Whereas cloud computing is becoming a growing trend due to its remote availability and simple services, control and data management concerns are invited by its reliance on outside providers. Despite their best effort to provide accessibility and efficiency, cloud systems may inadvertently expose themselves to threats due to their complexity and large distribution. Traditional IDSs are skewed toward familiar attack patterns, which limits their capacity to identify unexpected behavior. IDSs are often employed to audit activity rather than to prevent attacks actively. Although ML is commonly proposed as a means of improving threat detection, it might not always yield stable performance in various situations or real-time reliability. Furthermore, the extent to which ML and IDS can be integrated is still limited, and thus the issues involved in converging automation and adaptive security mechanisms are far from being completely addressed.

The objective of this work is to develop an ensemble-driven intrusion detection model that integrates ML techniques into traditional IDS for enhancing cloud computing environment security. It increases the detection accuracy of known and unknown attacks, reduces false alarms, and supports timely and adaptive reactions to novel attack patterns.

The remainder of this paper is structured as follows: section 2 contains background, which presents an introduction to cloud computing, IDSs, and ML. Section 3 discusses related work. Section 4 offers an overview of the proposed model, including the NSL-KDD dataset, feature selection methods, classifiers employed, performance metrics for assessment, and experimental framework. Section 5 contains the results and analysis. Finally, section 6 concludes conclusion and future work.

Cloud computing is the delivery of information technology services such as servers, storage, databases, software, network management, and artificial intelligence (AI) via the Internet. It allows for speedier, more flexible, and cost-effective solutions than traditional techniques, resulting in improved productivity and efficiency. There are several sorts of cloud infrastructures, each with its design and development. Various models, varieties, and services have evolved to meet distinct demands. Cloud services may be delivered in three ways: public, private, and ensemble (Megouache et al., 2024; Zulifqar et al., 2021).

The public cloud is an example of a computer platform that restricts customization and is run by companies that supply standardized services to a large customer base. This group of services, which rely on shared infrastructure, includes Microsoft Azure, Google App Engine, and Amazon EC2. On the other hand, the private cloud is designed for internal usage only, sometimes at the expense of more stringent data isolation. It is appropriate for businesses with certain operating needs. Microsoft ECI data centers, Ubuntu Enterprise Cloud, and Amazon Virtual Private Cloud are a few examples. Organizations with similar operational objectives employ the community cloud, which divides duties among several users and might result in less consistent administration (Saran et al., 2022). New complexity brought forth by cloud technology’s ongoing development put both individuals’ and enterprises’ current security standards to the test. ML is being investigated as a supplemental technique to help conventional security measures rather than providing a comprehensive answer. Though their function is typically reactive rather than entirely preventive, ML techniques are frequently used in cloud systems to help identify trends and abnormalities (Nassif et al., 2021).

An IDS is a part of the system that watches data flow and system activities passively without taking any active action to prevent problems. It detects abnormalities that can indicate technical problems or unusual use rather than explicitly recognizing malicious activities or breaches. Instead of sending out instant notifications, it creates records when it notices specific trends. By providing insights rather than immediately addressing threats, an IDS indirectly enhances network security in contrast to proactive technologies like firewalls, antivirus software, or access control systems (Umar et al., 2024; Saranya et al., 2020). Instead of active defensive techniques like honeypots, IDSs are frequently coupled with monitoring tools. Network intrusion detection systems (NIDS) and host-based intrusion detection systems (HIDS) are the two broad categories of IDS development that are often recognized. With host-level modifications, HIDS mainly examines internal operations and file integrity on a single system. However, NIDS may miss host-specific events when monitoring network traffic. NIDS detection usually uses statistical deviations or pre-established procedures to identify anomalies (Useni et al., 2023). IDS may be broadly divided into two categories: pattern-matching and behavior-based. Although it may not always employ ML or DL algorithms, behavior-based IDSs rely on departures from predicted activities. However, instead of employing dynamic analysis, pattern-matching IDSs use a database of recognized behaviors to identify matches. Some IDS systems function in intervals rather than continuously, alerting users only when certain thresholds or circumstances are reached.

