Co esponding au ho : Mild ed Adwubi Bonsu
Copy igh © 2025 Au ho (s) e ain he copy igh o his a icle. This a icle is published unde he e ms o he C ea i e Commons A ibu ion Liscense 4.0.
Resilien IoT Secu i y: Ea ly Flood A ack De ec ion in IoT Ne wo ks Using GRU Deep
Lea ning Model
Mild ed Adwubi Bonsu * and Philip Akekudaga
College o Eme gency P epa edness, Homeland Secu i y and Cybe secu i y, Uni e si y a Albany, S a e Uni e si y o New
Yo k. USA
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 871-886
Publica ion his o y: Recei ed on 28 June 2025; e ised on 10 Augus 2025; accep ed on 12 Augus 2025
A icle DOI: h ps://doi.o g/10.30574/wja .2025.27.2.2897
Abs ac
Secu ing In e ne o Things (IoT) ne wo ks has become inc easingly c i ical as hei in eg a ion ac oss essen ial sec o s
con inues o expand. Among he mos p essing h ea s a e lood a acks, a o m o Dis ibu ed Denial o Se ice (DDoS)
ha o e whelms ne wo k esou ces and causes se ice deg ada ion. In his s udy, he de ec ion o lood a acks in IoT
en i onmen s is add essed using a deep lea ning model based on he Ga ed Recu en Uni (GRU) a chi ec u e. Wi hin
he scope o he analysis, he CICIoT2023 da ase , which e lec s ealis ic IoT a ic and a ack beha io , was employed
o aining and alida ion. The esul s ha e shown ha he lood a acks we e success ully de ec ed, and he model
achie ed an accu acy sco e o 0.98, wi h mode a e p ecision, ecall, and F1 sco es. In his way, lood a acks in IoT can
be iden i ied ea ly o mi iga e hei impac and enhance he esilience o IoT in as uc u e. This s udy con ibu es o
in elligen IoT secu i y by in eg a ing upda ed da ase s, sequen ial modeling, and empi ical e alua ion, es ablishing a
solid ounda ion o u u e esea ch in h ea de ec ion sys ems.
Keywo ds: In e ne O Things (IoT); IoT Secu i y; Dis ibu ed Denial o Se ice (DDOS); Deep Lea ning; Ga ed
Recu en Uni (GRU).
1. In oduc ion
The In e ne o Things (IoT) has become an in eg al pa o daily li e, ans o ming how we manage ou homes,
communica e, and ope a e ac oss a ious indus ies. Wi h IoT de ices playing a c ucial ole in sec o s such as
heal hca e, anspo a ion, ene gy, and sma homes, hey o e unp eceden ed con enience and e iciency. Howe e ,
as hei p esence g ows, so do he secu i y challenges associa ed wi h hei widesp ead in e connec i i y. Among he
mos c i ical h ea s o IoT ne wo ks a e lood a acks, a o m o Dis ibu ed Denial o Se ice (DDoS) a ack ha dis up s
he no mal unc ionali y o de ices and ne wo ks by o e whelming hem wi h illegi ima e a ic [1, 2]. Flood a acks
a ge ing IoT in as uc u e ha e escala ed in ecen yea s, p esen ing signi ican eal-wo ld implica ions. In 2022, he
numbe o IoT malwa e a acks wo ldwide eached 112.29 million, ma king an 87% yea -o e -yea inc ease om 2021.
By he ou h qua e o 2024, global DDoS a acks had isen o 512,000, up om 274,000 in he i s qua e o 2023
[3]. These inciden s o en esul in se ice dis up ions, da a loss, sys em down ime, and subs an ial epu a ional
damage, pa icula ly o o ganiza ions ha ely on eal- ime da a ansmission.
