oaches in T us Managemen
ML echniques ha e been ex ensi ely employed in IoT
us managemen . Fo ins ance, Shayes eh e al. [3] apply
Bayesian lea ning and Demps e -Sha e Theo y o de i e us
sco es om en i y and da a eliabili y. Algho aili and Ras-
sam [4] p esen a Mul i-C i e ia Decision-Making app oach
in eg a ed wi h a Deep Long Sho -Te m Memo y (LSTM)
model o e alua e us h ough packe loss and h oughpu
me ics, weigh ed by Shannon’s en opy. Resea che s ha e
also p oposed a Fede a ed Lea ning-based TM amewo k o
Indus ial IoT, whe e us is calcula ed ia a weigh ed sum [5].
Fu he mo e, Wang e al. [6] employ unsupe ised lea ning
and clus e ing o di e en ia e us wo hy ehicles in IoV,
using Di ec T us and Indi ec T us . Meanwhile, Ma e al. [7]
le e age LSTM o p edic de ice beha iou by inco po a ing
QoS, a measu e o ne wo k pe o mance, delays, secu i y, and
eedback, and ime-dependen ea u es.
Despi e p o iding aluable insigh s, hese ML-based me h-
ods o en depend on la ge da ase s, limi ing hei e ec i eness
in da a-sca ce IoT se ings. In addi ion, hey end o ocus on a
na ow ange o me ics, educing hei b oade applicabili y.
B. Non-Machine Lea ning App oaches in T us Managemen
Non-ML TM me hods ely on s a is ical o s uc u al ech-
niques. Alam e al. [2] p opose a TM me hod ha calcula es
us sco es using QoS and coope a ion me ics. La i [8]
in oduces a con ex -dependen app oach o social IoT, in-
eg a ing de ice capabili ies and sa is ac ion. Bampa sikos
e al. [1] p esen a wo-dimensional Ma ko chain model,
combined wi h MADM and a piecewise unc ion, o p edic
us e olu ion based on cybe isk, packe loss, and de ice u il-
isa ion. Bampa sikos e al. [9] also explo e his app oach wi h
Hype ledge Fab ic. This o e s a s ong s a is ical ounda ion
bu s uggles wi h dynamic h ea s due o i s s a ic na u e.
O he wo ks, such as Liu e al. [10] (i.e., Hidden Ma ko
Model o VANETs) and Bai e al. [11] (i.e., game heo y in
supply chains), p o ide s abili y ye lack adap abili y o new
o e ol ing scena ios.
C. Beyond he S a e o he A
The e iewed li e a u e highligh s limi a ions in bo h ML-
and non-ML-based TM app oaches. ML me hods equi e
la ge aining da ase s, po en ially diminishing e iciency [7],
whe eas non-ML app oaches end o be less adap able o
eme ging h ea s [1]. Ou p io Ma ko -based TM me hod p o-
ides a obus baseline by employing s a is ical modelling o
key pa ame e s such as cybe isk and QoS, bu i s p ede ined
ules impede esponsi eness in dynamic IoT en i onmen s [1].
The p oposed hyb id amewo k add esses hese issues by
in eg a ing ans e lea ning and FSL. T ans e lea ning le e -
ages p e- ained models on a syn he ic Ma ko chain da ase ,
ensu ing e iciency and enabling c oss-domain adap abili y
(e.g., om consume o indus ial IoT). Meanwhile, FSL
enhances his amewo k by equi ing minimal da a (only i e
samples), acili a ing apid us assessmen (e.g., XGBoos
in e ence a 0.9425 ms) wi h an a e age p edic ion e o
o 2.29. Unlike esou ce-in ensi e ML me hods, he solu ion
a oids la ge da ase dependencies, and unlike pu ely s a is ical
models, i emains adap able o e ol ing h ea s. Mo eo e ,
he s a is ical g ound u h mi iga es poisoning a acks by
building upon he Ma ko ounda ion while esol ing i s
inhe en limi a ions. Consequen ly, he p oposed amewo k
signi ican ly ad ances consume IoT secu i y and scalabili y.
III. ALGORITHM DESIGN AND LOGIC
The p oposed amewo k o us sco e p edic ion in IoT
de ice onboa ding le e ages ans e lea ning and ew-sho
lea ning echniques o add ess he pe asi e challenge o
spa se da a in consume elec onics en i onmen s. Building
on a p e ious wo k which employed a 2D Ma ko chain
model o simula e us dynamics [1], his s udy in oduces a
no el hyb id ML and s a is ical app oach o p edic ing us
sco es in new IoT de ices. By enabling secu e, eal- ime on-
boa ding in esou ce-cons ained consume IoT con ex s, such
as wea able heal h de ices and au onomous ehicle sys ems,
his me hod ensu es he apid us assessmen necessa y o
ope a ional sa e y and secu i y. This sec ion elucida es he con-
cep ual design logic o he p oposed algo i hm and p o ides
a ma hema ically g ounded desc ip ion o i s implemen a ion,
including ea u e enginee ing, model aining, and p edic ion
blending.
A. Design Logic
The p oposed us sco e p edic ion algo i hm is d i en by
h ee p ima y challenges in consume IoT de ice onboa ding:
3
(i) he sca ci y o labeled da a o new de ices, (ii) he
need o apid us assessmen o acili a e eal- ime decision-
making, and (iii) he equi emen o esilience agains secu-
i y h ea s (e.g., da a poisoning, ze o-day a acks). To add ess
hese issues, he algo i hm employs a wo-phase s uc u e ha
combines ans e lea ning ( o exploi p io knowledge) wi h
FSL ( o adap e icien ly o limi ed onboa ding da a).
