Fuzzy computation on ensemble deep network for the performance of vampire attack detection model in WSN

In 2017, Kumar [18] have developed an intelligent vampire attack detection approach, Parno, Luk, Gaustad, and Perrig (PLGP), to provide security in WSN. Previous attack detection mechanisms were only focused on the detection of DoS, and this work has been especially designed for detecting vampire attacks. The vampire attacks affected the performance of the system during the transmission of information in WSN. The resource depletion attacks in the routing protocol layer quickly debilitated the battery power of the nodes, and hence, the networks were disabled permanently. Therefore, the detection of attacks was important before routing in WSN.

In 2022, Juneja et al.[19] have suggested a twofold method for detecting vampire attacks. Here, two mechanisms as the cooperation trust mechanism and attack migration were integrated by policy gradient-deep reinforcement learning (PG-DRL). Secure hop was selected by the proposed PG-DRL approach, while vampire attacks were presented. The detection ratio of the proposed technique was higher than the existing schemes, and the lifetime of the network has also been enhanced by this approach.

In 2022, Alkwai et al.[20] have proposed a probabilistic fuzzy-based routing protocol with an authentication strategy for detecting the vampire attacks, where data optimization was performed using a hybrid clustering approach. The growing types of vampire assaults were detected through fuzzy-based chain rule with the help of probability formulas. The network security at the time of routing has been performed using the authentication routing protocol. The experimental outcome has proved that the proposed mechanism attained better effectiveness according to throughput, control overhead, energy consumption, and latency.

In 2014, Anoopa and Sudha [21] have offered an intelligent vampire attack detection and control scheme in WSN. The resource depletion attacks were explored at the routing protocol layer, which permanently disabled the networks in the application layer. From the empirical analysis, the developed vampire attack detection and control method saved the ad-hoc wireless nodes from power drainage, and secure packet forwarding has been carried out in WSN.

In 2020, Sajan and Jasper [22] have proposed a secure atom search routing (SASR) strategy for the prevention of vampire attacks which provided higher security during data transmission. The global optimization problems were solved, and the balance between exploitation and exploration was improved by this SASR methodology. The effectiveness of the SASR method has been validated with the previous algorithms based on measures like network lifetime, trust, energy, delay, throughput, trust detection rate, packet delivery ratio, and communication cost. The lifetime of the network has also been improved by this suggested approach.

In 2017, Rathish and Rajaram [23] have developed a clustering-based attack detection mechanism during message forwarding in the network, where the attack was mitigated by discarding the attacked file before it passed into the destination node. In the first aspect, the cluster was organized, and then cluster member grouping was performed. Finally, the cluster head was selected for the detection of vampire attacks in the network. The malicious nodes were determined based on notification messages, and the detected attacks were mitigated through a filtering mechanism. The proposed model improved the detection rate, increased the packet transmission, and reduced overhead, packet loss rate, and energy consumption than the prior models.

In 2011, Vasserman et al.[24] have explored resource depletion attacks that permanently disabled networks at the layer of the routing protocol. There is no specific relationship between the vampire attacks to the protocols, but they relied on the properties related to classes of routing protocols. The detection has been carried out by sending messages to the network. The mitigation strategies have also been explained to ensure secure data forwarding in WSN.

In 2015, Mulla et al.[25] have developed a vampire attack detection and control strategy in wireless networks. The secure packet forwarding mechanism has been implemented, which keeps the network from harm and danger due to these vampires at the packet forwarding level.

In 2021, Lansky et al. [38] have designed a novel research work to verify the efficiency of the deep learning-based intrusion detection models. In most of the research work, intrusion detection procedures were employed to identify the intrusions effectively.

In 2014, Kim et al. [39] have implemented a hybridized intrusion detection mechanism using a decision tree and SVM. The developed hybridized technique was good in reducing the time complexity issues at the training phase and accomplished better outcomes.

In 2021, Srikaanth and Nagarajan [40] have proposed an efficient vampire attack detection technique through fuzzy rough set (FRS) with routing procedures. The FRS was better at quantifying the characteristics of mobile node characters and detected the vampire attacks. FRS-derived probabilistic variable precision-based mitigation technique (FRS-PV-MT) was employed to overcome the problems while detecting the attacks by minimizing the latency.

