Sleep-wakeup based secure multipath routing in wsn using fennec fox optimized deep learning framework

Abstract

Wireless Sensor Networks (WSNs) consist of spatially distributed sensor nodes that monitor and transmit data to the base station. Wireless sensor networks are widely used in environmental monitoring, smart cities, healthcare, and disaster management applications. However, the major disadvantage of WSNs is their excessive energy consumption, which shortens their lifetime and limits their data integrity. Traditional methods face redundant retransmissions and collision problems which lead to congestion and fast energy consumption. To overcome these problems, a novel Secure Clustering and Sleep-Wakeup based Energy Efficient Routing using Fennec fox optimized deep learning (SCS-EEF) framework has been proposed to reduce energy usage and increase network lifetime in WSN. Fuzzy C Means based Balanced Iterative Reducing and Clustering Using Hierarchies (Fuzzy-BIRCH) clustering is employed in the proposed framework to improve cluster formation and reduce communication costs. To enhance energy efficiency, Fennec Fox Optimization (FFO) is used for optimal cluster head selection (CHS) with fair energy distribution across nodes and to reduce premature node failures. An energy-saving dynamic sleep-wakeup schedule is proposed to eliminate redundant transmissions. The hybrid Temporal Convolutional Network-based Bidirectional Gated Recurrent Unit (TCN-BiGRU) network predicts multipath routes through the classification of data into emergency and non-emergency. The proposed framework reduces energy consumption by 39.3%, 29.2%, and 26.1% over HBWCO, IBORSDFFNL, and EER-CGHHOA, respectively. Furthermore, the proposed model reduces latency by 10.7%, 9.42%, and 7.4% over HBWCO, IBORSDFFNL, and EER-CGHHOA, respectively.

Introduction

WSNs are self-organizing, self-healing, and self-configurable networks that comprise homogeneous or heterogeneous sensor nodes¹. These nodes sense environmental factors such as temperature, humidity, pressure, vibration, and lighting in various locations, including forests, near active volcanoes, and cold regions^2,3. However, network lifespan and latency are critical factors in WSN, which are handled by clustering, duty cycling, and routing. Zone-based, hexagonal, and grid-based designs are the best ways to arrange sensor nodes since random deployment could lead to excessive energy usage^4,5.

In the case of wireless sensor networks, clustering is one of the most effective solutions for energy conservation as it involves the collection of data from several nodes instead of just one. Majority of the traditional clustering techniques make use of bio-inspired optimization methods, including Mayfly Optimization (MFO)⁶, Genetic Algorithm (GA)⁷, and Particle Swarm Optimization (PSO)⁸. These methods determine the suitable cluster head (CH) based on different criteria like remaining energy, distance, and latency. Choosing the right CH improves the communication quality and trustworthiness which in turn prolongs the network’s lifetime. Another major aspect of WSN for energy saving is duty cycling, which manages the schedule of sensor nodes and assigns the slots to them. The sensor nodes carry out their sensing activities according to their schedule in order to prolong the network lifetime ⁹.

Nonetheless, routing is still an indispensable operation in Wireless Sensor Networks (WSNs) that helps the transmission of data from sensor nodes to the Base Stations (BS). To achieve the reliable data delivery and at the same time to cut down the energy consumption, energy-efficient routing protocols are very much necessary¹⁰. The routing procedure, in turn, allows the transfer of the monitored data from the cluster head (CH) to either the base station or the sink in a WSN. Out of numerous paths, the most suitable ones are selected so as to lessen the routing failures that are the result of exceedingly long delays. In a cluster-based routing protocol, data is sent through the shortest and multi-hop optimal routing path¹¹. Multi-hop routing makes it possible for nodes to keep an eye on and detect any malicious activities happening among their fellow nodes in the WSN.

In existing studies, deep learning techniques have been increasingly used to enhance routing in WSNs by predicting reliable paths, classifying traffic types, and adapting to dynamic network conditions. Models such as Deep Q-Networks (DQN)¹², Deep Feedforward Neural Networks (DFNN)¹³, and hybrid optimization–deep learning frameworks have demonstrated improvements in packet delivery, latency reduction, and energy-efficient data aggregation^14,15. However, existing approaches have several limitations, including static or non-adaptive cluster head selection, unbalanced energy consumption, redundant retransmissions, and high latency in data delivery. While many optimization-based methods converge slowly and rarely scale efficiently in large dynamic networks, deep learning-based models lack integration with clustering and duty-cycling mechanisms, causing increased energy consumption¹⁶. To overcome these issues, a novel Secure Clustering and Sleep-Wakeup based Energy Efficient Routing using Fennec fox optimized deep learning (SCS-EEF) framework is proposed to reduce energy usage and increase network lifetime in WSN. The objectives of the proposed work are as follows,

• The key goal of this work is to develop an integrated energy-efficient and secure multipath routing framework for WSNs that increases network lifetime and reduces energy consumption.

• The proposed method uses Fuzzy-BIRCH to generate balanced and non-redundant clusters, which effectively reduces latency in data delivery.

• A dynamic sleep–wakeup scheduling algorithm is employed to eliminate redundant transmissions, reduce congestion and collisions, balance traffic load, and extend overall network lifespan.

• The Fennec Fox Optimization (FFO) is utilized for adaptive cluster head selection, ensuring balanced energy consumption, reducing premature node failures, and extending network lifetime.

• The proposed technique incorporates a hybrid TCN-BiGRU network that predicts paths that are both reliable and energy-efficient for the transmission of emergency as well as non-emergency messages thereby increasing the life of the network.

• The efficiency of the proposed framework is assessed using several metrics, namely network lifetime, packet delivery ratio, latency, throughput, residual energy, and energy consumption.

The following describes the remaining section of the suggested framework: The literature survey is explained in Sect. "Literature survey". Section "Proposed methodology" explains the proposed SCS-EEF approach. In Sect. "Result and discussion", the results are analysed and discussed. In Sect. “Conclusion”, the future work and conclusion described.

Literature survey

In this section, a comprehensive review of recent works on energy-efficient clustering, node scheduling, and secure routing in WSN is presented.

In 2022 Guo, H., et al.,¹⁶ proposed an adaptive dual-mode energy-saving routing method based on deep Q-networks (DQN) to improve the sustainability of rechargeable WSN (RWSN). The simulation findings demonstrate that the flexible modification of the routing mode allows the approach to significantly enhance energy efficiency, which results in a noticeable extension of the network’s lifespan. With restricted data, the accuracy of routing mode selection achieves 95%.