IDS systems are divided into two types based on their response mechanism:

Active IDS detects threats but also takes preventative steps such as blocking suspicious traffic to avoid prospective assaults.

ML is a subset of AI that allows systems to learn and develop automatically based on experience rather than explicit programming. In IDSs, ML algorithms identify intrusions in huge datasets more quickly. ML algorithms are often classified into three categories. Supervised, Unsupervised, Semi-supervised (Saranya et al., 2020). Supervised ML approaches use labeled datasets to train algorithms to spot patterns of normal and aberrant behavior. Common intrusion detection algorithms include LR, Gaussian NB, RF, MLP Classifier, KNN, DT, AdaBoost, XGBoost, and LightGBM. Unsupervised ML approaches use unlabeled data and focus on detecting deviations from usual behavior. Clustering comparable data points and detecting abnormalities is accomplished using algorithms such as K-means and Gaussian Mixture Model (GMM). Semi-supervised ML algorithms use labeled and unlabeled data to increase detection accuracy. They can detect anomalies that are similar to tagged cases, even if they do not completely match. These strategies attempt to improve the performance of anomaly-based IDSs, hence leading to greater computer network security (Adhikari and Bal, 2023; Nassif et al., 2021; Parameswarappa et al., 2023).

The KDDcup99 dataset was generated for the Third International Knowledge Discovery and Data Mining Tools Competition, which took place in 1999 in connection with KDD-99, the Fifth International Conference on Knowledge Discovery and Data Mining. This dataset contains over 5 million training samples and over 2 million testing samples. It also has a high number of duplicate entries and unbalanced class distributions (Umar et al., 2024). The NSL-KDD dataset is based on the KDD Cup ‘99 dataset and solves flaws identified in the original, such as duplicated entries in the training set and duplicate records in the test set (Protić and Stanković, 2023). The NSL-KDD dataset has 41 attributes: three are categorical, four are binary, and the remaining 34 are continuous. The training set includes 23 traffic types, whereas the testing set has 30. The assaults in this dataset are divided into four categories: DOS, probing, U2R, and R2L. The features are divided into three categories: (1) fundamental features, (2) content-based features, and (3) traffic-related features (Alkadi et al., 2023).

Data preparation is critical in converting raw information into a format that enhances the performance of ML models. Raw data frequently contains missing, null, or inconsistent values, as well as unused or duplicated fields that provide no useful information. To solve this, the process starts by identifying and eliminating inconsistencies in the data to ensure it is clean. The following stages describe how the dataset was prepared for ML applications:

The preparation begins with the load _data function, which imports the text file and converts it into a pandas DataFrame with easy-to-understand column names.

Choosing the pertinent attributes comes next after data analysis. By determining the most important variables, feature selection improves the accuracy of ML models, which is why it is so important in intrusion detection (Vibhute et al., 2024). The process of feature subset selection aims to improve classification performance by selecting a smaller set of features from a larger pool. While some researchers describe feature subset selection (FSS) as a technique for decreasing the size of the feature set, others view it as a means of improving classifier accuracy. FSS is thought to be a method for finding the best feature subsets that, by removing unnecessary and duplicated features, preserve the most important information in a dataset (Alhayali et al., 2021). The most significant characteristics are found using feature selection approaches, which also remove redundant or superfluous data and increase the dataset’s overall effectiveness (Vibhute et al., 2024). In the beginning, the study employed all of the features without using any feature selection techniques. Following that, two filter-based feature selection strategies were presented: Mutual Information and Gain Information. These methods rank the features according to their scores, which assess each feature’s importance and pertinence to the class label. The characteristics with the highest ranking are then used to create the IDS. Filter approaches have the advantage of being computationally efficient because they do not need IDS training (Alalhareth and Hong, 2023). Filters are employed in the popular feature selection method known as Gain Information. It reduces the impact of irrelevant data by ranking characteristics according to their significance. By doing so, it discovers the characteristics that give the most information to a certain class, improving the model’s overall performance (Kurniabudi et al., 2020). In Gain Information (GI), the quantity of information is primarily assessed using the entropy and conditional entropy concepts. Firstly, a discrete random variable’s entropy, “𝑋, is defined as follows in Equation 1”: (1) H ( X ) = − ∑ x i ∈ X p ( x i ) log ( p ( x i ) )