In 2016, Mirjalili and Lewis [43] have designed an efficient optimization technique named whale optimization algorithm (WOA), which was good in tackling the optimization-related issues and offered comparatively higher performance efficiency than other schemes.

In 2023, Alotaibi and Ilyas [45] have designed an efficient intrusion detection model to enhance the security of the network. Here, ensemble machine learning techniques such as LR, RF, KNN, and DT were suggested. Moreover, a stacking procedure was suggested to perform the final validation process.

In 2023, Hossain and Islam [46] have proposed an intrusion detection model by considering ensemble machine learning techniques. Ensemble techniques utilized for the validation were AdaBoost, gradient boosting, RF, and stacking. Here, correlation analysis was suggested for acquiring the significant features to detect the intrusions in the network.

In 2023, Ismail et al. [47] have recommended a lightweight ensemble machine learning technique for detecting intrusions in WSN. Ensemble machine learning procedures utilized for the validation were stacking, bagging, and boosting. In the validation, the developed technique effectively reduced the false alarm rate and detected the respective attack in minimal time.

In 2019, Vinayakumar et al. [50] have suggested a deep neural network (DNN) to identify the efficiency of the developed scheme in detecting various attacks. The developed scheme was good in detecting various attacks and verified their performance over dynamic conditions.

In 2025, Karakaya [44] have proposed a hybridized technique by integrating fuzzy logic and ensemble learning models for improving the overall security of the network. Here, multiple classifiers were suggested to enhance the detection efficiency and also network flexibility was improved through fuzzy logic.

In 2025, Sivakumar et al. [48] have recommended the recurrent crossover-based dynamic differential (RC-DD) for detecting various attacks in WSN. Here, the localization was carried out using KNN among the sensor nodes. Then, multiple attacks presented in the WSN were identified using an RNN by reducing the localization errors.

In 2025, Sharma et al. [49] have designed a hybridized technique by fusing bidirectional gated recurrent unit (BiGRU) and attention mechanism (ATTN), which helped to identify the intrusions in WSN. Next, the parameters in BiGRU-ATTN were tuned using enhanced black widow optimization (EBWO) and achieved a superior outcome than others.

An efficient optimization technique named E-PFOA is employed in this research work to enhance the vampire attack detection efficiency. The most important features from the input data are selected optimally by the proposed E-PFOA. Here, the relief score in the network is enhanced, which supports selecting the optimal features. The E-PFOA is the advanced version of PFOA [26], which is designed by updating the random parameters according to the fitness related to the best and worst.

A new vampire attack detection model EDN is implemented in this research work for the identification of attacks in the WSN. The developed EDN is an ensemble approach, which is designed by considering different deep learning networks like TCN, RNN, GRU, and LSTM. Using an ensemble vampire attack detection model for the WSN helps to handle the network from complex and dynamic threats and also collects the required patterns by adapting the dynamic changes. Moreover, it supports reducing the false positive rates for attaining better decision-making.