In 2022 Narayan, V. and Daniel, A.K.,¹⁷ suggested a method for improving coverage and fixing gaps that effectively addresses the issues of overlapping and coverage issues in the network through different stages, including starting the network, forming clusters, choosing cluster heads, and managing sleep and active periods. The performance results from simulations of the method compared to LEACH, TEEN, SEP, and DEEC routing protocols demonstrate a longer stable period for the network and a better overall network lifespan.

In 2023, Biswas, K., et al.,¹⁸ developed an Energy Efficient Secure Multipath (EESM) routing system for efficiently creating routes and transferring data packets between SNs and base stations. In regards to energy efficiency, the suggested protocol beats Secure and Reliable Multipath Routing (SRMR), Secure and Energy Efficient Multipath (SEEM), and Reliable and Multipath Encounter Routing (RMER), extending network lifetime by 37% and improving throughput by 6%.

In 2023 Rani, S.S., and Sankar, K.S., et al.,¹⁹ proposed an Enhanced Buffalo Optimized Route Selective Deep Feed Forward Neural Learning (EBORSDFFNL) approach for energy-efficient data collection in WSN. Simulation outcomes indicate that the suggested model achieved reduced energy usage and minimized end-to-end latency for data collection, as well as improved packet delivery rates in WSNs.

In 2023, Fanian, F., and M.K. Rafsanjani,²⁰ presented a calibrated fuzzy metaheuristic clustering routing method (CFMCRS) for on-demand WRSN. In addition to resource saving, the proposed CFMCRS manages responsibility and energy allocation inside and between nodes. As a result, the proposed approach surpassed the other methods significantly in terms of application requirements and enhanced assessment aspects. Experimental results were assessed using ANOVA and post hoc analyses.

In 2023, Dong et al.,²¹ proposed a new flexible backhaul design that can adjust to varying traffic flows. The quantitative findings demonstrated that the combined power optimization and mixed backhaul structure can boost overall network throughput by 18% when compared to the existing optimized fixed designs. Additionally, the suggested approach can lower energy usage by 30% and improve user satisfaction quality by 24.5% in relation to user distribution trends.

In 2024 Jayachandran, J. and Devi, K.V.,²² suggested a combined method called Energy Efficient Routing using Cluster-based Genetic Harris Hawkeye Optimization Algorithm (EER-CGHHOA) for improving the efficiency of border monitoring WSNs, particularly in very spread-out networks. The findings showed that this method used 11. 36% less energy, successfully sent 97. 02% of packets to the main station, and lower network delays by 6. 25% compared to other methods looked at in this research.

In 2025 Senthil, G.A., et al.,²³ introduced a new Hybrid Beluga Whale-Coati Optimization (HBWCO) method to enhance energy-efficient data transfer. The HBWCO strategy has reached an exceptional reliability rate of 0.948 and a maximum throughput of 3496. As a result, the HBWCO method presented an effective option for ensuring dependable data transmission and routing.

In 2025 Akram, M., et al.,²⁴ suggested an Energy-Efficient Machine Learning-based Clustering and Routing (EEMLCR) method for Wireless Sensor Networks (WSNs). This method makes it possible to form clusters as well as to choose the routing paths. The results of the experiments showed that EEMLCR is more effective than LEACH and also its multi-hop variations, namely DMHT LEACH and EDMHT LEACH. EEMLCR after 600 rounds in networks with 400 nodes showed significant improvements in major performance metrics.

In 2025 Bourebia, N.E.H. and Bourebia, S.,²⁵ proposed a clustering-based sleep scheduling mechanism (RMIS) to reduce unnecessary transmissions and enhance the energy efficiency of the network. Besides that, RMIS employed the Chinese Whisper technique for clustering to facilitate the formation of closely-knit and distinct groups depending on the distance and the transmission range. The simulated outcomes are in line with the performance of the suggested protocol compared to other methods concerning the network lifetime, the period of stability, energy efficiency, and the number of active nodes.

Existing approaches still have significant drawbacks, including low packet delivery ratios, slow routing decisions, and congestion problems. These limitations shorten the lifespan and reliability of networks in dynamic, large-scale settings. The next section will discuss a novel SCS-EEF framework that has been proposed to address these issues by optimizing energy usage and extending network lifetime.

Proposed methodology

This section proposes a novel SCS-EEF framework to improve network lifetime and lower energy consumption in WSNs. Firstly, data is sampled from nodes observing the distributed sensor network in a vast area. Sensor nodes are clustered with the Fuzzy-BRICH method for best clustering while using a sleep-wakeup mechanism to save energy, which is one of the features of the sensor nodes. By employing the Fennec Fox Optimization algorithm in the CHS Phase for carrier selection through the assessment of factors such as distance, delay, energy, and trust, reliability and efficiency are elevated. In the Decision and Routing Phase, parameters of the network like node density, bandwidth, traffic patterns, and energy availability are given to a TCN-BiGRU network that distinguishes the data into emergency and non-emergency paths and also picks the top-ranked route having the lowest hop counts. At last, the Data Transmission Phase is there for the secure facilitation of packets which were selected paths. The design of the SCS-EEF technique is shown in Fig. 1.

Fig. 1

Architecture of the SCS-EEF Method.

Clustering via fuzzy-BIRCH algorithm

Initially, the sensor nodes are organized into clusters using a Fuzzy-BIRCH algorithm to manage the network congestion. In BIRCH, a cluster is defined by its Cluster Features (CFs), and the hierarchical structure of clusters is displayed using a CF tree. In BIRCH clustering, the cluster centroid {$\overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\rightharpoonup}$}} {X} _{1}$} is obtained using Eq. (1).

$$\overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\rightharpoonup}$}} {X} _{0} = \frac{{\sum _{{\text{i}}}^{{\text{N}}} \overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\rightharpoonup}$}} {X} _{i} }}{{\text{N}}}$$

(1)

By choosing the number of clusters, the processed data is separated into discrete subgroups according to specific CFs. Cluster tags are then applied to these subsets to cluster them in an energy-constrained manner. The Fuzzy C-Means (FCM) divides n vectors into k groups and initializes the affiliation matrix (S) by calculating the clustering center of each group through fuzzy partitioning to minimize the objective function. The class center vector and the affiliation matrix are represented in Eq. (2)-(3).