Here, a specific outcome of the random variable 𝑋 is indicated by 𝑥𝑖, and the probability that 𝑥𝑖 will occur among the potential values of 𝑋 is shown by P(𝑥𝑖). According to another discrete random variable 𝑌, the conditional entropy of “𝑋 is defined as follows in Equation 2”: (2) H ( X / Y ) = − ∑ y i ∈ Y p ( y i ) ∑ x i ∈ X P ( x i / y i ) log ( P ( x i / y i ) )

The chance of seeing 𝑦𝑗 from 𝑌 is represented by 𝑃(𝑦𝑗) in this instance, but the conditional probability of seeing 𝑥𝑖 from 𝑋 provided that 𝑦𝑗 from 𝑌 has happened is shown by 𝑃(𝑥𝑖|𝑦𝑗). The Gain Information (GI) between 𝑋 and 𝑌 quantifies the amount of information communicated between both variables as follows in Equation 3: (3) IG ( X , Y ) = H ( X ) − H ( X / Y ) = ∑ x i ∈ X ∑ y i ∈ Y P ( x i , y i ) log ( P ( x i , y i ) P ( x i ) p ( y i ) )

(𝑥𝑖, 𝑦𝑗) refers to the joint probability of 𝑥𝑖 and 𝑦𝑗 occurring together (Zhang et al., 2024). Mutual Information (MI)-based feature selection is an approach for dimensionality reduction that works independently of classifiers. It seeks to address the difficulty of picking a meaningful collection of characteristics (Liu and Motani, 2022). Mutual information feature selection is a popular strategy for improving the effectiveness of IDSs. It assesses the link between each characteristic and the class label, deciding which features have the highest mutual information values (Alalhareth and Hong, 2023). The study used an RF classifier, which is an ensemble-based ML technique that integrates numerous DTs. To limit the danger of overfitting and improve the model’s generalizability, each DT in the forest is built separately from a randomly selected subset of the training data and features. The RF algorithm’s ultimate output is selected by a majority vote among all DTs. Each tree in the forest votes for the class of the input data point and the class that receives the most votes is chosen as the prediction (Ali et al., 2023). As shown in Table 1, multiple feature selection approaches discover critical properties for the IDS. The whole feature set has 41 characteristics, however, approaches like manual information acquisition and RF selection emphasize essential aspects like “src_bytes,” “dst_host_serror_rate,” and “count.” These features serve to improve the IDS model’s performance in identifying unusual network traffic.

Support Vector Machine (SVM): is commonly regarded as one of the most successful algorithms for binary classification, notably in IDSs (IDS), where transactions are classified as normal or invasive (Aldallal and Alisa, 2021). The primary goal of SVM is to identify a hyperplane within an n-dimensional feature space that maximizes the separation margin between classes. One of the key advantages of SVM is its ability to work effectively with smaller training datasets, as it only relies on a few support vectors to define the hyperplane. However, SVM performance can be negatively impacted by noise near the hyperplane (Alotaibi and Rassam, 2023).

Gradient boosting (GB): is an ensemble learning method that combines many decision trees to increase prediction accuracy. This method builds decision trees progressively, with each node making a binary choice. The model’s performance is progressively improved as each tree fixes the mistakes of the one before it. GB is an effective technique for challenging jobs because of its iterative approach, which produces forecasts that are more accurate (Boldini et al., 2023).

XGBoost: To improve efficiency and performance, XGBoost is a sophisticated and optimized variant of gradient boosting. It enhances the approach and system design of conventional GB models. Parallel processing, distributed computing, out-of-core execution, and cache optimization are some of the characteristics that XGBoost integrates to provide quicker processing and more accurate convergence to the global minimum. Its speed and accuracy are increased by these enhancements, which enable it to handle massive amounts of data across several devices (Thapa et al., 2020).

