IAES Inter national J our nal of Articial Intelligence (IJ-AI) V ol. 14, No. 6, December 2025, pp. 4913 ∼ 4922 ISSN: 2252-8938, DOI: 10.11591/ijai.v14.i6.pp4913-4922 ❒ 4913 Comparati v e e v aluation of machine lear ning models f or intrusion detection in WSNs using the IDSAI dataset Mansour Lmkaiti 1 , Houda Moudni 2 , Hicham Mouncif 1 1 LIMA TI Laboratory , F aculty of Polydisciplinary , Uni v ersity Sultan Moulay Slimane, Beni-Mellal, Morocco 2 TIAD Laboratory , F aculty of Sciences and T echnology , Uni v ersity Sultan Moulay Slimane, Beni-Mellal, Morocco Article Inf o Article history: Recei v ed Aug 27, 2024 Re vised Oct 24, 2025 Accepted No v 8, 2025 K eyw ords: Gradient boosting IDSAI Intrusion detection system Logistic re gression Machine learning Random forest W ireless sensor netw orks ABSTRA CT This paper pro vides com parati v e assessment of three lightweight machine learning (ML) models (logistic re gression (LR), random forest (RF), and gradient boosting (GB)), which are emplo yed to detect intrusions in wireless sensor netw orks (WSNs) using the IDSAI dataset. The goal is to determine the most ef fecti v e and deplo yable classier within the constraints of WSN resources. In order to pre v ent data leakage and report accurac y , precision, recall, F1-score, and recei v er operating characteristic-area under the curv e (R OC-A UC) with mean ± SD, we implement strati ed 5-fold cross v alidation with in fold preprocessing. The results indicate that RF pro vides the most optimal generalization and o v erall performance (accurac y 0 . 9994 ± 0 . 0001 , precision 0 . 9995 ± 0 . 0001 , recall 0 . 9994 ± 0 . 0001 , F1-score 0 . 9994 ± 0 . 0001 , R OC–A UC 0 . 9998 ± 0 . 0000 ). RF is closely follo wed by GB (accurac y 0 . 9990 ± 0 . 0001 , precision 0 . 9995 ± 0 . 0001 , recall 0 . 9985 ± 0 . 00 01 , F1-score 0 . 9990 ± 0 . 0001 , R OC-A UC ≈ 1 . 0000 ). LR demonstrates limitations in linearly o v erlapping classes, as e videnced by its high precision b ut reduced recall (accurac y 0 . 9167 ± 0 . 0010 , precision 0 . 9829 ± 0 . 0002 , recall 0 . 8481 ± 0 . 0018 , F1-score 0 . 9105 ± 0 . 0011 , R OC–A UC 0 . 9707 ± 0 . 0001 ). In order to e v aluate deplo yability , we characterize the inference throughput on a modest PC: LR ∼ 6 . 5 × 10 5 samples/s, GB ∼ 2 . 2 × 10 5 samples/s, and RF ∼ 1 . 3 × 10 5 samples/s, indicating a tiered intrusion detection system (IDS) (LR at sensors, RF at cluster -heads, and GB at the g ate w ay). W e also addre ss the potential dangers of o v ertting that may arise from the cleanliness of the dataset and pro vide a roadmap for future v alidation on a more di v erse set of traf c. The research establishes a baseline for lightweight IDS in actual WSNs that is deplo yable and reproducible. This is an open access article under the CC BY -SA license . Corresponding A uthor: Mansour Lmkaiti LIMA TI Laboratory , F aculty of Polydisciplinary , Uni v ersity Sultan Moulay Slimane Beni-Mellal, Morocco Email: lamkaitimansour@gmail.com 1. INTR ODUCTION This paper i ntroduces a methodical approach that uses cutting-edge machine learning (ML) [1] algorithms to thoroughly assess the ef fecti v eness of intrusion detection systems (IDS) [2]. Ensuring strong netw ork security is crucial in the quic k l y changing c ybersecurity landscape of today , which is mark ed by an increase in c yberthreats and the widespread inte gration of internet of things (IoT) de vices [3]. IDS [4] are essential for protecting netw orks because the y k eep an e ye on traf c patterns and spot possible harmful acti vity . J ournal homepage: http://ijai.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
