TELK OMNIKA T elecommunication, Computing, Electr onics and Contr ol V ol. 23, No. 4, August 2025, pp. 1128 ∼ 1136 ISSN: 1693-6930, DOI: 10.12928/TELK OMNIKA.v23i4.26989 ❒ 1128 EdgeShield: a r ob ust and agile cybersecurity ar chitectur e f or the inter net of medical things Anass Misbah 1 , Anass Seb bar 2 , Imad Hadi 1 1 IPIM Laboratory , National School of Applied Sciences, Sultan Moulay Slimane Uni v ersity , Beni Mellal, Morocco 2 International Uni v ersity of Rabat, ESIN, TICLab, Rabat, Morocco Article Inf o Article history: Recei v ed Feb 14, 2025 Re vised May 13, 2025 Accepted May 26, 2025 K eyw ords: Data processing Dimensionality reduction Feature selection Federated learning Internet of medical things security Intrusion detection systems Principal component analysis ABSTRA CT W e present EdgeShield, a lightweight pipeline that str eamlines internet of medi- cal things (IoMT) traf c analysis by pairing aggressi v e dimensionality-reduction with federated model aggre g ation. It emplo ys systematic preprocessing, ad- v anced feature selection, and rob ust sampling to reduce computational o v er - head while enhancing performance. Through feature engineering techniques such as principal component analysis (PCA), tar geted feature selection, and em- bedding methods, EdgeShield reduces dataset dimensionality by 96%, enabling near real-time detection and pre v ention of c yber attacks on resource-constrained edge de vices. T o ha rden the IoMT perimeter , EdgeShield trains ten lightweight edge models in just 54s and mer ges their paramet ers into a single global clas- sier with ne gligibl e e xtra delay . This method requires no additional training or predictions, thus accelerating deplo yment. Additionally , by using a com- pact dataset with v e top-performing features and PCA with tw o components, EdgeShield consistently achie v es accurac y le v els e xceeding 99.2% for indi vid- ual edge models and the consolidated global model. W ith a b uilt-in continuous impro v ement loop, EdgeShield dynamically adapts to emer ging data patterns and operational conditions, dri ving substantial adv ancements in IoMT ecosys- tem management. This approach deli v ers both rapid machi ne learning model deplo yment and rob ust c yber at tack detection, illustrating its potential to re v olu- tionize IoMT security and ele v ate healthcare data inte grity . This is an open access article under the CC BY -SA license . Corresponding A uthor: Anass Misbah IPIM Laboratory , National School of Applied Sciences, Sultan Moulay Slimane Uni v ersity Beni-Mellal 23000, Morocco Email: anassmisbah@gmail.com 1. INTR ODUCTION Internet of medical things (IoMT) de vices no w enable clinicians to monitor patients continuously and in real time, openi n g ne w doors for personalised care [1], [2]. Despite of fering considerable benets such as impro v ed diagnostic accurac y and cost-ef fecti v e care, the highly connected IoMT infrastructure also introduces major challenges. Chief among these are data security threats, limited computational resources on edge de vices, and increasing v olumes of medical information that must be processed promptly [3], [4]. Cyberattacks on IoMT systems can jeopardize patient pri v ac y , disrupt vital services, and e v en endanger li v es [5], [6]. In response to these challenges, we propose EdgeShield, a comprehensi v e frame w ork designed to streamline IoMT data processing while enhancing c yber re silience. This proposed approach is applied to CICIoMT2024 dataset [7]. EdgeShield inte grates systematic preprocessing, adv anced feature selection (FS), J ournal homepage: http://journal.uad.ac.id/inde x.php/TELK OMNIKA Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 1129 controlled sampling, and a no v el model aggre g ation technique to bolster performance and ef cienc y . Central to its design is an aggressi v e dimensionality reduction pipeline: tar geted FS methods and principal component analysis (PCA) [8]-[10] together reduce the dataset’ s dimensionality by approximately 96%, thereby lo wering the computational b urden and enabling real-time analysis on resource-constrained edge de vices. T o further strengthen c yber defense, EdgeShi eld trains 10 edge models in a mere 54 