Trust-Aware Secure Routing in Wireless Sensor Networks Using HSCRNet-Based Trust Prediction and FAABO Optimization

2.1 Related works

In 2022, Bin-Yahya et al. [18] initiated the routing protocol for enhancing the trust and energy efficiency in WSN using the adaptive algorithm. The new framework was developed to tackle the general routing attack. The filtering mechanism and the adaptive penalty factor were used to determine the values and volatilization. Based on the cluster head, the optimal secure routing was performed by the adaptive algorithm.

In 2020, Xu et al. [19] introduced the heuristic-based secure routing protocol for achieving efficient performance in the WSN network. The AI was utilized by the designed protocol to determine the intellectual and reliable scheme. The developed model achieved greater security to protect the data transmission with less complexity. The outcome from the simulation demonstrated that the designed protocol enhanced the network efficiency in dynamic scenarios.

In 2023, Yang et al. [20] presented the system using a quad-tree structure to minimize the complexity in the network management performance. The location and ID were determined to construct the authentication sensor using the encryption algorithm. This model avoids illegitimate sensor nodes to provide high security. The event metrics and time were considered to perform the cluster head selection to enhance communication. The generative algorithm was further used to secure the multi-path routing based on the hop count.

In 2024, Siddiq and Ghazwani [21] suggested the arithmetic encoding-based approach for the energy-efficient and secure WSN. The sensor node initialized after the WSN setup using the standard algorithm. The significant node feature was selected continuously to enhance the lifetime. The higher throughput achieved by the implemented model was revealed by the experimental findings.

In 2023, Salim et al. [22] adopted energy efficient and stable cluster selection using the advanced protocol in the WSN network. The random number generation was performed to stabilize the proposed work, which includes the CH selection using the node coverage. The finding revealed the energy-efficient performance achieved in the WSN by the adopted approach.

In 2023, Verma et al. [23] developed an innovative framework using Q-learning for secure routing. The optimized Q-learning and the unclonable function were utilized by the algorithm to ensure a reliable transmission path. The performance of the implemented algorithm was analyzed by conducting a performance analysis, which enhanced the filtering, and packet delivery ratio.

In 2024, Ahmed et al. [24] designed secure data transmission using the adaptive clustering model. The information was gathered initially based on the randomly placed nodes, which ensures the node's fair energy dissipation. The initiated protocol selected the cluster head and determined the intrusion detection to identify the presented malicious nodes in it. The malicious nodes were distinguished from the legal nodes based on the interference approach.

In 2022, Feng et al. [25] generated the multi-hop routing model in WSN using an advanced algorithm. The data transmission and the cluster head selection were included in this work for achieving the multi-hop routing. For the cluster head, the sensor nodes were sent the information, and it was further transmitted through the base station for the optimal hop selection. The trust model was introduced to perform the security-aware multi-hop routing.

Table 1. Comparative analysis of existing secure routing approaches in Wireless Sensor Networks (WSNs)

Note: WSNs: Wireless Sensor Networks; IDS: Intrusion Detection System; CH: Cluster Head; PDR: Packet Delivery Ratio; QoS: Quality of Service; IoT: Internet of Things.

Table 1 demonstrates that existing studies have significantly improved secure routing performance through trust management, clustering, optimization, and machine-learning-based approaches. However, most methods focus on either trust estimation or routing optimization independently. Furthermore, limited attention has been given to simultaneously handling multiple routing attacks while optimizing trust, throughput, delay, packet delivery ratio, scalability, and energy consumption. These limitations motivate the development of the proposed HSCRNet–FAABO framework.

2.2 Research gaps and challenges

Although several trust-aware and secure routing approaches have been developed for WSNs, significant challenges remain unresolved. Existing methods primarily focus on either trust evaluation or routing optimization independently, resulting in limited overall network performance. Most routing schemes fail to simultaneously optimize trust, throughput, scalability, packet delivery ratio, delay, and energy consumption. Furthermore, several existing trust models rely on static trust estimation mechanisms that cannot effectively capture temporal node behavior variations.

