Secure and Reliable Intrusion Detection System for Cyberattack Detection in IoT-Edge-Enabled Cyber-Physical Systems

The fast development of the Internet of Things (IoTs) has significantly reformed the landscape of modern cyber-physical systems, enabling seamless connectivity among smart devices, sensors, edge nodes, and cloud infrastructure [1]. IoT-enabled devices, ranging from consumer wearables to industrial sensors and autonomous vehicles, continuously generate and transmit large volumes of continuous-stream information for intelligent processing and actionable insights [2]. To manage this data efficiently and minimize latency, edge computing has emerged as a fundamental standard by enabling localized data analysis closer to the source [3]. When local computational capacity is exceeded, edge devices seamlessly offload tasks to cloud platforms, facilitating large-scale processing and long-term data storage [4]. This hierarchical data flow across IoT, edge, and cloud layers ensures system-wide responsiveness and scalability in areas like healthcare, smart city, and industrial applications.

Despite the operational benefits of this integrated architecture [5], IoT systems remain highly vulnerable to cyber threats, particularly at the communication layers bridging IoT-Edge and Edge-Cloud nodes [6]. IoT devices possess limited storage, computing power, battery capacity, and processing capabilities; consequently, they present an attractive attack surface for adversaries [7]. Figure 1 illustrates various cyber threats, including domain-based spoofing, Distributed Denial-of-Service (DDoS), and multi-layer link attacks [8].

Figure 1. Denial-of-Service (DoS) attack

Cyberattacks such as Man-in-the-Middle (MitM) intrusions, Domain Name System (DNS) spoofing, and DDoS attacks [8], as shown in Figure 1, exploit these vulnerabilities to disrupt services, compromise data integrity, and endanger user privacy [9]. Among these, network-level attacks such as link-flooding DDoS pose a substantial risk by overwhelming edge nodes with malicious traffic, leading to degraded performance and system outages [10].

In recent years, numerous intrusion detection systems (IDSs) [11] based on machine learning and deep learning models have been developed to enhance IoT network security [12, 13]. Machine learning (ML)-based approaches typically rely on handcrafted traffic features, whereas deep learning (DL)-based techniques employ hierarchical neural architectures to capture complex attack patterns. However, existing ML/DL-based IDS suffer from several critical limitations, including high computational overhead unsuitable for resource-constrained IoT-edge environments, poor adaptability to evolving and zero-day attacks, sensitivity to class imbalance in real-world traffic datasets, and delayed threat detection caused by static learning mechanisms [14, 15]. These challenges result in increased misclassification rates and reduced reliability, particularly at the dynamic IoT-Edge-Cloud interface, thereby limiting their effectiveness in real-time cyber-physical systems [16, 17].

From the above discussion, it is evident that current ML and DL-based intrusion detection frameworks lack adaptive learning capability, computational efficiency, and real-time responsiveness, especially under dynamic and large-scale IoT-edge deployments. Existing approaches are often unable to efficiently balance detection accuracy with low latency and scalability, highlighting the need for a lightweight yet adaptive intrusion detection framework capable of addressing evolving attack behaviors at the network and communication layers.

In addressing these shortcomings, this work proposes a new intrusion detection framework, namely Gradient Tree Multi-Layer Perceptron–Aware Secure and Reliable Security (GTMLP-SRS), designed to enhance cybersecurity in IoT-edge-enabled cyber-physical systems. The GTMLP-SRS model integrates gradient-boosted decision trees with a multi-layer perceptron to effectively capture both linear and non-linear patterns in network traffic behavior. By incorporating statistical monitoring and adaptive learning mechanisms, the framework dynamically responds to evolving attack strategies, ensuring timely and accurate threat detection with minimal computational overhead.