seconds, consol- idating them into a unied global model with near -zero addit ional latenc y . By le v eraging a compact dataset of v e top-performing features and tw o PCA components, both indi vidual edge models a nd the aggre g at ed global model consistently achie v e accurac y rates e xceeding 99.2%. This syner gy of rapid dimensionality reduction, real-time analytics, and rob ust aggre g ation pa v es the w ay for more secure, adapti v e, and ef cient IoMT ecosys- tems. The remainder of this paper is structured as follo ws. Section 2 pro vides an o v ervie w of related w ork in IoMT security and dimensionality reduction approaches. Section 3 outlines the EdgeShield methodology and its k e y components, while section 4 concludes with a discussion of the results, open challenges, and directions for future research. 2. METHOD 2.1. Related w ork This section highlights the role of PCA in di v erse elds, the adv antages of combining PCA with FS, and related frame w orks that address security and ef cienc y in IoMT systems. 2.1.1. PCA in v arious domains Across disciplines from po wer-grid monitoring to image recognition, researchers rely on PCA to compress high-dimensional data while preserving most of the signal. F or e xample, the w ork of [8] combined PCA with support v ector machines (SVM)-assisted neural netw orks and cut state-of-char ge prediction error for LiFePO4 batteries to a v ery lo w v alue, while another study used PCA alongside gradient tree boosting to predict biocrude oil yields with near -perfect correlation [9]. Similarly , PCA-dri v en dimensionality reduction enhanced freshness detection of fruits and v e getables [10], optimized data aggre g ation in IoT netw orks [11], and impro v ed feature learning in a deep k ernel PCA frame w ork [12]. Its broad applicability—spanning coal outb urst prediction [13], lithium e xploration [14], and hotel occupanc y forecasting [15] underscores PCA ’ s v ersatility in boosting model accurac y and interpretability . 2.1.2. Combining FS with PCA FS remo v es less informati v e attrib utes from high-dimensional datasets, thereby reducing noise and impro ving model performance [16]. When combined with PCA, FS ensures that only the most rele v ant fea- tures contrib ute to the principal components, yielding a more ef cient representation of the data. F or instance, a mut ual-information-based technique has been proposed for better intrusion detection in IoMT systems [17], whereas distance correlation has been applied to random forest (RF) models to handle high-dimensional, non- linear data [18]. Empirical e vidence further supports the syner gy of PCA and FS for impro v ed accurac y in classication tasks [19]-[21], especially i n resource-constrained en vironments lik e IoMT . Moreo v er , carefully orchestrating FS before PCA has been sho wn to enhance clustering and reduce computational o v erhead [22]. 2.1.3. Similar appr oaches in IoMT security Numerous frame w orks address IoMT security from dif ferent angles. F or e xample, a pri v ac y-a w are system inte grates Federated Learning and eXplainable AI for intrusion detection [23], while another focuses on machine learning algorithms to safe guard IoT -based healthcare data in real time [24]. IoMT -SAF intro- duces an ontological method to systematically e v aluate de vice vulnerabilities and recommend tail ored defense strate gies [25]. Although these solutions mitig ate v arious security risks, most still f ace challenges in handling lar ge, heterogeneous datasets and ensuring lo w-latenc y performance—critical f actors for IoMT en vironments. The proposed EdgeShield frame w ork (detailed in subsequent sect ions) aims to ll these g aps by le v eraging an ef cient dimensionality reduction pipeline, model aggre g ation with minimal o v erhead, and high-accurac y threat detection. 2.2. Main appr oach EdgeShield pro v i des a streamlined yet ef fecti v e frame w ork for processing internet of medical things (IoMT) data, balancing high accurac y with lo w computational o v erhead on resource-constrained edge de vices. Edg eShield: a r ob ust and a gile cyber security ar c hitectur e for the internet of medical things (Anass Misbah) Evaluation Warning : The document was created with Spire.PDF for Python.