$${\text{C}}_{\text{j}}=\frac{{\sum }_{\text{i}=1}^{\text{N}}{\text{s}}_{\text{ij}}^{\text{m}}.{\text{x}}_{\text{i}}}{{\sum }_{\text{i}=1}^{\text{N}}{\text{s}}_{\text{ij}}^{\text{m}}}$$

(2)

$${\text{s}}_{\text{ij}}=\frac{1}{{\sum }_{\text{k}=1}^{\text{c}}{\left(\frac{\Vert {\text{x}}_{\text{i}}-{\text{c}}_{\text{j}}\Vert }{{\text{x}}_{\text{i}}-{\text{c}}_{\text{k}}}\right)}^{\frac{2}{\text{m}-1}}}$$

(3)

where m denotes the number of clusters for clustering, $i,$ and $j$ denote the affiliation of matrix $i$ concerning class cluster $j$. The objective function ${\text{J}}_{\text{m}}$ is computed in Eq. (4) as:

$${\text{J}}_{\text{m}}=\sum_{\text{i}=1}^{\text{N}}\sum_{\text{j}=1}^{\text{c}}{\text{s}}_{\text{ij}}^{\text{m}}{\Vert {\text{x}}_{\text{i}}-{\text{c}}_{\text{j}}\Vert }^{2}$$

(4)

where, ${\text{c}}_{\text{j}}$ denotes the center of class cluster $j$, ${\Vert {\text{x}}_{\text{i}}-{\text{c}}_{\text{j}}\Vert }^{2}$ is the Euclidean distance among the $j$ ^th selected node and the $i$ ^th node. The Fuzzy Partition Coefficient (FPC) measures the degree of fuzzy overlap between clusters. It ranges from 0 to 1, indicating energy-efficient clusters with higher values. Therefore, the Fuzzy-BIRCH clustering algorithm enhances the clustering efficiency by grouping the optimal nodes for minimizing latency in WSNs.

Sleeping and waking node selection

Balanced clusters are formed using the Fuzzy-BIRCH algorithm, but redundant transmissions is challenging due to overlapping sensing regions among nodes. To overcome this issue, a sleep–wakeup algorithm is used to conserve energy and extend network lifetime. Overlapping data is gathered from neighboring and overlapping nodes as a result of their spatiotemporal data correlation properties. To maximise network lifespan and decrease energy usage, the overall traffic should be reduced by eliminating redundant data. During sleep, the wakeup approach reduces the retransmission of the same message, which increases energy efficiency. Therefore, the sensor nodes remain in sleep mode to conserve network energy. Once the wake-up nodes have been activated, they cover the entire network region. When a wake-up node runs out of energy, the sensor node goes to sleep.

A network is shown as a collection of wireless nodes with the coordinates $n= \{{n}_{1}; {n}_{2}; {n}_{3}\}$, which are all able to interact with the BS directly and are located within each other’s sensing ranges. Therefore, along the sensing range ${S}_{r}$, each node performs accurate sensing. Moreover, there is no need to spread similar data to the BS for analysis more than once. Each node n_i also conducts the periodic tasks ${P}_{x}=\{{SE}_{x}, {C}_{x},{O}_{x}\}$, where ${SE}_{x}$ is the sensing task (data collecting), ${C}_{x}$ is the computation task, and ${O}_{x}$ is the communication task. The subset of nodes in each cluster that will complete the periodic tasks ${P}_{x}$ for the current round are selected after the clusters have been formed, with the remaining Sleeping Node (SN). The selected Waking Node (WN) provides complete monitoring over the sensing range during this round. Each node chooses a time slot (${t}_{x}$) for SN and WN, that is proportional to its current weights, which are denoted by Eqs. (5), (6), and (7), respectively.

$${W}_{1}\left({t}_{x}\right)=\frac{1}{{En}_{i}(T)}$$

(5)

$${W}_{2}\left({t}_{x}\right)=\frac{{S}_{r}}{En({t}_{x})}$$

(6)

$${W}_{3}\left({t}_{x}\right)=\frac{D({t}_{x}-BS)}{En({t}_{x})}$$

(7)

This indicates that nodes with lower weights are more likely to activate. Then, each node waits for the time slot before finding whether or not to remain awake for the current round. If a node chooses to become a guardian node while waiting for the time period (${t}_{x}$) to finish during the current round, the neighbouring nodes send a WAKING signal. Nodes with lower weights are more likely to wake up, while others remain asleep, minimizing redundant transmissions. To quantitatively assess the effect of the sleep–wakeup scheduling system, a mathematical model of total energy consumption was used for each node. Each sensor node expends energy for data transmission, reception, and aggregation during communication rounds. The total energy consumed by a node ${n}_{i}$ is computed by using Eq. (8),

$${E}_{total}\left({n}_{i}\right)={E}_{tx}\left({n}_{i}\right)+{E}_{rx}\left({n}_{i}\right)+{E}_{agg}\left({n}_{i}\right)$$

(8)

where ${E}_{tx}$ and ${E}_{rx}$ are the energy consumed for transmission and reception, respectively. The energy required for data transmission and reception are defined as:

$${E}_{tx}\left(l,d\right)=l{E}_{elec}+l{\varepsilon }_{amp}{d}^{\gamma }, {E}_{rx}\left(l\right)= l{E}_{elec}$$

(9)

where $l$ indicates data packet size (bits), $d$ denotes transmission distance, ${E}_{elec}$ represents the energy per bit consumed by transmitter, and ${\varepsilon }_{amp}$ indicates the transmit amplifier constant.

$${E}_{res}\left({n}_{i},r\right)={E}_{init}\left({n}_{i}\right)-({E}_{tx}+{E}_{rx})$$

(10)

where ${E}_{init}\left({n}_{i}\right)$ is the initial energy of node ${n}_{i}$ and ${E}_{res}\left({n}_{i},r\right)$ is its residual energy after $r$ rounds. This energy model helps the best CH selection in the next stage and offers the quantitative foundation for assessing node-level energy consumption.