4914 ❒ ISSN: 2252-8938 Ho we v er , depending on the detection methods utilized and the caliber of the training and e v aluation datasets, IDS ef cac y might v ary greatly [4]. Our study suggests an or g anized strate gy that includes se v eral crucial steps to fully address these issues: careful dataset preparation [4], stringent feature selection and engineering procedures, e xtensi v e model training and e v aluation techniques, reliable c ross-v alidation procedures , and in-depth performance analysis. Our research intends to impro v e the accurac y and dependability of IDS implementations by utilizing the v ariety of real-w orld intrusion scenarios captured in the IDSAI dataset [4]. IDS models are de v eloped and e v aluated using ML [1] algorithms, including gradient boosting (GB), random forest (RF) [5], and logistic re gression (LR) [6], in order to impro v e their ability to ef fecti v ely detect and mitig ate security breaches [7]. By using this systematic and empirical research, our study aims to of fer detailed insight into the adv antages and disadv antages of ML based IDS [8], [9] techniques. W e help de v elop more e xible and ef fecti v e security measures suited to the comple x dynamics of modern netw orks and the changing terrain of c yberthreats by critically assessing the performance of dif ferent algorithms ag ainst benchmark datasets and a range of attack scenarios [10]. Pre viously , classical classiers were frequently e v aluated on synthetic or restricted IoT datasets without taking into account genuine wireless sensor netw ork (WSN) constraints. This w ork addresses that lacuna by introducing a rob ust statistical v alidation (mean ± SD, 95% condence interv al), assessing computational feasibility at the sensor , cluster -head, and g ate w ay le v els, and benchmarking three interpretable, lightweight models on the IDSAI dataset. The k e y contrib utions of this study are as follo ws. First, an e v aluation approach that can be replicated for lightweight IDS benchmarking using the IDSAI dataset is proposed. Second, LR, RF , and GB are inte grated in a hierarchical IDS architecture for scalable WSN security . Third, statistical tests, including Friedman and W ilcoxon, are used to v alidate the rob ustness of the model. F ourth, computational footprint and inference throughput are included in the deplo yment analysis. Finally , a plan for upcoming v alidation with a v ariety of scenarios is outlined. 2. RELA TED W ORK Much research has been done in the eld of computer security [8], especially in WSNs, to address the changing challenges posed by security threats [9]. Numerous strate gies for impro ving WSN security , such as intrusion detection, encryption methods and secure routing protocols, ha v e been e xamined in earlier research [10], [11]. The creation and assessment of IDS [8] designed especially for WSNs constitutes a substantial eld of study [12]. By k eeping an e ye on netw ork traf c and spotting unusual acti vity suggesti v e of malicious acti vity , these systems are essential in detecting and pre v enting security breaches [7] within WSNs. Numerous st udies ha v e used a v ariety of datasets and e v aluation metrics t o assess the ef fecti v eness of IDS in WSNs [13], [14]. These tests seek to determine ho w well IDS identify dif ferent kinds of assaults, such as routing attacks, data manipulation and denial-of-service attacks [14], [15]. Researchers ha v e shed important light on the adv antages and disadv antages of current IDS in WSNs by compar ing v arious to common datasets and attack scenarios. Additionally , research has focused on creating lightweight security measures that are suited for WSN de vices with limited resources [16]. These safe guards are designed to reduce ener gy usage and computational o v erhead while of fering strong defense ag ainst security risks [16]. T o address the particular security issues presented by WSNs, methods lik e ener gy-ef cient k e y management techniques, se cure routing protocols, and lightweight cryptograph y ha v e been de v eloped [12]. Additionally , research has look ed into ho w to incorporate cutting-edge technologies lik e machine intelligence and blockchain into WSN security designs [16]–[20]. While ML [18], [19], [21], [22] methods pro vide for adapti v e and autonomous intrusion detection capabilities, blockchain [17]-based techniques pro vide decentralized and tamper -resistant mechanisms for safe guarding WSN data and transactions. Recent research has in v estig ated the use of anomaly-based methods (e.g., one-class support v ector machine (SVM), isolation forest) and long short-term memory (LSTM), as well as