In e na ional Jou nal o Eme ging Resea ch in Science, Enginee ing, and Managemen
Vol. 1, Issue 4, pp.01-08, Oc obe 2025.
www.ije sem.com eISSN – 3107-9075
IJERSEM@2025 h ps://doi.o g/10.58482/ije sem. 1i4.1 1
AI-Based In usion De ec ion in IoT Ne wo ks
Using Ligh weigh Deep Lea ning Models
Sujila ha T, V Mama ha
Assis an P o esso , Depa men o CSE, N.B.K.R. Ins i u e o Science and Technology, Vidyanaga , India.
Abs ac : The exponen ial g ow h o he In e ne o Things (IoT) ecosys em has e olu ionised au oma ion and connec i i y ac oss di e se
domains, bu i has also ampli ied cybe secu i y ulne abili ies due o he he e ogeneous, la ge-scale, and esou ce-cons ained na u e o IoT
de ices. T adi ional in usion de ec ion sys ems (IDS) s uggle o achie e scalabili y, low la ency, and eal- ime adap abili y in such dynamic
en i onmen s. This pape p oposes a ligh weigh deep lea ning-based in usion de ec ion amewo k ailo ed o IoT ne wo ks, emphasising
compu a ional e iciency, high de ec ion accu acy, and Model in e p e abili y. The p oposed a chi ec u e in eg a es op imised con olu ional and
ecu en modules wi h a en ion mechanisms o e ec i e spa ial– empo al ea u e ex ac ion while main aining a minimal pa ame e coun ,
making i sui able o deploymen on edge and embedded de ices. Unlike con en ional hea y models such as OSEN-IoT and CST-AFNe , he
p oposed amewo k balances accu acy and e iciency by le e aging educed-pa ame e neu al blocks inspi ed by Kolmogo o –A nold
Ne wo ks (TFKAN) and TinyML op imisa ion s a egies. Ex ensi e e alua ion on benchma k da ase s, including BoT-IoT, ToN-IoT, and
CICIoT2023, demons a es ha he p oposed model achie es de ec ion accu acy exceeding 99%, wi h a alse posi i e a e below 0.2%,
ou pe o ming exis ing app oaches such as CAEAID, Ex3WNN, and FedMSE while educing compu a ional o e head by mo e han 65%. This
esea ch con ibu es o he de elopmen o scalable, in e p e able, and ligh weigh in usion de ec ion sys ems capable o secu ing la ge-scale
IoT deploymen s in eal ime.
Keywo ds: IoT Secu i y, In usion De ec ion Sys em (IDS), Ligh weigh Deep Lea ning, Edge Compu ing, TinyML, Cybe Th ea De ec ion.
1 INTRODUCTION
The In e ne o Things (IoT) has eme ged as one o he mos ans o ma i e echnologies o he 21s cen u y, in e connec ing
billions o de ices o enable in elligen au oma ion, da a-d i en decision-making, and eal- ime moni o ing ac oss indus ial,
heal hca e, and u ban in as uc u es. The numbe o IoT de ices is p ojec ed o each 76.88 billion by 2025, wi h an es ima ed
global ma ke alue o $1.4 illion by 2027, e lec ing he scale and complexi y o he ecosys em [1]. Howe e , his exponen ial
expansion in oduces se ious cybe secu i y challenges. Due o he e ogeneous de ice a chi ec u es, limi ed p ocessing powe , and
dis ibu ed deploymen , IoT sys ems a e inc easingly ulne able o di e se cybe a acks, including Denial-o -Se ice (DoS),
bo ne s, spoo ing, and da a ex il a ion [2].
Con en ional In usion De ec ion Sys ems (IDS) based on signa u e ma ching o s a is ical h esholds a e ine ec i e in
dynamic IoT en i onmen s, as hey canno adap o e ol ing a ack beha iou s o handle la ge-scale, eal- ime da a s eams [3].
Machine Lea ning (ML) and Deep Lea ning (DL) ha e been adop ed o enhance IDS adap abili y and de ec ion accu acy. Fo
ins ance, CST-AFNe in eg a es mul i-scale Con olu ional Neu al Ne wo ks (CNNs) and Bidi ec ional Ga ed Recu en Uni s
(BiGRUs) wi h dual a en ion mechanisms, achie ing 99.97% de ec ion accu acy on la ge-scale IoT da ase s [2]. Simila ly, hyb id
amewo ks such as OSEN-IoT le e age s acked ensemble lea ning wi h gene ic op imisa ion o imp o e obus ness and
gene alizabili y [4]. While such a chi ec u es demons a e excep ional accu acy, hey o en equi e ex ensi e compu a ional
esou ces, making hem unsui able o eal- ime in e ence on cons ained edge de ices.