T adi ional secu i y solu ions, such as signa u e-based in usion de ec ion sys ems and basic i ewalls, ha e p o en
inadequa e in he ace o hese e ol ing a ack pa e ns. No ably, many legacy sys ems a e no op imized o IoT
en i onmen s' dynamic, esou ce-cons ained, and he e ogeneous na u e [4]. As a esul , he e is an u gen need o
mo e adap i e and in elligen de ec ion me hods. This s udy aims o p o ide a solu ion o imp o e he e iciency o he
de ec ion o lood a acks in IoT en i onmen s wi h deep lea ning using he Ga ed Recu en Uni (GRU) algo i hm,
which is e ec i e in cap u ing empo al dependencies in sequen ial da a. The model is ained and e alua ed using he
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 871-886
872
CICIoT2023, a comp ehensi e da ase cu a ed by he Canadian Ins i u e o Cybe secu i y [5], which includes a di e se
se o simula ed IoT a ic, encompassing bo h benign and malicious ac i i ies. The s udy has h ee esea ch objec i es:
o imp o e he accu acy in de ec ing lood a acks in he IoT en i onmen , de elop an op imal deep lea ning model
capable o de ec ing comp omises in secu i y wi hin he IoT en i onmen , and le e age he po en ial o deep lea ning
o imp o e bo h he alse posi i e and ue posi i e a e me ics. The esea ch adop s a igo ous expe imen al
me hodology g ounded in deep lea ning p inciples, o assess he model's pe o mance. Key pe o mance me ics, such
as accu acy, p ecision, ecall, F1-sco e, and ROC-AUC, a e used o e alua e he model's e ec i eness in de ec ing lood
a acks. This s udy add esses he ollowing key esea ch ques ions; 1. How can he e iciency o a deep lea ning-based
model be imp o ed in he de ec ion o lood a acks in an IoT con ex ? 2. Wha a e he key pa ame e s ha a e o be
conside ed in de eloping a deep lea ning model o enhance i s applicabili y in iden i ying secu i y comp omises in eal-
wo ld IoT sys ems? 3. How can he p oposed deep lea ning model be designed o imp o e i s a e o ue posi i es while
main aining a low a e o alse ala ms in lood a ack de ec ion? This s udy con ibu es o imp o ing p oac i e h ea
de ec ion sys ems in IoT en i onmen s. The indings a e expec ed o p o ide aluable insigh s o he deploymen o
mo e obus secu i y mechanisms in IoT sys ems, pa icula ly hose ulne able o DDoS- ela ed dis up ions.
Figu e 1 Numbe o DDoS A acks Wo ldwide om 1s Qua e 2023 o 4 h Qua e 2024 (Sou ce: S a is a)
Among he a ious lood a ack and DDoS de ec ion app oaches de eloped in p e ious s udies, se e al signi ican issues
pe sis . Key challenges include he ime equi ed o iden i y a acks, de ec ion accu acy, and he ealism o he app oach.
These challenges o en depend on he ype o da ase and he ea u es selec ed o ep esen he a ack classes. A e iew
o he li e a u e e eals ha many s udies used ou da ed, small, o imbalanced da ase s, which hinde ed he models'
abili y o e ec i ely iden i y ce ain ypes o a acks. Addi ionally, some solu ions sac i iced accu acy o speed o
execu ion. These challenges a e p ima ily due o he da ase s used o ain deep lea ning models. T aining on a mo e
cu en , eal- ime da ase could imp o e he model's abili y o de ec a acks in eal-wo ld scena ios. These limi a ions
highligh he need o u he in es iga ion and he de elopmen o op imal solu ions o enhance he e iciency o lood
a ack de ec ion. Deep lea ning echniques ha e shown p omising esul s, bu se e al issues mus be add essed. Many
s udies ained, es ed, and alida ed hei models on small da ase s, which may no accu a ely e lec eal-wo ld
condi ions. Fu he mo e, some s udies did no add ess he compu a ional complexi y o hei models o p o ide
adequa e in e p e a ion o hei indings, which a e c i ical o p ac ical implemen a ion. This s udy aims o add ess
hese challenges by imp o ing bo h he alse ala m a e and de ec ion accu acy using a mo e ecen and eal- ime
da ase .
The con ibu ion o his wo k is h ee old. Fi s , i con ibu es o he exis ing body o li e a u e by o e ing a compa a i e
assessmen o deep lea ning echniques ailo ed o he de ec ion o lood-based a acks in IoT ne wo ks. The s udy
p esen s quan i iable pe o mance me ics, including accu acy, ecall, and F1-sco e, unde ealis ic condi ions, he eby
o e ing a e e ence poin o u u e expe imen al eplica ion and op imiza ion. Second, he s udy p o ides a model-
d i en pe spec i e on in eg a ing sequen ial lea ning in o in usion de ec ion sys ems. By illus a ing how GRU-based
models can be uned o pa e n ecogni ion in noisy and he e ogeneous IoT a ic, he esea ch ad ances he
me hodological ounda ion o low-o e head, high-accu acy de ec ion in cons ained ne wo k en i onmen s. Las ly,
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 871-886
873
his wo k yields p ac ical alue o he b oade cybe secu i y communi y. Ne wo k enginee s, sys em a chi ec s, and
egula o y s akeholde s can u ilize he indings o in o m design choices in IoT ne wo k de ense a chi ec u es, es ablish
baseline de ec ion capabili ies, and align wi h eme ging s anda ds o secu e de ice in e ope abili y. The s udy hus
con ibu es no only o academic inqui y bu also o applied e o s aimed a s eng hening he esilience o nex -
gene a ion IoT in as uc u es.
