Jou nal o Theo e ical and Applied In o ma ion Technology
31s May 2025. Vol.103. No.10
© Li le Lion Scien i ic
ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195
4210
DIFT-VAR: A DYNAMIC MULTI-LAYER FRAMEWORK
FOR DEVICE FINGERPRINTING OF IDENTICAL DEVICES
MANOJ KUMAR VEMULA1, KAILA SHAHU CHATRAPATI 2
1Resea ch Schola , Depa men o CSE, JNTUH, Hyde abad, India
2P o esso , Depa men o CSE, JNTUH, Hyde abad, India
E-mail: 1manojkuma [email protected], 2shahujn[email p o ec ed]om
ABSTRACT
De ice inge p in ing is a powe ul echnique o iden i ying de ices in an IoT en i onmen , o e ing
mul iple ad an ages such as enhanced secu i y h ough de ice au hen ica ion, imp o ed ne wo k
managemen by moni o ing de ice beha io s, and anomaly de ec ion o iden i ying unau ho ized o
comp omised de ices. The majo i y o ecen inge p in ing schemes conside a he e ogeneous de ice
en i onmen and use di e en machine lea ning echniques o iden i y de ices using ne wo k a ic, signal-
le el in o ma ion, adio equency cha ac e is ics, e c. Howe e , inge p in ing de ices o he same make
and model is a signi ican challenge in mode n IoT en i onmen s, whe e many de ices o en sha e iden ical
ha dwa e and so wa e con igu a ions. Exis ing echniques canno eliably di e en ia e iden ical de ices as
hey lack su icien da a. This pape p oposes a no el app oach o De ice Iden i ica ion and Finge p in ing
wi h Time-Va ian Adap i e Recogni ion (DIFT-VAR) based on mul i-laye , ime- a ying ea u e
ex ac ion. We cons uc dynamic inge p in s ha uniquely iden i y each de ice by moni o ing and using
ea u es such as p obe eques beha io , clock skew, anspo laye cha ac e is ics, and adio signal me ics
o e ime. We u ilize machine lea ning algo i hms such as Random Fo es s o classi y de ices based on
hese dynamic inge p in s. We u he p opose he use o dynamic ime wa ping (DTW) o ea u e
alignmen and classi ica ion. Expe imen al esul s demons a e he e icacy o ou app oach in
dis inguishing iden ical de ices wi h an accu acy o o e 97% using s anda d machine lea ning me ics.
Keywo ds: De ice inge p in ing, IoT Secu i y, Dynamic ime wa ping (DTW), Time- a ian ea u e
ex ac ion, Machine lea ning o IoT secu i y.
1. INTRODUCTION
An IoT (In e ne o Things) de ice is a physical
objec embedded wi h senso s connec ed o he
in e ne . This physical de ice di e s om any
ypical compu e whose p ima y unc ionali y is no
compu ing. IoT de ices connec ed o he in e ne
ha e many ad an ages, including con enience,
com o , sa e y, secu i y, eliabili y, e c. The IoT
ma ke is a apidly expanding indus y ha co e s a
wide ange o applica ions ac oss di e en business
e icals, om e ail o heal h, anspo ,
manu ac u ing, en e ainmen e c. Ensu ing he
secu i y o IoT de ices is essen ial o hei
widesp ead adop ion. Due o de ice cons ain s,
adi ional secu i y solu ions o con en ional
compu ing pa adigms canno be di ec ly applied o
IoT. Addi ional challenges include scalabili y,
ope a ing en i onmen s, di e se de ice
a chi ec u es, pla o ms, and p o ocols wi hin he
IoT ecosys em. No ably, S a is a p ojec s ha by he
end o 2024, o e 50 billion IoT de ices will be
connec ed o he in e ne . P o ocols, a chi ec u es,
and pla o ms employed by hese de ices a y
g ea ly, and wi h a sho p oduc de elopmen li e-
cycle, he numbe o endo s manu ac u ing hese
p oduc s is nume ous. These challenges and IoT's
open ope a ing en i onmen pose se e al secu i y
challenges. Au hen ica ing a de ice be o e
p o iding i wi h ne wo k access is he i s line o
de ense o p o ec agains secu i y a acks in he
cybe wo ld. E en hough many c yp og aphic
schemes ailo ed o he IoT en i onmen ha e been
de eloped, hese schemes (excep hose ha use
PKI ce i ica es) a e ulne able o node o ge y o
impe sona ion a acks, whe e he iden i y o he
o he legi ima e de ices is employed [1], wi h
sec e -keys (as secu i y c eden ials) being he mos
popula way o au hen ica ing a de ice, weak
passwo ds ollowed by unpa ched de ices wi h
epo ed secu i y ulne abili ies subjec IoT de ices
o comp omise. Once comp omised, hese de ices
Jou nal o Theo e ical and Applied In o ma ion Technology
31s May 2025. Vol.103. No.10
© Li le Lion Scien i ic
ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195
4211
can be exploi ed o leak con iden ial in o ma ion,
used as launch pads o launch la ge-scale a acks
( o example, he ecen Mi ai Bo ne ), a ge
c i ical in as uc u e, e c. The isk o a acks is
signi ican ly heigh ened as de ices a e suscep ible
o hacking, comp omising, and e e se enginee ing,
as well as inadequa e secu i y managemen
mechanisms in wi eless ne wo k sys ems. Thus,
mul iple app oaches a e equi ed o secu e IoT
de ices, anging om design conside a ions o
moni o ing o po en ial a acks and implemen ing
e ec i e mi iga ion s a egies.
In a ne wo k, de ices a e usually iden i ied using
IP add esses, MAC add esses, de ice se ial
numbe s, e c. The limi a ion o hese iden i ie s is
ha hey a e suscep ible o spoo ing. The abili y o
spoo add esses o mul iple legi ima e de ices
allows he a acke o launch mo e sophis ica ed
a acks wi hou being de ec ed. Recen ly, de ice
inge p in ing has eme ged as an al e na i e
app oach o iden i y de ices, in which unique
cha ac e is ics o a de ice a e used o gene a e
de ice signa u es and used o iden i y de ices [6].
The p emise is o ex ac de ice ea u es o o
de i e pa e ns h ough communica ions wi h he
de ices. Di e en ne wo k laye s can con ibu e o
his ea u e se , which can help in de eloping a
inge p in o he de ice. De ice inge p in ing can
be used no only o iden i y de ices, bu also o
au hen ica e hem, p o ided he inge p in is
unique [2]. Fu he mo e, a ew wo ks [3], [4], [5]
ha e also employed inge p in ing o de ec hidden
ea esd oppe s like a hidden came a, a ne wo k
ea esd oppe , o any such passi e wi eless de ices
o p o ec p i acy. Also, iden i ying a de ice ype
helps in moni o ing he beha io o a pa icula
de ice, he eby di e en ia ing he anomalous
beha io om no mal beha io . I also helps in
isola ing malicious nodes by main aining an asse
lis . Unique inge p in s employed o
au hen ica ion p e en he de ice iden i ies om
being o ged.
The majo s eps in ol ed in he inge p in ing
p ocess a e: (1) iden i ying he ele an ea u es, (2)
ex ac ing and modelling hem, and (3) iden i ying
he de ice. Employing sui able machine lea ning
algo i hms o achie e he abo e has made de ice
inge p in ing an e ec i e echnique o add ess he
unique secu i y challenges posed by IoT ne wo ks.
The exis ing inge p in ing echniques can be
ca ego ized based on he ea u es employed o
iden i ying de ices: o example, ne wo k-based
inge p in ing echniques ely on ne wo k a ic
pa e ns [4], Wi-Fi based echniques use medium
access con ol (MAC) sub-laye in o ma ion [5],