1130 ❒ ISSN: 1693-6930 This sec tion details each step of the pi peline—ranging from optional preprocessing to nal model aggre g a- tion—and e xplains ho w the frame w ork’ s components t together . 2.2.1. Ov erall framew ork Figure 1 of fers an o v ervie w of EdgeShield’ s w orko w . Data rst enters a (potentially optional) pre- processing phase. From there, the system applies FS techniques, performs controlled sampling, and e x ecutes PCA to reduce dimensionality . Multiple lightweight models (simulating IoMT edge de vices) are trained on these reduced datasets, and their trained parameters are aggre g ated in a fog or cloud en vironment to form a rob ust global model. Finally , the aggre g ated model returns to edge de vices, along with updated security rules, closing the continuous impro v ement loop. Figure 1. EdgeShield global architecture: from optional data preprocessing through FS, sampling, PCA-based reduction, local model training at the edge, and aggre g ation into a global model 2.2.2. Data pr epr ocessing Although CICIoMT2024 is generally deli v ered in a well-structured format, data preprocessing may still be rele v ant when w orking with ra w or di v erse IoMT data sources. The k e y steps are: - Cleaning: remo v es noise, outliers, or inconsistent entries. - Scaling: standardizes feature magnitudes (e.g., via z-score normalization). - Normalization: maps features onto a uniform scale, often [0 , 1] . - Harmonization: aligns disparate data from multiple sources, ensuring coherent structure. In man y IoMT applications, these steps signicantly reduce data heterogeneity , thereby impro ving subsequent modeling. F or CICIoMT2024, these tasks are minimal because the dataset arri v es precleaned. TELK OMNIKA T elecommun Comput El Control, V ol. 23, No. 4, August 2025: 1128–1136 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 1131 2.2.3. F eatur e selection Ne xt, we identif y and retain only the most inuential features—a crucial step to reduce computational cost without harming accurac y . Figure 2 compares four primary methods: - RF feature importance: e v aluates ho w often each feature contrib utes to reducing impurity across decision trees as sho wn in Figure 2(a). - Mutual information (MI): quanties the shared information between a gi v en feature and the tar get label as sho wn in Figure 2(b). - Gradient boosting (GB): uses feature importance metri cs deri v ed from boosting iterations as sho wn in Figure 2(c). - PCA loadings: determines contrib utions of original features to principal components (though PCA is typi- cally an unsupervised method) as sho wn in Figure 2(d). (a) (b) (c) (d) Figure 2. Feature importance rankings from four selection approaches: (a) RF , (b) MI, (c) GB, and (d) PCA loadings (do wn-right) consistently highlight inter -arri v al time (IA T) and related netw ork beha vior features as top contrib utors to attack detection T able 1 further compares their accurac y and computational cost on CICIoMT2024. Although mutual information can yield slightly higher accurac y , it demands signica n t ly more runtime. RF strik es a practical balance, consistently achie ving near -optimal results. W e ultimately k eep the top v e features from whiche v er method performed best under time/accurac y constraints (in man y e xperiments, this w as RF feature importance). T able 1. Comparison of FS methods on CICIoMT2024 FS method T op 10 acc. T ime (T op 10) T op 5 acc. T ime (T op 5) RF importance 0.9972 23m 9.9s 0.9960 10m 30s PCA loadings 0.8920 66m 61s 0.8326 39m 33s Gradient boosting (10 k) 0.9793 12m 39s 0.9789 26.8s Mutual info 0.9979 1010m 22.9s 0.9976 96m 29.2s Edg eShield: a r ob ust and a gile cyber security ar c hitectur e for the internet of medical things (Anass Misbah) Evaluation Warning : The document was created with Spire.PDF for Python.