autoencoders, for the purpose of intrusion detection in IoT/WSN. Although these methods are frequently precise, the y typically necessitate substantial rening and higher compute and ener gy b udgets. W e concentrate on the de v elopment of interpretable and lightweight model s that are appropriate for decentralized WSN nodes. In the future, we will in v estig ate the inte gration of on-node lightweight classiers with g ate w ay-le v el deep feature e xtraction. T able 1 summarizes recent IDS studies in WSN and IoT en vironments. Int J Artif Intell, V ol. 14, No. 6, December 2025: 4913–4922 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 4915 T able 1. Comparati v e summary of recent IDS studies in WSNs and IoT en vironments Study Dataset Model Accurac y (%) Main limitation Dharini et al . [9] WSN-LEA CH XGBoost 98.7 High computational cost Meenakshi and Karunkuzhali [10] IoT -Custom GAN-V AE 99.2 Comple x training Ajmi et al . [16] Hardw are IDS CNN 96.8 Not ener gy-ef cient This study IDSAI RF / GB / LR 99.9 Dataset simplicity 3. TYPES OF CYBERSECURITY A TT A CKS WSNs are at risk from a number of frequent assaults, s uch as ARP spoong, SYN/A CK ooding, and ICMP Echo oods, which can o v erwhelm nodes and reroute communication channels. Brute-force SSH attempts and UDP port scans tak e use of service a ws, while distrib uted denial-of-service (DDOS) attacks use a lot of netw ork capacity by making a lot of requests at once. These assaults demonstrate the need for ef fecti v e and portable IDS that can identify anomalous acti vity at se v eral netw ork tiers. As sho wn in Figure 1, the main dataset parameters and attack classes are illustrated. Figure 1. Dataset parameters 4. METHODOLOGY Using the IDSAI dataset [4], we emplo yed a systematic e v aluation procedure in this study to e v aluate the ef fecti v eness of IDS models based on ML algorithms [20], [23]. Dataset preparation, feature engineering and selection, model training, cross-v alidation, and performance e v aluation are the v e primary processes of the methodology . The hierarchical IDS architecture emplo yed in this in v estig ation is depicted in Figure 2, with RF functioning at the cluster -head le v el, LR at the sensor layer , and GB at the g ate w ay for retraining and v alidation. 4.1. Dataset pr eparation The IDSAI dataset, which replicates actual c yberattacks in WSNs, we emplo yed. The IDSAI datas et w as originally introduced by Fernando et al . [4], and it has been widely used for e v aluating IDS in IoT en vironments. Labeled traf c cases from a range of attack methods, including DDoS, ARP spoong, and port scanning, are including this collection. The IDSAI dataset comprises more than eighty thousand labeled o ws that encompass both standard traf c and a v ariety of at tack cate gories, including DDoS, ARP spoong, SYN ooding, and port scanning. In order to maintain the distrib ution of labels across folds, we emplo yed stratied sampling and v eried class proportions. V ariance analysis re v ealed lo w noise and partial feature redundanc y ,which may result in prominent metrics being inated. Consequently , we pro vide fold-wise statistics and e xplicitly address o v ertting risks in sections 5.1 and 5.2. 4.2. F eatur e selection and engineering W e used lter -based feature se lection techniques, such a s mutual information and v ariance thresholding, to increase model ef cienc y and decrease o v ertting. Based on their contrib ution to classication Compar ative e valuation of mac hine learning models for intrusion detection in ... (Mansour Lmkaiti) Evaluation Warning : The document was created with Spire.PDF for Python.