clock-skew based me hods use sligh a ia ions in
clock o each de ice[8], me hods based on
elec omagne ic emana ions (EME) use unique
signals om de ice componen s [7], and adio
equency (RF) based echniques use physical laye
in o ma ion like signal- o-noise a io and o he
simila ac o s [8]. All hese echniques ely on
some ype o machine lea ning algo i hm o classi y
and iden i y de ices.
Howe e , he majo challenge is o di e en ia e
de ices ha a e iden ical in e ms o ha dwa e and
so wa e con igu a ions. As he majo i y o he IoT
p oduc s a e mass-manu ac u ed, hey end o be o
iden ical con igu a ion, which makes he issue o
de ice iden i ica ion mo e challenging [9].
1.1. Mo i a ion And P oblem S a emen
Typically, IoT de ices deployed o a speci ic
applica ion, such as in sma homes, indus ial
moni o ing, o heal hca e en i onmen s, a e
comp ised o de ices ha may be iden ical in e ms
o ha dwa e and so wa e con igu a ions. The
limi a ions o s a ic ne wo k iden i ie s c ea e a
signi ican challenge in iden i ying and
di e en ia ing hese de ices once hey a e
ope a ional wi hin he same ne wo k. He e, he ask
is o uniquely inge p in IoT de ices o he same
make and model wi hou elying on EME, which
o en equi es specialized ha dwa e and is
imp ac ical in la ge-scale deploymen s. E en
hough he de ices a e iden ical in bo h ha dwa e
and so wa e, hey s ill exhibi sub le di e ences in
ne wo k beha io , communica ion iming, and
usage pa e ns due o na u al a ia ions in
manu ac u ing, en i onmen al ac o s, and in e nal
p ocessing s a es. The majo i y o he exis ing
wo ks ocused on he he e ogeneous ype o de ices
and ha e p oduced signi ican de ice p o iling and
au hen ica ion app oaches. Mo eo e , we highligh
he wo k by collec ing minu e di e ences in
signi ican pa ame e s wi h he ne wo k is in
ope a ion om he a ious homogeneous de ices o
add ess he au hen ica ion issues. In addi ion o his,
he p oposed model has p oduced angible esul s in
e ms o inge p in ing iden ical de ices. As he
model ocused on ime se ies da a, which is
collec ed wi h a ia ion in ime slo s.
This wo k aims o design a no el
inge p in ing me hod o cap u e and analyze hese
sub le, na u ally occu ing a ia ions in de ice
beha io o uniquely iden i y each de ice. This
inge p in ing me hod should in eg a e mul i-laye
ea u es o iden i y and inge p in de ices.
Jou nal o Theo e ical and Applied In o ma ion Technology
31s May 2025. Vol.103. No.10
© Li le Lion Scien i ic
ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195
4212
1.2 Con ibu ions
This pape makes he ollowing con ibu ions:
 We in oduce a no el mul i-laye app oach o
inge p in ing iden ical de ices, le e aging
ime- a ying ea u es ac oss he MAC,
anspo , and physical laye s.
 A dynamic ime wa ping (DTW)-based
algo i hm is p oposed o empo al alignmen
and classi ica ion o de ice inge p in s.
 We demons a e h ough ex ensi e expe imen s
ha ou app oach can success ully di e en ia e
de ices o he same make and model, using
eal-wo ld IoT da a.
2. RELATED WORKS
De ice iden i ica ion and inge p in ing
a e key o imp o e he secu i y pos u e in an IoT
en i onmen whe e de ices a e suscep ible o
comp omise. In his pape , we p esen di e en
wo ks in his di ec ion o iden i ying IoT de ices,
ocusing on hei e ec i eness, echnical de ails,
s eng hs, and limi a ions. One o he simple
app oaches o iden i ying de ices is based on
de ice signa u es. These me hods c ea e ules based
on known de ice beha io s, which can be e ec i e
o iden i ica ion [10], [11]. Howe e , hese
echniques equi e cons an upda es o adap o new
de ices and h ea s. Al e na i ely, he ne wo k
a ic gene a ed by IoT de ices can be used o
lea n unique pa e ns ha can be used o classi y
and iden i y de ices. Mul iple such wo ks use
ne wo k a ic o de ice iden i ica ion [12], [13],
[14], [15], [16]. Fo example, he au ho s o [12]
use machine lea ning-based me hods o iden i y IoT
de ices h ough ne wo k a ic analysis om a es
se o nine de ices. The model achie es 99.28%
accu acy in dis inguishing de ices u ilizing ne wo k
ea u es. I pe o ms session-le el analysis and uses
a mul is age classi ie o speci ic IoT
iden i ica ion.
Simila ly, he wo k p esen ed in [13] also
uses di e en ne wo k a ic pa e ns such as ypes
o a ic (e.g. di e en p o ocols), packe sizes, he
equency o packe s along wi h hei s a is ical
measu es, packe in e -a i ing imes, e c. o
iden i y de ices using machine lea ning algo i hms
and hen apply p ede ined secu i y measu es
ollowing a secu i y policy. In [14], he au ho s
add ess he challenge o managing a la ge numbe
o IoT de ices in a la ge-scale IoT en i onmen
such as a sma ci y. The p esen ed wo k uses
ne wo k a ic analy ics o cha ac e ize and
moni o IoT de ice beha io by collec ing a ic
aces om a di e se se o IoT de ices ha include
sma came as, sma ligh s, and heal h moni o s.
Fea u es such as da a a es, ac i i y cycles, and
signaling pa e ns we e employed o i s
dis inguish IoT om non-IoT a ic and hen
iden i y speci ic IoT de ices wi h mo e han 95%
p ecision. O he wo ks based on ne wo k
cha ac e is ics, like [15], use gene ic algo i hms o
ea u e selec ion and di e en machine lea ning
algo i hms o de ice classi ica ion. On he o he
hand, he au ho s in [16] add ess he p oblem o
da a imbalance in de ice classi ica ion. All o he
abo e-men ioned app oaches use di e en ne wo k
cha ac e is ics and ailo hei ML app oaches o
de ice classi ica ion and iden i ica ion.
O he app oaches o de ice iden i ica ion
a e ei he RF-based, clock-skew based, o
specialized ha dwa e-based echniques ha ha e
hei espec i e ad an ages and limi a ions. RF-
based de ice iden i ica ion schemes le e age he
unique cha ac e is ics o he adio signals, such as
signal's ampli ude, equency, and phase
cha ac e is ics emi ed by IoT de ices, hen apply
ML algo i hms o c ea e a dis inc inge p in o
each de ice. The au ho s o [17], [18] p esen a
su ey o such app oaches. In [19], he au ho s
p opose o exploi he unique cha ac e is ics ound
in he ene gy spec um o ansmi e u n-on
ansien signals. These ansien signals con ain
unique, ha dwa e-speci ic a ia ions ha se e as
inge p in s o de ice iden i ica ion. The me hod
ex ac s he ene gy dis ibu ion ac oss di e en
equency componen s om hese ansien s, which
helps o dis inguish be ween de ices, e en a low
signal- o-noise a ios. In [20], he au ho s p esen
an app oach ha con e s he ime se ies da a in o
images ins ead o di ec ly using aw signals o
ex ac ing s a is ical ea u es om hem. This
ans o ma ion allows he use o well-es ablished
image-p ocessing echniques and machine lea ning
models ha a e pa icula ly powe ul o pa e n
ecogni ion. App oaches based on clock skew ake
ad an age o minu e di e ences in he in e nal
clock a es o de ices o c ea e unique iden i ie s
[21], [8]. This is done by moni o ing he ne wo k