The primary research objectives of this work are summarized as follows:

• To design an adaptive intrusion detection framework capable of accurately identifying network- and communication-layer attacks in IoT-edge-enabled cyber-physical systems.
• To integrate gradient tree boosting with a multi-layer perceptron for enhanced feature representation and classification of complex attack patterns.
• To incorporate real-time statistical monitoring to improve adaptability against evolving and previously unseen cyber threats.
• To achieve low-latency and computationally efficient intrusion detection suitable for resource-constrained IoT-edge environments.
• To validate the effectiveness of the proposed GTMLP-SRS framework through benchmark evaluations and comparative analysis with existing ML and DL-based IDS methods.

A scalable and computationally efficient intrusion detection framework is thus proposed, enhancing secure and reliable communication between IoT, edge, and cloud layers. By integrating GTMLP-SRS into the IoT-edge-based architecture, the proposed solution strengthens the overall system security, ensuring resilient and uninterrupted communication across cyber-physical environments.

Section 2 reviews relevant literature on ML and DL-based intrusion detection methods. Section 3 presents the GTMLP-SRS model architecture deployed at the IoT-edge layer. Section 4 discusses experimental results and comparative evaluations. Section 5 concludes the paper with key findings and future research directions.

3.2 Gradient tree multi-layer perception-based secure model

Let the time-dependent behavior of IoT-collected observations be denoted by a dataset $X$, defined as Eq. (1) and here, ${{x}_{t}}$ represents the feature vector at time step $t$, and ${{y}_{t}}\in \left\{ 0,1 \right\}$, where ${{y}_{t}}=1$ indicates a detected attack and ${{y}_{t}}=0$ denotes normal behavior. The GTMLP-SRS model $F$ performs a mapping from input features $X$ to output labels $y$, defined as Eq. (2).

$X=\left\{ \left( {{x}_{1}},{{y}_{1}} \right),\left( {{x}_{2}},{{y}_{2}} \right),\ldots \left( {{x}_{t}},{{y}_{t}} \right) \right\}$                         (1)

$F=f\left( X \right)=\underset{k=1}{\overset{K}{\mathop \sum }}\,{{\theta }_{k}}{{h}_{k}}\left( X \right)$                                      (2)

In this expression, $K$ is the number of decision trees within the ensemble, ${{\theta }_{k}}$ is the assigned weight to the ${{k}^{th}}$ tree, and ${{h}_{k}}\left( X \right)$ is the output of that tree for input $X$. The model’s training objective is to minimize a composite loss function composed of binary cross-entropy loss and a regularization term to avoid overfitting in Eq. (3). The first term measures the prediction error between the actual and predicted classes, while the second term introduces a regularization penalty to prevent overfitting. Minimizing this loss function enables the model to learn optimal parameters for accurate intrusion detection. The attack probability ${{p}_{t}}$ is estimated (between 0 and 1) using a sigmoid activation function in Eq. (4). This probabilistic output indicates the likelihood that a given network traffic instance belongs to the attack class, making it suitable for binary intrusion detection tasks. The regularization function $\Omega \left( {{h}_{k}} \right)$ for the ${{k}^{th}}$ tree is expressed as in Eq. (5).

$L=\underset{t=1}{\overset{N}{\mathop \sum }}\,\left[ {{y}_{t}}\log \left( {{p}_{t}} \right)+\left( 1-{{y}_{t}} \right)\log \left( 1-{{p}_{t}} \right) \right]+\underset{k=1}{\overset{K}{\mathop \sum }}\,\Omega \left( {{h}_{k}} \right)$                         (3)

${{p}_{t}}=\sigma \left( f\left( {{x}_{t}} \right) \right)=\frac{1}{1+{{e}^{-f\left( {{x}_{t}} \right)}}}$                       (4)

$\Omega \left( {{h}_{k}} \right)=\gamma N+\frac{1}{2}\lambda {{\left| \left| {{w}_{k}} \right| \right|}^{2}}$                               (5)