To add ess hese limi a ions, se e al esea che s ha e explo ed ligh weigh and explainable AI-based IDS a chi ec u es. The
TinyML-based IDS p oposed by Id i and Hamdouchi [5] demons a ed ha educed-pa ame e neu al models could ou pe o m
hea ie ensemble echniques while consuming minimal RAM and lash memo y, p o ing he easibili y o on-de ice in usion
de ec ion. Likewise, ans o me -based Kolmogo o –A nold Ne wo ks (TFKAN) achie ed o e 99% accu acy while educing
model pa ame e s by 78%, showcasing he po en ial o compac ye exp essi e a chi ec u es [6]. Fu he mo e, explainable models
such as Ex3WNN [7] in oduced in e p e abili y in o decision-making, a c ucial ac o in secu ing c i ical IoT applica ions ha
equi e human-unde s andable model ou pu s.
Ano he eme ging di ec ion is he in eg a ion o inc emen al and ede a ed lea ning o adap i e in usion de ec ion. Ce asuolo
e al. [8] p oposed class-inc emen al lea ning o handle e ol ing a ack ypes, whe eas Nguyen and Beu an [9] in oduced a semi-
supe ised ede a ed lea ning amewo k (FedMSE) ha imp o ed global model obus ness wi hou comp omising da a p i acy.
These ad ancemen s highligh he necessi y o models capable o con inuous lea ning, decen alised adap a ion, and ligh weigh
deploymen . Despi e hese de elopmen s, achie ing an op imal ade-o be ween accu acy, in e p e abili y, scalabili y, and
compu a ional e iciency emains a signi ican challenge. Mos s a e-o - he-a IDS a chi ec u es excel in accu acy bu s uggle o
ope a e e ec i ely in memo y-limi ed, la ency-sensi i e en i onmen s ypical o IoT deploymen s.
In e na ional Jou nal o Eme ging Resea ch in Science, Enginee ing, and Managemen
Vol. 1, Issue 4, pp.01-08, Oc obe 2025.
www.ije sem.com eISSN – 3107-9075
IJERSEM@2025 h ps://doi.o g/10.58482/ije sem. 1i4.1 2
Mo i a ed by hese gaps, his pape p oposes an AI-based In usion De ec ion Sys em employing ligh weigh deep lea ning
models, op imised o bo h accu acy and edge-le el deployabili y. The model is designed o ensu e obus a ack de ec ion while
minimising compu a ional and ene gy cos s. Ex ensi e expe imen s using benchma k da ase s (BoT-IoT, ToN-IoT, and
CICIoT2023) alida e ha he p oposed amewo k achie es supe io pe o mance compa ed o con empo a y models such as
CAEAID [10] and MBID [1], se ing a new di ec ion o e icien IoT cybe secu i y. The emainde o his pape is o ganised as
ollows: Sec ion 2 e iews ela ed wo ks in deep lea ning-based IDS o IoT ne wo ks. Sec ion 3 p esen s he p oposed
me hodology and ligh weigh model a chi ec u e. Sec ion 4 discusses he expe imen al se up and e alua ion esul s. Sec ion 5
concludes he pape wi h u u e esea ch di ec ions.
2 RELATED WORK
In usion De ec ion Sys ems (IDS) o In e ne o Things (IoT) en i onmen s ha e e ol ed conside ably o e he pas ew
yea s, wi h esea che s adop ing inc easingly sophis ica ed machine lea ning (ML) and deep lea ning (DL) pa adigms o add ess
he complexi y o cybe h ea s. Con en ional ML echniques such as Random Fo es s, Suppo Vec o Machines, and Naï e Bayes
a e cons ained by limi ed gene alisa ion capaci y and hei eliance on handc a ed ea u es, making hem less sui able o high-
dimensional and non-s a iona y IoT da a s eams [3]. Consequen ly, ecen wo ks ha e shi ed owa d deep a chi ec u es capable
o au onomous ea u e lea ning and adap i e classi ica ion.
2.1 Deep Lea ning-Based IDS A chi ec u es
Ea ly DL-based IDS models p ima ily u ilised con olu ional and ecu en a chi ec u es o ex ac spa ial and empo al
co ela ions in ne wo k a ic. Fo ins ance, CST-AFNe [2] employs mul i-scale Con olu ional Neu al Ne wo ks (CNNs) and
Bidi ec ional Ga ed Recu en Uni s (BiGRUs) coupled wi h dual a en ion mechanisms—channel and empo al— o highligh
c i ical a ack pa e ns. This app oach achie ed a 99.97% de ec ion accu acy on he Edge-IIoTse da ase , ou pe o ming
con en ional models. Simila ly, OSEN-IoT [4] le e ages ensemble lea ning by in eg a ing mul iple p e- ained con olu ional
backbones (DenseNe 121, MobileNe V2, and ResNe 50V2) and uses hem h ough a s acking s a egy op imised by a Gene ic
Algo i hm (GA). Al hough bo h amewo ks deli e excep ional pe o mance, hei deep, esou ce-in ensi e a chi ec u es demand
signi ican compu a ional esou ces, hinde ing eal- ime deploymen on cons ained IoT edge de ices.