4916 ❒ ISSN: 2252-8938 performance, we k ept the most discriminati v e feat ures that were pertinent to intrusion detection. This choice enhances interpretability , reduces redundanc y , and e xpedites training without compromising the quality of detection. Figure 2. Hierarchical IDS architecture for WSN 4.3. Model training W e chose three popular ML classiers: RF [23], LR [24], and GB [25]. L2 re gulari zation w as used to train the LR model. Grid search w as used to determine the in v erse of re gularization strength (C). 4.4. Mathematical o v er view The LR model minimizes the ℓ 2 -re gularized ne g ati v e log-lik elihood: J ( θ ) = − 1 m m X i =1 h y i log h θ ( x i ) + (1 − y i ) log 1 − h θ ( x i ) i + λ ∥ θ ∥ 2 2 , (1) with h θ ( x ) = 1 1+exp( − θ ⊤ x ) . RF aggre g ates T decision trees { f t } T t =1 by majority v ote, ˆ y = mo de f 1 ( x ) , . . . , f T ( x ) . GB b uilds an additi v e model F M ( x ) = P M m =1 η h m ( x ) , where h m are shallo w trees tted stage-wise to the ne g ati v e gradients of the loss, and η is the learning rate. Grid search selected h yperparamet ers to balance bias and v ariance. 4.5. Cr oss-v alidation and generalization assessment During model e v aluation, we used stratied 5-fold cross-v alidation to guarantee the rob ustness and generalizability of our ndings. In order to monitor training beha vior and identify an y possible o v ert ting or undertting tendencies, learning curv es were created. All preprocessing stages (feature selection, scaling when applicable, and model tting) were e x ecuted within each training fold of the stratied 5-fold cross-v alidation pipeline. T est folds were concealed until the nal scoring in order to pre v ent optimistic bias. 4.6. P erf ormance metrics Precision, recall, F1-s core, accurac y , and recei v er operating characteri stic-area under the curv e (R OC-A UC) are common classication metrics that we used to e v aluate the performance of the models. Prediction probability histograms, R OC curv es, and precision-recall curv es were used to display these metrics, which were calculated for e v ery model. Ev ery outcome w as e xamined in light of the models interpretability and usefulness for IDS deplo yment in WSNs [26]. Int J Artif Intell, V ol. 14, No. 6, December 2025: 4913–4922 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 4917 5. RESUL TS OF D A T ASET T o achie v e our research objecti v es, we propose a systematic methodology for e v aluating IDS performance with ML algorithms. Figure 3 illustrates the precision-recall curv es for LR, RF , and GB models, highlighting that GB and RF maintain near -perfect precision across almost the entire recall range, whereas LR sho ws a noticeable drop in performance as recall increases. Figure 4 demonstrates that while LR suf fers from o v ertting, RF generalizes ef fecti v ely with gro wing data. GB impro v es v alidation scores whil e pro viding a balanced performance. Figure 3. Precision-Recall curv e for LR, RF , and GB models Figure 4. Learning curv es (train vs. CV) sho wing: LR o v erts as data gro ws (g ap persists), RF maintains high and stable generalization, and GB impro v es v alidation steadily—aligning with the metric ranking in T able 1 Compar ative e valuation of mac hine learning models for intrusion detection in ... (Mansour Lmkaiti) Evaluation Warning : The document was created with Spire.PDF for Python.
4918 ❒ ISSN: 2252-8938 Figure 5 illustrates the learning trajectories of the three models. RF demonstrates a high de gree of generalization, with v alidation scores that continue to impro v e as the training set e xpands. GB closely matches RF and enhances v alidation performance, whereas LR o v erts in late re gimes (training v alidation), which e xplains its lo wer recall and F1-score. Figure 5. Learning curv es conrm RF’ s superior generalization; GB is a close second; LR e xhibits persistent train–v alidation g ap (o v ertting), consistent with its lo wer recall and F1-score 5.1. Pr ediction pr obability distrib ution Figure 6 sho ws that ensemble models (RF and GB) gi v e more condent predictions, whil e LR sho ws more uncertainty in its probability estimates. T abl e 2 summarizes the performance of the three classiers (mean ± SD o v er 5 strat ied folds). RF achie v es the best o v erall performance across all metrics; GB is a close, well-balanced runner -up. LR attains high precision b ut noticeably lo wer recall, leading to a lo wer F1-score than the ens emble methods. T o pre v ent leakage, all preprocessing is conducted within each training fold, and these interv als are calculated o v er v e stratied folds. W e report 95% condence interv als together with fold-wise means and standard de viations for all metrics o v er v e stratied folds. F or accurac y , RF attains the highest score ( 0 . 