1) Few-Sho T ans e Lea ning: Concep and Ra ionale:
Few-sho ans e lea ning syn hesises he bene i s o ans e
lea ning and FSL o achie e e ec i e and lexible us sco e
p edic ion. T ans e lea ning elies on a p e- ained model,
ini ially ained on a la ge, he e ogeneous da ase . We u ilise
syn he ic da a om a 2D Ma ko chain model [1], p o iding a
obus ounda ion o us p edic ion. Few-sho lea ning hen
ine- unes his model using a small numbe o onboa ding
samples (e.g., 5 samples) o adap o he speci ic cha ac e is ics
o new IoT de ices. This hyb id app oach ensu es ha he
model can gene alise om subs an ial p io knowledge while
adap ing apidly o spa se, eal- ime da a, a c i ical equi e-
men o consume elec onics applica ions.
Se e al ac o s mo i a ed his app oach. Fi s , consume
IoT onboa ding aces se e e da a limi a ions om esou ce
cons ain s, p i acy egula ions, o he no el y o eme ging
de ices (e.g., nex -gene a ion heal h wea ables o au onomous
ehicle modules). Few-sho lea ning add esses his issue by
equi ing only a small numbe o samples, he eby mi iga -
ing da a collec ion o e head. Second, eal- ime ope a ion in
consume IoT en i onmen s demands swi us e alua ion.
T ans e lea ning accele a es his p ocess by p o iding a obus
s a ing poin by educing aining ime, while ew-sho ine-
uning ensu es apid adap a ion wi hou equi ing comp ehen-
si e e aining. Thi d, he app oach ein o ces secu i y and
obus ness by le e aging a la ge, di e se aining co pus. This
p e- aining p omo es esilience agains noise and a acks,
while ew-sho ine- uning ailo s he model o speci ic de ice
a ibu es, mi iga ing o e i ing isks on small da ase s.
This me hodology yields mul iple ad an ages. I o e s
e iciency by educing da a and compu a ional demands; a
pi o al ac o o esou ce-cons ained IoT de ices. I p o ides
adap abili y by enabling he model o accommoda e new
de ices using minimal samples; an impo an capabili y o
dynamic consume IoT ecosys ems. I also enhances accu acy
and s abili y by employing a blended p edic ion s a egy,
achie ing high pe o mance (e.g., a ine- uned Random Fo es
(RF) yields R2= 0.8891). Finally, i enhances secu i y ia
eliance on s a is ical g ound u h and blending p edic ions,
educing da a poisoning in ew-sho samples, a common
secu i y conce n in IoT con ex s.
B. Algo i hm Desc ip ion
The implemen a ion o he algo i hm o malises his design
logic in ma hema ical e ms. The amewo k p oceeds in
h ee s ages: ans e lea ning using a Ma ko -based s a is ical
g ound u h, ew-sho ine- uning wi h limi ed onboa ding
da a, and blended p edic ion o obus decision-making. Fig-
u e 1 illus a es he wo k low, while Algo i hm 1 p o ides
pseudocode. The ma hema ical ep esen a ion o each s age
is p esen ed below.
Load P io Da a
(Dp io )
Fea u e Enginee ing
(Gene a e X, Calc y, No malize)
T ans e Lea ning
(T ain Mbase, E alua e)
Few-Sho Fine-Tuning
(Collec Donboa d, Fine-Tune M ine)
P edic o New De ice
(P ep ocess Xnew, Blend)
End
O line ans e lea ning
On-de ice ew-sho ine- uning
Fig. 1. End- o-end wo k low o he p oposed hyb id us -sco e engine.
The uppe band (g ey) ep esen s o line ans e lea ning; he lowe band
(blue) depic s on-de ice ew-sho ine- uning. Thei con e gence shows how
each in e ence blends long- e m knowledge wi h eal- ime adap a ion. The
inal s age includes la ency e alua ion (Eq. 10), whe e onboa ding delay is
decomposed in o iden i y, communica ion, ledge , and in e ence componen s
o ensu e compliance wi h he 500 ms eal- ime a ge .
1) S a is ical G ound T u h and Fea u e Rep esen a ion:
Each IoT de ice iis desc ibed by a ea u e ec o
xi={Ci, Ri, Si, Pi, RSi, RepSi, OSi, P kgi},(1)
whe e Ci, Rideno e CPU and RAM u ilisa ion, Siis he
secu i y le el, Pi he packe loss a io, RSi, RepSi he isk
and epu a ion s a es, and (OSi, Pkgi) he in e ed ages o
he ope a ing sys em and ins alled packages. The p oblem is
o lea n a eg ession unc ion
:R8→[0,100], (xi)≈Ti,
whe e Tiis he us sco e ha quan i ies he de ice’s eliabili y
du ing onboa ding.
a) G ound T u h T us Sco e: The s a is ical us sco e
is gene a ed ia he piecewise unc ion:
T=
max(0,min(100,100 −20R−10C+
+4S+ 2OSi+ 1Pi−10P)),i RS ≥4
min(100,80 −20R−10C+
+4S+ 2OSi+ 1Pi−10P),i RS ≤2
max(20,min(100,70 −20R−10C+
+4S+ 2OSi+ 1Pi−10P)),o he wise
(2)
Equa ion 2 balances nega i e con ibu ions om u ilisa ion
and packe loss wi h posi i e con ibu ions om secu i y and
eshness. The coe icien s (e.g., 20, 10, 4) a e empi ically
de i ed om domain knowledge and p io analysis [1].