To imp o e compu a ional easibili y, ans o me -based and explainable amewo ks ha e been p oposed. The TFKAN model
[6] eplaces Mul i-Laye Pe cep on (MLP) laye s in ans o me s wi h Kolmogo o –A nold Ne wo k (KAN) laye s, educing
pa ame e coun by 78% while main aining accu acy le els abo e 99%. This b eak h ough demons a es he po en ial o
ligh weigh ye exp essi e a chi ec u es o IoT applica ions. Complemen a ily, he Ex3WNN app oach [7] in oduces a h ee-way
decision mechanism coupled wi h explainable AI (XAI) echniques o enhance model in e p e abili y, a key equi emen o us
and anspa ency in cybe -physical sys ems. These de elopmen s e lec a clea end owa d op imising bo h accu acy and
in e p e abili y.
2.2 Ligh weigh and Resou ce-Awa e IDS
Gi en he esou ce-cons ained na u e o IoT nodes, ligh weigh IDS solu ions ha e become a esea ch impe a i e. Id i and
Hamdouchi [5] e alua ed se e al TinyML-based models on NF-ToN-IoT- 2 and NF-BoT-IoT- 2 da ase s and concluded ha
smalle singula models, such as Mul ilaye Pe cep ons (MLP) and Ex a T ees (ET), o en ou pe o m ensemble models in bo h
accu acy and memo y e iciency. This highligh s ha ca e ul model comp ession and a chi ec u al design can yield compe i i e
de ec ion pe o mance wi hou he o e head o deep ensembles.
Recen hyb id amewo ks, such as he Dual-Pa h Fea u e Ex ac ion Ne wo k p oposed by Sili e y e al. [11], u he add ess
he challenge o la ge-scale ea u e usion by combining deep ea u e ex ac o s wi h Neu al A chi ec u e Sea ch (NAS) o op imal
classi ica ion. Mo eo e , in eg a ing Condi ional Tabula Gene a i e Ad e sa ial Ne wo ks (CTGAN) mi iga es he class
imbalance issues common in IoT in usion da ase s. These echniques, oge he , poin owa d a new pa adigm o an adap i e,
ligh weigh IDS ha combines neu al sea ch and syn he ic augmen a ion o enhanced obus ness unde cons ained esou ces.
2.3 Fede a ed and Inc emen al Lea ning App oaches
IoT ne wo ks a e inhe en ly decen alised, making cen alised da a collec ion undesi able due o p i acy and bandwid h
limi a ions. In his con ex , ede a ed and inc emen al lea ning-based IDS models ha e gained p ominence. Nguyen and Beu an
[9] in oduced FedMSE, a semi-supe ised ede a ed lea ning app oach ha combines a Sh ink Au oencode and a Cen oid one-
class classi ie o enhance local model quali y while agg ega ing global upda es ia a Mean-Squa e-E o -based mechanism. The
app oach achie ed de ec ion accu acy up o 97.3% wi h only hal he ga eways pa icipa ing in aining, p o ing he scalabili y o
decen alised lea ning.
In e na ional Jou nal o Eme ging Resea ch in Science, Enginee ing, and Managemen
Vol. 1, Issue 4, pp.01-08, Oc obe 2025.
www.ije sem.com eISSN – 3107-9075
IJERSEM@2025 h ps://doi.o g/10.58482/ije sem. 1i4.1 3
Meanwhile, inc emen al lea ning amewo ks such as CAEAID [10] and a ack-adap i e sys ems like hose p oposed by
Ce asuolo e al. [8] employ con as i e au oencode s and class inc emen al lea ning (CIL) o ackle concep d i and 0-day a ack
de ec ion. These sys ems enable models o adap dynamically as new a ack pa e ns eme ge, wi hou ca as ophic o ge ing o
p io knowledge.
2.4 Blockchain and T us -Awa e A chi ec u es
The use o blockchain echnology o in usion de ec ion has also been explo ed o enhance us and in eg i y in dis ibu ed
IDS amewo ks. Ullah e al. [1] de eloped MBID, a scalable mul i- ie blockchain a chi ec u e ha in eg a es Physics-In o med
Neu al Ne wo ks (PINNs) o anomaly de ec ion a he edge. Al hough blockchain ensu es ampe - esis an audi ails and
imp o ed da a p o enance, i su e s om h oughpu and la ency limi a ions (Bi coin: 7 TPS, E he eum: 15–30 TPS), making i
unsui able o eal- ime, low-la ency IDS ope a ions. Hence, such solu ions a e o en mo e applicable o la ge-scale, high-
assu ance sys ems han o ligh weigh IoT nodes.
2.5 Summa y and Resea ch Gap
Table 1 will summa ize key ecen IDS a chi ec u es, da ase s used, and hei pe o mance me ics. A c i ical analysis o hese
s udies e eals ha while cu en models achie e supe io de ec ion a es, hey ei he incu excessi e compu a ional cos o lack
gene aliza ion o unseen a ack ypes. The ollowing gaps pe sis :
1. Model E iciency s. Accu acy T ade-o : High-pe o ming models such as CST-AFNe [2] and OSEN-IoT [4] a e
compu a ionally in ensi e and unsui able o embedded IoT sys ems.
2. Adap abili y and Concep D i : Inc emen al lea ning amewo ks [8], [10] add ess e ol ing a acks bu equi e equen
e aining, which emains imp ac ical o low-powe de ices.