The s udy is o ganized in o i e sec ions. Sec ion 1. in oduces he s udy, p o iding an o e iew o he esea ch p oblem
and ques ions. I also p esen s he need o e ec i e de ec ion o lood a acks in IoT ne wo ks. Sec ion 2. p esen s he
backg ound o he s udy, e iews he ele an li e a u e, and summa izes exis ing ela ed wo k. Sec ion 3. p esen s he
concep ual amewo k and me hodology employed in his s udy, including he esea ch design, da a collec ion
p ocedu es, p e-p ocessing s a egies, model a chi ec u e, aining and es ing p o ocols, and pe o mance e alua ion
c i e ia. Sec ion 4. p esen s he esul s and discusses he indings, while Sec ion 5. concludes he pape by discussing
he s udy’s limi a ions and di ec ions o u u e esea ch.
2. Ma e ials And Me hods
This sec ion p esen s he me hodology used in he s udy, including he esea ch design, concep ual amewo k, da a
collec ion, p epa a ion and p e-p ocessing, c oss- alida ion, sugges ed model c ea ion, aining and es ing p ocesses,
pe o mance e alua ion, and me hodological o e iew a e all co e ed in de ail. Figu es 4 and 9 show Py hon code used
o he expe imen s.
2.1. Resea ch Design
This s udy's expe imen al esea ch design en ails he c ea ion and assessmen o a deep lea ning-based model. The
pu pose o he s udy is o de elop a supe io model based on a Ga ed Recu en Uni (GRU) ha enhances he a e a
which lood a acks in IoT a e de ec ed.
2.2. Concep ual F amewo k
The concep ual amewo k encompasses he a ious s ages o he esea ch p ocess. I includes da a acquisi ion and
desc ip ion, da a p epa a ion and p e-p ocessing, de elopmen o he p oposed model, aining and es ing p ocedu es,
c oss- alida ion, and pe o mance me ics. These componen s o m he ounda ion o he de elopmen and e alua ion
o he p oposed op imized neu al ne wo k model.
Figu e 2 Concep ual F amewo k
2.3. Da ase Acquisi ion
The model p oposed in his s udy was ained and es ed using he CICIoT2023. This da ase was c ea ed o ep esen
as closely as possible, eal-wo ld DDOS a ack scena ios, especially hose o lood a acks. I con ains a balanced se o
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 871-886
874
se en ca ego ies o DDoS a acks in he IoT con ex . The CICIoT2023 is publicly a ailable on he Canadian Ins i u e o
Cybe secu i y websi e[5]. In Figu e 3, a comple e b eakdown o he CICIoT2023 wi h he a ious ypes o classes is
p esen ed.
Sou ce: Canadian Ins i u e o Cybe secu i y websi e
Figu e 3 The comple e da ase b eakdown.
Ne o e al. [5] se up se e al de ices ha imi a e a eal-wo ld ins alla ion o IoT de ices and se ices and con igu e
a ic moni o s on hem o cap u e a ack da a. Each a ack in ol es a unique expe imen ha in ol es all ele an
de ices. In he end, he coun o each o he hi y- h ee ca ego ies o a ack is illus a ed as seen in Figu e 3. I is clea
om his g aph ha he au ho s ga he ed an ex ensi e amoun o lood a acks, making his da ase an ideal choice o
aining he deep lea ning model.
2.4. Da a P epa a ion, Cleaning, and P e-p ocessing
Da a p e-p ocessing is essen ial o p epa ing aw da a o deep lea ning models, especially when he da a is incomple e
o inconsis en . In his s udy, he da ase was i s cleaned by emo ing i ele an , edundan , o e oneous en ies,
handling missing alues h ough emo al, and disca ding ou lie s o in ini e alues o ensu e da a in eg i y. Once
cleaned, he da a was ans o med and no malized o i he Ga ed Recu en Uni (GRU) model's inpu equi emen s.