4
Algo i hm 1 IoT De ice Onboa ding T us Sco e P edic ion
1: Inpu : P e-gene a ed Ma ko chain da a (Dp io ), onboa ding
samples (Donboa d), de ice ea u es (Xnew)
2: Ou pu : P edic ed us sco e (T inal) o a new de ice
3: Load P io Da a: Load Dp io =
{RepS a e, RiskS a e, T us Sco es} om CSV iles.
4: Fea u e Enginee ing:
5: a. Gene a e syn he ic ea u es X=
{C, R, S, P, RS, RepS, OSi, Pi}.
6: b. Calcula e us sco es yusing Eq. 2.
7: c. No malize Xusing MinMaxScale o ob ain Xscaled.
8: T ans e Lea ning (Base T aining):
9: a. Spli Xscaled, y in o aining (80%) and es (20%) se s.
10: b. T ain base models Mbase ={RF, k-NN, XGBoos }by
minimising Eq. 3.
11: c. E alua e Mbase on es se , compu ing me ics (MSE, R2,
MAE, MAPE, RMSE, in e ence ime).
12: Few-Sho Fine-Tuning:
13: a. Collec Donboa d ={Xonboa d, yonboa d}wi h Ksamples.
14: b. Impu e missing ea u es in Xonboa d wi h aining se
means, no malize o Xonboa d,scaled.
15: c. Fine- une models M ine =
{RF ine, k-NN ine, XGBoos ine}by minimising Eq. 5.
16: d. Blend p edic ions wi h Eq. 6:
T inal =α·Mbase(X es ) + (1 −α)·M ine(X es ).
17: e. E alua e blended p edic ions on es se , compu ing me ics.
18: P edic ion o New De ice:
19: a. P ep ocess Xnew: impu e missing ea u es, no malize o
Xnew,scaled.
20: b. P edic Tbase =Mbase(Xnew,scaled),T ine =
M ine(Xnew,scaled).
21: c. Compu e T inal =α·Tbase + (1 −α)·T ine pe Eq. 6.
22: d. Re u n T inal.
23: Sys em La ency E alua ion:
24: Compu e end- o-end onboa ding la ency Lacco ding o Eq. 10,
decomposing iden i y (LSSI), messaging (LMQT T ), ledge
(LDLT ), and in e ence (Lin ) componen s.
Impo an ly, Eq. 2 also encodes domain knowledge by
penalising esou ce sa u a ion (Ci,Ri) and packe loss (Pi)
wi h ela i ely high weigh s, since hese di ec ly deg ade
de ice eliabili y. Posi i e con ibu ions a e d awn om se-
cu i y le el (Si) and so wa e eshness (OSi,Pkgi), which
imp o e esilience and educe ulne abili y. The bounding
unc ions (min,max) ensu e ha sco es emain wi hin [0,100],
p e en ing ex eme luc ua ions due o any single ea u e. This
design e lec s p ac ical IoT conside a ions (i.e., no de ice is
en i ely un us wo hy i i s co e beha iou is s able, and no
de ice is ully us wo hy i c i ical esou ces a e exhaus ed).
b) Ma ko Dynamics o S a es: As i was de ailed in [1],
he isk (RS) and epu a ion (RepS) s a es e ol e o e ime
acco ding o a wo-dimensional Ma ko chain model [1]. Le
p( )deno e he join dis ibu ion o (RS, RepS)a ime . The
s a e ansi ions ollow:
p( +1) =p( )P,
whe e Pis he join ansi ion ma ix. This cap u es empo al
co ela ions in de ice beha iou , e.g., a de ice wi h declining
epu a ion is mo e likely o emain in low- us s a es, while
a de ice wi h s ong his o ical epu a ion ends o pe sis in
highe - us s a es.
c) Fea u e Sampling: The emaining ea u es a e
s ochas ically sampled o e lec he e ogeneous de ice cha ac-
e is ics. CPU and RAM u ilisa ion a e d awn om unca ed
uni o m dis ibu ions,
Ci∼ U(0.2,0.8), Ri∼ U(0.1,0.9),
secu i y le el is sampled om a ca ego ical dis ibu ion biased
owa ds medium secu i y,
Si∼Ca ego ical({1,2,3,4,5}, pS), pS= (0.2,0.5,0.3),
packe loss ollows a Be a dis ibu ion skewed owa ds low-
loss egimes,
Pi∼Be a(α= 2, β = 8),
and OS/package ages a e sampled uni o mly as in e ed alues
o e lec eshness:
OSi, Pkgi∼ U{1,2,3,4,5}.
d) Syn he ic Da ase .: The labelled da ase is hen
D={(xi, Ti)}N
i=1, N = 1000,
wi h Ticompu ed om Eq. 2. This se up encodes domain
knowledge such as esou ce sa u a ion (Ci, Ri) and packe loss
(Pi) ecei e hea y penal ies, while highe secu i y le els (Si)
and upda ed so wa e (OSi, Pkgi) imp o e us . The esul ing
da ase p o ides bo h us ed and un us ed de ices in ealis-
ic p opo ions (app oxima ely 64% us ed, 36% un us ed),
ensu ing balanced aining o ans e lea ning and obus
e alua ion.
2) T ans e Lea ning Phase:The base models a e ained
on Dby minimising mean squa ed e o :
b= a g min
∈F
1
N
N
X
i=1
( (xi)−Ti)2,(3)
whe e b∈ { RF, kNN, XGB}. RF p o ides obus ness, k-
NN simplici y, and XGBoos high in e ence e iciency. These
models encode ans e able p io s.