1132 ❒ ISSN: 1693-6930 2.2.4. Data sampling and PCA Random sampling, since CICIoMT2024 is fully labeled a n d already representati v e, we sample 100– 1000 ro ws to manage computational o v erhead. This approach balances data di v ersity with speed, especially on edge de vices where memory and processing po wer are scarce . Acti v e learning [26] though benecial for unlabeled or partially labeled datasets is not mandatory here. Dimensionality reduction o v er PCA, PCA rotates the original ax es so that the resulting components are orthogonal and capture descending amounts of v ariance. By retaining just the top components (tw o in our e xperiments), we drastically lo wer dimensionality while preserving k e y v ariance. Figures 3(a) and (b), Figures 4(a) and (b) sho w ho w PCA consistently enhances accurac y across dif ferent sample sizes, particularly when combined with a carefully chosen subset of features (top v e). (a) (b) Figure 3. Sampling size vs accurac y of (a) all features and (b) all features + PCA. PCA stabilizes performance, especially as the dataset size v aries from 100 to 1000 (a) (b) Figure 4. Sampling size vs accurac y using: (a) top v e features and (b) top v e + PCA. This further reduces noise and can approach near -perfect detection rates 2.2.5. Edge model training and perf ormance T o simulate ten IoMT edge de vices, we split the dataset into ten balanced subsets, each retaining approximately 1% of labeled instances for each attack type and benign traf c. This design addresses real- w orld IoMT scenarios where each de vice sees only a fraction of o v erall traf c. W e then train lightweight RF (or XGBoost) classiers on these subsets. Figure 5 sho ws that despite minimal training time (under a minute in total), each local model achie v es strong accurac y . TELK OMNIKA T elecommun Comput El Control, V ol. 23, No. 4, August 2025: 1128–1136 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 1133 Figure 5. Accuracies of the ten local edge models (each subset co v ers 1% of dataset labels). The uniform performance underscores the representati vity of each sub-dataset 2.2.6. Global model construction and aggr egation Once trained, each local model’ s parameters are transferred to a fog or cloud node for further e v alua- tion and consolidation into a single RF ensemble. T able 2 outlines four sub-model e xperiments on a 5,000-ro w sample. Combining PCA with the top v e features (Exp. 4) yields a high accurac y of 0.9984, slightly outper - forming the all-features approach. T able 2. Comparison of Sub-model e xperiments (5,000-ro w subset, train/test) Exp. Features PCA Model Sub-models Meta -acc. 1 All feats No RF 10 0.9664 2 All feats Y es RF 10 0.9978 3 All feats Y es XGB 10 0.9915 4 T op 5 Y es RF 10 0.9984 Aggre g ation mechanics, EdgeShield assembles the global model by mer ging the decision trees learned on each de vice into one RF ensemble. Because no retraining is needed, the step adds virtually zero latenc y while raising accurac y , as the consolidated classier benets from the di v ersity of local edge models. Pri v ac y is maintained, as ra w data remains on each de vice, only the learned parameters (tree estimators) are shared. 2.2.7. Global model v alidation and adjustment After aggre g ation, the meta-model is subjected to v arious data range tests (e.g., 100 to 1000 sample s) to c on rm stability and generalization. Figures 6(a) and (b), Figures 7(a) and (b) depict ho w PCA usage, FS strate gy , and chosen model type (RF vs XGB) af fect nal accurac y . W e observ e that PCA + top- v e FS can surpass 99% accurac y under multiple congurations. A summary is gi v en in T able 3. (a) (b) Figure 6. Aggre g ated RF meta-model: (a) with PCA vs (b) without PCA using all features. PCA consistently yields higher and more stable accurac y Edg eShield: a r ob ust and a gile cyber security ar c hitectur e for the internet of medical things (Anass Misbah) Evaluation Warning : The document was created with Spire.PDF for Python.
1134 ❒ ISSN: 1693-6930 (a) (b) Figure 7. The RF-based approach can slightly outperform XGB in certain scenarios: (a) XGBoost with PCA on all features and (b) RF with PCA on top-5 features T able 3. Global meta-model accurac y results for v arious data ranges Exp. Acc. Range (200–400 samples) (400–600 samples) (+600 samples) 1 0.92–0.98 > 0.95 > 0.95 > 0.97 2 0.95–0.99 > 0.98 > 0.98 Up to 0.999 3 0.95–0.99 > 0.98 > 0.98 Up to 0.998 4 0.9825–0.9990 > 0.9925 > 0.9925 > 0.9990 2.2.8. Continuous impr o v ement EdgeShield incorporates a self-learning loop to sustain reliability in e v olving IoMT en vironments. As ne w data accumulates, each edge de vice updates its local model, whose parameters are then mer ged into an upgraded global classier . This iterati v e o w aids in: - Adaptation: coping with emer ging attack patterns or changing traf c beha viors. - Scalability: supporting more de vices or e xpanded datasets as IoMT netw orks gro w . - Ef cienc y: preserving pri v ac y by ne v er sharing ra w data and only e xchanging tree estimators. 