packe s sen by de ices o e ime o calcula e he
de ia ion o each de ice's clock om a e e ence
clock. Then, a unique p o ile is c ea ed based on he
consis en skew pa e ns obse ed. A su ey o
such app oaches is p esen ed in [18]. Mo e
ecen ly, a clock skew-based de ice iden i ica ion
scheme combined wi h a emo e a es a ion
p o ocol has been p esen ed in [22] o class-I IoT
de ices (de ices wi h andom access memo y less
han 10 KB and code size less han 100 KB). The
majo challenges wi h he clock skew-based
Jou nal o Theo e ical and Applied In o ma ion Technology
31s May 2025. Vol.103. No.10
© Li le Lion Scien i ic
ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195
4213
app oaches a e ha hey need con inuous
moni o ing o e an ex ended pe iod o ime and a e
a ec ed by ne wo k la ency and ji e .
Apa om he a o emen ioned de ice
iden i ica ion app oaches, ew o he wo ks conside
a hyb id app oach [23], ha is o combine mul iple
de ice cha ac e is ics and o he uses de ice senso s
o inge p in ing [24]. Howe e , none o he abo e
app oaches conside mul iple iden ical de ices wi h
he same ha dwa e and i mwa e in hei es se up
and e alua e he de ice iden i ica ion and
inge p in ing app oaches. In his wo k, we p opose
a no el app oach o De ice Iden i ica ion and
Finge p in ing wi h Time-Va ian Adap i e
Recogni ion (DIFT-VAR) ha is based on mul i-
laye , ime- a ying ea u e ex ac ion. we cons uc
dynamic inge p in s ha uniquely iden i y each
de ice by moni o ing and using ea u es such as
p obe eques beha io , clock skew, anspo laye
cha ac e is ics, and adio signal me ics o e ime.
Al e na i ely, o he wo ks ha e explo ed he issue
o inge p in ing in a he e ogeneous en i onmen .
Fo example, he wo k p esen ed in [25] p esen s a
me hod o classi y IoT de ice ypes (e.g., came as,
ou e s, p in e s) using Shodan me ada a. I
emphasizes on he e ec i eness o using cu a ed
me ada a and machine lea ning o accu a e
iden i ica ion o IoT de ice ypes. The pape [26]
explo es he use o ICMP and IP imes amp
esponses o dis inguish be ween physical And oid
de ices and i ual machines. Designed o de ec
VM-based malwa e e asion, he s udy e eals
consis en iming disc epancies ha se e as
passi e indica o s o he execu ion en i onmen .
The pape in [27] employs inge p in s o IOT
de ices. I is ne wo k beha io -based, ex ac ing
ea u es om he ne wo k, T anspo , and
applica ion laye s, and has also employed neu al
ne wo ks o classi ica ion. Employed neu al
ne wo k o classi y IoT de ices by u ilizing 3
s ages, i.e, de ice ype, endo , and p oduc . He has
employed a glo e and a Bi-LSTM o he
applica ion laye da a. De ice beha io , howe e ,
a ies as when i mwa e upda es. The pape [28]
u ilizes he MRFE deep lea ning app oach o
ecognize IoT de ices using RF inge p in s. The
model imp o es he iden i ica ion accu acy wi h
mul i-dimensional ea u es. The da a se was
acqui ed in noise noise- ee en i onmen . The
model uses a inge p in -ampli ying laye , h ee-
channel inpu , an A en ion mechanism, and
esidual connec ions and ully connec ed laye s.
The model depends mo e on he beha io o he RF
da a. Howe e , he pe o mance o he model d ops
o below 8dB o SNR
3. DIFT-VAR: THE PROPOSED
FINGERPRINTING TECHNIQUE
De ice iden i ica ion in a he e ogeneous
IoT en i onmen is compa a i ely simple as he
de ices exhibi su icien ly a ied cha ac e is ics
ha can be cap u ed ia ne wo k, RF, de ice
beha io , o clock skew, and o he simila
app oaches. Howe e , in an IoT en i onmen whe e
mul iple de ices sha e he same de ice a chi ec u e
in e ms o bo h ha dwa e and i mwa e, he issue
o de ice iden i ica ion becomes a challenge. The
in e ac ion o hese de ices in a ne wo k exhibi s
almos simila pa e ns ha a e ha d o dis inguish,
especially when he de ices a e ope a ional on he
same applica ion. A ypical use-case scena io is a
su eillance sys em whe e a se o sma secu i y
came as can belong o a single manu ac u e . The
su eillance applica ion necessi a es ha all de ices
epo su eillance da a ha can exhibi simila
communica ion pa e ns. In such a scena io, he key
is o exploi sub le a ia ions in hei beha io ha
a ise om ha dwa e impe ec ions, ne wo k
condi ions, and en i onmen al ac o s. In his wo k,
we p opose a no el me hod o p ocessing he da a
by le e aging mul i-laye da a usion combined
wi h ime- a ying ea u e ex ac ion o c ea e
unique de ice inge p in s o e ime. The co e idea
is o design a mul i-laye , ime- a ying inge p in
Cons uc ion me hod. Ra he han elying on a
s a ic ea u e se , he p oposed app oach ocuses on
he empo al e olu ion o ea u es ac oss mul iple
laye s (MAC laye , adio cha ac e is ics, anspo
laye , and iming da a) o di e en ia e de ices. The
no el y is in dynamically combining empo al
ends and sub le a ia ions in ne wo k and iming
beha io o e ime, a he han s a ic snapsho s o
ea u es.
The p oposed app oach is named DIFT-
VAR, whe e a ime- a ying mul i-laye
inge p in ing app oach ha uses ea u es ac oss
he MAC, anspo , and iming laye s. Ou
app oach consis s o h ee main componen s: mul i-
laye ea u e ex ac ion, ime- a ying ea u e
agg ega ion, and classi ica ion using dynamic ime
wa ping (DTW). The uniqueness lies in he
p ocessing me hod, speci ically in di e en ia ing
iden ical de ices based on a empo al dependency
analysis using dynamic ime wa ping (DTW) and
ecu en neu al ne wo ks (RNN) wi h an emphasis
on sequence p ocessing.
Jou nal o Theo e ical and Applied In o ma ion Technology
31s May 2025. Vol.103. No.10
© Li le Lion Scien i ic
ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195
4214
3.1 Mul i-Laye Fea u e Fusion and Tempo al
Va ia ion T acking.
As iden ical de ices may exhibi sligh
ha dwa e-induced a ia ions o e ime due o
en i onmen al ac o s (like dis ance om he Wi-Fi
access poin ), small a ia ions in he in e nal clock,
o p ocessing speed di e ences ha a ise om
manu ac u ing impe ec ions, we do no ely on a
ixed da ase o ea u es collec ed a one poin in
ime. We conside he e olu ion o ea u es o e
ime and use da a ac oss mul iple laye s (MAC,
adio, anspo , and iming) o c ea e a highly
unique inge p in .
A he MAC laye , we conside he MAC add ess
and he ecei ed signal s eng h indica o (RSSI) o
p obe eques s. The MAC add ess is conside ed a
basic iden i ie . E en hough i can easily be
spoo ed, i is conside ed a s a ing poin . We
conside he Wi-Fi communica ion among de ices
whe e de ices pe iodically send p obe eques s o
ne wo k disco e y and e-associa ion. Le
( )
k
i
T
ep esen he in e al be ween he
i- h
and
(i+1)- h
p obe eques s o he de ice
k
. To compa e
de ices, we cons uc a ime-se ies ec o and use
Dynamic Time Wa ping (DTW) o calcula e he
simila i y be ween wo ime se ies. ( ) ( )
( ( , ) :
k m
ST T
( ) ( ) ( ) ( )
1
( , ) min ( , )
n
k m k m
i j
i
DTW d T T