The op imisa ion in Eq. 3 p oduces base models b ha
se e as p io s: hey lea n a gene ic mapping om he ea u e
space o us sco es ac oss a wide ange o de ices. T aining
on syn he ic da a allows hese models o cap u e gene al us
dynamics wi hou elying on sca ce eal-wo ld onboa ding
samples. Random Fo es p o ides obus ness agains noise
and non-linea ea u e in e ac ions, k-NN o e s sensi i i y o
local a ia ions in ea u e space and XGBoos con ibu es high
accu acy wi h millisecond-le el in e ence ime. Impo an ly,
each model is ained and e alua ed independen ly, no as an
ensemble, o allow sys ema ic compa ison o s eng hs and
weaknesses unde spa se onboa ding condi ions.
3) Few-Sho Fine-Tuning Phase:Fo onboa ding, only K
samples a e a ailable pe de ice:
D ew ={(xj, Tj)}K
j=1, K ∈ {5,10,20}.(4)
The ine- uned model is ained wi h educed complexi y o
a oid o e i ing:
= a g min
∈F
1
K
K
X
j=1
( (xj)−Tj)2.(5)
5
The da ase D ew con ains only Ksamples pe de ice,
wi h K∈ {5,10,20}chosen o e lec ealis ic onboa ding
condi ions whe e only a hand ul o in e ac ions a e a ailable
be o e he de ice mus be us ed o ejec ed. The op imisa ion
in Eq. 5 adap s he base model o hese limi ed samples. To
educe a iance unde such small K, he ine- uned models a e
delibe a ely es ic ed in complexi y (e.g., ewe ees in RF,
shallowe dep h in XGBoos ). Unlike ypical N-way K-sho
classi ica ion se ings, ou o mula ion is eg ession hence,
each sample p o ides a con inuous us sco e label and K
indica es he numbe o a ailable eg ession pai s. This ensu es
he model adap s wi hou o e i ing o a hand ul o poin s.
4) Blended P edic ion:To mi iga e he isk o o e i ing in
ew-sho scena ios while ensu ing adap abili y, he amewo k
employs a blended p edic ion s a egy ha combines he base
and ine- uned models. The inal us sco e o a de ice is
gi en by:
T inal(x) = α b(x) + (1 −α) (x), α ∈[0,1],(6)
whe e bis he base p edic o ( ained on he la ge syn he ic
da ase ) and is he ine- uned p edic o (adap ed o K
onboa ding samples).
Equa ion 6 o malises he ade-o be ween obus ness and
adap abili y: he base model b ypically yields low- a iance
bu some imes biased p edic ions o new de ices, while he
ine- uned model adap s o de ice-speci ic beha iou bu
exhibi s highe a iance unde small K. Blending balances
hese e ec s, implemen ing a bias– a iance comp omise.
The ma hema ically op imal solu ion weigh α ∗ can be
ob ained by minimising he alida ion mean squa ed e o
(MSE). Le eb= b(x)−Tand e = (x)−Tdeno e
he esiduals o he base and ine- uned models wi h espec
o he g ound u h T. The expec ed blended e o is:
L(α) = E(αeb+ (1 −α)e )2.(7)
Expanding and di e en ia ing wi h espec o αgi es:
α ∗ =S −Sb
Sb+S −2Sb
,(8)
whe e Sb=E[e2
b],S =E[e2
], and Sb =E[ebe ]. In he
special case o unco ela ed e o s (Sb ≈0), his educes o:
α ∗ ≈S
Sb+S
,(9)
which in ui i ely assigns mo e weigh o he model wi h
smalle e o .
In p ac ice, we compu e α ∗ on a alida ion spli o he ew-
sho onboa ding da a. Fo nume ical s abili y, α ∗ is clipped o
[0,1] and egula ised when he denomina o o Eq. 8 is small.
When Kis e y small (e.g., 5), we also e alua e α ia g id
sea ch o con i m consis ency wi h he analy ical solu ion.
Empi ically, α ∗ ≈0.8ac oss mos expe imen s, e lec ing
he dominance o he base model unde spa se condi ions
while s ill le e aging he co ec i e signal om he ine-
uned model. This con ex combina ion ensu es s able us
es ima ion wi hou sac i icing adap abili y, yielding obus
p edic ions o secu e IoT de ice onboa ding.
5) La ency Model:End- o-end onboa ding la ency is de-
composed as:
L=LSSI +LMQT T +LDLT +Lin
whe e LSSI ≈50 ms, LMQT T ≈20 ms, LDLT ≈200 ms,
and Lin <7ms. Fo Ksequen ial samples:
L o al(K)≈K·(LSSI +LMQT T +LDLT ) + Lin (10)
Ba ching op imisa ions educe LDLT pe sample o mee
he 500 ms a ge .
Eq. 10 decomposes end- o-end onboa ding la ency in o ou
measu able componen s: iden i y gene a ion (LSSI ), commu-
nica ion (LMQT T ), dis ibu ed ledge logging (LDLT ), and
in e ence ime (Lin ). Among hese, in e ence la ency is
consis en ly below 7 ms and he e o e negligible compa ed
o he ∼200 ms pe - ansac ion DLT cos . The decomposi ion
makes he bo leneck explici and mo i a es ba ching: agg e-
ga ing Bonboa ding e en s pe ledge ansac ion educes he
amo ised LDLT o app oxima ely 200/B ms pe de ice, a
c i ical op imisa ion o mee he 500 ms eal- ime equi emen .
This sys em-le el model hus alida es he easibili y o he
amewo k unde deploymen condi ions.
Algo i hm 1 p esen s he pseudocode implemen a ion o
hese s ages, di ec ly co esponding o he ma hema ical ame-
wo k de ined by Equa ions 2–10.