Another limitation is that many routing protocols are evaluated only under generic attack conditions without explicitly considering blackhole, wormhole, and Sybil attacks. In addition, optimization-based routing algorithms frequently suffer from premature convergence and reduced exploration capability when network size increases. To address these limitations, this work proposes the HSCRNet model for temporal trust prediction and the FAABO algorithm for multi-objective trust-aware secure routing optimization in WSNs.

4.1 Temporal Convolution Network

TCN [30] is a kind of convolution network that has three different modules in its structure called dilated residual and casual convolution. By using these convolution modules, the long-term features can be integrated. Here the result sequences $\widehat{k}_0, \ldots, \widehat{k}_T$ are predicted using the initial sequence of input $z_0, \ldots, z_T$. The prediction is generated by the input sequence model by utilizing Eq. (1).

$\hat{k}_0, \ldots, \hat{k}_T=f\left(z_0, \ldots, z_T\right)$   (1)

Here the model temporal sequence is evaluated based on the function f. The expected loss is minimized by the implementation of this function between the predicted and the actual result. The significant process of the TCN structure is to generate the outcome that corresponds to the length of the input and solve the leakage based on the presented casual convolution. The filter is adopted by the network to determine the larger region, where the length can be expanded. The original filter is expanded by adding the zero padding and it is formulated using Eq. (2).

$F(v)=(Z * f)(v)=\sum_{i=0}^{m-1} f(i) \cdot Z_{v-d . i}$   (2)

Here the term $d$ indicates the dilation factor, $l$ defines the convolution filter. The term $v-d$ is the earlier index measure. Further, the number of layers is analyzed to evaluate the complete coverage based on the receptive kernel derived from Eq. (3).

$n=[(m-1) /(l-1)]$   (3)

Here, n is considered to be the number of layers. There are multiple filters in each layer for performing the feature extraction.

4.2 Dataset preparation and training procedure

The trust prediction dataset was generated from simulated WSN environments containing trusted and malicious nodes. Each sample consisted of eight attributes: node position, long link indicator, bit error rate (BER), number of links, traffic volume, lambda value, total transmission length, and modulation type. All numerical features were normalized to the range [0,1] using min-max normalization. Trusted labels were assigned to nodes exhibiting packet forwarding ratios greater than 90% and normal communication behavior. Malicious labels were assigned to nodes performing packet dropping, route manipulation, or identity spoofing activities. A sliding time window of 20 communication rounds was used for feature extraction. The dataset was divided into 70% training samples, 15% validation samples, and 15% testing samples. Model training was performed using the Adam optimizer with a learning rate of 0.001 and a batch size of 32.

4.3 Capsule network

CapsNet [30] is a type of Artificial Neural Network (ANN) that is mainly developed to perform the duplication in the biological neural organization. One of the significant components in the network structure is the capsule layer. The main performance of this capsule layer is to affine the transformation with squashing and weighted sum. The affine transformer is more useful in determining the real orientation based on the feature obtained by the convolutional layer. The expression of the prediction vector is computed using Eq. (4).

$\hat{g}_{z \mid j}=k_{j z} \cdot g_j$   (4)

Here, the term $\hat{g}_{z \mid j}$ is the predictive vector with the weight matrix $k_{j z}$. The resultant vector from the capsule is represented as $g_j$. The evaluation of the weighted vector is determined using Eq. (5).

$w_z=\sum n_{j, k} \hat{g}_{z \mid j}$   (5)

Here, the weighted sum is indicated as $w_z$ with the coefficient $n_{j, z}$. The result is evaluated using Eq. (6).

$g_j=\frac{\left\|y_j\right\|^2 y_j}{1+\left\|y_j\right\|^2\left\|y_j\right\|}$   (6)

Here, $y_j \in E b_{E x}$ implies the input with the resultant representation $g_j \in w t_2$. Based on this, the system can evaluate the spatial relationship for the given input.