T T
whe e
( , )
d
 
I is a dis ance me ic ( ypically
Euclidean dis ance).
The ime in e als be ween consecu i e
p obe eques s can a y sligh ly due o in e nal
ha dwa e di e ences and he iming p ecision o
each de ice. The p obe eques s a e analyzed o
suppo ed da a a es, SSID, and capabili ies. E en
iden ical de ices can ha e sub le di e ences based
on i mwa e o en i onmen al adap a ions. The
iming o p obe eques s is used o gene a e a
empo al pa e n o each de ice, which is hen
compa ed using a dynamic ime-wa ping app oach
o iden i y sub le di e ences in eques in e als.
The RSSI alues a e cap u ed o packe s
be ween he de ices and he Wi-Fi access poin .
These alues a e collec ed con inuously since
physical placemen and an enna di e ences cause
a ia ions. Fo signal- o-noise a io (SNR), he
signal quali y ela i e o he backg ound noise is
measu ed. The di e ences in he de ice's RF
componen s can esul in sligh bu de ec able
a ia ions in SNR. Also, as each de ice may ha e
sligh a ia ions in ansmission powe , i can lead
o di e ences in RSSI o e ime. We se up
mul iple access poin s o measu e how equen ly
he de ices swi ch channels when nume ous access
poin s a e a ailable, conside ing hei channel
u iliza ion capabili ies.
Le
( ) ( ) ( ) ( )
1 2
[ , , ,
]
k k k k
n
 RSSI be
he RSSI eadings o he de ice
k
. Fo
di e en ia ion, calcula e he a iance o e a sliding
window o size
w :
( ) ( ) 2 ( )
1 1
1 1
( ) , whe e
w w
k k k
RSSI i RSSI RSSI i
i i
w w
  
 
  
 
To cap u e changes o e ime, gene a e an
agg ega ed ea u e ec o o each ime window:
( ) (1) (2) ( )
[ , , , ]
k m
RSSI RSSI RSSI RSSI
  