IV. USE CASE AND PERFORMANCE EVALUATION
This sec ion p esen s a p ac ical use case o he p oposed
us sco e p edic ion algo i hm. I de ails he onboa ding
scena io o IoT de ices in a consume elec onics con ex . An
expe imen al se up e alua es he implemen a ion o he algo-
i hm. The pe o mance assessmen , suppo ed by quan i a i e
me ics and isual aids, illus a es he algo i hm’s e ec i eness
and sheds ligh on i s po en ial o eal-wo ld applica ions,
including sma heal h de ices and au onomous ehicles.
A. Use Case and Expe imen al Se up
The ocus o his use case is on he secu e onboa ding and
egis a ion o a heal h- ela ed sma de ice (e.g., a i ness
acke ) wi hin an IoT ne wo k, whe e a apid and eliable
e i ica ion o de ice c eden ials is essen ial o sa e y and
ope a ional us . The onboa ding sequence, as illus a ed in
Figu e 2, s a s wi h he IoT de ice gene a ing a Decen al-
ized Iden i ie (DID) managed by an Sel -So e eign Iden i y
(SSI) managemen sys em, adhe ing o he W3C DID s an-
da d [12]. The SSI clien hen p o isions c yp og aphic keys,
o e ing h ee oo s o us : Physical Unclonable Func ion
(PUF), T us ed Execu ion En i onmen (TEE), and Hype -
ledge A ies [13] as a allback, p o iding lexibili y ac oss
di e se ha dwa e pla o ms.
Upon success ul DID c ea ion and publica ion o he co e-
sponding DID Documen , he SSI Manage acknowledges eg-
is a ion. The IoT de ice hen eques s issuance o a Ve i iable
C eden ial (VC) embedding i s unique c yp og aphic oo p in .
This c eden ial is signed and he de ice is pai ed wi h i s
owne , who au hen ica es he DID using hei own keys.
Subsequen ly, he IoT de ice ansmi s us da a and i s signed
6
Fig. 2. Secu e onboa ding message low. The diag am highligh s (a) c yp og aphic binding o de ice iden i y ia DID/SSI and (b) ampe -e iden logging o
he compu ed us sco e o he DLT.
DID o he T us Managemen Se e (TMS) using MQTT. The
TMS compu es an ini ial us sco e, hen eco ds his sco e
and DID o he Dis ibu ed Ledge Technology (DLT) ne wo k
ia a gRPC in e ace. The DLT chaincode secu ely logs he
onboa ding e en , wi h acknowledgemen s p opaga ed back o
he TMS, ensu ing he onboa ding comple es wi hin he a ge
la ency budge .
The ull onboa ding message low, including c yp og aphic
binding ia SSI and ampe -e iden logging o he DLT, is
shown in Figu e 2.
To e alua e he p oposed us managemen amewo k, we
simula e a es bed o 50 he e ogeneous IoT de ices (including
wea ables and ehicle senso s) using an In el i7 CPU, 16 GB
RAM, and Ubun u 20.04. Syn he ic da ase s a e gene a ed
using he 2D Ma ko chain-based s a is ical us modelling
app oach om [1], whe e de ice us e olu ion is a unc ion o
bo h epu a ion and isk s a es. Speci ically, he da ase consis s
o 1000 syn he ic de ice samples, each comp ising ea u es
such as CPU u ilisa ion (C), RAM u ilisa ion (R), secu i y
le el (S), packe loss (P), isk s a e (RS), epu a ion s a e
(RepS), and in e ed OS/package ages (OSi,Pi). T us sco es
(T) a e compu ed o each de ice acco ding o he piecewise
s a is ical unc ion de ined in Eq. 2.
The use o a Ma ko chain and MADM-d i en syn he ic
simula ion [1] enables he gene a ion o ea u e ec o s and
us sco es spanning a wide spec um o de ice beha iou s
and us wo hiness. By explici ly sampling om a ange
o isk, epu a ion, and pe o mance s a es, he simula ion
p o ides a da ase ha is bo h su icien ly la ge and balanced
o model aining. Consequen ly, no addi ional o e sampling,
unde sampling, o pos -p ocessing echniques a e needed o
mi iga e da a imbalance. When bina ising he us sco e (e.g.,
us ed i T≥70), he da ase includes app oxima ely 64%
us ed and 36% un us ed de ices, ensu ing ai ep esen a ion
o bo h classes.
In his o mula ion, esou ce usage (R,C,P) nega i ely
impac s us , while secu i y and de ice/so wa e eshness (S,
OSi,Pi) con ibu e posi i ely. The isk s a e h esholds ensu e
ha de ices wi h high isk a e penalised, while low- isk s a es
yield highe base us .
Fo each expe imen , 80% o he syn he ic da ase is used o
aining he base models (Random Fo es , k-NN, XGBoos ),
and 20% o e alua ion. Realis ic onboa ding scena ios a e
emula ed using K={5,10,20} ew-sho samples pe de ice.
All ea u e alues a e sampled using Py hon 3.10, NumPy,
and pandas, wi h unca ed no mal o be a dis ibu ions o
CPU, RAM, and packe loss, and ca ego ical assignmen s
o secu i y, isk, and epu a ion s a es based on empi ically
obse ed de ice beha io s [1]. The g ound u h us sco e o
each de ice is gene a ed using he s a is ical unc ion de i ed
om he Ma ko chain and MADM-based model. These
alues se e as he e e ence labels o supe ised aining and
e alua ion o all machine lea ning models conside ed in his
s udy. Model p edic ions a e assessed agains hese Ma ko -
based us sco es o de e mine accu acy and e o me ics
h oughou ou expe imen s.