Cluster head selection using fennec fox algorithm (FFO)

The sleep–wakeup scheduling method helps to limit redundant transmissions and thus, save energy; nevertheless, the proper functioning of the network depends to a large extent on the right choice of CHs. For that purpose, the FFO is employed to select the best CHs by examining parameters like distance, delay, energy, and trust, thus, making the system more reliable and efficient. The Fennec Fox Optimization (FFO) algorithm was selected for Cluster Head (CH) selection because of its efficient balance between exploration and exploitation through its dual behavioural phases, namely digging (local search) and escaping (global diversification). Unlike Particle Swarm Optimization (PSO) or Grey Wolf Optimization (GWO), which may stagnate early, FFO adaptively tunes step size based on prey–predator energy modelling, ensuring fast and stable convergence. Computationally, FFO has linear update equations requiring only arithmetic operations, leading to O(P·m·I) time complexity (P = population, m = variables, I = iterations), lower than swarm-based models that require pairwise position updates (O(P²)). The flowchart of the FFO-based CH selection process is demonstrated in Fig. 2.

Fig. 2

Flowchart of FFO.

Initialization

According to the proposed scheme, the FFO algorithm picks each fennec fox as a candidate CH solution. The location of a fox in the search space denotes the decision variables, i.e. residual energy, intra-cluster distance, delay, trust value, and communication cost. A population matrix is constructed where each fox represents a possible CH configuration for the wireless sensor network. At the beginning, these foxes are placed randomly in the search space, thus providing diverse candidate solutions for CHS. The description of each fox is considered as a solution vector and the whole population moves through the search space together to find the best set of cluster heads by utilizing Eq. (11).

$${X}_{i}:{x}_{i,j}={lb}_{j}+r\times \left({ub}_{j}-{lb}_{j}\right),i=\text{1,2},\cdots ,N,j=\text{1,2},\cdots ,m$$

(11)

where ${X}_{i}$ represents the $i$ th fennec fox, ${x}_{i,j}$ signifies its jth dimension (which refers to the decision variable), N indicates the overall count of fennec foxes, m denotes the total count of decision variables, r indicates random value within the range of [0, 1], ${lb}_{j}$ represents the minimum limit, and ${ub}_{j}$ refers to the maximum limit of the jth decision variable.

Phase 1: Digging prey behavior in the sand (local search)

The fennec fox will be the initial member of the M group clusters. The fennec fox relies on its exceptional hearing to locate the prey that It initially digs a hole, then buries itself in the sand before eventually capturing its prey. By replicating this approach, FFO may improve its ability to Using local searches, you can get closer to the best answer overall. The location of the ${x}^{th}$ fox is denoted by ${A}_{x}=\left({a}_{x1},{a}_{x2},\dots ,{a}_{xn}\right)$, and the fitness is the value in this area is ${T}_{x}$. It is believed that there are M fennec foxes. A local radius S and around the center ${A}_{x}$ The neighborhood is searched. The first phase location is updated using Eqs. (12) through (13):

$${A}_{x,y}^{H1}={A}_{x,y}+(2.k-1){S}_{x,y}$$

(12)

$${S}_{x,y}=\alpha \left(1-\frac{i}{I}\right){A}_{x,y}$$

(13)

$${A}_{x}=\left\{\begin{array}{c}{A}_{x}^{H1}, {T}_{x}^{H1}<{T}_{x} \\ {A}_{x}, else\end{array}\right.$$

(14)

where ${A}_{x}^{H1}$ is the location of the x-th fennec fox in the first phase update, ${A}_{x,y}^{H1}$ is the position of its y-th dimension, and ${T}_{x}^{H1}$ is its fitness value. An is a constant set to 0.2, k denotes the interval [0,1], and ${S}_{x,y}$ is its neighborhood radius.

Phase 2: Escaping predators (global search)

The fennec fox may avoid predators that are chasing it thanks to its great speed and quick shifts in course. Global search is based on the The fennec fox’s method of escape. This strategy improves FF’s exploration by allowing it to avoid the local optimum area and find the global ideal area. capabilities. The second phase position update model is shown here:

$${A}_{x}^{rand}:{A}_{x,y}^{rand}={A}_{q,y},q\in \left\{\text{1,2},3,\dots , N\right\}, x=\text{1,2},\dots , N$$

(15)

$${A}_{x,y}^{H2}=\left\{\begin{array}{c}{A}_{x,y}+s\left({A}_{x,y}^{rand}-J{A}_{x,y}\right), {T}_{x}^{rand}<{T}_{x}\\ {A}_{x,y}+s\left({A}_{x,y}-{A}_{x,y}^{rand}\right), else\end{array}\right.$$

(16)

$${A}_{x}=\left\{\begin{array}{c}{A}_{x}^{H2}, {A}_{x}^{H2}<{T}_{x} \\ {A}_{x}, else\end{array}\right.$$

(17)

where ${A}_{x,y}^{rand}$ indicates the x-th fennec fox’s position in the y-th dimension, ${T}_{x}^{rand}$ is the fitness value at that location, and ${A}_{x}^{rand}$ is the goal point of the x-th fennec fox escaping. According to the second phase update, the x-th fennec fox’s position is ${A}_{x}^{H2}$, its position in the y-th dimension is ${A}_{x,y}^{H2}$, and its fitness value is ${A}_{x}^{H2}$ . J depicts a random number chosen from the set {1, 2}, and s is a random number between [0, 1]. A node is selected as a CH if it maximizes residual energy and minimizes intra-cluster distance:

$$CH=arg\underset{{X}_{i}}{\text{max}}\left({w}_{1}\frac{{E}_{res}\left({n}_{i},r\right)}{{E}_{max}}-{w}_{2}\frac{{d}_{avg}}{{d}_{max}}-{w}_{3}{T}_{i}-{w}_{4}{D}_{i}\right)$$

(18)

where, ${E}_{res}\left({n}_{i},r\right)$ indicates the residual energy of the node $i$, ${d}_{avg}$ indicates the average distance between the CH and its cluster members, ${T}_{i}$ denotes the trust value of the node $i$, ${D}_{i}$ depicts the delay associated with the node $i$ and ${w}_{1}, {w}_{2}, {w}_{3}, {w}_{4}$ are the weighting factors balancing energy, distance, trust, and delay.

The fitness function $F({X}_{i})$ evaluates the suitability of each candidate CH node, combining multiple factors:

$$F\left({X}_{i}\right)=\alpha \frac{{E}_{res}\left({n}_{i},r\right)}{{E}_{max}}+\beta \frac{1}{{d}_{avg}}+\gamma {T}_{i}+\delta \frac{1}{{D}_{i}}$$

(19)

where, $\alpha ,\beta ,\gamma ,\delta$ are normalized weights $(\alpha +\beta +\gamma +\delta = 1)$, Higher fitness indicates more suitable CH nodes. The FFO algorithm updates candidate CH positions using Eqs. (9)-(14), and selects the node with the highest fitness in each iteration. The FFO algorithm effectively identifies the most suitable cluster head nodes by considering factors such as residual energy, distance, delay, trust value, and communication cost. As a result, energy consumption is reduced, packet delivery ratio is improved, and overall network lifetime is increased.