4.4 Training data generation and feature engineering

The trust prediction dataset was generated from simulated WSN environments containing both trusted and malicious nodes. The dataset was created by varying the network size from 50 to 200 sensor nodes under different communication conditions and attack scenarios. Each node record consisted of eight attributes, namely node position, long-link indicator, bit error rate (BER), number of links, traffic volume, lambda value, total transmission length, and modulation type. Prior to model training, all numerical features were normalized into the range [0,1] using min-max normalization to ensure stable convergence of the deep learning model. Trusted labels were assigned to nodes exhibiting normal packet forwarding behavior and forwarding ratios greater than 90%, whereas malicious labels were assigned to nodes performing packet dropping, route manipulation, selective forwarding, Sybil, or blackhole attack activities. A sliding temporal window containing 20 communication rounds was employed to capture sequential node behavior patterns. The generated dataset was divided into 70% training samples, 15% validation samples, and 15% testing samples. The HSCRNet model was trained using the Adam optimizer with a learning rate of 0.001 and a batch size of 32. The training process continued until convergence was achieved with minimum validation loss. The extracted temporal features from the TCN were subsequently provided to CapsNet for trusted-node classification. The diagram of the proposed HSCRNet for trusted node prediction is shown in Figure 3.

image025.jpg

Figure 3. Detailed Hybrid Serial Cascaded Residual Network (HSCRNet) architecture for trusted and malicious node classification using communication and network features

6.1 Experimental setup

The proposed framework was implemented and evaluated using MATLAB R2020a in a simulation environment covering an area of 1000 m × 1000 m. The network consisted of 50, 100, 150, and 200 sensor nodes randomly deployed within the simulation area, with the sink node positioned at the center (500 m, 500 m). Each sensor node was initialized with an energy of 2 J, and a First-Order Radio Energy Model was employed to estimate energy consumption during wireless communication. The transmission range of each node was set to 100 m, while the packet size was fixed at 512 Bytes. To evaluate the robustness of the proposed security framework, different network attack scenarios were considered by introducing 10%, 20%, and 30% malicious nodes implementing Blackhole, Wormhole, and Sybil attacks. The simulations were executed for 1000 rounds, and a random seed value of 42 was used to ensure reproducibility of the experimental results. Performance evaluation was conducted using metrics such as Packet Delivery Ratio (PDR), Throughput, End-to-End Delay, Packet Loss Ratio (PLR), Energy Consumption, Network Lifetime, Detection Accuracy, Precision, Recall, and F1-Score. All simulations were carried out on a system equipped with an Intel® Core™ i7 processor, 16 GB RAM, and Windows 10 (64-bit) operating.

6.2 Evaluation metrics

Accuracy, mean absolute error and mean absolute scaled error are evaluated using Eqs. (15), (16) and (17).

$A c y=\frac{T k+T s}{T k+T s+F k+F s}$   (15)

Here the term $T k$ and $T s$ denoted the true positive and negative. Similarly, the term $F k$ and $F s$ specifies the false positive and false negative respectively.

$M A E=\frac{\sum_{i=1}^z\left(u_i-s_i\right)}{k}$   (16)

$M A S E=\frac{1}{k} \sum_{i=1}^z\left(\frac{s_i-u_i}{s_i}\right)$   (17)

Mean percentage error, root mean square error and symmetric mean absolute percentage error demonstrated using Eqs. (18), (19) and (20).

$M E P=\frac{100 \%}{k} \sum_{i=1}^z\left(\frac{s_i-u_i}{s_i}\right)$   (18)

$R M S E=\sqrt{\frac{\sum_{i=1}^z\left(u_i-s_i\right)^2}{k}}$   (19)

$S S M A P E=\frac{100}{k} \sum_{i=1}^k \frac{\left|u_i-s_i\right|}{\left(\frac{\left|u_i\right|+\left|s_i\right|}{2}\right)}$   (20)

Here the observed and the predicted value is represented as $u$ and $s$ respectively.

6.3 Convergence calculation of Fitness Augmented African Bison Optimization algorithm

Determining the convergence for the implemented FAABO algorithm in performing the optimal routing differentiated from the existing conventional algorithms for different node variations is defined in Figure 5. The major goal of the designed FAABO is to achieve scalable and reliable routing in the WSN network. It helps in the secure data transmission based on the trusted path selection. Here the number of iterations varied to perform the comparison, which ranges between 0 to 100. From the graphical determination, the initiated FAABO algorithm attained better convergence than 70% of GOA, 35% of YSGA, 32.5% of CWO, and 30% of ABOA. The obtained result defines the better convergence attained by the FAABO algorithm.