9994 ± 0 . 0001 ; 95% CI [0.9994, 0.9995]), GB is close ( 0 . 9990 ± 0 . 0001 ; [0.9989, 0.9991]), while LR is lo wer ( 0 . 9167 ± 0 . 0010 ; [0.9158, 0.9175]). LR’ s lo wer recall leads to a lo wer F 1-score than the ensemble methods. In order to e v aluate the performance disparities among models, a non-parametric Friedman test w as implemented on fold-wise F1-scores. The results indicated that at least one model performed dif ferently , as e videnc ed by the signicant aggre g ate dif ference ( χ 2 (2) = 10 . 00 , p = 0 . 0067 ). The post-hoc W ilcoxon signed-rank tests with Holm correction re v ealed no signicant dif ference between RF and GB ( p adj > 0 . 05 ), thereby conrming that both ensemble models obtain consistently high performance. This statistical consistenc y emphasizes the reliability and rob ustness of the data, thereby bolstering the credibility of the comparati v e frame w ork. 5.2. Ov ertting consideration The unusually high v alues of precision and recall (near 1.0) require critical consideration. These may be due to: i) well-separated class boundaries in the IDSAI dataset, ii) feature redundanc y or lo w noise, and iii) lack of real-w orld di v ersity in attack v ectors. W e mitig ated o v ertting risks through 5-fold stratied Int J Artif Intell, V ol. 14, No. 6, December 2025: 4913–4922 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 4919 cross-v alidation and comparison across multiple metrics and plots. Nonetheless, future w ork will inte grate more challenging datasets to further assess generalizability . Figure 6. Prediction probability distrib utions: RF/GB yield condent, well-separated posteriors; LR sho ws broader uncertainty , consistent with its reduced recall T able 2. Classication performance (mean ± SD o v er 5 stratied folds) Model Accurac y Precision Recall F1-score R OC–A UC LR 0.9167 ± 0.0010 0.9829 ± 0.0002 0.8481 ± 0.0018 0.9105 ± 0.0011 0.9707 ± 0.0001 RF 0.9994 ± 0.0001 0.9995 ± 0.0001 0.9994 ± 0.0001 0.9994 ± 0.0001 0.9998 ± 0.0000 GB 0.9990 ± 0.0001 0.9995 ± 0.0001 0.9985 ± 0.0001 0.9990 ± 0.0001 1.0000 ± 0.0000 Lo w label noise and well-separated classes are corroborated by the nearly a wless metrics. In spite of this, learning curv es indicate disparities in c apacity (LR o v ertting in late re gimes v ersus RF stability). There are plans for future v alidation on a wider range of traf c, such as unseen de vices, and blended protocols, to stress-test generalization. The computational footprint of each model is summarized in T able 3. T raining is performed of ine; inference reects on-de vice cost in deplo yment. Interpretation: LR is the optimal choice for embedded ltering at sensor nodes due to its ability to generate the quick est inference ( ∼ 6 . 5 × 10 5 samples/s). RF achie v es a rob ust accurac y-cost trade-of f ( ∼ 1 . 3 × 10 5 samples/s) that is suitable for cluster -heads, while GB remains viable at the g ate w ay with competiti v e inference speed ( ∼ 2 . 2 × 10 5 samples/s) despite being more e xpensi v e to train. These on-de vice inference costs are the primary constraint for real-w orld deplo yment, as training is conducted of ine. T able 3. Computational footprint on a modest PC Model T raining time (s) Inference on 9 . 98 × 10 5 samples (s) Throughput (samples/s) LR 823.40 1.54 6 . 48 × 10 5 RF 125.02 7.48 1 . 33 × 10 5 GB 460.33 4.47 2 . 23 × 10 5 6. DISCUSSION OF THE RESUL TS The results demonstrate the strong capabilities of ML algorithms in detecting intrusions in WSNs. RF achie v es the strongest o v erall performance; GB is a close, well-balanced alternati v e. LR attains high precision b ut lo wer recall, leading to a lo wer F1-score than the ensemble methods. Ho we v er , learning curv es re v eal o v ertting as dataset size increases, limiting its generalization. GB is a well-balanced, close alternati v e to RF , Compar ative e valuation of mac hine learning models for intrusion detection in ... (Mansour Lmkaiti) Evaluation Warning : The document was created with Spire.PDF for Python.