Secure multipath routing

After identifying the best cluster heads with the FFO algorithm, the suggested approach predicts communication routes that are safe and stable by employing a hybrid TCN-BiGRU network for data classification into emergency or non-emergency classes, and predicts the most energy-efficient multipath routes. The TCN-BiGRU model is trained to predict emergencies in a WSN by detecting both the temporal dependencies and the sequential patterns in the node-level data. The input to the model is time-series data for each node of length 10, where the features are residual energy (E), Node Degree, Hop Count, Link Reliability, Trust, Traffic Rate, Distance to Cluster Head (DistanceToCH), and Queue Length. Each time series is tagged as either an Emergency or Non-Emergency event. The model is being trained offline at the Base Station with the Adam optimizer of learning rate 0.001, for 50 epochs, batch size of 64, and a dropout rate of 0.3 for the regularization. The cross-entropy loss function is used for the classification task, and the early stopping is employed to prevent the training from going on unnecessarily. At the time of deployment, the trained model will do the inference at Cluster Heads or edge devices, thus enabling the event detection and response in real-time. In the Decision and Routing Phase, network parameters are fed to a TCN-BiGRU network, which classifies data into emergency and non-emergency paths and selects the best-rank route with minimal hop counts. Figure 3 demonstrates the Architecture of TCN-BiGRU.

Fig. 3

Architecture of TCN-BiGRU.

Routing-related information from CHs and member nodes, including residual energy, hop count, distance to sink, link dependability, and trust value, is fed into the TCN. TCN maintains temporal causality by using causal convolution to make sure that routing decisions at time t are solely based on current and previous states. The standard 1-D causal convolutional layer for a 1-D input $h\in {\mathbb{R}}^{S}$ and a filter $c:\left\{0,\dots ,k-1\right\}\to {\mathbb{R}}$ is defined as follows in Eq. (20) and (21):

$$C\left({h}_{s}\right)=\left(h*c\right)\left(s\right)=\sum_{x=0}^{k-1}{c}_{x}{h}_{s-x}$$

(20)

$$\widetilde{os}=\left(C\left({h}_{1}\right),C\left({h}_{2}\right),\dots ,C\left({h}_{S}\right)\right)$$

(21)

where $C\left({h}_{s}\right)$ denotes the convolutional operation, k is the convolutional kernel size, and $\widetilde{os}$ is the resulting output sequence. To effectively capture the temporal relationships of routing parameters such as residual energy, hop count, trust level, and distance to sink, a Temporal Convolutional Network (TCN) is used in the proposed routing framework. The TCN captures these dependencies by combining dilated convolution (DC) with causal convolution. In detail, the DC introduces a dilation factor that skips certain input elements, thus enabling the model to learn long-range dependencies in routing behavior without increasing the filter size. This allows the network to evaluate both short-term and long-term path stability, which is an indispensable feature for selecting reliable and energy-efficient multipath routes. In particular, the DC operation combined with causal convolution in the ith layer can be represented as (22).

$$C\left({h}_{s}\right)=\left(h*{g}_{i}c\right)\left(s\right)=\sum_{x=0}^{k-1}{c}_{x}{h}_{s-{g}_{i}x}$$

(22)

$$\widetilde{os}=\left(C\left({h}_{1}\right),C\left({h}_{2}\right),\dots ,C\left({h}_{S}\right)\right)$$

(23)

where ${g}_{i}$ which can be adjusted to ${2}^{i-1}$ is the ${i}^{th}$ layer’s dilation factor. The past direction is indicated by $s-{g}_{i}x$. A TCN layer is represented by Eq. (18), and TCN is created by stacking several TCN layers.

The output feature map generated by the TCN layer is provided as input to the BiGRU layer. While TCN captures the routing features, including such as residual energy, hop count, distance to sink, and trust values, the BiGRU processes these sequences bidirectionally to capture both past and future routing dynamics. This helps in accurately predicting stable and secure multipath routes. The BiGRU uses Gated Recurrent Units (GRUs), which are a simplified variant of LSTMs, to overcome issues such as gradient vanishing in traditional RNNs. GRUs efficiently retain long-term dependencies while reducing computational complexity, making them well-suited for dynamic WSN environments. The forward and backward GRU layers work in parallel to model routing sequences from both directions, and their combined output represents a more comprehensive temporal encoding.

Algorithm 1

SCS-EEF for Secure Multipath Routing in WSN.

Algorithm 1 demonstrate the proposed SCS-EEF framework for Secure Multipath Routing in WSN. Formulas (24), (25), (26), and (27) are applied to compute the hidden state ${h}_{s}$, which ultimately contributes to the final path classification.

$${i}_{s}={\varphi }({G}_{\text{q}}{y}_{s}+{V}_{\text{q}}{h}_{\text{s}-1}+{n}_{\text{q}})$$

(24)

$${\text{a}}_{s}={\varphi }({G}_{\text{a}}{y}_{s}+{V}_{\text{a}}{h}_{\text{s}-1}+{n}_{\text{a}})$$

(25)

$${\widetilde{h}}_{s}=\text{tanh}({G}_{\text{f}}{\text{y}}_{s}+{V}_{h}\left({h}_{\text{s}-1}\otimes {i}_{s}\right)+{n}_{s})$$

(26)

$${h}_{s}=\left(1-{\text{a}}_{s}\right)\otimes {h}_{\text{s}-1}+{\text{a}}_{s}\otimes {\widetilde{h}}_{s}$$

(27)

Here, ${y}_{s}$ denotes the input vector, and ${h}_{s}$ provides the output vector of the GRU. At time $s$, the input vector ${\text{y}}_{s}$ and the hidden state ${h}_{\text{s}-1}$ are fed as inputs to the GRU network, which generates the output ${h}_{s}$. The Sigmoid function is represented by the symbol $\varphi$ to store the information GRU neural networks. The reset and update gates are ${i}_{s}$ and ${\text{a}}_{s}$, respectively and the elementwise production is ⊗ . Furthermore, the candidate’s assumed state at the time $s$ is represented by ${\widetilde{h}}_{s}$. The forward and backward hidden layers make up the BiGRU structure. Two symmetric hidden layer state vectors are created by giving each data pattern to both the forward and backward directions of the GRU network. The fused representation of the input can be obtained by fusing bidirectional vectors, as shown below Eq. (28):

$${H}_{s}=[\overrightarrow{{H}_{s}}\oplus \overleftarrow{{H}_{s}}]$$

(28)

This combined representation is passed to a fully connected layer followed by the SoftMax classifier. The classifier predicts whether a given data packet should follow a high-rank path (low hop count, suitable for emergency data) or a medium-rank path (normal hop count, suitable for non-emergency data). The TCN-BiGRU network predicts reliable and energy-efficient paths for both emergency and non-emergency messages while enhancing network lifetime.