In the developed EDN-based vampire attack detection model, different techniques like TCN, RNN, GRU, and LSTM are used for the implementation. RNN is good at collecting the temporal dependencies and also recognizing the most significant patterns over time sequences. It has the ability to process the information presented in the prior steps and also quickly process the sequential data. Yet, RNNs are prone to vanishing gradient issues and also take more time to learn the required details from the sequences. Moreover, it has limited memory with low processing speed and also their training procedures are complicated. Thus, to rectify these issues LSTM is employed, which is good at handling the longer-term dependencies for remembering the prior information and also effectively manages the information flow. In addition, it effectively maintains the flexibility of the network over different variable lengths and supports to maintain the time series forecasting. However, their implementation expenses are higher, and they also process the data samples slowly. Network complication is increased due to the enormous memory requirement while processing the large samples. In some cases, they are prone to overfitting issues while training procedures are executed and also lead to poor generalization among the unseen data. Hence, to tackle these issues, GRU is suggested to use in the network as it has quick training efficiency. The architecture of GRU is simple and also good in handling large dependencies by maintaining proper training to address the vanishing gradient issues. Moreover, it effectively reduces the overfitting problems by maintaining higher stability in the network. Yet, they are lagging in collecting the most suitable patterns with higher understanding rates. In some cases, they are limited to interpretability. It wasn’t able to maintain better balancing in different classes, and also managing the memory is complex. In order to rectify all these problems, TCN is considered which is capable of minimizing the training time and also effectively maintains the flexibility while collecting the longer-range dependencies. This technique supports parallel processing with a proper training procedure. It is suitable to use in real-time detection by avoiding information loss. Yet, their real-world performance is low and also complicated to identify the relationship among the events. Moreover, it needs to tackle generalization issues under a wide range of conditions. Hence, it is important to tackle several issues that arise takes place in the prior technique, a new ensemble technique EDN is implemented by considering the strengths of RNN, LSTM, GRU, and TCN, which supports accomplishing better vampire attacks in WSN.

In the developed EDN-based vampire attack detection model, optimally selected features \({\text{Fe}}_{f}^{{{\text{op}}}}\) are used as the input. Here, the inputs are given to RNN, LSTM, GRU, and TCN. Next, the 1st set of attack detected scores is attained from RNN, 2nd set of attack detected outcomes is obtained from LSTM, 3rd set of attack detected outcomes is acquired from GRU and 4th set of attack detected scores is gathered from TCN. Next, the attained four sets of detected scores are forwarded to the fuzzy ranking phase. Here, a fuzzy ranking procedure is used to execute the vampire attack detection procedure and accomplish the vampire attack detection outcome for WSN.

The developed EDN is designed by considering RNN, LSTM, GRU, and TCN. RNN is good in processing sequential information, and also it is good at remembering prior information. But, the developed EDN is efficient at accomplishing better vampire attack detection efficiency by tackling the overfitting issues. Moreover, it maintains higher robustness over the noisy information and also effectively reduces the errors. LSTM is better at handling the gradient vanishing issues and improves the network adaptability. The EDN is good at maintaining network flexibility and also capable of handling higher-dimensional information. When compared to individual learning models, the detection performance of EDN is greatly enhanced for complex classes. The variance and bias of this EDN are lower and achieve higher vampire attack detection efficiency under the varying environmental factors. The hidden threats are easily detected through EDN, and it is recommended for detecting the vampire attacks.

The basic network details of RNN, LSTM, GRU, and TCN are utilized for constructing the ensemble network for detecting the vampire attacks are detailed as follows.

RNN [27]: RNN is developed based on ANN, and it has the ability to deal with temporal data for providing better detection outcomes. The input vectors are directly mapped to the output vector for processing diverse input sizes. Based on the history of input, it produces output at the end. Input, hidden, and output are the three layers present in RNN. The optimally selected feature \({\text{Op}}_{n}\) is given as the input, \(r_{n}\) is the hidden state value, and \({\text{Ts}}_{n}\) is the output. These three variables are determined based on the time step. Previously hidden state values are helpful for the determination of the output value. The hidden state function is mathematically represented in Eq. (15).

The output state values are provided in Eq. (16).

The weight matrix is denoted by the term \(W\),\(\varsigma\) is the nonlinearity function, and the term \({\text{bi}}_{r}\) as well as \({\text{bi}}_{{{\text{ts}}}}\) is the hidden state and output state biases. The mapping between functions is carried out in the hidden layer to solve the gradient issues. Finally, it produces a prediction score about the vampire attacks in WSN.

LSTM [27]: The gradient vanishing problems are solved using LSTM. Here, the memory cell is presented to learn the information from the inputs and store the information in short as well as long-term memory. The optimally selected feature \({\text{Op}}_{n}\) is given as input. The previous memory cell state input and hidden state value are represented by the terms \(m_{n - 1}\) and \(r_{n - 1}\), respectively.