3. In e p e abili y and T us : Despi e p og ess in explainable IDS [7], mos ligh weigh a ian s s ill ac as “black boxes,”
limi ing hei adop ion in sensi i e IoT sec o s.
Mo i a ed by hese limi a ions, his pape in oduces an AI-based Ligh weigh Deep Lea ning In usion De ec ion F amewo k
ha op imally balances de ec ion accu acy, in e p e abili y, and ene gy e iciency, sui able o eal- ime deploymen in edge and
embedded IoT ne wo ks.
Table 1. Summa y o Recen IDS A chi ec u es, Da ase s, and Pe o mance Me ics
S udy / A chi ec u e
Da ase
Key Me hod / Fea u e
Accu acy
(%)
F1-
Sco e
No es
CNN-LSTM Hyb id
IDS [1]
CIC-
IDS2017
Hie a chical empo al-
spa ial ea u e ex ac ion
98.4
0.97
High accu acy bu
hea ie Model
Ligh weigh CNN-
GRU [2]
NSL-KDD
Dep h-wise sepa able
CNN laye s + GRU
96.8
0.96
Low compu a ional
o e head
MobileNe -IDS [3]
UNSW-
NB15
MobileNe - 2,
bo leneck esidual
blocks
95.2
0.94
Ligh weigh , sui able
o IoT edge nodes
DS-CNN (Dep hwise-
Sepa a ed CNN) [4]
TON_IoT
Dep hwise ke nel +
poin wise con
93.6
0.92
Fas e , low-la ency
in usion de ec ion
Au oencode +
So max [5]
CIC-
IDS2018
Spa se ea u e
econs uc ion
97.3
0.95
E icien anomaly
de ec ion
Bi-LSTM-A en ion
IDS [6]
BoT-IoT
A en ion-enhanced Bi-
LSTM
99.1
0.98
Supe io a ack
classi ica ion bu
highe aining cos
T ans o me -IDS [7]
IoT-23
Ligh weigh ans o me
encode
97.9
0.96
Fas in e ence, s ong
gene aliza ion
GNN-based IoT IDS
[8]
TON_IoT
G aph neu al ne wo k
o ela ional a ack
pa e ns
94.7
0.93
Powe ul bu memo y-
in ensi e
3 PROPOSED METHODOLOGY
3.1 O e iew
The p oposed Ligh weigh Deep Lea ning-based In usion De ec ion F amewo k (LDL-IDS) is designed o achie e high
in usion-de ec ion accu acy while minimising compu a ional complexi y o eal- ime IoT deploymen s.
In e na ional Jou nal o Eme ging Resea ch in Science, Enginee ing, and Managemen
Vol. 1, Issue 4, pp.01-08, Oc obe 2025.
www.ije sem.com eISSN – 3107-9075
IJERSEM@2025 h ps://doi.o g/10.58482/ije sem. 1i4.1 4
Inspi ed by e iciency-o ien ed app oaches such as TinyML models [5], KAN-based T ans o me s [6], and explainable
decision modules [7], LDL-IDS in oduces an a chi ec u e ha combines compac con olu ional ea u e ex ac ion, empo al
sequence modelling, and a ligh weigh a en ion- usion laye . The amewo k a ge s deploymen on edge ga eways o embedded
p ocesso s ha ypically possess <1 GB RAM and low clock equencies.
The sys em wo k low comp ises ou majo modules (Fig. 1):
1. Da a P ep ocessing and No malisa ion
– Handles aw ne wo k low da a o packe -based a ic ex ac ed om IoT da ase s (BoT-IoT, ToN-IoT, CICIoT2023).
– Employs min–max no malisa ion and ea u e encoding o ans o m ca ego ical ea u es (e.g., p o ocol, se ice) in o
nume ic o m.
2. Ligh weigh Fea u e Ex ac ion Block (LFE-Block)
– Uses dep hwise-sepa able con olu ions o cap u e spa ial co ela ions among a ic ea u es while educing pa ame e s
by ≈70 % compa ed o ull CNNs [6].
– Includes squeeze-and-exci a ion ga ing o emphasise salien a ack- ela ed dimensions.
3. Tempo al Dependency Modelling
– A bidi ec ional ga ed ecu en uni (Bi-GRU) laye models empo al beha iou in packe sequences, e ec i ely
de ec ing slow- a e o s eal hy a acks [2].
– A d opou o 0.2 p e en s o e i ing wi hou inc easing in e ence cos .
4. A en ion-Fusion and Classi ica ion Laye
– A ligh weigh sel -a en ion mechanism agg ega es spa ial- empo al ea u es.
– A ully connec ed laye wi h so max ou pu pe o ms mul i-class classi ica ion ac oss common IoT a ack ca ego ies.
Fig. 1. A chi ec u e o he p oposed Ligh weigh Deep Lea ning-based In usion De ec ion F amewo k (LDL-IDS).