Here, $N$ denotes the overall leaf node considered within a tree, ${{w}_{k}}$ is the weight vector for leaf nodes, and $\gamma $, $\lambda $ are hyperparameters control model complexity to regulate the magnitude of the Multi-Layer Perceptron (MLP) weights ${{w}_{k}}$. This regularization mechanism helps reduce model overfitting and enhances generalization by penalizing overly complex models. The GTMLP-SRS model learns optimal splits and weight assignments to balance prediction accuracy and generalizability. To further improve the detection accuracy and competence in attaining generalizability of the GTMLP-SRS model, an MLP is employed to optimize the hyperparameters and facilitate ensemble model construction. The MLP, acting as a meta-learner, is designed to fine-tune key constraints like estimators $K$, learning rate, tree size, and regularization coefficients $\gamma $ and $\lambda $ defined in Eq. (6). The MLP architecture comprises an input-layer consistent to the feature dimensionality with respect to $X$, then, it has multiple hidden-layers composed of ReLU activations, and an output layer with sigmoid activation to map the final prediction probability. The loss function is designed using binary cross-entropy function for performing the training process of MLP networks through backpropagation algorithm and an adaptive learning rate optimizer. To identify the optimal configuration of the GTMLP-SRS model, a grid-based cross-validation (GridSearchCV) approach is employed. This technique systematically explores combinations of hyperparameters across a predefined search space:

$H=\left\{ max\_depth,learnin{{g}_{rate}},K,\gamma ,\lambda  \right\}$                          (6)

In Eq. (6), each candidate set of hyperparameters in $H$ is evaluated which includes key parameters such as max_depth and learning_rate for the gradient tree component, the number of neurons $K$ in the MLP, and the regularization parameters $\gamma $and $\lambda $ using $k$-fold cross-validation (CV) mechanism. A k-fold cross-validation strategy systematically evaluates all possible combinations of these parameters to identify the configuration that provides optimal detection performance. In performing CV, the data considered during training is equally divided into $k$ segments. The GTMLP-SRS is iteratively trained on $k-1$ folds; the feature weights are optimized and validated with remaining data. The GTMLP-SRS model training performance is studied considering various receiver operating characteristic metrics and are recorded for each configuration. Upon completion, the combination yielding the highest cross-validated performance is selected, and the corresponding GTMLP-SRS ensemble classifier is reconstructed using the optimal parameters. This MLP-driven grid search mechanism ensures that the final model is both well-calibrated and robust against overfitting while maintaining high sensitivity to temporal and statistical shifts in attack behavior.

The MLP employed in the proposed system is a fully connected feedforward neural network that approximates nonlinear mappings between the input IoT behavioral feature space and the binary output space $\left\{ 0,1 \right\}$, indicating normal or attack behavior. Let the input feature vector at time $t$ be ${{x}_{t}}\in {{R}^{d}}$, where $d$ is the feature dimension. The MLP consists of $L$ layers, including $L-1$ hidden-layers and one output-layer. The transformation of the input through the network is defined recursively as follows in Eqs. (7)-(9):

${{z}^{\left( l \right)}}={{W}^{\left( l \right)}}{{a}^{\left( l-1 \right)}}+{{b}^{\left( l \right)}},~~l=1,2,\ldots ,L$                    (7)

${{a}^{\left( l \right)}}={{\phi }^{\left( l \right)}}\left( {{z}^{\left( l \right)}} \right),~~\text{for}~l=1,2,\ldots ,L$                                (8)

${{\hat{y}}_{t}}=\sigma \left( {{z}^{\left( l \right)}} \right)=\frac{1}{1+{{e}^{-{{z}^{\left( L \right)}}}}},~~\text{for}~l=1,2,\ldots ,L$                        (9)