The overall workflow of the proposed method is demonstrated in Fig. 4.

Fig. 4

Proposed method flowchart.

Result and discussion

In this section, the effectiveness of the proposed method is assessed and compared with existing systems such as HBWCO ²³, IBORSDFFNL ¹⁹, and EER-CGHHOA ²².

Implementation setup

The proposed SCS-EEF method was evaluated using the well-known NS-3.26 simulator. Ubuntu 20.04.1 LTS OS, 8 GB of RAM, and an i9-9820X 3.30 GHz CPU were the parameters of the PC utilized for this experiment. Table 1 demonstrates the Simulation parameters of the proposed method.

Table 1 Simulation parameters.

Dataset description

A dataset of 10,000 samples was generated using NS-3 simulations for node counts of 200, 400, 600, 800, and 1000. Each sample consists of a time-series sequence of length 10, containing the features residual energy (E), node degree, hop count, link reliability, trust, traffic rate, distance to the cluster head (DistanceToCH), and queue length. The labels indicate whether an event is an emergency or a non-emergency. The dataset was split into training, validation, and testing sets with a ratio of 70:15:15. Model training was performed offline at the Base Station using the Adam optimizer with a learning rate of $1\times {10}^{-3}$, over 50 epochs, with a batch size of 64, dropout of 0.3, and cross-entropy loss. Early stopping was applied to prevent overfitting. After training, the model performs inference at the Cluster Head or edge devices, enabling real-time detection of emergency events without centralized computation.

An energy-efficient routing is performed by using TCN-BiGRU, which intelligently identifies the most reliable paths for energy-efficient data transmission and communication in a WSN environment. The simulation output shown in Fig. 5 demonstrates that the proposed framework provides robust communication in a WSN environment. Table 2 demonstrates the Computational complexity analysis of each stage of the framework.

Fig. 5

Simulation Output of the proposed framework.

Table 2 Computational complexity analysis.

Figure 6 shows the SCS-EEF framework’s adaptation in a wireless sensor network over 1000 rounds. In Phase 1, around 40 nodes are active with minimal traffic. Phase 2 increases this to about 50 nodes as density and traffic rise. Phase 3 boosts active nodes to 120 to handle emergency traffic and ensure reliability. Finally, Phase 4 reduces active nodes to around 55 as traffic decreases. Overall, the SCS-EEF framework balances energy efficiency and network reliability, achieving significant energy savings during peak periods.

Fig. 6

Analysis of sleep–wake scheduling dynamically adapts to traffic load.

Figure 7 illustrates the sleep–wakeup algorithm selects low-weight (high-energy) nodes as Wake Nodes (WNs) while others sleep to conserve energy. Multiple Cluster Heads (CHs) are elected among active nodes using FFO. As shown in the graph, green nodes are active WNs, red nodes are sleeping, and blue stars are CHs. Connectivity links indicate that active WNs always form a connected subgraph, ensuring reliable multi-hop routing to CHs/BS. Coverage analysis confirms all regions remain monitored, with sleeping nodes reactivated adaptively when local coverage drops. This demonstrates that the sleep–wakeup strategy preserves network connectivity and sensing coverage while balancing energy efficiency.

Fig. 7

Analysis of the sleep-wakeup algorithm.

Performance metrics

The parameters, such as energy consumption, delay, network lifetime, and throughput, are used to assess the efficiency of the SCS-EEF approach, which is explained as follows.

Network lifetime

The lifespan of sensor nodes is the main factors to be considered when measuring the efficiency of the WSNs. It is the time from the very first deployment of the network through which the nodes are capable of carrying the required sensing and communication tasks.

Energy consumption

Energy Efficiency is measured as the summation of total energy used by each sensor node during clustering, sleep–wake scheduling, and secure multipath routing for transmitting a specific number of information packets to the whole network. The energy consumption is computed utilizing Eq. (29).

$$\text{Energy Efficiency}=\sum_{\text{i}\in \text{N}}\text{TEC}({\text{n}}_{\text{i}})$$

(29)

where, TEC stands for the overall energy usage of the nodes.

Average delivery delay

The average delivery delay ($\overline{\text{D} }$) is computed by measuring the delivery delay of each packet transmitted by all nodes and then averaging these values across the entire network as shown in Eq. (30):

$$\overline{\text{D} }=\frac{\sum_{\text{i}=1}^{\text{N}}\text{ D}\left({\text{ n}}_{\text{i}}\right)}{\text{N}}$$

(30)

Throughput

Throughput (${T}_{p}$) is the number of bits sent to the BS over the WSN. Bits per second are used as the unit of measurement for throughput.

$${T}_{p}=\frac{Received packet size}{stop time-start time}$$

(31)

$${\varvec{A}}{\varvec{c}}{\varvec{c}}{\varvec{u}}{\varvec{r}}{\varvec{a}}{\varvec{c}}{\varvec{y}}\boldsymbol{ }(AC)=\frac{TrP+TrN}{FaN+TrP+FaP+TrN}$$

(32)

Precision (PR): It is stated as the proportion of precisely anticipated positive observations to total positive observations.

$$PR=\frac{TrP}{TrP+FaN}$$

(33)

$${\varvec{Recall}} \, \left( {{\varvec{RC}}} \right) = \frac{TrP}{{TrP + FaN}}$$

(34)

$$F1 score (F1S): =2\times \frac{PR.RC}{PR+RC}$$

(35)

where $TrP$ indicates True positive, $FaN$ indicates false negative, $FaP$ indicates false positive and $TrN$ indicates true negative respectively.