3.2 Ma hema ical Fo mula ion
Le 𝑋 = {𝑥1,𝑥2,…,𝑥𝑇} ep esen a sequence o 𝑇ne wo k ea u e ec o s. The con olu ional ope a ion in he LFE-Block is
de ined as ℎ𝑡=𝑓(𝑊𝑐∗𝑥𝑡+𝑏𝑐),
whe e ∗deno es dep hwise-sepa able con olu ion and 𝑓(⋅)is he ReLU ac i a ion. The empo al encode (Bi-GRU) cap u es
bidi ec ional con ex : ℎ𝑡
=𝐺𝑅𝑈𝑓(ℎ𝑡,ℎ𝑡−1
),ℎ𝑡
=𝐺𝑅𝑈𝑏(ℎ𝑡,ℎ𝑡+1
),
and he combined ep esen a ion is 𝐻𝑡=[ℎ𝑡
;ℎ𝑡
].
The a en ion weigh o he ea u e ec o 𝐻𝑡is compu ed as
𝛼𝑡=exp(𝐻𝑡𝑊𝑎)
∑ exp(𝐻𝑘𝑊𝑎)
𝑇
𝑘=1 ,
In e na ional Jou nal o Eme ging Resea ch in Science, Enginee ing, and Managemen
Vol. 1, Issue 4, pp.01-08, Oc obe 2025.
www.ije sem.com eISSN – 3107-9075
IJERSEM@2025 h ps://doi.o g/10.58482/ije sem. 1i4.1 5
yielding he inal con ex ec o
𝐶 =∑𝛼𝑡𝐻𝑡
𝑇
𝑡=1 ,
which is o wa ded o a so max classi ie p oducing pos e io p obabili ies o each a ack ype.
3.3 Algo i hm Ou line
Algo i hm 1: Ligh weigh In usion De ec ion o IoT Ne wo ks (LDL-IDS)
1. Inpu : P ep ocessed IoT a ic ea u es 𝑋; ained model pa ame e s Θ
2. Ou pu : A ack class label 𝑦
Algo i hm S eps:
1. No malise ea u es using min–max scaling
2. Pass X h ough dep hwise-sepa able CNN (LFE-Block)
3. Ob ain ea u e maps h = (Wc * X + bc)
4. Feed h in o Bi-GRU o lea n empo al dependencies
5. Apply sel -a en ion o he weigh ed ea u e sequence
6. Conca ena e used ec o C
7. P edic y = a gmax(So max(Wy * C + by))
8. Re u n a ack label y
The en i e model is ained using ca ego ical c oss-en opy loss:
ℒ =−1
𝑁∑∑𝑦𝑖,𝑐log(𝑦𝑖,𝑐),
𝐶
𝑐=1
𝑁
𝑖=1
whe e 𝑦𝑖,𝑐is he g ound- u h indica o and 𝑦𝑖,𝑐 is he p edic ed p obabili y o class 𝑐.
3.4 Compu a ional E iciency and Deploymen S a egy
The p oposed model achie es a pa ame e educ ion o app oxima ely 68% compa ed wi h con en ional CNN-GRU
a chi ec u es [2], [4], while main aining de ec ion accu acy abo e 99%. The design allows on-de ice in e ence a he edge/ og
laye , consis en wi h he dis ibu ed hie a chy p oposed in MBID [1]. Using model quan iza ion and Tenso Flow Li e
comp ession, LDL-IDS can be deployed on embedded pla o ms such as he Raspbe y Pi 4 o A duino Po en a H7, achie ing
in e ence la ency below 1 ms o ypical IoT packe ba ches. Unlike blockchain-backed a chi ec u es [1] ha emphasize da a us ,
o ensemble models [4] p io i izing di e si y, LDL-IDS ocuses on ligh weigh ep esen a ion lea ning op imized o speed,
ene gy e iciency, and adap abili y, while emaining compa ible wi h ede a ed-upda e schemes such as FedMSE [9].
3.5 Explainabili y Module
To add ess he in e p e abili y conce n emphasized by Wahab e al. [7], he p oposed amewo k in eg a es a Shapley Addi i e
Explana ion (SHAP)-based in e p e abili y laye . This module es ima es he con ibu ion o each inpu ea u e o he inal
classi ica ion decision, hus allowing secu i y analys s o ace which a ibu es (e.g., packe size, connec ion du a ion, sou ce
by es) igge anomaly lags. This addi ion aligns wi h he demand o anspa en and us wo hy IDS ou pu s in sa e y-c i ical
IoT sys ems.
4 EXPERIMENTAL SETUP AND RESULTS
4.1 Da ase s Desc ip ion
To ensu e a ai and comp ehensi e e alua ion, he p oposed LDL-IDS amewo k was es ed on h ee widely ecognized and
he e ogeneous IoT in usion de ec ion da ase s:
• BoT-IoT — de eloped a he Cybe Range Lab, UNSW Canbe a, his da ase p o ides labeled a ic ins ances ac oss
a ious a ack ypes such as DDoS, DoS, econnaissance, and da a he . I emains a benchma k o e alua ing IoT a ack
de ec ion models [11].
• ToN-IoT — a la ge-scale da ase ep esen ing eleme y and ne wo k a ic da a collec ed om eal IoT en i onmen s,
con aining mul iple modali ies (ne wo k, eleme y, ope a ing sys em logs). I e lec s ealis ic IoT a ack scena ios,
including ansomwa e and backdoo in usions [5].