4920 ❒ ISSN: 2252-8938 which obtains the strongest o v erall performance. The F1-score is l o wer than that of the ensemble methods due to the f act that LR achie v es high precision b ut a reduced recall. GB of fered balanced performance, with high condence in predictions and competiti v e scores across all metrics. In order to minimize f alse ne g ati v es in IDS, precision-recall curv es v erify that RF and GB ha v e good precision e v en as recall rises. Additionally , prediction probability dist rib utions demonstrate that, in contrast to LR, RF , and GB of fer more certain classications . The near -perfect metrics may be indicati v e of minor o v ertting or dataset simplicity , despite the highly encouraging results. Therefore, these disco v eries should be v eried in future research using datasets that are more intricate and di v erse. In general, GB pro vides a balanced performance, LR is appropriate for straightforw ard scenarios, and RF remains the most scalable and dependable option for intrusion detection in real WSNs 7. CONCLUSION AND DEPLO YMENT INSIGHTS W e recommend a hierarchical IDS, which consists of LR at sensor nodes (ne gligible latenc y), RF at cluster -heads (rob ust aggre g ation), and GB at the g ate w ay (v alidation and periodic retraining). This approach minimizes communication o v erhead and concentrates hea vier computation in areas where resources are less constrained. Using the IDSAI dataset, this study of fered a systematic assessment of ML-based IDS in WSNs. W e illustrated the benets of each model by contrasting RF , GB, and LR . On IDSAI, RF e xhibited superior generalization and stability , while GB w as a close, well-balanced second. The F1-score w as lo wer than that of the ensemble methods due to the f act that LR achie v ed high precision b ut a reduced recall. The ndings highlight ho w cruc ial it is to choose ML models based on the particular deplo yment conte xt, whether that conte xt is one of scalability , accurac y , or interpretability . In order to increase IDS reliability , the study also e xamined o v ertting concerns and the requirement for realistic, di v erse attack data. T o sum up, our results conrm the importance of ML in protecting WSNs and recommend that more sophisticated and adaptable methods that can manage dynamic and di v erse IoT settings be e xplored in future studies. In order to further enhance IDS e xibility in dynamic IoT conte xts, future research will concentrate on mer ging federated and h ybrid learning methodologies. A CKNO WLEDGMENTS The authors w ould lik e to thank the LIMA TI Laboratory and the F aculty of Polydisciplinary at Uni v ersity Sultan Moulay Slimane for their scientic guidance, technical resources, and continuous support during the preparation of this w ork. FUNDING INFORMA TION The author(s) recei v ed no nancial support for the research, authorship, and/or publication of this article. A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the Contrib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib utions, 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 Mansour Lmkaiti ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Houda Moudni ✓ ✓ ✓ ✓ ✓ ✓ ✓ Hicham Mouncif ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ 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 Int J Artif Intell, V ol. 14, No. 6, December 2025: 4913–4922 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 4921 CONFLICT OF INTEREST ST A TEMENT Authors state no conict of interest. D A T A A V AILABILITY The data underpinning the results of this study are accessible from the corresponding author upon reasonable request. REFERENCES [1] C. S. W . Ng, M. N. Amar , A. J. Ghahf arokhi, and L. S. Imsland, “ A surv e y on the application of machine learning and metaheuristic algorithms for intelligent proxy modeling i n reserv oir simulation, ” Computer s & Chemical Engineering , v ol. 170, Feb . 2023, doi: 10.1016/j.compchemeng.2022.108107. [2] S. T abbassum and R. K. P athak, “Ef fecti v e data trans mission through ener gy-ef cient clustering and fuzzy-Based IDS routing approach in WSNs, ” V irtual Reality & Intellig ent Har dwar e , v ol. 6, no. 1, pp. 1–16, Feb . 