 F
The RSSI and SNR a e collec ed as ime-
se ies da a. This in o ma ion is p ocessed using
sequence models o iden i y he d i in signal
s eng h o noise condi ions, p o iding unique
iden i ie s o each de ice ega ding hei physical
connec i i y o he ne wo k.
We conside he TCP window size a he
anspo laye le el and moni o he changes
du ing da a exchanges. TCP sequence numbe s a e
obse ed du ing he handshake and communica ion
phases. The di e ence be ween sequence numbe s,
o he sequence numbe gaps, can a y depending
on how he de ice's ne wo k s ack handles
e ansmissions and acknowledgmen s.
Le he TCP sequence numbe s be
ep esen ed as
( ) ( ) ( ) ( )
1 2
,[ , , ]
k k k k
n
seq seq seq SEQ
Using DTW o alignmen o sequence
numbe s o iden i y sub le de ice-speci ic
a ia ions, i is ep esen ed as ollows:
( ) ( ) ( ) ( )
pa hs
1
( , ) min
n
k m k m
i j
i
DTW seq seq

 

SEQ SEQ
‖ ‖
We conside he ound- ip ime (RTT) o
TCP connec ions and consis en ly measu e he RTT
be ween he de ices and a se e on he ne wo k.
I s alue a ies based on clock skew and p ocessing
speed di e ences. The TCP sequence numbe and
window size a e modeled as ea u es con ibu ing o
ne wo k beha io di e ences. The RTT is ea ed
as empo al da a and is con inuously moni o ed o
build a clock skew p o ile o e ime.
The RTT is ep esen ed as
( )
k
i
RTT o
de ice
k
. Using he ime-a e aged mean and
a iance:

Jou nal o Theo e ical and Applied In o ma ion Technology
31s May 2025. Vol.103. No.10
© Li le Lion Scien i ic
ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195
4215
( ) ( ) ( ) ( ) ( ) 2
1 1
1 1
, ( )
n n
k k k k k
RTT i RTT i RTT
i i
RTT RTT
n n
  
 
  
 
Clock skew
C( )
is he de ia ion o he
clock o e ime. The clock skew is measu ed by
sending ICMP pings o each de ice and eco ding
he ound- ip ime (RTT). The d i o e ime is
calcula ed o p oduce a clock skew alue ha is
unique o each de ice. On he o he hand, he in e -
a i al ime (IAT) o packe s, pa icula ly du ing
high ac i i y, is measu ed. The clock skew unc ion
o he de ice
k
can be ep esen ed as:
( )
k k k
C
 
 
whe e
k

is he d i a e and
k

ep esen s
he ini ial o se . Compa ing clock skews o wo
de ices in ol es compu ing he di e ence:
,
( ) | ( ) ( ) |
k m k m
C C C
  
Di e ences in clock d i and packe
p ocessing speed esul in de ec able a ia ions in
packe iming. The clock skew and IAT a e
analyzed as ime-se ies ea u es, p o iding a unique
empo al signa u e o each de ice.
3.2 Time-Va ying Fea u e Agg ega ion
The p oposed app oach, algo i hm 1,
conside s ea u es collec ed o e ime a he han
s a ic ea u es collec ed a a single ime ins ance.
This enables us o cap u e empo al pa e ns in
ne wo k beha io ha a e sub le bu consis en
ac oss iden ical de ices. We show ha ime- a ying
pa e ns o en e eal mino di e ences due o
en i onmen al o ha dwa e ac o s. To achie e ha ,
we con e key ea u es in o ime se ies da a and
apply ime se ies analysis echniques like mo ing
a e ages, Fou ie ans o ms, and wa ele analysis
o cap u e empo al ends in ea u e beha io . The
me hod chosen is dependen on he ea u e ype.
Fo example, o analyze he RSSI alues ha
luc ua e due o en i onmen al condi ions, we
conside mo ing a e ages o smoo h ou he signal
s eng h a ia ions.
On he o he hand, o analyze p obe
eques in e als and suspec ha a de ice sends
eques s a egula in e als, we apply a Fou ie
ans o m ha will highligh his cyclical pa e n.
The Fou ie spec um will peak a he
co esponding equency, hus allowing he
di e en ia ion o de ices based on hei pe iodic
beha io . Fo ea u es such as clock skew ha
exhibi a long- e m d i sp ead ac oss wi h
occasional a ia ions due o in e nal p ocessing
delays, we employ wa ele analysis ha allows us
o cap u e bo h he g adual changes and he sudden
shi s in iming beha io .
Algo i hm-1 Mul i-Laye De ice Finge p in ing
wi h Tempo al Sequence Analysis
1. Inpu :
2. Se o
N
de ices 1 2
{ , , , }
N
D D
D
3. Fo each de ice
i
D
, collec ime-se ies
da a
i
X :
4. MAC laye ea u es
i
MAC
5. Radio cha ac e is ics ,
i i
RSSI SNR
6. T anspo laye ea u es
, ,
i i i
TCPSeq RTT TCPWnd
7. Timing ea u es
,
i i
P obe In e al Clock Skew
8. Fea u e Ex ac ion:
9. Fo each de ice
i
D ,
ex ac ime-se ies
ea u es:
10. MAC laye ea u es:
,
i i
MAC P obe In e al
11. Radio cha ac e is ics: ,
i i
RSSI SNR
12. T anspo laye ea u es:
, ,
i i i
TCPSeq RTT TCPWnd
13. Timing ea u es:
i
Clock Skew
14. Time-Se ies Alignmen Using DTW:
15. Fo each de ice
i
D
and ea u e
i
X
,
pe o m DTW on sequences o :
16. ,
i i
P obe In e al
TT
R
17. Compu e he DTW dis ance o align iming
a ia ions:
18.
( )
DTW i
S X
19. Tempo al Modeling Using LSTM:
20. T ain an LSTM on ime-se ies da a
Xi
o
each ea u e:
21. , , ,
i i i i
RSSI RTT TCPSeq Clock Skew
22. Ou pu a sequence embedding
( )
LSTM i
E X
o each de ice
23. Fea u e Fusion:
24. Fuse all ea u es in o a ea u e ec o
Fi
:
, , ,
( ), ( ),
, ,
i i i
i DTW i LSTM i
i i i
MAC RSSI SNR
F S X E X
TCPSeq RTT TCPWnd
 
 

 
 