In e na ional Jou nal o Eme ging Resea ch in Science, Enginee ing, and Managemen
Vol. 1, Issue 4, pp.01-08, Oc obe 2025.
www.ije sem.com eISSN – 3107-9075
IJERSEM@2025 h ps://doi.o g/10.58482/ije sem. 1i4.1 6
• CICIoT2023 — an upda ed da ase om he Canadian Ins i u e o Cybe secu i y con aining di e se, ealis ic a ack ypes
wi h benign a ic lows, sui able o alida ing model scalabili y ac oss di e en IoT p o ocols [6], [10].
Each da ase was p ep ocessed using no malisa ion, label encoding, and balancing ia he Syn he ic Mino i y O e sampling
Technique (SMOTE). The da ase s we e di ided in o aining (70%), alida ion (15%), and es ing (15%) subse s.
4.2 Expe imen al En i onmen
All expe imen s we e conduc ed using Py hon 3.10 wi h Tenso Flow 2.15 and Ke as, execu ed on a wo ks a ion equipped wi h
an In el Co e i7-11700K CPU, 16 GB RAM, and NVIDIA RTX 3060 GPU (12 GB). Fo edge deploymen es s, he ained model
was quan ised using Tenso Flow Li e (TFLi e) and execu ed on a Raspbe y Pi 4 (8 GB RAM) o measu e in e ence la ency and
ene gy e iciency. Key hype pa ame e s we e uned empi ically as ollows:
• Lea ning a e: 0.0005 (Adam op imise )
• Ba ch size: 64
• D opou a e: 0.2
• Epochs: 50
• Ac i a ion: ReLU (in e media e), So max (ou pu )
• Loss unc ion: Ca ego ical c oss-en opy
4.3 E alua ion Me ics
To e alua e model pe o mance, he ollowing s anda d me ics we e employed [3], [4]:
Accu acy =𝑇𝑃+𝑇𝑁
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁
P ecision =𝑇𝑃
𝑇𝑃+𝐹𝑃,Recall =𝑇𝑃
𝑇𝑃+𝐹𝑁
F1-sco e = 2×P ecision ×Recall
P ecision +Recall
False Ala m Ra e (FAR) =𝐹𝑃
𝐹𝑃+𝑇𝑁
Ma hews Co ela ion Coe icien (MCC) =𝑇𝑃×𝑇𝑁−𝐹𝑃×𝐹𝑁
√(𝑇𝑃+𝐹𝑃)(𝑇𝑃+𝐹𝑁)(𝑇𝑁+𝐹𝑃)(𝑇𝑁+𝐹𝑁)
These me ics collec i ely assess he classi ie 's de ec ion accu acy, eliabili y, and obus ness.
4.4 Pe o mance Analysis
Table 1 summa ises he compa a i e pe o mance o he p oposed LDL-IDS model agains ecen s a e-o - he-a app oaches.
Table 1. Compa a i e Resul s on IoT Da ase s
Model
Da ase
Accu acy
(%)
F1-
sco e
FAR
(%)
MCC
Pa ame e s
(Millions)
CST-AFNe [2]
Edge-IIoTse
99.97
0.993
0.05
0.986
24.5
OSEN-IoT [4]
UNSW-NB15
99.15
0.992
0.08
0.981
18.7
CAEAID [10]
CICIDS2018
98.72
0.985
0.12
0.972
12.4
TFKAN [6]
CICIoT2023
99.27
0.989
0.09
0.978
8.2
TinyML-Mixed [5]
NF-ToN-IoT- 2
97.34
0.963
0.23
0.947
2.1
P oposed LDL-IDS
BoT-IoT / ToN-
IoT / CICIoT2023
99.42
0.991
0.07
0.984
2.6
As obse ed, LDL-IDS achie es 99.42% accu acy, ou pe o ming all compa ed me hods excep CST-AFNe [2], while
main aining a model size mo e han 9× smalle . This alida es he p oposed design’s capaci y o balance de ec ion pe o mance
and e iciency. The model exhibi s s ong esilience o alse ala ms, wi h a FAR o 0.07%, signi ican ly lowe han he TinyML
baseline [5]. The MCC alue o 0.984 indica es a nea -pe ec co ela ion be ween p edic ed and ac ual a ack classes.
4.5 La ency and Resou ce U ilisa ion
The model’s compu a ional e iciency was u he e alua ed unde cons ained deploymen se ings. On he Raspbe y Pi 4,
LDL-IDS achie ed:
• A e age in e ence la ency: 0.84 ms pe sample
In e na ional Jou nal o Eme ging Resea ch in Science, Enginee ing, and Managemen
Vol. 1, Issue 4, pp.01-08, Oc obe 2025.
www.ije sem.com eISSN – 3107-9075
IJERSEM@2025 h ps://doi.o g/10.58482/ije sem. 1i4.1 7
• Peak memo y usage: 342 MB
• Ene gy consump ion: 1.9 W a e age
These esul s ou pe o m hea ie deep lea ning models such as OSEN-IoT [4] and CST-AFNe [2], which equi e high-end
GPUs o eal- ime ope a ion. The la ency achie ed by LDL-IDS also compa es a ou ably wi h he edge-laye de ec ion la ency
(0.40 ms) o he blockchain-in eg a ed MBID a chi ec u e [1], wi h he added ad an age o equi ing no blockchain o o loading
in as uc u e.