3. RESUL TS AND DISCUSSION EdgeShield compresses IoMT traf c into just v e discriminati v e features and tw o principal com- ponents, allo wing each of ten edge de vices to nish training in under a mi nute yet still e xceed 99% local accurac y; when these lightweight RF models are federated, meta-accurac y climbs to 0.998 while pri v ac y is preserv ed because ra w data ne v er lea v e the de vice. The combination of aggressi v e feature pruning and PCA, rather than the specic base learner , dri v es the g ain, yielding performances that remain abo v e 95% e v en on 200-ro w subsets and scale to near perfect detection as sample size gro ws. Compared with recent feder - ated IoMT intrusion-detection systems that top out at 95–98% accurac y [23], [24], EdgeShield impro v es the state-of-the-art by 1–4 percentage points and compresses end-to-end processing time from hours to minutes, pro viding a practical, real-time and pri v ac y-a w are defence for resource-constrained medical de vices. 4. CONCLUSION By uniting f ast FS, tw o-component PCA and a pri v ac y-preserving aggre g ation of ten edge classi- ers, EdgeShield cuts compute load yet still surpasses 99% detection accurac y . Its modular design allo ws for continuous updates, making it well-suited to adapt as ne w threats emer ge or IoMT de vice beha viors shift. Look- ing ahead, future w ork could e xplore inte grating e xplainable AI (XAI) techniques to increase transparenc y in decision-making, applying more adv anced sampling or acti v e learning strate gies to reduce labeling costs, and e xtending model aggre g ation approaches be yond RF for impro v ed scalability . Such de v elopments stand to fur - ther strengthen EdgeShield’ s role as a secure, ef cient, and adaptable frame w ork in ne xt-generation healthcare systems. TELK OMNIKA T elecommun Comput El Control, V ol. 23, No. 4, August 2025: 1128–1136 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 1135 A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the Contrib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib u- tions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu Anass Misbah √ √ √ √ √ √ √ √ √ √ √ √ Anass Sebbar √ √ √ √ √ √ Imad Hadi √ √ √ √ C : C onceptualization I : I n v estig ation V i : V i sualization M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administration V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting CONFLICT OF INTEREST ST A TEMENT Authors state no conict of interest. D A T A A V AILABILITY The data that support the n di ngs of this study are a v ailable from the corresponding author , [A Mis- bah], upon reasonable request. REFERENCES [1] M. Sig ala, A. Beer , L. Hodgson, and A. O’Connor , “Big data for measuring the impact of tourism economic de v elopment pro- grammes: A process and quality criteria frame w ork for using big data, ” Big Data and Inno vation in T ourism, T r avel, and Hospitality: Mana g erial Appr oac hes, T ec hniques, and Applications , pp. 57–73, 2019, doi: 10.1007/978-981-13-6339-9 4. [2] G. Nguyen et al., “Machine Learning and Deep Lea rning frame w orks and libraries for lar ge-scale data mining: a surv e y , ” Articial Intellig ence Re vie w , v ol. 52, no. 1, pp. 77–124, Jun. 2019, doi: 10.1007/s10462-018-09679-z. [3] R. Dwi v edi, D. Mehrotra, and S. Chandra, “Potential of Internet of Medical Things (IoMT) applications in b uilding a smart health- care system: A systematic re vie w , ” J ournal of Or al Bi olo gy and Cr aniofaci al Resear c h , v ol. 12, no. 2, pp. 302–318, 2022, doi: 10.1016/j.jobcr .2021.11.010. [4] J. P . A. Y aacoub, H. N. Noura, O. Salman, and A. Chehab, “Ethical hacking for IoT : Security issues, challenges, solutions and recommendations, ” Internet of Things and Cyber -Physical Systems , v ol. 3, pp. 280–308, 2023, doi: 10.1016/j.iotcps.2023.04.002. [5] Gerodimos , L. Maglaras, M. A. Ferrag, N. A yres, and