Logistic regression (LR): A classification method for categorical outcome prediction, logistic regression may be applied to both binary and multi-class classification applications. It uses the logistic function to determine the likelihood of an event happening, with values ranging from 0 to 1. Usually, a threshold of 0.5 is used to differentiate between two classes: values below 0.5 are categorized as class 0, whereas those over 0.5 are classified as class 1. F(x) = 1/(1 + e^-x) is the logistic sigmoid function, which is used to convert the input into a number between 0 and 1 that indicates the likelihood of a specific result. The wider application of the logistic sigmoid function in multi-class classification situations is not the same as this method (Somogyi, 2021).

Naïve Bayes (NB): is a variation of Bayes’ Theorem in which the qualities are assumed to be extremely independent of one another. It is a classification approach based on Bayes’ probability theory, with the assumption that the presence of one characteristic does not affect the chance of another (Devidas and Adesh, 2021). The Naïve Bayes method uses conditional probability and the premise that characteristics are independent. The classifier assigns the sample to the class with the highest probability after calculating the conditional probabilities for each class for each input (Useni et al., 2023).

Decision Tree (DT): For classification problems, the DT method is frequently utilized. It organizes data in a tree structure, with classifications determined by decisions made at every level. The branches show the results of those tests, the leaf nodes show the final classification, and each non-terminal node denotes a test or decision point.

Random Forest (RF): For both classification and regression tasks, RF, an ensemble learning technique, is employed. During training, it creates numerous DTs and forecasts the class using the majority vote from each tree. This strategy seeks to decrease overfitting and underfitting by averaging predictions to strike a balance between bias and variance (Devidas and Adesh, 2021).

Ensemble Model: This ensemble model guarantees thorough and accurate threat detection by combining the various strengths of the individual classifiers to improve classification accuracy through the use of Voting Hard and Stacking with DT, RF, and GB.

A number of essential criteria will be used to assess threat detection algorithms’ performance. These measurements will provide quantitative insights into how well the models recognize and respond to threats in real-time settings (Akinbolaji, 2023).

Accuracy: is the percentage of correctly categorized cases, which include both normal and pathological data points. as follows in Equation 4: (4) Accuracy = TP + TN TP + TN + FP + FN

Precision: refers to the fraction of forecasted threats that are valid. It is the proportion of accurately recognized anomalies to total presented anomalies. as follows in Equation 5: (5) Precision = TP TP + FP

Recall: The proportion of genuine positive records (e.g., true anomalies) that are accurately identified. It calculates the ratio of detected attacks to total attacks. as follows in Equation 6: (6) Recall = TP TP + FN

The F1 Score: is the harmonic mean of accuracy and recall, providing a single statistic for evaluating the model’s overall efficacy. as follows in Equation 7: (7) F 1 − Sccore = Recall ∗ Precision Recall + Precision

True Negative (TN) represents accurately anticipated normal cases.

False Negative (FN) indicates incidents that were mistakenly anticipated as normal.

True Positive (TP) represents successfully anticipated abnormal incidences.

False Positive (FP) indicates incidents that were wrongly anticipated as abnormal (John et al., 2024; Vibhute et al., 2024; Megouache et al., 2024).

All of the study’s experiments were carried out on a top-tier workstation running Windows 10 Pro and equipped with an AMD Ryzen 95,950X CPU (3.7 GHz) and 64 GB of RAM. This setup was intended to guarantee that repeated tests with various feature sets and numerous ensemble classifiers would go well. This configuration is not representative of situations with restricted resources, such as low-cost systems or edge devices. Future research will evaluate the effectiveness of the IDS on these platforms and investigate real-time optimization techniques including pruning, quantization, and compression.

Ensemble models outperform individual classifiers in measures like accuracy, precision, recall, and F1-score when using all available features as shown in Table 1.