2024, doi: 10.1016/j.vrih.2022.10.002. [3] B. Suresh and G. S. C. Prasad, “ An ener gy ef cient secure routing scheme using LEA CH protocol in WSN for IoT netw orks, ” Measur ement: Sensor s , v ol. 30, Dec. 2023, doi: 10.1016/j.measen.2023.100883. [4] G.-P . Fernando, A.-A. H. Brayan, A. M. Florina, C.-B. Liliana, A.-M. H. -Gabriel, and T .-S. Reinel, “Enhancing intrusion detection in IoT communications through ML model generalization with a ne w dataset (IDSAI), ” IEEE Access , v ol. 11, pp. 70542–70559, 2023, doi: 10.1109/A CCESS.2023.3292267. [5] S. Nieland, R. Oostendorp, M. Heinrichs, and R. Cyg anski, “T ransferability analysis of user groups in t ra v el beha viour surv e ys using a random forest classication model, ” T r ansportation Resear c h Pr ocedia , v ol. 76, pp. 81–95, 2024, doi: 10.1016/j.trpro.2023.12.040. [6] B. K olukisa, B. K. Dedeturk, H. Hacilar , and V . C. Gungor , “ An ef cient netw ork intrusion detection approach based on logistic re gression model and parallel art icial bee colon y algorithm, ” Computer Standar ds & Interfaces , v ol. 89, Apr . 2024, doi: 10.1016/j.csi.2023.103808. [7] N. Balakrishnan, A. Rajendran, D. Pelusi, and V . Ponnusamy , “Deep belief netw ork enha nced intrusion detection system to pre v ent security breach in the internet of things, ” Internet of Things , v ol. 14, Jun. 2021, doi: 10.1016/j.iot.2019.100112. [8] T . Nandy , R. Md Noor , R. K olandaisamy , M. Y . I. Idris, and S. Bhattacharyya, “ A re vie w of security attacks and intrusion detection in the v ehicular netw orks, ” J ournal of King Saud Univer sity - Computer and Information Sciences , v ol. 36, no. 2, Feb . 2024, doi: 10.1016/j.jksuci.2024.101945. [9] N. Dharini, J. Katira v an, S. D. M. Priya, and S. V . A. Sne ghaa, “Intrusion detection in no v el WSN-Leach Dos attack dataset using machine learning based boosting algorithms, ” Pr ocedia Computer Science , v ol. 230, pp. 90–99, 2023, doi: 10.1016/j.procs.2023.12.064. [10] B. Meenakshi and D. Karunkuzhali, “Enhancing c yber security in WSN using optimized self-attention-based pro visional v ariational auto-encoder generati v e adv ersarial netw ork, ” Computer Standar ds & Interfaces , v ol. 88, Mar . 2024, doi: 10.1016/j.csi.2023.103802. [11] D. A. J. Rajan and E. R. Nag anathan, “T rust based anon ymous intrusion detection for cloud assisted WSN-IO T , ” Global T r ansitions Pr oceedings , v ol. 3, no. 1, pp. 104–108, Jun. 2022, doi: 10.1016/j.gltp.2022.04.022. [12] S. Md Zin, N. B. Anuar , M. L. M. Kiah, and I. Ahmedy , “Surv e y of secure multipath routing protocols for WSNs, ” J ournal of Network and Computer Applications , v ol. 55, pp. 123–153, Sep. 2015, doi: 10.1016/j.jnca.2015.04.018. [13] R. Y ada v , I. Sreede vi, and D. Gupta, “ Augmentation in performance and security of WSNs for IoT applications using feature selection and classication techniques, ” Ale xandria Engineering J ournal , v ol. 65, pp. 461–473, Feb . 2023, doi: 10.1016/j.aej.2022.10.033. [14] M . R. Kadri, A. Abdelli, J. B. Othman, and L. Mokdad, “Surv e y and classication of Dos and DDos attack detection and v alidation approaches for IoT en vironments, ” Internet of Things , v ol. 25, Apr . 2024, doi: 10.1016/j.iot.2023.101021. [15] T . T . Lai, T . P . T ran, J. Cho, and M. Y oo, “DoS attack detection using online learning techniques in wireless sensor netw orks, ” Ale xandria Engineering J ournal , v ol. 85, pp. 307–319, Dec. 2023, doi: 10.1016/j.aej.2023.11.022. [16] H. Ajmi et al ., “Ef cient and lightweight in-memory com puting architecture for hardw are security , ” J ournal of P ar allel and Distrib uted Computing , v ol. 190, Aug. 2024, doi: 10.1016/j.jpdc.2024.104898. [17] M . Kharjana, F . H. Pohrmen, S. C. Sahana, and G. Saha, “Blockchai n-based k e y management system in named data netw orking: A surv e y , ” J ournal of Network and Computer Applications , v ol. 220, No v . 2023, doi: 10.1016/j.jnca.2023.103732. [18] L. Eck e, M. Magdolen, S. Jaquart, R. Andre, and P . V ortisch, “ A case study of checking national household tra v el surv e y data with machine learning, ” T r ansportation Resear c h Inter disciplinary P er spectives , v ol. 24, Mar . 