I. Kantza v elou, “IoT : Communication protocols and security threats, ” Internet of Things and Cyber -Physical Systems , v ol. 3, pp. 1–13, 2023, doi: 10.1016/j.iotcps.2022.12.003. [6] M. L. Hernandez-Jaimes, A. Martinez-Cruz, K. A. Ram ´ ırez-Guti ´ errez, and C. Fere grino-Uribe, “ Articial intelligence for IoMT security: A re vie w of intrusion detection systems, attacks, datasets and Cloud–F og–Edge architectures, ” Internet of Things (Nether - lands) , v ol. 23, 2023, doi: 10.1016/j.iot.2023.100887. [7] S. Dadkhah, E. C. P . Neto, R. Ferreira, R. C. Molokwu, S. Sade ghi, and A. Ghorbani, “CICIoMT2024: Attack V ectors in Healthcare de vices-A Multi-Protocol Dataset for Assessing IoMT De vice Security , ” J ournal of Computer Communications , 2024. [8] C. M ehta, A. V . Sant, and P . Sharma, “SVM-assisted ANN model with principal component analysis based dimensionality reduction for enhancing state-of-char ge estimation in LiFePO4 batteries, ” e-Prime - Advances in Electrical Engineering , Electr onics and Ener gy , v ol. 8, 2024, doi: 10.1016/j.prime.2024.100596. [9] T . Katongtung, T . Onsree, K. Y . T ippaya w ong, and N. T ippaya w ong, “Predicti on of biocrude oil yields from h ydrothermal lique- f action using a gradient tree boosting machine approach with principal component analysis, ” Ener gy Reports , v ol. 9, pp. 215–222, 2023, doi: 10.1016/j.e gyr .2023.08.079. [10] Y . Y uan and X. Chen, “V e getable and fruit freshness detection base d on deep features and principal component analysis; V e getable and fruit freshness detection, ” Curr ent Resear c h in F ood Science , v ol. 8, 2024, doi: 10.1016/j.crfs.2023.100656. [11] Baj pai, H. V erma, and A. Y ada v , “Optimizing data aggre g ation and clustering in Internet of thi ngs netw orks using principal compo- nent analysis and Q-learning, ” Data Science and Mana g ement , v ol. 7, no. 3, pp. 189–196, 2024, doi: 10.1016/j.dsm.2024.02.001. [12] F . T onin, Q. T ao, P . P atrinos, and J. A. K. Suyk ens, “Deep K ernel Principal Component Analysis for multi-le v el feature learning, ” Neur al Networks , v ol. 170, pp. 578–595, 2024, doi: 10.1016/j.neunet.2023.11.045. [13] C. F an, X. Lai, H. W en, and L. Y ang, “Coal and g as outb urst prediction model based on principal component analysis and impro v ed support v ector machine, ” Geohazar d Mec hanics , v ol. 1, no. 4, pp. 319–324, 2023, doi: 10.1016/j.ghm.2023.11.003. [14] S. Esmaeiloghli, A. Lima, and B. Sade ghi, “Lithium e xploration tar geting through rob ust v ariable selection and deep anomaly detection: An inte grated application of sparse principal component analysis and stack ed autoencoders, ” Geoc hemistry , 2024, doi: 10.1016/j.chemer .2024.126111. [15] D. Contessi, L. V i v erit, L. N. Pereira, and C. Y . Heo, “Decoding the future: Proposing an interpretable machine learning model for Edg eShield: a r ob ust and a gile cyber security ar c hitectur e for the internet of medical things (Anass Misbah) Evaluation Warning : The document was created with Spire.PDF for Python.
1136 ❒ ISSN: 1693-6930 hotel occupanc y forecasting using principal component analysis, ” International J ournal of Hospitality Mana g ement , v ol. 121, 2024, doi: 10.1016/j.ijhm.2024.103802. [16] Guyon I and Elissef A, “ An Introduction to V ariable and Feature Selection, ” J ournal of mac hine learning r esear c h , v ol. 3, pp. 1157–1182, 2003. [17] M . Alalhareth and S. C. Hong, “ An Impro v ed Mutual Information Feature Selection T echnique for Intrusion Detection Systems in the Internet of Medical Things, ” Sensor s , v ol. 23, no. 10, 2023, doi: 10.3390/s23104971. [18] S. Ratnasing am and J. Mu ˜ noz-Lopez, “Distance Correlation-Based Feature Selection in Random F orest, ” Entr opy , v ol. 25, no. 9, 2023, doi: 10.3390/e25091250. [19] V . Rupapara, F . Rust am, A. Ishaq, E. Lee, and I. Ashraf, “Chi-Square and PCA Based Feature Selection for Diabetes De- tection with Ensemble Classier , ” Intellig ent A utomation & Soft Computing , v ol. 36, no. 2, pp. 1931–1949, 2023, doi: 10.32604/iasc.2023.028257. [20] T ri pathi and P . Rani, “PCA-Based Feature Selection and Hybrid Classication Model for Spe ech Emotion Recognition, ” Lectur e Notes in Networks and Systems , v ol. 703 LNNS, pp. 347–353, 2023, doi: 10.1007/978-981-99-3315-0 26. [21] P . A. Vysakh and P . Mayank, “Solar Flare Prediction and Feature Selection Using a Light-Gradient-Boosting M achine Algorithm, ” Solar Physics , v ol. 298, no. 11, 2023, doi: 10.1007/s11207-023-02223-5. [22] J .