No maliza ion s anda dizes he da a, ensu ing ha all ea u es a e on a compa able scale, which helps imp o e model
e iciency and accu acy. The p e-p ocessing s eps we e execu ed using Py hon lib a ies such as Pandas and NumPy. Key
asks in ol ed:
S anda d Scaling: Each ea u e was scaled o ha e a mean o 0 and a s anda d de ia ion o 1, op imizing ac i a ion
unc ions like sigmoid and anh, which pe o m bes wi h scaled inpu s. This ensu es imp o ed model con e gence.
Label Mapping: Ca ego ical da a in he CICIoT2023 was con e ed in o nume ical o m by assigning a unique in ege o
each ca ego y. This ans o ma ion enabled he neu al ne wo k o p ocess he da a and make p edic ions.
Da a Con e sion: The ex ac ed ea u es and labels we e con e ed in o NumPy a ays, making he da ase compa ible
wi h he p oposed GRU model o aining.
2.5. Model Building and T aining
The p oposed ecu en model is ained wi h he Tenso Flow amewo k. The model was also ained and alida ed
using K- old c oss- alida ion. A e scaling and mapping he labels o be used, he model is de eloped. A sequen ial
neu al ne wo k is de ined wi h a GRU laye o he model. This laye e u ns a linea sequence o da a and has 64 uni s.
A e his laye , a ba ch no maliza ion laye is added o no malize he ou pu om he i s laye , hence ensu ing ha
he aining p ocess emains s able. A new GRU laye o 32 uni s ha e u ns a single ou pu wi h a ba ch no maliza ion
laye is hen added o he model's a chi ec u e. A ully connec ed dense laye wi h a "so max" ac i a ion unc ion is
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 871-886
875
now applied since he model will classi y mul i-classes. Now, he k- old alida ion p ocess is ini ialized o i e wi h he
shu le op ion se o ue. The hype pa ame e s, which a e he lea ning a e, ba ch size, and he numbe o epochs, o
aining he model a e speci ied. The epoch speci ies how many imes he aining p ocess is o be i e a ed. A hi d
dimension is added o change he shape o he model. The model is ained wi h he Ke as amewo k. I has o do wi h
compiling he model, de ining i s me ics, loss, and op imize , and aining he model wi h he aining and alida ion
da a. In he compila ion phase, he loss, op imize , and me ics o he model a e con igu ed. The Spa se Ca ego ical
C oss-en opy is he ypical loss unc ion used in his s udy. Fo he op imize , he Adam algo i hm is used on he
speci ied lea ning a e. This algo i hm adjus s he lea ning a e du ing aining. Du ing e alua ion and aining, he
accu acy me ic will be de e mined and communica ed. The model is now ained using he aining and alida ion da a
(X_ ain, X_ al) and i s co esponding a ge labels (y_ ain, y_ al). The pe o mance o he model on he aining and
alida ion da a is assessed a e each epoch, keeping a eco d o he accu acy me ic.
2.5.1. The Ga ed Recu en Uni
Ga ed Recu en Uni s (GRUs) a e designed o cap u e model dependencies in sequen ial da a.
Wi h sequen ial da a, e e y inpu depends on he ones be o e i . So, GRUs ha e a hidden s a e
(h ) ha ex ac s in o ma ion om he p e ious ime s ep, o upda ing his s a e a each ime
s ep. GRUs a e composed mainly o wo ga es which a e he Upda e (z ) and Rese ga es ( )
which decides how much o he pas in o ma ion is o be passed along and hose ha a e o be o go en, espec i ely.
The new hidden s a e (h ) is a combina ion o he p e ious hidden s a e
(h -1) and a candida e hidden s a e (~h ) whe eas he ~h is a weigh ed combina ion o he
p e ious hidden s a e (h -1) and he cu en inpu (x ). These weigh s a e gi en by he ese ga e. The ma hema ical
exp essions o how he candida e hidden s a e (~h ), upda e ga e (z ), and ese ga e ( ) a e calcula ed a e shown
below:
In he exp essions, he W and U a e he weigh ma ices, (ʘ) deno es elemen -wise mul iplica ion and (σ) is he sigmoid
ac i a ion unc ion. The inal hidden s a e (h ) is hen de i ed om combining he h -1 and he ~h , which a e weigh ed
by he upda e ga e.