5.png

Figure 5. Convergence estimation of the designed Fitness Augmented African Bison Optimization (FAABO) algorithm differentiated from other conventional algorithms in terms of “(a) Node 50, (b) Node 100, (c) Node 150, and (d) Node 200”

6.4 Statistical calculation of the Fitness Augmented African Bison Optimization algorithm

The statistical estimation of the initiated FAABO algorithm differentiated from the other optimization approaches based on different statistical measures is explained in Table 3. Here number of nodes varied to achieve the result to evaluate the effectiveness of the algorithm. The functionality of the FAABO is accessed by the statistical analysis in terms of the optimal path selection in the WSN network. The recommended FAABO attained a better value than 27.7% of GOA, 3.2% of YSGA, 8.7% of CWO, and 6.7% of ABOA. The outcome from the result defines the better functionality achieved by the FAABO than other existing optimization models.

Table 3. Statistical determination of suggested Fitness Augmented African Bison Optimization (FAABO) algorithm compared with other traditional algorithms

Note: FAABO: Fitness Augmented African Bison Optimization; ABOA: African Bison Optimization Algorithm; GOA: Grasshopper Optimization Algorithm; YSGA: Yellow Saddle Goatfish Algorithm; CWO: Carpet Weaver Optimization; SD: Standard Deviation.

6.5 Performance comparison of HSCRNet-based trust prediction

Performance assessment of the presented HSCRNet model for the trusted node prediction compared with other similar approaches and it expressed in Figure 6. Here, the determination is performed by varying the nodes on different performance measures. This analysis helps in evaluating the performance of HSCRNet in classifying the nodes in the WSN based on the trusted and malicious nodes. From the graphical calculation, the presented HSCRNet attained better accuracy than 4% of Deep Neural Network (DNN), 3% of Long Short-Term Memory (LSTM), 2% of TCN, and 1% of CapsNet. Due to this result, the presented model attained a better result than other approaches.

6_1.png

6_2.png

6.6 Overall comparison of HSCRNet-based trust prediction

The overall comparison of the designed HSCRNet-based trusted node prediction differentiated from the other baseline model expressed in Table 4. Here the number of nodes varied to perform the analysis, along with the different optimizers to evaluate the given models. From the comparison the initiated HSCRNet attained the minimum MEP of 78% of DNN, 59% of LSTM, 35% of TCN, and 12% of CapsNet for the number of nodes 50 while determining with the adam optimizer. Based on the overall comparison the HSCRNet has a better classification result than the other existing approaches.

Table 4. Overall comparison of the suggested HDCRNet for trust prediction differentiated from other techniques

Note: DNN: Deep Neural Network; LSTM: Long Short-Term Memory; TCN: Temporal Convolutional Network; CapsNet: Capsule Network; HSCRNet: Hybrid Serial Cascaded Residual Network; MEP: Mean Error Percentage; SMAPE: Symmetric Mean Absolute Percentage Error; MASE: Mean Absolute Scaled Error; MAE: Mean Absolute Error; RMSE: Root Mean Square Error.

Each experiment was repeated 20 times using different random seeds. The reported values correspond to mean ± standard deviation.

A paired t-test at a significance level of 0.05 confirmed that the performance improvements achieved by HSCRNet and FAABO were statistically significant compared with baseline methods.

To confirm the effectiveness and efficiency of the HSCRNet-FAABO framework in providing trustworthy and reliable routing solutions for WSNs, all the presented experiments have been conducted 20 times with different random seeds under identical simulation environments. All performance measures have been calculated using mean ± standard deviation (SD). Furthermore, the statistical significance tests, based on a paired t-test at α = 0.05, have been used to compare the proposed scheme against the baseline algorithms. As can be seen, all improvements made using HSCRNet and FAABO are statistically significant (p < 0.05).