Comparison analysis

The comparative analysis is performed between the suggested technique and the existing methods such as HBWCO ²¹, IBORSDFFNL ²², and EER-CGHHOA ²³. The efficiency of the proposed framework is validated using several metrics, namely network lifetime, PDR, latency, throughput, residual energy, and energy consumption.

Figure 8illustrates the network lifetime (NL) comparison of the SCS-EEF model with HBWCO, IBORSDFFNL, and EER-CGHHOA. Figure 8(a) denotes the NL for the number of nodes from 0 to 250. Figure 8(b) illustrates the NL for the number of nodes from 250 to 500. Figure 8(c) displays the NL when the number of nodes is 500–750. Figure 8(d) illustrates the NL when the number of nodes is 750–1000. The findings show that the SCS-EEF model surpasses existing frameworks in terms of network lifetime. Existing approaches lack efficient clustering, scheduling, and routing integration, resulting in higher energy consumption and reduced network longevity. Overall, the suggested approach achieves a higher network lifetime than HBWCO, IBORSDFFNL, and EER-CGHHOA.

Fig. 8

Network lifetime comparison (a) 0–250 nodes, (b) 250–500 nodes, (c) 500–750 nodes, and (d) 750–1000 nodes.

Figure 9 illustrates the Packet Delivery Ratio (PDR) comparison of the proposed model with HBWCO, IBORSDFFNL, and EER-CGHHOA for different node densities. Figure 9(a) denotes the PDR when the number of nodes is 0–250. Figure 9(b) illustrates the PDR when the number of nodes is 250–500. Figure 9(c) displays the PDR when the number of nodes is 500–750. Figure 9(d) illustrates the PDR when the number of nodes is 750–1000. For 1000 nodes, the proposed model maintains a PDR of 97%, while HBWCO, IBORSDFFNL, and EER-CGHHOA achieve only 91%, 88%, and 85%, respectively.

Fig. 9

Packet Delivery Ratio (PDR) comparison (a) 0–250 nodes, (b) 250–500 nodes, (c) 500–750 nodes, and (d) 750–1000 nodes.

Figure 10 demonstrates the comparison of energy consumption for the SCS-EEF model and the existing HBWCO, IBORSDFFNL, and EER-CGHHOA techniques under different node densities. Figure 10(a) depicts the energy consumption for the number of nodes ranging from 0 to 250. Figure 10(b) depicts the energy consumption for the number of nodes from 250 to 500. Figure 10(c) displays the energy consumption when the number of nodes is 500–750. Figure 10(d) depicts the energy consumption as the number of nodes is 750–1000. The SCS-EEF model achieves lower energy consumption, which demonstrates longer network lifetime. At 1000 nodes, the proposed approach consumes 85 J, whereas the existing HBWCO, IBORSDFFNL, and EER-CGHHOA consume 140 J, 120 J, and 115 J, respectively. This shows clearly that using the proposed technique leads to a great decrease in energy usage, hence lengthening the network lifetime and improving energy efficiency as contrasted to the base techniques. Based on the outcomes, the proposed framework lowers energy consumption by 39.3%, 29.2%, and 26.1% over HBWCO, IBORSDFFNL, and EER-CGHHOA, respectively.

Fig. 10

Energy Consumption comparison (a) 0–250 nodes, (b) 250–500 nodes, (c) 500–750 nodes, and (d) 750–1000 nodes.

Figure 8 illustrates the comparison of alive nodes between the SCS-EEF model and the existing HBWCO, IBORSDFFNL, and EER-CGHHOA techniques under varying rounds. Figure 11(a) denotes the alive nodes as the number of nodes from 0 to 250. Figure 11(b) illustrates the alive nodes for the number of nodes from 250 to 500. Figure 11(c) displays the alive nodes for the number of nodes from 500 to 750. Figure 11(d) illustrates the alive nodes when the number of nodes is 750–1000. The SCS-EEF model achieves a higher number of alive nodes throughout the simulation, whereas the existing methods experience a faster decline. The alive nodes for 500 rounds show that the SCS-EEF model sustains 430 nodes alive, while HBWCO, IBORSDFFNL, and EER-CGHHOA attains 350, 330, and 300 nodes, respectively.

Fig. 11

Alive Nodes comparison (a) 0–250 nodes, (b) 250–500 nodes, (c) 500–750 nodes, and (d) 750–1000 nodes.

Figure 12 compares the latency of the proposed model with HBWCO, IBORSDFFNL, and EER-CGHHOA under different node densities. Specifically, Fig. 12(a) shows latency for 0–250 nodes, Fig. 12(b) for 250–500 nodes, Fig. 12(c) for 500–750 nodes, and Fig. 12(d) for 750–1000 nodes. As expected, latency increases with the number of nodes due to higher routing complexity and network traffic. Despite this, the proposed model consistently achieves the lowest latency across all scenarios. For instance, at 500 nodes, the proposed model records an average latency of 0.75 s, while HBWCO, IBORSDFFNL, and EER-CGHHOA reach 0.90 s, 0.88 s, and 0.85 s, respectively. At 1000 nodes, the proposed model maintains a latency of 1.25 s, compared to 1.40 s, 1.38 s, and 1.35 s for HBWCO, IBORSDFFNL, and EER-CGHHOA. Overall, the proposed model reduces latency by approximately 10.7%, 9.42%, and 7.4% relative to HBWCO, IBORSDFFNL, and EER-CGHHOA, demonstrating its efficiency under varying network loads.

Fig. 12

Latency comparison (a) 0–250 nodes, (b) 250–500 nodes, (c) 500–750 nodes, and (d) 750–1000 nodes.

Figure 13 illustrates the throughput comparison of the proposed framework with HBWCO, IBORSDFFNL, and EER-CGHHOA under different node densities ranging from 0 to 1000 nodes. Figure 13(a) denotes the throughput when the number of nodes is 0–250. Figure 13(b) illustrates the throughput when the number of nodes is 250–500. Figure 13(c) displays the throughput when the number of nodes is 500–750. Figure 13(d) illustrates the throughput when the number of nodes is 750–1000. On all four scenarios, the proposed model achieves higher throughput, which demonstrates the ability to sustain reliable data transmission even as the network size increases. In 250 nodes, the proposed method achieves a throughput of 2600 bps, whereas existing HBWCO, IBORSDFFNL, EER-CGHHOA achieves 2100 bps, 2050 bps, and 1900 bps.