-S. Dessureault and D. Massicotte, “Feature selection or e xtraction decision process for clustering using PCA and FRSD, ” Mac hine Learning , 2021. [23] A. Si-ahmed, M. A. Al-Garadi, and N. Boustia, “Explainable Machine Learning-Based Security and Pri v ac y Protection Frame w ork for Internet of Medical Things Systems, ” Crypto gr aphy and Security , Mar . 2024. [24] S . Pirbhulal, N. Pombo, V . Felizardo, N. Garcia, A. H. Sodhro, and S. C. Mukhopadh yay , “T o w ards machine le arning enabled secu- rity frame w ork for iot-based healthcare, ” Pr oceedings of the International Confer ence on Sensing T ec hnolo gy , ICST , v ol. 2019–De- cember , 2019, doi: 10.1109/ICST46873.2019.9047745. [25] F . Alsubaei, A. Ab uhussein, V . Shandilya, and S. Shi v a, “IoMT -SAF: Internet of Medical Things Security Assessment Frame w ork, ” Internet of Things (Netherlands) , v ol. 8, 2019, doi: 10.1016/j.iot.2019.100123. [26] Z. Zhao, Y . Shang, J. W u, and Y . Y an, “Dataset Quantization with Acti v e Learning Based Adapti v e Sampling, ” in Lectur e Notes in Computer Science , 2025, pp. 346–362. doi: 10.1007/978-3-031-73027-6 20. BIOGRAPHIES OF A UTHORS Anass Misbah recei v ed the Engineering de gree from the ´ Ecole Nationale Sup ´ erieure des Ing ´ enieurs de Bour ges, France, in 2008. He has been w orking in softw are engineering e v er since and is currently pursuing the Ph.D. de gree at the IPIM Laboratory , National School of Applied Sciences, Sultan Moulay Slimane Uni v ersity , Morocco. His research interests include c ybersecurity , Internet of Things (IoT), machine learning, deep learning, and federated learning. His contrib utions in the current study encompass the design of adv anced machine learning algorithms for IoT netw ork security , the de v elopment of secure data transmission protocols for res ource-constrained IoT de vices, and the application of deep/federated learning techniques to enhance intrusion detection and threat mitig ation. He can be contacted at email: anassmisbah@gmail.com. Anass Seb bar recei v ed his Ph.D. Adv anced detection and pre v ention technique to mitig ate threats in multi-domain Softw are-Dened Netw orking using supervised machine learning models from the The National Higher School of Computer Science and Systems Analysis (ENSIAS- UM5) & International Uni v ersity of Rabat (UIR), Morocco. As part of his doctoral research, he completed their research at the Netw ork, Information, and Computer Security (NICS) Lab at the Uni v ersity of M ´ alag a, Spain. Before joining UIR as an Assis tant Professor and head of the Cybersecurity depart- ment at the School of Computer Science and Digital Engineering, he w ork ed as a research engineer on the E-T ourism project a t TICLab, UIR. He is an IEE EXtreme Ambassador and has contrib uted to or g anizing se v eral academic e v ents, including the Internat ional W orkshop on Emer ging Netw orks and Communications and the International Conference on W ireless and Mobile Communications. His research interests include c ybersecuri ty , softw are-dened netw orking, communicati on netw orks, the Internet of Things (IoT) & IoMT , and blockchain security . He has published his w ork in IEEE/A CM conferences, academic journals, and security and communication conferences. He can be contacted at email: anass.sebbar@uir .ac.ma. Imad Hadi is a professor (since 2009) at the national school of applied sciences Khouribg a. He gets accredi tation to supervise research in 2013. Before coming to Khouribg a, he completed man y programs: Ph.D. de gree (2005) in Applied Mathematics at the National institute of applied sciences L yon (INSA), and the MS de gree (2008) in Softw are Engineering at the school of mines saint-etienne. His research focuses te xt mining, IO T , data i nte gration and big data. He can be contacted at email: i.hadi@usms.ma. TELK OMNIKA T elecommun Comput El Control, V ol. 23, No. 4, August 2025: 1128–1136 Evaluation Warning : The document was created with Spire.PDF for Python.