 
Jou nal o Theo e ical and Applied In o ma ion Technology
31s May 2025. Vol.103. No.10
© Li le Lion Scien i ic
ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195
4216
25. Classi ica ion:
26. T ain a Random Fo es classi ie using he
ea u e ec o s
i
F
27. Au hen ica ion:
28. Gi en an unknown de ice
j
D
, ex ac he
ea u e ec o
j
F
29. Classi y
j
F
using he ained classi ie o
au hen ica e
j
D .
30. Ou pu : De ice classi ica ion and unique
inge p in o au hen ica ion.
Fu he , we employ a sliding window
app oach whe e he key ea u es a e calcula ed
dynamically wi hin ime windows a he han
s a ic ea u e ec o s, hus c ea ing a empo al
inge p in wi h dynamic windows. Fo a ime-
a ying ea u e ( )
( )
k
X
, we ex ac ea u es using
a sliding window app oach o size
w
.
( ) ( )
1
( ) ( )
w
k k
window
i
X X i
w



Fo each ime window, we cap u e and
summa ize a ia ions in key ea u es like clock
skew, RSSI, TCP beha io , and p obe eques
iming. Fo each o he de ice, he agg ega e
ea u es om mul iple laye s wi hin a dynamic ime
window c ea es a composi e inge p in . We
employ a weigh ed app oach o gi e mo e
impo ance o laye s ha exhibi g ea e a iabili y
be ween de ices. Fea u es calcula ed o each
sliding window a e used o gene a e dynamic
inge p in s ha emphasize a iabili y o e ime.
The key componen s o he p oposed a chi ec u e
a e shown in Figu e 1.
3.3 Mul i-Resolu ion Fea u e Ma ching
To uniquely di e en ia e de ices, he
p oposed app oach pe o ms a mul i- esolu ion
analysis, whe e ea u es a e compa ed a di e en
le els o g anula i y. To begin wi h, o each o he
de ices, we c ea e bo h sho - e m and long- e m
ea u e p o iles. The sho - e m p o iles cap u e
immedia e luc ua ions, while long- e m p o iles
ack cumula i e beha io o e an ex ended pe iod
o ime. Then, we aim o di e en ia e de ices based
on a combina ion o sho - e m and long- e m
p o iles using echniques like dynamic ime
wa ping (DTW) o align and compa e ime se ies
da a, accoun ing o sligh ime shi s o phase
di e ences.
We de ine he sho - e m
( )
k
S and long-
e m
( )
k
L
ea u e ec o s o each de ice
k
as
ollows:
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )
1 2 1 2
[ , , , ], [ , , , ]
k k k k k k k k
n m
s s s l l l
   S L
DTW is used o measu e he simila i y be ween
empo al sequences o di e en de ices. DTW
allows o alignmen o ime-se ies da a ha migh
be ou o phase bu s ill ep esen simila pa e ns.
We apply DTW o align and compa e he de ice
p o iles using:
( ) ( ) ( ) ( )
( , ) ( , )
k m k m
DTW DTWS S L L
3.4 Fea u e Fusion and Classi ica ion
Fo di e en ia ing and iden i ying de ices
we employ a ea u e usion app oach. In his
me hod, we combine ea u es om all laye s
(MAC, adio, anspo , and iming) o c ea e a
comp ehensi e ea u e ec o o each de ice. Fo
any ea u e ec o , le
( )
k
F
ep esen he ea u e
ec o o de ice
k
, con aining ea u es om
MAC, adio, anspo , and iming laye s:
( ) ( ) ( ) ( ) ( ) ( )
[ , , , , ]
k k k k k k
MAC adio anspo iming LSTM
F F F F hF
The weigh ed usion o ea u es, whe e weigh
i
w
co esponds o he a iabili y o each ea u e is
ep esen ed as:
( ) ( )
, 1
k k
composi e i i i
i i
F w F wi h w
 
 
Nex , we employ he Random Fo es
classi ie s o ain on he used ea u e se . The
LSTM ou pu ( empo al analysis) is also included
in he ea u e se , p o iding addi ional empo al
con ex o he classi ie . The used ea u es a e used
o c ea e a unique inge p in o each de ice. The
aining o he Random Fo es classi ie on he
used ea u e ec o s
( )
k
F
is ep esen ed as ollows,
whe e
ˆ
y
is he p edic ed class:
( )
ˆ
( )
k
y Classi ie F
The combined use o ime-se ies alignmen
(DTW), empo al sequence modeling (LSTM), and
laye -speci ic ea u es makes he p oposed
app oach unique.
Jou nal o Theo e ical and Applied In o ma ion Technology
31s May 2025. Vol.103. No.10
© Li le Lion Scien i ic
ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195
4217
3.5 De ice Iden i ica ion
The p oposed app oach iden i ies de ices
by collec ing da a o e an ex ended pe iod and
c ea ing a de ice p o ile. Con inuous da a
collec ion in ol es con inuously ga he ing ea u es
ac oss all ele an laye s o main ain bo h sho -
e m and long- e m p o iles o each de ice. In he
ma ching and classi ica ion p ocess, DTW is
applied o align newly collec ed ime-se ies da a
wi h s o ed p o iles. An LSTM model is used o
in e empo al dependencies in eal- ime ea u es.
The aligned da a and LSTM ou pu s a e hen used
and classi ied using a ained Random Fo es
classi ie o au hen ica e he de ice. In he decision-
making s age, au hen ica ion is de e mined based
on he DTW alignmen sco e and he Random
Fo es classi ica ion. I is ep esen ed as ollows:
De ice au hen ica ion is based on compa ing he
DTW alignmen sco e DTW
Sco e
and classi ica ion
ou come:
ˆ
DTW ue
Au hen ica e i Sco e and y y