Fig. 13

Throughput comparison (a) 0–250 nodes, (b) 250–500 nodes, (c) 500–750 nodes, and (d) 750–1000 nodes.

Figure 14 shows the residual energy comparison of the SCS-EEF method with existing HBWCO, IBORSDFFNL, and EER-CGHHOA approaches. Figure 14(a) denotes the residual energy when the number of rounds is 0–250. Figure 14(b) depicts the residual energy when the number of rounds is 250–500. Figure 14(c) indicates the residual energy when the number of rounds is 500–750. Figure 14(d) illustrates the residual energy for rounds from 750 to 1000. In all cases, the proposed scheme obtains higher residual energy than the existing approaches, demonstrating balanced energy consumption and reduced failure node exhaustion. At 500 rounds, the proposed model achieves 45% residual energy, whereas HBWCO, IBORSDFFNL, and EER-CGHHOA obtain 38%, 32%, and 28%, respectively.

Fig. 14

Residual Energy comparison (a) 0–250 rounds, (b) 250–500 rounds, (c) 500–750 rounds, and (d) 750–1000 rounds.

Figure 15 shows the comparative convergence performance of the proposed FFO and existing PSO, GWO, and ACO algorithms. The convergence analysis of the proposed FFO demonstrates that it converges significantly faster than conventional metaheuristic techniques. In WSNs, rapid convergence is essential for ensuring real-time routing adaptation and minimizing computational delay. As illustrated in Fig. 15, the proposed FFO achieves quicker and smoother convergence compared to Grey Wolf Optimization (GWO), Particle Swarm Optimization (PSO), and Ant Colony Optimization (ACO). The FFO rapidly stabilizes within fewer than 65 iterations, whereas PSO, GWO, and ACO require considerably more iterations to reach their optimal fitness values. The faster convergence of FFO can be attributed to its dual-phase search mechanism, which dynamically balances exploitation and global exploration. This adaptive balance prevents premature stagnation and ensures efficient discovery of optimal cluster-head configurations. These results confirm that FFO provides an optimal, stable, and computationally efficient solution for energy-aware cluster-head selection and secure routing in WSN environments.

Fig. 15

Convergence comparison of metaheuristic algorithms.

Table 3 demonstrates the performance comparison of Packet Delivery Ratio, Throughput, and Latency parameters between the proposed and existing methods for 200, 400, 600, 800 and 1000 nodes.

Table 3 Packet Delivery Ratio, Throughput, and Latency Performance Comparison.

Table 4 shows an ablation study of the SCS-EEF framework, evaluating different component combinations on energy consumption, packet delivery ratio (PDR), latency, and network lifetime. The full configuration—combining Sleep–Wake scheduling, Fuzzy-BIRCH clustering, Fennec Fox Optimization, and TCN-BiGRU routing achieves the best energy consumption of 0.62, the highest PDR of 97%, the lowest latency of 38.2 ms, and longest network lifetime of 1880 rounds. Removing any component leads to higher energy use, lower reliability, increased latency, and shorter lifetime, highlighting the importance of all components to the framework’s performance.

Table 4 Ablation study of the proposed method.

Figure 16 compares the computational complexity of the SCS-EEF framework with baseline models like PSO-CH, FFO-CH, BIRCH, Fuzzy-BIRCH, TCN, BiGRU, and BiLSTM. The SCS-EEF framework, combining Fuzzy-BIRCH clustering, Fennec Fox Optimization for cluster head selection, and a TCN-BiGRU network, shows the lowest complexity. Traditional methods like BIRCH and PSO-CH require more resources, while deep learning models like BiGRU and BiLSTM have moderate complexity. This highlights the SCS-EEF framework’s efficiency for energy-constrained wireless sensor networks, ensuring reliable and secure data transmission.

Fig. 16

Time Complexity Comparison of the proposed and baseline models.

Table 5 illustrates the convergence comparison of FFO with PSO, GWO, and ACO. It can be observed that FFO stabilizes within fewer than 65 iterations, while PSO, GWO, and ACO require 120, 105, and 130 iterations, respectively. Moreover, FFO achieves a higher fitness value of 0.91 with lower computational overhead (1.58 s). These results confirm that the proposed FFO provides an optimal, stable, and computationally efficient solution for energy-aware cluster head selection in WSNs.

Table 5 Comparative Performance of FFO with Existing Metaheuristic Algorithms.

Table 6 compares the performance of the proposed TCN-BiGRU, baseline DQN, and DFNN using metrics like accuracy, precision, recall, and F1-score. The TCN-BiGRU model achieve an accuracy of 0.98 and an F1-score of 0.945, indicating its effectiveness in classification. The DQN model shows an accuracy of 0.88 and an F1-score of 0.86. The DFNN model has the lowest performance, with an accuracy of 0.78 and an F1-score of 0.75. Overall, TCN-BiGRU stands out as the most robust option for reliable predictions. Table 7 illustrates the Control Overhead Comparison with Existing Protocols.

Table 6 Performance Comparison.

Table 7 Control Overhead Comparison with Existing Protocols.

Conclusion

This research introduced a novel SCS-EEF framework for energy-efficient and secure communication in WSNs. The use of Fuzzy-BIRCH clustering effectively mitigates the drawbacks of uneven cluster formation and poor scalability in traditional clustering techniques by generating balanced and non-redundant clusters. The FFO-based cluster head (CH) selection addresses the limitations of static or random CH assignment by ensuring optimal energy utilization, thereby reducing premature node failures and extending network lifetime. Experimental Findings confirm that the SCS-EEF model consistently surpasses HBWCO, IBORSDFFNL, and EER-CGHHOA in terms of throughput, energy consumption, network lifetime, and PDR. Based on the outcomes, the proposed framework lowers energy consumption by 39.3%, 29.2%, and 26.1% over HBWCO, IBORSDFFNL, and EER-CGHHOA, respectively. Moreover, the proposed model attains a latency of 1.25 s for 1000 nodes, whereas HBWCO, IBORSDFFNL, and EER-CGHHOA achieve 1.40 s, 1.38 s, and 1.35 s, respectively. However, the proposed method is limited to simulation-based evaluations and focuses primarily on packet drop resilience for security. Future work will involve real-world deployment to validate practicality, designing lightweight deep-learning strategies for resource-constrained nodes, and extending the security model to handle diverse attacks such as Sybil, wormhole, and jamming, thereby making the framework adaptable to next-generation IoT and 6G-enabled WSN environments.