Random Forest

Table 4 compares the performance of several classifiers using accuracy, precision, recall, and F1-score. RF, DT, and XGBoost all produced near-perfect results, with 99.99% accuracy and perfect precision, recall, and F1-score values of 1.00. NB has a much lower accuracy (78%), as well as worse precision, recall, and F1-score. SVM and LR both performed well, with accuracy rates of 99.26 and 97%, respectively, while retaining good precision, recall, and F1 scores. GB likewise produced outstanding results, with 99.82% accuracy and flawless precision, recall, and F1-score.

(DT+ RF+ GB) Voting hard

The ensemble model, which combines Decision Tree, Random Forest, and Gradient Boosting with hard voting, produced flawless classification results. It achieved an accuracy of 1.00, with precision, recall, and F1-score all at 1.00. This demonstrates strong predictive potential for the test results. As shown in Table 5 and Figure 6.

b. (DT+ RF+ GB) Stacking

The stacking ensemble model, which combines Decision Tree, Random Forest, and Gradient Boosting, likewise performed perfectly. It scored 1.00 for accuracy, precision, recall, and F1-score. This implies that stacking effectively harnessed the strengths of each base model. as shown in Table 5 and Figure 7.

Table 6 shows the performance of an ensemble model that combines DT, RF, and GB with the Voting Hard and stacking methods. The ensemble models performed flawlessly across all criteria, with accuracy, precision, recall, and an F1 score of 1.00.

The performance of Individual Classifiers by Gain Information and Manual Information, as shown in Table 1 is evaluated based on accuracy, precision, recall, and F1-score, the evaluation of different classifiers and ensemble models reveals the advantages and disadvantages of each model for certain classification tasks. In order to improve overall performance, this section covers both single classifiers and ensemble models, which integrate many classifiers.

XGBoosting

The GB model performed excellently with an accuracy of 99.89%. It achieved perfect values of 1.00 for accuracy, recall, and F1-score, indicating flawless classification of both positive and negative cases. This model outperformed all other classifiers, as shown in Table 7.

g. Logistic Regression

The performance of each classifier is displayed in Table 9, with particular attention paid to accuracy, precision, recall, and F1-score. Table 9 compares the performance of multiple classifiers, revealing that RF, DT, XGBoost, and GB all obtained near-perfect results, with 99.97% or 99.89% accuracy and flawless precision, recall, and F1 scores. In contrast, NB had an extremely low accuracy of 0.63%, as well as poor precision, recall, and F1 scores. SVM and LR performed moderately, with accuracy rates of 88.09 and 72.0%, respectively.

(DT+ RF+ GB) Voting hard

In order to attain flawless detection performance across all measures, ensemble models incorporate multiple classifiers, as shown in Table 11.

Table 11 shows the performance of ensemble models that combine DT, RF, and GB with the Voting Hard and Stacking approaches. The ensemble models performed flawlessly across all measures, including accuracy, precision, recall, and an F1-score of 1.00.

The performance of Individual Classifiers by RF-Selected Features, as shown in Table 1 evaluated based on accuracy, precision, recall, and F1-score of separate classifiers and ensemble models are compared in this investigation. The objective is to evaluate the performance of each model separately and the classification skills of ensemble models, which combine many classifiers.

Decision Tree

GB showed poor performance with an accuracy of 49.35%, precision of 0.48, recall of 0.29, and F1-score of 0.29. This underperformance may be due to overfitting or a mismatch with the dataset’s characteristics, like data imbalance or feature correlation, as shown in Table 12.

e. Random Forest

XGBoost achieved 98.79% accuracy, with precision, recall, and F1-score of 0.76, 0.74, and 0.74, respectively. Although strong, its performance was slightly lower than RF, possibly due to hyperparameter settings or the dataset’s structure, as shown in Table 12.

g. Support Vector Machine

SVM achieved 92.40% accuracy but with poor precision (0.39), recall (0.27), and F1-score (0.29). Its low performance suggests it struggles with the dataset’s complexity, possibly due to kernel selection or insufficient tuning, as shown in Table 12 and Figure 16.

Table 13 summarizes the performance of different classifiers, emphasizing accuracy, precision, recall, and F1 score.