In particular, HSCRNet outperformed DNN (96.94 ± 0.56%, paired t-test = 〈<0.001), LSTM (97.32 ± 0.48%, p-value < 0.001), TCN (97.76 ± 0.41%, p-value = 0.002), and CapsNet (98.21 ± 0.34%, p-value = 0.014) in terms of trusted node classification accuracy. It is noteworthy that, on average, the difference in performance reached 2%, which means that HSCRNet demonstrates superior capabilities in detecting malicious nodes and avoiding security breaches caused by them.

At the same time, FAABO showed better performance compared to GOA (89.71 ± 1.18%, p-value < 0.05), YSGA (91.64 ± 1.04%, p-value <0.05), CWO (93.52 ± 0.88%, p-value < 0.05), and ABOA (95.84 ± 0.73%, p-value <0.05). The average improvements reached 2-5%, demonstrating the superior capabilities of FAABO in providing trustworthy and reliable routing solutions.

Finally, to verify the stability and robustness of the framework, the HSCRNet–FAABO solution was tested in three attack modes: blackhole, wormhole, and sybil attacks. In this experiment, PDR, TCA, and latency have been chosen as performance criteria. Thus, the results for Blackhole attack mode include 96.8 ± 0.42% PDR, 98.2 ± 0.31% trust accuracy, and 0.25 ± 0.01 s delay; for Wormhole attack mode – 95.7 ± 0.47% PDR, 97.8 ± 0.36% trust accuracy, and 0.28 ± 0.02 s delay; and for Sybil attack mode – 94.9 ± 0.53% PDR, 97.3 ± 0.41% trust accuracy, and 0.31 ± 0.02 s delay. A one-way ANOVA test was used to investigate the impact of the type of attacks on the results of the experiments. No statistically significant differences (P = 0.087, p-value > 0.05 for trust prediction accuracy and P = 0.094 > 0.05) across the considered attack scenarios.

6.7 Comparative analysis of the Fitness Augmented African Bison Optimization algorithm for the optimal routing mechanism

The performance estimation of the presented FAABO algorithm for the optimal route selection in the WSN network compared with the other existing algorithms regarding several nodes and density is defined in Figures 7 and 8. The main aim of this comparative analysis is to determine the routing efficiency based on the secured data transmission. Based on the graphical illustration, the implemented FAABO algorithm achieved the effective result of 80% of GOA, 75% of YSGA, 67% of CWO, and 25% of ABOA for taking the number of nodes as 50. This shows the effectiveness of the algorithm in performing the optimal route selection.

7_1.png

7_2.png

Figure 7. Comparative estimation of the implemented Fitness Augmented African Bison Optimization (FAABO) algorithm in functioning optimal routing compared with other optimization models in terms of “(a) Availability, (b) Delay, (c) Energy consumption, (d) PDR, (e) Scalabiltiy, (f) Throughput, and (g) Trust”

2.png

3.png

Figure 8. Comparative determination of the suggested FAABO algorithm in performing optimal routing differentiated with other models based on number of density in terms of “(a) Availability, (b) Delay, (c) Energy consumption, (d) PDR, (e) Scalability, (f) Throughput, and (g) Trust”

The proposed framework maintained stable routing performance under blackhole, wormhole, and Sybil attack conditions. HSCRNet successfully identified malicious behavior, while FAABO selected alternative trusted paths, preventing severe packet loss in Table 5.

Table 5. Performance evaluation of the proposed HSCRNet–FAABO framework under different routing attack scenarios

Note: PDR: Packet Delivery Ratio; Blackhole: Blackhole Attack; Wormhole: Wormhole Attack; Sybil: Sybil Attack; HSCRNet: Hybrid Serial Cascaded Residual Network; FAABO: Fitness-Augmented African Bison Optimization.

6.8 Computational complexity analysis

The complexity of HSCRNet is O(N×T×F), where N represents the number of nodes, T the temporal sequence length, and F the feature dimension. The routing complexity of FAABO is O(P×I×D), where P denotes population size, I the number of iterations, and D the routing dimension. Although the proposed model introduces additional computation compared with conventional routing protocols, trust prediction is executed periodically at the sink node, minimizing sensor-node overhead.