Eq. (9) represents the final output layer, where the sigmoid function converts the network output into a probability value between 0 and 1. This probability ${{\hat{y}}_{t}}$ indicates whether a given network traffic instance is classified as normal or malicious, enabling binary intrusion detection in the GTMLP-SRS framework. In these equations: ${{a}^{\left( l \right)}}={{x}_{t}}$ is the input layer, ${{W}^{\left( l \right)}}\in {{R}^{{{n}_{l}}\times {{n}_{l-1}}}}$ and ${{b}^{\left( l \right)}}\in {{R}^{{{n}_{l}}}}$ are the weighted matrices and bias term at the layer $l$, ${{\phi }^{\left( l \right)}}\left( \cdot \right)$ is the activation process considering layer $l$; typically ReLU are used in hidden layers: $\phi \left( z \right)=\text{max}\left( 0,z \right)$, $\sigma \left( \cdot  \right)$ denotes the sigmoid function used at the output layer for binary classification. The MLP is trained using a dataset $\left\{ \left( {{x}_{t}},{{y}_{t}} \right) \right\}_{t=1}^{N}$, and the optimization aim are to reduce the loss by employing binary cross-entropy process using below minimization Eq. (10):

 ${{\mathcal{L}}_{MLP}}=-\frac{1}{N}\underset{t=1}{\overset{N}{\mathop \sum }}\,\left[ {{y}_{t}}\log \left( {{{\hat{y}}}_{t}} \right)+\left( 1-{{y}_{t}} \right)\log \left( {{{\hat{y}}}_{t}} \right) \right]$                     (10)

Eq. (10) defines the binary cross-entropy loss function used to train the MLP. In this formulation, ${{y}_{t}}$ represents the true label of the ${{t}^{th}}$ training sample, while ${{\hat{y}}_{t}}$ denotes the predicted probability output by the MLP. The loss function measures the discrepancy between the predicted and actual labels across all $N$ samples in the dataset. By minimizing this loss during training, the MLP learns model parameters that improve classification accuracy for distinguishing between normal and malicious network traffic in the proposed GTMLP-SRS framework. During training, gradient descent or a variant such as Adam is used to update the weights ${{W}^{\left( l \right)}}$ and biases ${{b}^{\left( l \right)}}$ to minimize ${{\mathcal{L}}_{MLP}}$. Once trained, the MLP model is used to score and select optimal hyperparameter configurations in the search space $H$ using performance metrics computed using GridSearchCV, where each hyperparameter tuple is evaluated through cross-validation. By integrating MLP in this optimization loop, the GTMLP-SRS model benefits from automatic, data-driven tuning of its structure and training dynamics, improving both attack prediction accuracy and generalization performance across varying IoT environments.

3.3 Secure and reliable security model for Gradient Tree Multi-Layer Perceptron

Given the evolving nature of attack patterns, the model incorporates a statistical adaptation mechanism to detect behavioral drift in real-time. A high divergence value indicates a significant deviation, prompting an update of the time window $W$. The window is divided into two segments: a stable historical segment $S$ with sub-window ${{S}_{sub}}$, and a recent testing segment $T$, where $\mid S\mid \gg \mid T\mid $. To statistically validate behavioral changes [36], a two-sample t-test is conducted in Eq. (11):

$test-t=\frac{\bar{S}-\bar{T}-\left( {{\mu }_{1}}-{{\mu }_{2}}-\delta  \right)}{{{\sigma }_{w}}\sqrt{\left( \frac{1}{{{n}_{1}}}+\frac{1}{{{n}_{2}}} \right)}}\sim t\left( {{n}_{1}}+{{n}_{2}}-2 \right)$                       (11)

Eq. (11) represents a two-sample t-test used to statistically verify whether there is a significant behavioral change between historical and recent network traffic patterns. In this formulation, $\bar{S}$, $\bar{T}$ denote the mean values of the stable historical segment and the recent testing segment, respectively. The parameters ${{n}_{1}}$and ${{n}_{2}}$ represent the sample sizes of these two segments, ${{\sigma }_{w}}$ indicates the pooled standard deviation, ${{\mu }_{1}}$, ${{\mu }_{2}}$ are population means. The term $\delta $ accounts for an allowable threshold between the two segments. The computed test statistic follows a t-distribution with $\left( {{n}_{1}}+{{n}_{2}}-2 \right)$ degrees of freedom. A high-test statistic value indicates a statistically significant deviation in recent behavior, thereby triggering an update of the time window $W$ for adaptive intrusion detection. If the test indicates significant behavioral change, the window is reset ($W=\varnothing $) and the model is retrained from the updated observation ${{x}_{1}}={{x}_{2}}$. Otherwise, the model remains unchanged. Once trained, the GTMLP-SRS model predicts threats in new IoT input streams ${{X}_{new}}$ as in Eq. (12).