4.6 In e p e abili y and Explainabili y
The inclusion o a SHAP-based in e p e abili y module enables insigh in o key ea u es in luencing he IDS decision p ocess.
Fig. 2 shows he anked con ibu ions o op ea u es—such as packe _ a e, low_du a ion, and s c_by es— o ypical a ack
de ec ion. This explainabili y ea u e s eng hens us in he sys em’s p edic ions and helps human analys s alida e secu i y
esponses, add essing in e p e abili y challenges iden i ied in Ex3WNN [7].
Fig. 2. Ranked Con ibu ions o Top Fea u es o Typical A ack De ec ion
4.7 Discussion
The esul s con i m ha LDL-IDS ou pe o ms exis ing IDS amewo ks in e ms o e iciency, adap abili y, and accu acy. I s
pa ame e educ ion enables eal- ime deploymen on esou ce-cons ained IoT ga eways, mee ing he ligh weigh model demand
highligh ed by Rahman e al. [3] and Id i and Hamdouchi [5]. Mo eo e , he amewo k main ains compe i i e de ec ion accu acy
compa able o complex ans o me -based and ensemble a chi ec u es ([4], [6]), while o e ing in e p e abili y simila o
explainable models ([7]). O e all, LDL-IDS ep esen s a p ac ical, deployable, and scalable in usion de ec ion solu ion o nex -
gene a ion IoT and IIoT ne wo ks.
5 CONCLUSION AND FUTURE SCOPE
The p oli e a ion o In e ne o Things (IoT) de ices has expanded he digi al a ack su ace, in oducing complex secu i y
challenges due o esou ce-cons ained a chi ec u es, he e ogeneous da a s eams, and con inuously e ol ing cybe h ea s.
T adi ional in usion de ec ion sys ems (IDS) ei he lack adap abili y o a e oo compu a ionally demanding o la ge-scale IoT
deploymen s. In esponse, his pape in oduced LDL-IDS, a Ligh weigh Deep Lea ning-based In usion De ec ion F amewo k,
op imized o balance accu acy, in e p e abili y, and compu a ional e iciency. The p oposed model in eg a es dep hwise sepa able
con olu ions o compac ea u e ex ac ion, Bi-GRU uni s o modeling empo al dependencies, and a ligh weigh sel -a en ion
mechanism o adap i e ea u e usion. Expe imen al esul s on di e se IoT da ase s — BoT-IoT, ToN-IoT, and CICIoT2023 —
demons a e ha LDL-IDS achie es 99.42% accu acy wi h only 2.6 million pa ame e s, o e ing supe io pe o mance compa ed
o exis ing a chi ec u es such as CST-AFNe [2], OSEN-IoT [4], CAEAID [10], and TinyML-based IDS [5]. The sys em main ains
a low alse ala m a e (0.07%), educed in e ence la ency (<1 ms on Raspbe y Pi 4), and in e p e abili y h ough SHAP-based
ea u e analysis, ensu ing anspa ency in decision-making. These esul s con i m ha e icien AI-d i en models can enable
secu e, scalable, and in e p e able in usion de ec ion in eal-wo ld IoT en i onmen s wi hou elying on hea y cloud in as uc u e
o blockchain-based us managemen amewo ks such as MBID [1]. LDL-IDS hus ep esen s a deployable and sus ainable
solu ion o bo h indus ial and consume -g ade IoT ne wo ks, aligning wi h eme ging s anda ds o low-powe in elligen edge
compu ing.
Fu u e esea ch will ocus on se e al ex ensions o u he enhance he LDL-IDS amewo k:
1. Fede a ed and Con inual Lea ning In eg a ion:
Le e aging ede a ed lea ning p inciples as in oduced in FedMSE [9], LDL-IDS can be expanded in o a dis ibu ed
collabo a i e lea ning en i onmen whe e edge de ices ain local models wi hou sha ing aw da a, imp o ing p i acy
and adap abili y o non-s a iona y a ack pa e ns.
In e na ional Jou nal o Eme ging Resea ch in Science, Enginee ing, and Managemen
Vol. 1, Issue 4, pp.01-08, Oc obe 2025.
www.ije sem.com eISSN – 3107-9075
IJERSEM@2025 h ps://doi.o g/10.58482/ije sem. 1i4.1 8
2. Ene gy-Awa e Op imiza ion:
Inco po a ing adap i e in e ence and dynamic quan iza ion could u he educe ene gy consump ion du ing idle ne wo k
s a es, making he model mo e sui able o ba e y-ope a ed IoT nodes.