The o e all pipeline, including model aining, ine- uning,
and pe o mance e alua ion, is implemen ed in Py hon 3.9
wi h sciki -lea n 1.0.2, XGBoos 1.5.0, and Hype ledge Fab ic
2.2. The us sco es gene a ed se e as g ound u h labels o
all supe ised lea ning expe imen s.
In line wi h ew-sho lea ning e alua ion p ac ices, we adop
a K-sho eg ession app oach in ou expe imen al design.
He e, Kdeno es he numbe o labelled onboa ding samples
pe de ice used o model ine- uning and e alua ion, wi h
K={5,10,20}. Unlike adi ional classi ica ion-based ew-
7
TABLE I
PERFORMANCE METRICS FOR BASE AND FINE-TUNED MODELS
(FINE-TUNED: 5 ONBOARDING SAMPLES)
Model MSE R² MAE In e ence Time (ms)
RF (Base) 0.9904 0.9917 0.6796 2.9747
k-NN (Base) 18.3057 0.8460 2.6887 0.3256
XGBoos (Base) 0.4436 0.9963 0.4782 0.4660
RF (Fine-Tuned) 13.1759 0.8891 3.6297 3.4938
k-NN (Fine-Tuned) 33.0026 0.7223 5.7440 3.4716
XGBoos (Fine-Tuned) 20.5551 0.8270 4.5347 0.9425
sho lea ning (N-way K-sho ), ou ask is o mula ed as a
eg ession p oblem o con inuous us sco e p edic ion, so
he concep o “N-way” (numbe o classes) does no apply.
Ins ead, we analyse he model’s abili y o apidly adap and
gene alise om a small se o onboa ding da a poin s pe
de ice, closely mi o ing eal-wo ld IoT onboa ding scena ios
whe e only limi ed ea u e-label pai s a e a ailable a un ime.
B. Pe o mance Analysis
The pe o mance o he us sco e p edic ion algo i hm is
e alua ed using quan i a i e me ics and isual aids, ocusing
on accu acy, s abili y, and in e ence ime ac oss he simula ed
IoT ne wo k. The analysis compa es RF, k-Nea es Neighbo s
(k-NN), and XGBoos models in bo h base and ine- uned con-
igu a ions, le e aging he ew-sho ans e lea ning app oach.
The p ima y me ics include MSE, R-squa ed (R2), MAE,
Mean Absolu e Pe cen age E o (MAPE), Roo Mean
Squa ed E o (RMSE), and in e ence ime. These me ics
a e compu ed on he es se o base models and he blended
p edic ions o ine- uned models. Table I summa ises he
esul s o he base models and o he ine- uned models using
i e onboa ding samples, a e aged o e 10 uns o accoun o
a iabili y.
Table I shows ha XGBoos achie es he highes R2
(0.9963) and lowes MSE (0.4436) among base models,
indica ing excellen p edic i e accu acy on he p e- ained
da ase . A e ine- uning wi h only i e onboa ding samples,
all models expe ience an inc ease in MSE and a dec ease in
R2, consis en wi h he expec ed isk o o e i ing in ex eme
ew-sho egimes. Fo example, RF main ains a s ong R2o
0.8891 bu MSE inc eases o 13.18, while k-NN exhibi s he
g ea es sensi i i y o limi ed da a, wi h R2d opping o 0.72
and MSE ising o 33.00. These esul s e lec he inhe en
ade-o in ew-sho lea ning: minimal da a can limi p ecision
bu enable apid adap a ion and pe sonalisa ion.
To u he analyse he impac o onboa ding sample size,
we sys ema ically a ied he numbe o samples used o ine-
uning. Table II epo s he pe o mance o he ine- uned
models as he onboa ding sample size inc eases om 5 o
10 and 20. As shown, inc easing he numbe o onboa ding
samples subs an ially imp o es bo h MSE and R2 o RF and
XGBoos : o RF, MSE d ops om 13.18 (5 samples) o 5.07
(20 samples), while R2inc eases om 0.89 o 0.96. Simila ly,
XGBoos eaches an MSE o 4.45 and R2o 0.96 wi h 20
samples. k-NN emains less obus in he ew-sho egime,
hough i also bene i s om addi ional da a.
Random Fo es k-NN XGBoos
−20
−15
−10
−5
0
5
10
15
20
Residuals (P edic ion - G ound T u h)
Base Model Residuals
Random Fo es k-NN XGBoos
−20
−15
−10
−5
0
5
10
15
Residuals (P edic ion - G ound T u h)
Fine-Tuned Model Residuals
Fig. 3. Residual e o dis ibu ion o base s. ine- uned models (5 onboa d-
ing samples). Random Fo es main ains s able e o dis ibu ion wi h a igh
IQR, while k-NN exhibi s wide a iabili y and mo e ou lie s, e lec ing highe
sensi i i y o spa se da a. XGBoos shows mode a e dispe sion, balancing
accu acy and obus ness.
These indings demons a e ha , while o e i ing can occu
when e y ew samples a e a ailable o ine- uning, he hyb id
ew-sho ans e lea ning app oach becomes obus as mo e
onboa ding da a is p o ided. Fo eal-wo ld deploymen s, we
ecommend using a leas 10 onboa ding samples o ensu e
eliable us sco e calib a ion and gene alisa ion. This end
is consis en wi h s a is ical lea ning heo y, which s a es ha
inc eased sample size leads o educed empi ical isk and
imp o ed gene alisa ion pe o mance.
Fig. 4 isually compa es he R2and MSE alues o each
model in bo h base and ine- uned (5-sample) con igu a ions.