Table 13 shows the performance evaluations of several classifiers. RF and DT produced reasonably high accuracy (98.83 and 98.71%, respectively), but with modest precision, recall, and F1 scores. XGBoost also fared well, with 98.79% accuracy, although it had lesser precision and recall than RF and DT. NB performed poorly, with an accuracy of just 33% and low precision, recall, and F1 scores. SVM and LR fared rather well, with accuracy values of 92.40 and 90.56%, respectively, but with low precision and recall. GB had the poorest result, with an accuracy of 49.35% and similarly low precision, recall, and F1 score.

(DT+ RF+ GB) Voting hard

The Voting Hard ensemble model, consisting of DT, RF, and GR, achieved 0.99 accuracy. However, its precision (0.71), recall (0.68), and F1-score (0.68) were slightly lower than the stacking model, indicating room for improvement in precision and recall, as shown in Table 14 and Figure 17.

b. (DT+ RF+ GB) Stacking

The ensemble model, which used stacking to combine Decision Tree, Random Forest, and Gradient Boosting, obtained an accuracy of 0.99, a precision of 0.75, a recall of 0.68, and an F1 score of 0.70. This represents higher accuracy and F1-scores compared to the hard voting, stacking model, indicating room for improvement in precision and recall, as shown in Table 14 and Figure 18.

Table 15 evaluates the ensemble models, which integrate several classifiers to attain flawless detection performance across all criteria.

Table 15 shows the performance of the ensemble model that combines DT, RF, and GR with the Voting Hard and stacking approaches. The ensemble model with stacking outperforms hard voting in terms of precision (0.75 vs. 0.71), F1-score (0.70 vs. 0.68), and accuracy (0.99) while retaining recall (0.68). This shows that stacking gives a more balanced and effective mix of classifiers, increasing the model’s capacity to properly detect positive situations while maintaining overall accuracy.

The study used five-fold cross-validation on individual models, such as Random Forest (RF) and Gradient Boosting (GB), to evaluate and compare classifier performance. This ensured that the findings showed strong generalization and did not overfit to a particular data split (20%test_size, 80%train_size).

Furthermore, a hard Voting Classifier was assessed as compared to its base learners. While the differences between RF (t = 0.3780, p = 0.7055) and GB (t = 0.7746, p = 0.4386) were not statistically significant, the Voting Classifier performed much better than the Decision Tree (t = 4.0835, p < 0.0001), according to the results of the t-test. Given that it outperforms the strongest base models while enhancing overall resilience, this demonstrates the Voting ensemble’s resilience.

In addition, t-tests were used to choose features using Mutual Information and Information Gain in order to verify that the chosen features improved generalization by reducing dimensionality and greatly enhancing model performance.

SHAP values for the Random Forest and Decision Tree classifiers were calculated to interpret model options. In order to visualize the effects of features on both individual predictions and the overall relevance of features, SHAP provides both local and global interpretability.

According to the SHAP study, several variables were consistently more influential in both models, which is consistent with behaviors associated with documented network intrusions. This consistency provides information about possible dimensionality reduction techniques in addition to confirming the models’ validity.

While the Decision Tree model’s simple structure makes it easier to grasp., the Random Forest model outperformed it while allowing for insight through SHAP. SHAP’s ability to clearly explain feature contributions reduced the complexity of the ensemble, proving that accuracy and interpretability do not have to be mutually incompatible.

Figure 19 demonstrates that while service_auth is of little significance, network packet sizes (src_bytes, dst_bytes) and connection flags are the most important features for attack detection in the model. This illustrates how the model successfully identifies malicious activity by concentrating on network traffic patterns.

As seen in Figure 20: The graphic shows that while service authentication data (service_auth) makes very little contribution to attack detection, network packet metrics (src_bytes, dst_bytes) and connection flags dominate feature relevance. These outcomes demonstrate how well the model uses fundamental network traffic features to identify malicious behavior across all threat categories.

Table 16 compares several ML algorithms used to identify intrusions on the NSL-KDD dataset. Accuracy rates vary greatly based on the classifier and feature selection strategy employed.