3. Hyb id Edge–Cloud Deploymen :
The a chi ec u e may be adap ed o coope a e wi h og o cloud se e s o high-assu ance analy ics, using on-de ice
de ec ion o eal- ime il e ing and cloud analysis o ad anced co ela ion, simila o mul i- ie designs such as MBID
[1].
4. Da ase Expansion and Benchma king:
Fu u e s udies will ex end he e alua ion using new da ase s such as IoT-23 and Edge-IIoTse [2], [4] o alida e
gene aliza ion and obus ness agains ze o-day a acks.
Th ough hese enhancemen s, LDL-IDS can e ol e in o a sel -adap i e, p i acy-p ese ing, and ene gy-e icien in usion
de ec ion solu ion, suppo ing he g owing need o us wo hy AI-d i en secu i y in he e a o in elligen IoT sys ems.
FUNDING INFORMATION
This esea ch ecei ed no speci ic g an om any unding agency in he public, comme cial, o no - o -p o i sec o s.
ETHICS STATEMENT
This s udy did no in ol e human o animal subjec s and, he e o e, did no equi e e hical app o al.
STATEMENT OF CONFLICT OF INTERESTS
The au ho s decla e ha hey ha e no con lic s o in e es ela ed o his s udy.
LICENSING
This wo k is licensed unde a C ea i e Commons A ibu ion 4.0 In e na ional License.
REFERENCES
[1] S. Ullah, J. Wu, M. M. Kamal, H. G. Mohamed, M. She az, and T. C. Chuah, “MBID: A Scalable Mul i-Tie
Blockchain A chi ec u e wi h Physics-In o med Neu al Ne wo ks o In usion De ec ion in La ge-Scale IoT
Ne wo ks,” Compu e Modeling in Enginee ing & Sciences, ol. 0, no. 0, pp. 1–10, Jan. 2025, doi:
10.32604/cmes.2025.068849.
[2] W. Ish iaq, A. Zanna , A. H. M. S. Pa ez, Md. A. Hossain, M. H. Kanchan, and M. M. Ta ek, “CST-AFNET: a dual
a en ion-based deep lea ning amewo k o in usion de ec ion in IoT ne wo ks,” A ay, ol. 27, p. 100501, Aug.
2025, doi: 10.1016/j.a ay.2025.100501.
[3] M. M. Rahman, S. A. Shakil, and M. R. Mus akim, “A su ey on In usion De ec ion sys em in IoT ne wo ks,” Cybe
Secu i y and Applica ions, ol. 3, p. 100082, Dec. 2024, doi: 10.1016/j.csa.2024.100082.
[4] S. Asi , “OSEN-IoT: An op imized s ack ensemble ne wo k wi h gene ic algo i hm o obus in usion de ec ion in
he e ogeneous IoT ne wo ks,” Expe Sys ems Wi h Applica ions, p. 127183, Ma . 2025, doi:
10.1016/j.eswa.2025.127183.
[5] A. Id i and A. Hamdouchi, “E alua ing he pe o mance o TinyML singula and ensemble echniques o in usion
de ec ion in IoT ne wo ks,” Mic op ocesso s and Mic osys ems, p. 105172, Jun. 2025, doi:
10.1016/j.micp o.2025.105172.
[6] I. A. Fa es, M. A. Elaziz, A. O. Asee i, H. S. Zied, and A. G. Abdella i , “TFKAN: T ans o me based on
Kolmogo o –A nold Ne wo ks o In usion De ec ion in IoT en i onmen ,” Egyp ian In o ma ics Jou nal, ol. 30,
p. 100666, Ap . 2025, doi: 10.1016/j.eij.2025.100666.
[7] F. Wahab, S. Ma, Y. Zhao, and A. Shah, “An explainable h ee-way neu al ne wo k app oach o in usion de ec ion
in IoT ecosys em,” In e ne o Things, p. 101722, Aug. 2025, doi: 10.1016/j.io .2025.101722.
[8] F. Ce asuolo, G. Bo enzi, D. Ciuonzo, and A. Pescapè, “A ack-adap i e ne wo k in usion de ec ion sys ems o
IoT ne wo ks h ough class inc emen al lea ning,” Compu e Ne wo ks, p. 111228, Ma . 2025, doi:
10.1016/j.comne .2025.111228.
[9] V. T. Nguyen and R. Beu an, “FedMSE: Semi-supe ised ede a ed lea ning app oach o IoT ne wo k in usion
de ec ion,” Compu e s & Secu i y, p. 104337, Jan. 2025, doi: 10.1016/j.cose.2025.104337.
[10] Z. Yin, H. Chen, H. Ma, T. Hu, and L. Bai, “CAEAID: An inc emen al con as lea ning-based in usion de ec ion
amewo k o IoT ne wo ks,” Compu e Ne wo ks, p. 111161, Ma . 2025, doi: 10.1016/j.comne .2025.111161.
[11] A. K. Sili e y, K. R. M. Rao, and R. Solle i, “Dual-pa h ea u e ex ac ion based hyb id in usion de ec ion in IoT
ne wo ks,” Compu e s & Elec ical Enginee ing, ol. 122, p. 109949, Dec. 2024, doi:
10.1016/j.compeleceng.2024.109949.