The esul s con i m ha XGBoos achie es he highes ac-
cu acy in he base scena io, while all models expe ience a
educ ion in R2and an inc ease in MSE when ine- uned wi h
only i e onboa ding samples. This illus a es he ade-o
be ween adap abili y and gene alisa ion in ex eme ew-sho
egimes, as also e lec ed in Table I.
Figu e 3 p esen s he box plo s o esiduals o bo h base
and ine- uned models (5 onboa ding samples). Fo he base
models, Random Fo es exhibi s a igh in e qua ile ange
(IQR) o app oxima ely ±1.0, wi h mos esiduals close o
ze o and ew ou lie s, indica ing s able and accu a e p edic-
ions. A e ine- uning, he IQR o Random Fo es widens
o oughly ±3.0, consis en wi h he obse ed inc ease in
p edic ion a iabili y due o a limi ed ew-sho da a. The k-NN
model exhibi s a b oade base IQR (app oxima ely ±4.5), wi h
i s ine- uned IQR expanding o a ound ±6.0, accompanied by
se e al no iceable ou lie s, e lec ing i s highe sensi i i y o
small sample sizes and esul ing in g ea e p edic ion e o
a iance. XGBoos main ains a na ow base IQR (≈ ±0.7),
while i s ine- uned IQR inc eases o abou ±3.5, illus a ing
ha i emains obus in he base con igu a ion bu , like he
o he models, exhibi s mo e dispe sed esiduals a e ew-
sho adap a ion. O e all, hese dis ibu ions highligh Random
Fo es ’s obus ness and XGBoos ’s high p ecision unde base
aining.
The p edic i e capabili y o he models wi h i e onboa ding
samples pe de ice is illus a ed in Figu es 5, 6, and 7. Fo
Random Fo es , he p edic ed us sco es o a ep esen a i e
se o onboa ding de ices closely ack he g ound u h,
wi h p edic ions (85.89, 77.18, 86.96, 99.99, 99.31) aligning
well wi h calcula ed ue alues (85.25, 77.27, 85.92, 100.00,
100.00), and an a e age absolu e e o o 1.74. The k-NN
8
TABLE II
PERFORMANCE OF FINE-TUNED MODELS VS. NUMBER OF ONBOARDING SAMPLES
Samples Model MSE R2MAE MAPE (%) In e ence Time (ms) S d. Residuals
5
Random Fo es 13.18 0.889 3.14 3.48 3.33 3.12
k-NN 33.00 0.722 4.74 5.47 3.47 5.53
XGBoos 20.56 0.827 3.79 4.13 0.96 3.80
10
Random Fo es 6.29 0.947 2.07 2.45 3.49 2.50
k-NN 27.83 0.766 4.11 4.87 3.33 5.25
XGBoos 4.54 0.962 1.54 1.91 0.96 2.10
20
Random Fo es 5.07 0.957 1.54 1.89 3.70 2.23
k-NN 26.70 0.775 3.49 4.36 0.64 5.07
XGBoos 4.45 0.963 1.49 1.85 0.88 2.06
Random Fo es k-NN XGBoos
Models
0.0
0.2
0.4
0.6
0.8
1.0
R² Sco e
R² Compa ison: Base s Fine-Tuned Models
Base
Fine-Tuned
Random Fo es k-NN XGBoos
Models
100
101
MSE
MSE Compa ison: Base s Fine-Tuned Models
Base
Fine-Tuned
Fig. 4. Compa ison o R2and MSE o base and ine- uned models ( ine-
uned wi h 5 onboa ding samples).
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
Sample Index
50
60
70
80
90
100
T us Sco e
P edic ions s G ound T u h (Random Fo es )
G ound T u h
P edic ions
Fig. 5. Random-Fo es p edic ions on i e unseen onboa ding samples. The
e o ba s (±σ) emain wi hin ±3 us -sco e uni s, illus a ing he 2.29
a e age absolu e e o epo ed in Table I
p edic ions (82.66, 85.66, 77.48, 94.88, 88.67) show la ge
de ia ions om he co esponding g ound u h (81.37, 87.51,
74.82, 100.00, 71.37), wi h an a e age absolu e e o o
5.20, indica ing ins abili y unde ew-sho ine- uning. Fo
XGBoos , p edic ed alues (99.72, 83.63, 99.82, 99.32, 84.01)
also ollow he g ound u h (100.00, 85.36, 100.00, 100.00,
85.77) closely, yielding an a e age absolu e e o o 2.32.
These esul s alida e he algo i hm’s adap abili y in he ew-
sho se ing, wi h Random Fo es and XGBoos demons a ing
s ong gene alisa ion e en wi h minimal onboa ding da a,
while k-NN emains mo e sensi i e o he sample size.
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
Sample Index
50
60
70
80
90
100
T us Sco e
P edic ions s G ound T u h (k-NN)
G ound T u h
P edic ions
Fig. 6. k-NN p edic ions exhibi la ge de ia ions and wo clea ou -
lie s, consis en wi h i s highe pos - uning MSE (33.0). This con i ms ha
dis ance-based me hods s uggle wi h he high- a iance, low-sample egime.
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
Sample Index
50
60
70
80
90
100
T us Sco e
P edic ions s G ound T u h (XGBoos )
G ound T u h
P edic ions
Fig. 7. XGBoos balances accu acy and smoo hness. Al hough no as p ecise
as RF on iny samples, XGBoos a oids he k-NN ou lie s while main aining
less han 1ms in e ence, suppo ing i s ole as a ligh weigh allback.