 
I he compu ed sco e is below he es ablished
h eshold, he de ice is conside ed un ecognized o
spoo ed.
Fig. 1 The p oposed de ice iden i ica ion model
4. EXPERIMENTAL RESULTS
To analyze he pe o mance o he
p oposed de ice iden i ica ion and inge p in ing
algo i hm, we se up a ne wo k o 10 ESP32-
WROOM-32 mic ocon olle s, all unning he same
i mwa e and communica ing o e a Wi-Fi
ne wo k. Each de ice has he same se o senso s
( empe a u e, humidi y, ligh ), which epo da a
e e y 60 seconds. This expe imen aims o e alua e
he abili y o uniquely inge p in iden ical de ices
using mul i-laye ea u e ex ac ion and ime-se ies
p ocessing echniques. The expe imen will cap u e
Jou nal o Theo e ical and Applied In o ma ion Technology
31s May 2025. Vol.103. No.10
© Li le Lion Scien i ic
ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195
4218
he MAC laye , adio cha ac e is ics, anspo
laye , and iming ea u es, hen p ocess he da a
using Dynamic Time Wa ping (DTW) and LSTM-
based empo al modeling.
4.1 Ha dwa e Se -up
Table 1 p o ides a de ailed ESP32-
WROOM-32 mic ocon olle speci ica ion
summa y. The boo p ocess in ol es he i s -s age
boo loade , pa o he ROM. The second-s age
boo loade is s o ed in Flash and can be
cus omized. The pa i ion able de ines he lash
memo y layou , and he applica ion is pa o he
main i mwa e code. The ESP32 so wa e
de elopmen ki SDK, also known as he ESP-IDF
(Esp essi IoT De elopmen F amewo k), p o ides
a comp ehensi e se o ools, lib a ies, and
applica ion p og amming in e aces (APIs) ha
allow us o access and manipula e a ious ha dwa e
ea u es o he ESP32. The SDK p o ides APIs o
di ec access o ha dwa e ea u es like ADC
(Analog- o-Digi al Con e e ), ime s, and RF
componen s. This low-le el access allows o
p ecise measu emen s o ha dwa e cha ac e is ics.
The ESP-IDF includes a ull- ea u ed Wi-Fi and
Blue oo h s ack, enabling ine-g ained con ol o e
wi eless communica ions. This can be used o
implemen cus om ansmission pa e ns o analyze
ecep ion cha ac e is ics. The SDK is buil on
F eeRTOS, enabling p ecise iming con ol and ask
managemen . I can be le e aged o c ea e unique
beha io al pa e ns o measu e sys em esponse
imes accu a ely. Las ly, ESP-IDF p o ides
ad anced powe managemen capabili ies, allowing
o measu ing and con olling powe consump ion
pa e ns, which can a y be ween de ices.
The main CPU clock speed is up o 240
MHz wi h an in e nal 150 kHz RC oscilla o and an
ex e nal c ys al o 40 MHz.. The in e nal oscilla o
is 8 MHz wi h calib a ion. The in e nal RC
oscilla o and PLL ci cui s can a y sligh ly due o
manu ac u ing p ocesses, a ec ing hei exac
equencies. Also, empe a u e and ol age
luc ua ions can cause sub le clock d i pa e ns
unique o each de ice. The ESP32's abili y o
dynamically adjus clock speeds can be used o
c ea e unique beha io al pa e ns. The de ice
suppo s IEEE 802.11 b/g/n (2.4 GHz) p o ocols
wi h an adjus able ansmi powe o up o +20
dBm o a ecei e sensi i i y o up o -98 dBm. All
he de ices a e connec ed o he same Wi-Fi access
poin (TP-Link A che C7). The de ices use HTTP
as he communica ion p o ocol o ansmi senso
da a.
Table 1: Key Speci ica ions o Esp32-W oom-32.
Fea u e Speci ica ion
P ocesso
Dual
-
co e X ensa LX6, 32
-
bi
Clock Speed
Up o 240 MHz
RAM
520 KB SRAM
ROM
448 KB
Flash Memo y
4 MB (expandable o 16 MB)
Wi
-
Fi
802.11 b/g/n (2.4 GHz)
Blue oo h
4.2 BR/EDR and BLE
GPIO Pins
34 p og ammable
ADC
12
-
bi , up o 18 channels
DAC
2x 8
-
bi channels
Ha dwa e C yp o
AES, SHA, RSA, ECC
RTC
150 kHz in e nal oscilla o
Ex e nal C ys al
40 MHz
Powe
Consump ion
A e age 80mA
Ope a ing Vol age
3.0V o 3.6V
Ope a ing
Tempe a u e
-40°C o +85°C
4.2 Da a Collec ion and Da a P ocessing
In his expe imen , we collec da a om 10
iden ical ESP32-WROOM-32 mic o-con olle s.
Each de ice sends senso da a pe iodically. The
ollowing da a is collec ed as shown in he Table.2
The ea u es a e collec ed o e mul iple
sessions o ensu e su icien da a is cap u ed o
analysis. The a ia ions in he da a a e expec ed due
o he mino ha dwa e and en i onmen al
di e ences ac oss de ices, e en hough he de ices
a e iden ical.
The da a p ocessing pipeline is comp ised
o wo main s ages, i.e., ime-se ies alignmen using
DTW and empo al modeling using LSTM o
cap u e he empo al dependencies o each o he
de ice beha io . Time-se ies da a o ea u es like
P obe Reques In e als, RTT, and IAT a e
ex ac ed om each de ice. The aim o using DTW
is o align sequences o ime-se ies da a o accoun
o any a ia ion due o di e ences in clock skew,
p ocessing delays, o en i onmen al ac o s. Fo
each ea u e (e.g., P obe Reques In e al), we
compu e he DTW dis ance be ween each de ice’s
ime-se ies da a and a e e ence sequence. We hen
align he ime-se ies da a based on he minimum
DTW dis ance o ob ain a simila i y sco e
( ( )
DTW i
S X
o each de ice
i
X
. The ou come is a
se o aligned sequences ha cap u e he iming
a ia ions ac oss de ices.
To achie e empo al modeling using
LSTM, we conside he ime-se ies da a om
mul iple sessions o ea u es like RSSI, RTT, TCP
sequence numbe s, and clock skew as inpu . The
LSTM ne wo k con igu a ion has an inpu laye ,
whe e he inpu is he ime-se ies da a om each
Jou nal o Theo e ical and Applied In o ma ion Technology
31s May 2025. Vol.103. No.10
© Li le Lion Scien i ic
ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195
4225
s uden s, and o he s uden s (S3) Wo kshop,
e . S3 ’21. New Yo k, NY, USA: Associa ion
o Compu ing Machine y, 2021, p. 6–7.
[Online]. A ailable:
h ps://doi.o g/10.1145/3477087.3478381.
[8] T. Kohno, A. B oido, and K. Cla y, “Remo e
physical de ice inge -p in ing,” in 2005 IEEE
Symposium on Secu i y and P i acy
(SP’05),2005, pp. 211–225.
[9] J. Ku mi and R. Ma am, “De ice iden i ica ion
in io ne wo ks using ne wo k ace
inge p in ing,” in 2022 In e na ional
Con e ence on Sma Applica ions,
Communica ions and Ne wo king (Sma Ne s),
2022, pp. 1–6.
[10] V. Visoo i ise h, P. Saka in, J. Thongwilai,
and T. Choobanjong, “Signa u e-based and
beha io -based a ack de ec ion wi h machine
lea ning o home io de ices,” in 2020 IEEE
REGION 10 CONFERENCE (TENCON),2020,
pp. 829–834.
[11] N. Youse nezhad, A. Malhi, and K. F amling,
“Secu i y in p oduc li ecycle o io de ices: A
su ey,” Jou nal o Ne wo k and Compu e
Applica ions, ol. 171, p. 102779, 2020.
[Online]. A ailable:
h ps://www.sciencedi ec .com/science/a icle/p
ii/S1084804520302538.
[12] Y. Meidan, M. Bohadana, A. Shab ai, J. D.
Gua nizo, M. Ochoa,N. O. Tippenhaue , and
Y.Elo ici, “P o ilio : a machine lea ning
app oach o io de ice iden i ica ion based on
ne wo k a ic analysis,” in P oceedings o he
Symposium on Applied Compu ing,se .
SAC’17. New Yo k, NY, USA: Associa ion o
Compu ing Machine y, 2017, p. 506–509.
[Online]. A ailable:
h ps://doi.o g/10.1145/3019612.3019878.
[13] M. Mie inen, S. Ma chal, I. Ha eez, N.Asokan,
A.-R. Sadeghi,and S. Ta koma, “Io sen inel:
Au oma ed de ice- ype iden i ica ion o
secu i y en o cemen in io ,” in 2017 IEEE 37 h
In e na ional Con e ence on Dis ibu ed
Compu ing Sys ems (ICDCS), 2017, pp.2177–
2184.
[14] A. Si ana han, D. She a , H. H. Gha akheili
A.Rad o d, C. Wi-jenayake, A. Vishwana h,
and V. Si a aman, “Cha ac e izing and
classi ying io a ic in sma ci ies and
campuses,” in 2017 IEEE Con e ence on
Compu e Communica ions Wo kshops
(INFOCOMWKSHPS), 2017, pp. 559–564.
[15] A. Aksoy and M. H. Gunes, “Au oma ed io
de ice iden i ica ion using ne wo k a ic,” in
ICC 2019 - 2019 IEEE In e na ional Con e ence
on Communica ions (ICC), 2019, pp. 1–7.
[16] M. Mainuddin, Z. Duan, Y. Dong, S. Salman,
and T. Taami, “Io de ice iden i ica ion based
on ne wo k a ic cha ac e is ics,” in
GLOBECOM 2022 - 2022 IEEE Global
Communica ions Con e ence, 2022, pp.6067–
6072.
[17] L. Xie, L. Peng, J. Zhang, and A. Hu, “Radio
equency inge p in iden i ica ion o in e ne
o hings: A su ey,” Secu i y and Sa e y, ol.3,
p. 2023022, 2024.
[18] A. Jaganna h, J. Jaganna h, and P. S. P. V.
Kuma , “A comp ehensi e su ey on adio
equency ( ) inge p in ing: T adi ional
app oaches, deep lea ning, and open
challenges,” Compu e Ne wo ks, ol. 219, p.
109455, 2022. [Online]. A ailable:
h ps://www.sciencedi ec .com/science/a icle/p
ii/S1389128622004893.
[19] M. K¨ose, S. Tas¸cio˘glu, and Z. Tela a , “R
inge p in ing o io de ices based on ansien
ene gy spec um,” IEEE Access, ol. 7, pp. 18
715–18 726, 2019.
[20] G. Baldini, G. S e i, R. Giuliani, and C. Gen ile,
“Imaging ime se ies o in e ne o hings adio
equency inge p in ing,” in 2017 In e na ional
Ca nahan Con e ence on Secu i y Technology
(ICCST), 2017, pp. 1–6
[21] T. Kohno, A. B oido, and K. C. Cla y,
“Remo e physical de ice inge -p in ing,” IEEE
T ansac ions on Dependable and Secu e
Compu ing, Vol. 2, no. 2, pp. 93–108, 2005.
[22] C. Shang, J. Cao, T. Zhu, Y. Zhang, B. Niu, and
H. Li, “Cad a: A clock skew-based ac i e
de ice inge p in au hen ica ion scheme o
class-1 io de ices,” IEEE Sys ems Jou nal,
2024.
[23] H. Wang, D. Eklund, A. Op ea, and S. Raza,
“Fl4io : Io de ice inge p in ing and
iden i ica ion using ede a ed lea ning,” ACM
T ans. In e ne Things, ol. 4, no. 3, Jul. 2023.
[Online]. A ailable:
h ps://doi.o g/10.1145/3603257.
[24] H. Bojino , Y. Michale sky, G. Nakibly, and
D. Boneh, “Mobile de ice iden i ica ion ia
senso inge p in ing,” a Xi p ep in
a Xi :1408.1416, 2014.
[25] F. Z. Fag oud, H. Toumi, E. H. B. Lahma , K.
Ach aich, S. E. Filali, and Y. Baddi, “Connec ed
de ices classi ica ion using ea u e selec ion