2024, doi: 10.1016/j.trip.2024.101078. [19] D. Ma, X. Li, J. Liang, Z. W ang, and W . Y ang, “Distilling seed-assisted zeolite synthesis conditions by machine learning, ” Micr opor ous and Mesopor ous Materials , v ol. 339, Jul. 2022, doi: 10.1016/j.micromeso.2022.112029. [20] C. Ale x, G. Creado, W . Almobaideen, O. A. Alghanam, and M. Saadeh, “ A comprehensi v e surv e y for IoT security datasets taxonomy , classication and machine learning mechanisms, ” Computer s & Security , v ol. 132, Sep. 2023, doi: 10.1016/j.cose.2023.103283. [21] S. M. S. Bukhari, M. H. Zaf ar , M. A. Houran, Z. Qadir , S. K. R. Moosa vi, and F . Sanlippo, “Enhancing c ybersecurity in Edge IIoT netw orks: An asynchronous federated learning approach with a deep h ybrid detection model, ” Internet of Things , v ol. 27, Oct. 2024, doi: 10.1016/j.iot.2024.101252. [22] D. Karunkuzhali, K. P . Arunachalam, R. Ramamoorthi, and R. K. Kadu, “Cyber -ph ysical system for enhanced WSN-IoT security using spherical graph triple con v olutional neural netw ork with planet optimization algorithm, ” Pr o gr ess in Engineering Science , v ol. 2, no. 3, Sep. 2025, doi: 10.1016/j.pes.2025.100108. Compar ative e valuation of mac hine learning models for intrusion detection in ... (Mansour Lmkaiti) Evaluation Warning : The document was created with Spire.PDF for Python.
4922 ❒ ISSN: 2252-8938 [23] I. Saadi, A. Mustaf a, J. T eller , and M. Cools, “ A bi-le v el random forest based approach for estimating O-D matrices: Preliminary results from the Belgium national household tra v el surv e y , ” T r ansportation Resear c h Pr ocedia , v ol. 25, pp. 2566–2573, 2017, doi: 10.1016/j.trpro.2017.05.301. [24] M . O. M. Mohamm ed, “Pre v alence and risk f actors associated with under - v e years children diarrhea in Mala wi: Application of surv e y logistic re gression, ” Heliyon , v ol. 10, no. 7, Apr . 2024, doi: 10.1016/j.heliyon.2024.e29335. [25] A. Manoharan, K. M. Be g am, V . R. Apar o w , and D. Sooriamoorth y , “ Articial neural netw orks, gradient boosting and support v ector machines for electric v ehicle battery state estimation: A re vie w , ” J ournal of Ener gy Stor a g e , v ol. 55, Aug. 2022, doi: 10.1016/j.est.2022.105384. [26] M . Lmkaiti, I. Larhlimi, M. Lachg ar , H. Moudni, and H. Mouncif, “ Adv anced optimization of RPL-IoT protocol using ML algorithms, ” International J ournal of Advanced Computer Science and Applications , v ol. 16, no. 2, 2025, doi: 10.14569/IJ A CSA.2025.01602135. BIOGRAPHIES OF A UTHORS Mansour Lmkaiti is from Department of Computer Mathemati cs, F aculty of Polydisciplinary , Uni v ersity Sultan Moulay Slimane, Morocco. His domains of interests is high-performance computer systems and netw orks: t heory , machine learning algorithms; high performance in WSNs; and c ybersecurity in wireless sensor netw orks. He can be contacted at email: lamkaitimansour@gmail.com. Houda Moudni is currently w orking as theis an Assistant Professor at the National School of Business and Management, Sultan Moulay Slimane Uni v ersity , B ´ eni Mellal, Morocco. She recei v ed the Ph.D. de gree in Computer Sciences from the F aculty of Sciences and T echnology of Beni Mellal in 2019. She de v eloped a strong interest in computer netw orking. Her research w ork primarily focuses on securing routing protocols in mobile Ad Hoc netw orks (MANET), wireless sensor netw orks (WSN), and the internet of things (IoT). She can be contacted at email: h.moudni@usms.ma. Hicham Mouncif is from Department of Computer Mathematics, Uni v ersity Sultan Moulay Slimane, Morocco. He is currently w orking as the Professor at the Department of Mathematics and Informatics. His research interests include computer netw orking, communication engineering, and securing routing protocols in wireless sensor netw orks. His domains of interests is high-performance computer systems and net w orks: theory , machine learning algorithms; high performance in WSNs and c ybersecurity . He can be contacted at email: h.mouncif@usms.ma. Int J Artif Intell, V ol. 14, No. 6, December 2025: 4913–4922 Evaluation Warning : The document was created with Spire.PDF for Python.