The hidden state value from the previous state is reset by the forget gate, and it is no longer utilized in the current memory cell. The value added in the cell state is preferred by the write gate, and the input gates perform the writing process. Moreover, the input gate decides the amount of information needed to be added in the cell state. The functions related to input, hidden state, and cell state are given as following expressions.

The standard sigmoid activation function is denoted by \(\varsigma\), which ranges between \(\left[ {0,1} \right]\). No information is passed through in the gate when it lies between 0 and all the information is passed through the gate when it contains the value of 1. The term \(W\) indicates the weight matrices of its corresponding states, which are helpful for mapping cell state information to the previous states. These gates play a significant role in controlling the information. The newly updated cell state value by forgetting the previous cell state is given in Eq. (21).

The amount of value produced based on the new cell value is expressed in Eq. (22).

Finally, the vampire attack is detected by this LSTM model.

GRU [28]: It is slightly simpler than LSTM, and it consists of two gates, such as an update gate and a reset gate. The flow of information carried over the current state is controlled using the update gate, and the information to be ignored from the previous state is controlled through the reset gate. A huge amount of information is lost when the value of the update gate is larger. More information is ignored when the value of the reset gate is smaller. The frequent activation of reset gates accompanies short-term dependencies, and the frequent activation of update gates accompanies the long-term dependencies. The GRU gate function is represented in the following expressions.

The logistic activation function is represented by the term \(\varsigma\),\(op\) is the optimally selected feature applied to the GRU model,\(r_{n}\) is the previously hidden state value \(W_{{{\text{st}}}}\), \(W_{m}\),\(W_{\dddot r}\) and \(W_{{{\text{pu}}}}\) are the weight matrices of the reset gate, update gate, hidden state, and output gate. The matrix multiplication between two vectors is performed to provide the vampire attack detection outcome by the GRU.

TCN [29]: The bottleneck effect in recurrent models is solved by the TCN method. The fully CNN gives the same length of output sequence when compared to the input. The residual and dilated connections in the TCN provide efficient vampire detection outcomes. The optimally selected features are forwarded as the input to TCN. The detection ability of TCN is high when compared to other recurrent models. The receptive field of input is enlarged with the usage of dilated convolutions in TCN. The input sequence is represented as \({\text{op}} \in \Re^{n}\), the elements on the dilated convolution operation are indicated by \(G\)_, and the convolution filter is denoted as \(f:\left\{ {0,1,...,t - 1} \right\} \to \Re\). The TCN operation ends with an index of the sequence \(z\) given in Eq. (28).

The size of the convolution filter is represented as \(t\), the dilation factor is denoted as \(d\), and the index of the element based on the input is represented as \(z - d \cdot j\). The dilation factor is further improved with regard to the network depth. More suggested detection outcomes are provided with the support of this TCN model, and the interpretabilities of the TCN method are also high.

The structural depiction of the implemented EDN-based vampire attack detection scheme for WSN is given in Fig. 3.

Full size image

In the developed vampire attack detection model, the fuzzy ranking technique is employed in the attack detection procedures. Initially, the training procedure is carried out in the ensemble deep learning model and accomplished a predicted score. Next, the predicted score attained from each technique, such as RNN, LSTM, GRU, and TCN, is ranked according to its value. Then, the fuzzy ranking procedure is carried out to obtain the fuzzy weights for RNN, LSTM, GRU and TCN. Here, Rank 1 indicates the high membership value, Rank 2 represents the medium membership value, Rank 3 specifies the low membership value, and Rank 4 denotes the very low membership values for the ensemble technique, such as RNN, LSTM, GRU, and TCN. The fuzzy weights support obtaining more accurate decision-making and are also efficient in handling the uncertainties presented in the data to offer reliable outcomes. Generally, the fuzzy ranking is used widely to execute the multi-criteria decision-making in complex cases. Further, the fuzzy weighted aggregation procedure is carried out by considering the weights and the prediction score. In this phase, expert opinions as well as qualitative information are considered for better decision-making. Here, the fuzzy information displays the fuzzy ratings used for the alternatives. At the end, final scores are compared with the threshold value for multi-class prediction and provide accurate vampire attack detection outcomes by considering the higher values.