$F=f\left( {{X}_{new}} \right)=\underset{k=1}{\overset{K}{\mathop \sum }}\,{{\theta }_{k}}{{h}_{k}} · \left( {{X}_{new}} \right)$                            (12)

Eq. (12) describes the final decision function of the trained GTMLP-SRS model, where the output $F$ is computed as a weighted combination of $K$ learned components ${{h}_{k}}\left( {{X}_{\text{new}}} \right)$. Each component captures distinct behavioral patterns from the incoming IoT data stream ${{X}_{\text{new}}}$, and the corresponding weights ${{\theta }_{k}}$ reflect their relative importance in the final prediction. The probability of attack occurrence is estimated as in Eq. (13). An attack is flagged $\left( y=1 \right)$ if ${{p}_{attack}}$ surpasses a predefined threshold; otherwise, normal behavior is presumed.

${{p}_{attack}}=\sigma \left( f\left( {{X}_{new}} \right) \right)$                          (13)

Eq. (13) applies the sigmoid activation function to the aggregated output $f\left( {{X}_{\text{new}}} \right)$, converting it into a probability value ${{p}_{\text{attack}}}$ between 0 and 1. This probability represents the likelihood of an attack occurrence. If ${{p}_{\text{attack}}}$ exceeds a predefined threshold, the input is classified as malicious ($y=1$); otherwise, it is considered normal behavior. This probabilistic decision mechanism enables reliable and adaptive intrusion detection in real-time IoT-edge environments.

3.4 Flow diagram and algorithm

The GTMLP-SRS algorithm operates in six systematic stages to detect network and communication attacks in IoT-edge environments as shown in Algorithm 1. The complete flow of proposed work is given in Figure 2.

Figure 3. Flow diagram of proposed pipeline model

As shown in Algorithm 1 and Figure 3, initially, the model is trained on historical IoT data using a regularized log-loss function to prevent overfitting. Hyperparameters such as the number of estimators and tree complexity are optimized using a Multi-Layer Perceptron (MLP) with GridSearchCV for enhanced accuracy. As new data arrives, Kullback-Leibler divergence is computed to identify distributional changes in device behavior. If significant divergence is observed, a statistical t-test compares historical and new data to confirm behavioral shifts. Upon detecting a change, the model is retrained with the updated data window. For each new observation, the GTMLP-SRS computes the probability of an attack using a sigmoid-activated output. If this probability exceeds a predefined threshold, an attack is predicted. This adaptive learning framework ensures robust and dynamic detection of evolving threats in real-time IoT-edge environments. The proposed GTMLP-SRS algorithm integrates temporal behavior analysis, adaptive retraining, and dynamic thresholding to accurately detect evolving attack patterns. MLP-driven hyperparameter tuning further enhances model performance through ensemble optimization and grid-based validation.

5.3 Statistical significance and real-world impact discussion

To strengthen the comparative evaluation, a paired statistical significance test was conducted between GTMLP-SRS and the most competitive baseline models on both ToN-IoT and UNSW-NB15 benchmarks. As shown in Table 4, the obtained p values are consistently below 0.01, confirming that the observed performance improvements are statistically significant and not due to random variation.

Although the absolute accuracy improvement of 0.3–0.5% may appear marginal, its real-world impact is substantial in large-scale IoT deployments. In high-throughput CPS-IoT environments processing millions of packets daily, such an improvement can translate into thousands of additional attack instances correctly detected, significantly reducing missed intrusions and false negatives. This enhancement directly improves service availability, data integrity, and operational safety, particularly in mission-critical applications such as smart grids, healthcare monitoring, and industrial automation. The results therefore demonstrate both the statistical robustness and practical relevance of the proposed GTMLP-SRS framework.