C. Scalabili y and Real-Time Pe o mance
The us managemen algo i hm demons a es s ong scal-
abili y ac oss a 50-de ice ne wo k. Wi h he adop ion o
ba ch DLT onboa ding, he TMS now p ocesses eques s a
an a e age h oughpu o 137.7 de ices pe second ac oss he
Random Fo es , k-NN, and XGBoos models, as measu ed on
9
TABLE III
SCALABILITY AND PERFORMANCE METRICS ACROSS MODELS (BATCH
DLT ONBOARDING)
Model Th oughpu (de ices/s) To al La ency o 5 De ices (ms)
Random Fo es 106.77 745.41
k-NN 198.42 708.66
XGBoos 107.91 722.44
A e age 137.70 725.50
Random Fo es k-NN XGBoos
Models
0
25
50
75
100
125
150
175
200
Th oughpu (de ices/s)
TMS Th oughpu Ac oss Models
A e age (137.70 de ices/s)
Model Th oughpu
Fig. 8. Ba cha compa ing TMS h oughpu ac oss Random Fo es , k-NN,
and XGBoos models (ba ch DLT onboa ding), wi h he a e age h oughpu
(137.70 de ices/s) indica ed.
he simula ed es bed. Table III and Figu e 8 highligh he
model-wise a iabili y in h oughpu , wi h k-NN achie ing he
highes (198.4 de ices/s) and Random Fo es and XGBoos
p o iding consis en pe o mance abo e 100 de ices/s.
The onboa ding wo k low in ol es simula ed SSI ope a ions
(mean 53–63 ms pe de ice), MQTT ansmission (mean
20–35 ms pe de ice), and now a single ba ch DLT ansac ion
o each g oup o 5 de ices (app oxima ely 300ms o al). This
ba ching educes he cumula i e DLT la ency om o e 1000
ms (in he p e ious pe - ansac ion app oach) o well wi hin
he 500 ms eal- ime a ge se o indus ial IoT onboa ding.
The o al la ency o onboa ding 5 de ices now a e ages unde
750 ms o all models, wi h DLT con ibu ing only a ac ion o
he o al delay. This con i ms ha he p oposed sys em mee s
eal- ime equi emen s in p ac ical, scalable IIoT deploymen s.
Figu e 8 p esen s he h oughpu compa ison, showing ha
k-NN is as es in aw in e ence speed, bu all models deli e
eal- ime onboa ding pe o mance unde he ba ch-op imised
pipeline. The impac o de ice coun on h oughpu is u he
isualised in Figu e 9, con i ming s able scaling cha ac e is ics
o each model.
V. DISCUSSIONS
A. P oposed F amewo k Bene i s
The p oposed TM amewo k le e ages ans e lea ning and
FSL o add ess IoT de ice onboa ding challenges in consume
elec onics, o e ing no able bene i s in e iciency and adap -
abili y. T ans e lea ning u ilises p io knowledge om a la ge
syn he ic da ase (1000 de ices) gene a ed by a 2D Ma ko
chain model [1], enabling base models (RF, k-NN, XGBoos )
10 15 20 25 30 35 40 45 50
Numbe o De ices
80
100
120
140
160
180
200
Th oughpu (de ices/s)
Th oughpu s De ice Coun
Random Fo es
k-NN
XGBoos
Fig. 9. TMS h oughpu as a unc ion o onboa ding de ice coun o each
model.
o lea n gene al us pa e ns wi hou ex ensi e eal- ime
da a collec ion. This app oach educes compu a ional o e head
and aining ime, which is c i ical o esou ce-cons ained
IoT en i onmen s whe e da a sca ci y and p i acy conce ns
a e p e alen . Fo ins ance, p e- ained models achie e high
accu acy (e.g., XGBoos base R2o 0.9963), ensu ing apid
us assessmen wi hin eal- ime cons ain s.
Each sample in he syn he ic da ase consis s o eigh ea-
u es—CPU u ilisa ion, RAM u ilisa ion, secu i y le el, packe
loss, isk s a e, epu a ion s a e, in e ed OS age, and in e ed
package age—and is labeled by compu ing he us sco e ia
he piecewise unc ion in Eq. (2).
FSL enhances adap abili y by ine- uning hese models wi h
only a ew onboa ding samples (e.g., 5–20), add essing da a
sca ci y in dynamic IoT se ings. This enables he amewo k
o quickly adap o new de ices, such as sma heal h wea -
ables, wi h ine- uned RF R2 eaching 0.9574 (20 samples) and
p edic ion e o (MAE) as low as 1.54. The blended p edic-
ion app oach—whe e he op imal base/ ine- uned weigh s a e
empi ically de e mined o each model—balances obus ness
and adap abili y, minimising o e i ing isks as e idenced
by he alida ion e o cu e. Addi ionally, he amewo k’s
low in e ence imes (e.g., XGBoos a 0.9999 ms) and scala-
bili y (a e age h oughpu 126.9 de ices/s, peaking a 174.6
de ices/s o k-NN) make i sui able o la ge-scale, ime-
sensi i e applica ions. T ans e lea ning also enables c oss-
domain applicabili y, allowing he amewo k o adap o
con ex s like indus ial IoT o ehicula ne wo ks, enhancing
i s e sa ili y.
To quan i a i ely compa e he p oposed amewo k wi h ex-
is ing TM app oaches, Table IV summa ises key pe o mance
me ics ac oss ou models and s a e-o - he-a me hods. The
able highligh s he amewo k’s compe i i e in e ence imes
and base model accu acy, al hough ine- uned MSE alues
emain highe , e lec ing he inhe en ade-o in apid ew-
sho adap a ion.
The compa ison e eals ha ou amewo k excels in in-
e ence ime (e.g., unde 1 ms o XGBoos ) compa ed o
LSTM and Bi-LSTM, which equi e o de s o magni ude