Jou nal o Theo e ical and Applied In o ma ion Technology
31s May 2025. Vol.103. No.10
© Li le Lion Scien i ic
ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195
4226
wi h machine lea ning,” IAENG In e na ional
Jou nal o Compu e Science, ol. 49, no. 2,
2022.
[26] M. Noo a iza, K. Ishak, H. Maeda, M.
Shi a o i, T. Kinoshi a, and R. Uda,
“Cha ac e is ic pa e ns o imes amps om
and oid ope a ing sys em on mobile de ice and
i ual machine.” IAENG In e na ional Jou nal
o Compu e Science, ol. 43, no. 2, 2016.
[27] K. Yang, Q. Li, and L. Sun, “Towa ds
au oma ic inge p in ing o io de ices in he
cybe space,” Compu e Ne wo ks, ol. 148, pp.
318–327,2019. [Online]. A ailable:
h ps://www.sciencedi ec .com/science/a icle/p
ii/S1389128618306856.
[28] Qian Lu , Zaikai Yang , Hanlin Zhang “MRFE:
A Deep-Lea ning-Based Mul idimensional
Radio F equency Finge p in ing Enhancemen
App oach o IoT De ice Iden i ica ion IEEE
INTERNET OF THINGS JOURNAL, VOL.
11, NO. 18, 15 SEPTEMBER 2024 [Online]
A ailable:h ps://ieeexplo e.ieee.o g/s amp/s am
p.jsp? p=&a numbe =10556754