2.5.2. C oss- alida ion
This p ocess is ca ied ou o accu a ely es ima e he pe o mance o a deep lea ning model and i s abili y o be
gene alized. This is done o ake ca e o o e i ing. Fo his s udy, he K- old c oss- alida ion me hod was employed.
The da ase is spli in o i e subse s (K olds), and he model is epea edly ained and assessed on po ions o he
subse s. To ake ca e o bias, he da a is i s shu led be o e i is pa i ioned in o olds.
2.5.3. Model E alua ion
The ained model is now es ed on he alida ion da ase . Wi h he inpu ea u es o he alida ion da a (X_ al) and he
a ge labels o he alida ion da a (y_ al), he pe o mance o he model is calcula ed, e u ning he alida ion loss and
alida ion accu acy as i s esul s. Fo his s udy, we accumula e and keep ack o he di e en alida ion accu acy and
alida ion losses a each i e a ion o he c oss- alida ion.
2.6. Pe o mance Me ics
The ained model is es ed on he alida ion da ase , using he inpu ea u es (X_ al) and a ge labels (y_ al) o
calcula e alida ion loss and accu acy. These me ics a e acked ac oss i e a ions du ing c oss- alida ion. The model's
pe o mance is assessed using accu acy, p ecision, F1-sco e, and he ROC cu e. Key pe o mance indica o s include
ue posi i es (TP), ue nega i es (TN), alse posi i es (FP), and alse nega i es (FN). TP e e s o co ec ly p edic ed
posi i e da a poin s, TN o co ec ly p edic ed nega i e da a poin s, FP o inco ec posi i e p edic ions, and FN o
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 871-886
876
inco ec nega i e p edic ions. Accu acy, p ecision, F1-sco e, and ROC cu e a e me ics ha o e quan i a i e
e alua ions o how well he model can spo lood a acks.
• Accu acy sco e e e s o he a io o ue p edic ed labels o he o al numbe o labels. I measu es how
e icien ly he model pe o ms.
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (𝑇𝑃 + 𝑇𝑁)
(𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁) ---------- (1)
Sou ce: G.Ahmed,[39]
• P ecision ocuses on he numbe o he model's p edic ed ue posi i es ha a e eally, ue posi i es.
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 𝑇𝑃
(𝑇𝑃 + 𝐹𝑃) ----------- (2)
Sou ce: G.Ahmed, [39]
• Recei e Ope a ing Cha ac e is ic Cu e (AUC) is a me ic ha gi es a quan i a i e alue o he o e all
classi ica ion pe o mance a all h esholds by he model.
Sou ce: Yousu and Mi [35]
• Recall sco e p o ides a quan i a i e measu e o he p opo ion o ue posi i es p edic ed by he model agains
he ac ual posi i e cases.
𝑅𝑒𝑐𝑎𝑙𝑙 = 𝑇𝑃
(𝑇𝑃 + 𝐹𝑁) ---------- (3)
Sou ce: G.Ahmed,[39]
• F1-sco e p o ides an e alua ion o he pe o mance o he model by calcula ing he mean be ween p ecision
and ecall.
𝐹1 𝑆𝑐𝑜𝑟𝑒 = 2 ∗ (𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 ∗ 𝑅𝑒𝑐𝑎𝑙𝑙)
(𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙) ---------- (4)
Sou ce: B ownlee,[45]
The accu acy (1) o he model is ob ained by di iding he o e all numbe o he model's co ec ly p edic ed cases (TP +
TN) by he o al numbe o p edic ions (TN + TP + FP + FN) i made. Fo he p ecision (2) o he model, he ocus is se
only on he a io o co ec posi i e p edic ions (TP) ou o he o al posi i e p edic ions (TP + FP) by he model.
Rega ding he ecall me ic (3), he o al numbe o co ec ly p edic ed posi i e ins ances (TP) is di ided by he sum o
he numbe o co ec ly p edic ed posi i e ins ances and he numbe o w ongly p edic ed nega i e ins ances (TP + FN).
Las ly, he F1-sco e (4) p esen s quan i a i e da a on he balance be ween he model's p ecision (2) and ecall (3). A e
se e al expe imen s, he pe o mance o he model was measu ed based on accu acy, p ecision, ecall, ROC, and F1-
sco e.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 871-886
877
Figu e 4 Py hon Code showing model pe o mance pe epoch
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 871-886
878
Figu e 5 Calcula ion o he pe o mance me ics and con usion ma ix
Table 1 The pe o mance me ics o he model and hei co esponding alues.
Pe o mance Me ics
Value
Recall
0.61
P ecision
0.63
Accu acy
0.98
F1 sco e
0.61
3. Resul s and Discussion
The e alua ion o he p oposed me hodology, i s pe o mance, and how i compa es o o he ela ed wo ks in de ec ing
lood a acks show ha he ecall ob ained is 0.61, he p ecision ob ained was 0.63, he accu acy ob ained is 0.98, and
he F1 Sco e is 0.61. The p oposed model was es ed on he CICIoT2023, and he esul s o he expe imen s we e
analyzed. The Ga ed Recu en Uni (GRU) algo i hm was implemen ed along wi h Py hon, Ke as, and he Sklea n
lib a ies.
3.1. Model's Pe o mance on he CICIoT2023
The CICIoT2023 used o he expe imen s was he mos ideal. This da ase is new and an imp o emen on i s p e ious
e sions in e ms o size and gene alizabili y. Ne o e al.[5] employed an ex ensi e opology o eal-wo ld IoT de ices
o ob ain he da ase hence making i e y ealis ic and eal- ime. I is wo h no ing ha using such a da ase o aining
he model imp o es i s obus ness. Now, wi h his imp o ed le el o obus ness and, consequen ly, eliabili y, he model
can be used in eal-wo ld scena ios o add o he secu i y o IoT de ices. On aining and e alua ing he p oposed GRU-
based model on his imp o ed and ealis ic da ase o lood a ack de ec ion, i was obse ed ha he model pe o med
e y well in e ms o i s accu acy. This me hod could be he i s o se e al ha employ deep lea ning o de ec ing lood
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 871-886
879
a acks on such a eal- ime da ase . Based on he indings om Table 1, he model su e ed a li le in i s p ecision, ecall,
and F1-sco e. This may be because o he many classes i had o iden i y and co ec ly place. Howe e , i had a nea ly
pe ec sco e in accu acy. This means ha he model co ec ly p edic ed mos o he ins ances ou o he o al in he
da ase . In simple e ms, he model made mo e co ec p edic ions and hus enhanced he de ec ion o lood a acks in
he IoT en i onmen .
3.1.1. Model Valida ion Me ics
The pe o mance o he model in de ec ing he a ious lood a acks was e alua ed using accu acy and model loss. The
accu acy o he model measu ed how e ec i e he model's p edic ion was as compa ed o he ac ual da a. The loss
unc ion o he model was used o measu e i s op imali y. The loss unc ion also shows he le el o e o s in he aining
o alida ion o he model. Acco dingly, a g ea e loss unc ion deno es a model i e a ion ha unde wen poo model
op imiza ion, whe eas a lowe one deno es be e model op imiza ion. Figu es 6 and 7 show he model's aining and
alida ion accu acy and loss unc ion esul s, espec i ely.
Figu e 6 The aining and alida ion accu acy o he model
Figu e 7 The aining and alida ion loss o he model
In Figu e 6, i is seen ha he aining accu acy s eadily ises om he i s epoch o he second, sligh ly alls a he hi d,
and begins a s eady ise om he e o he ou h epoch. F om he e i sha ply ises o he se en h epoch and hen
main ains a s eady ise h ough o he en h epoch. The alida ion accu acy o he model gen ly ises om he i s epoch
o he hi d. I hen sha ply alls om his epoch o he ou h while main aining a s eady le el un il he i h epoch. The
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 871-886
886
[44] P ocopiou, A., Komninos, N., & Doulige is, C. (2019). Fo Chaos: Real ime applica ion DDoS de ec ion using
o ecas ing and chaos heo y in sma home IoT ne wo k. Wi eless Communica ions and Mobile Compu ing, 1–
14. h ps://doi.o g/10.1155/2019/8469410
[45] B ownlee, J. (2022, Augus 19). How o calcula e p ecision, ecall, F1, and mo e o deep lea ning models.
Re ie ed om h ps://machinelea ningmas e y.com/how- o-calcula e-p ecision- ecall- 1-and-mo e- o -deep-
lea ning-models/
[46] Pei, W., e al. (2019). A su ey on ne wo k in usion de ec ion echniques o DDoS a acks. Jou nal o Compu e
Ne wo ks and Communica ions.
