Enhanced Gated Sway Network and Hybrid Henon Encryption for Secured VANET Communication

El-Shafai et al. [18] developed an AI-based collective classifiers for identifying interference assaults in VANETs. The suggested framework combines machine learning and neural network classifiers to examine signal properties within VANET transmission pathways. Their ensemble classifier combines Random Forest, Extra Tree, and fine-tuned Convolutional Neural Network, achieving an impressive detection accuracy of 99.8125%, outperforming individual classifiers. This approach significantly enhances VANET security frameworks to counteract jamming assaults, strengthening the overall protection and dependability of VANET communication in smart city infrastructures. However, the model needs further validation in real-world dynamic VANET scenarios.

Bayan et al. [19] constructed a Deep Learning-driven intrusion recognition framework for detecting position falsification threats in VANETs. Their system employs Multi-Layer Perceptron (MLP) algorithm that considers RSSI aggregation of first-hop neighbors and Time Difference of Arrival (TDoA) as new detection features. Trained offline using the VeReMi dataset, the model can be deployed at a vehicle's Onboard Unit (OBU), reducing computational complexity and execution time. Their DL-IDS model demonstrates high accuracy and F1-score values, exceeding existing models by 2-7% with advantages in computational efficiency. However, the system may struggle with detecting complex hybrid attacks.

Suman et al. [20] proposed an Improved LeeNET (I-LeeNet) architecture to identify and mitigate various attacks including Botnet, sybil, DoS, wormhole, PortScan, Blackhole, and BruteForce. The architecture intelligently blends Convolutional Neural Networks (CNN) and Adaptive Neuro-Fuzzy Inference Systems (ANFIS) for real-time attack detection. Their approach includes KIDS module for known attack detection and UIDS module for learning previously unidentified attacks. Tested on three datasets (i-VANET, ToN-IoT, and CIC-IDS 2017), the proposed method achieved average accuracies of 97.21%, 97.75%, and 96.66% respectively, demonstrating promising real-time application potential. But the computational demands may challenge implementation in resource-limited environments.

Jabbar et al. [21] suggested the centrality relied clustering mechanism for the recognition of sybil and wormhole threats in VANET framework. The main idea is to maintain the network’s reliability by choosing the appropriate cluster head (CH) based on the centrality measures to cluster the vehicles for an effective data transmission. The findings demonstrated the excellent performance of the recommended model regarding network lifetime and computational cost. However, the suggested framework requires brighter light of analysis in deploying the intelligent system for the detection of multiple attacks.

Rafsanjani et al. [22] proposed unmanned aerial vehicles (UAV) in the VANET environment to detect the malicious vehicles. A vehicle routing unit (VRU) has been introduced as a method to direct the data, thereby mitigating the malicious vehicles. The proposed framework has improved the packet delivery ratio by 16% and detection ratio by 7% evaluated against the other methods. However, these methods fail to throw the deeper light of counterfeiting the attacks especially sybil and wormhole attacks.

Polat et al. [23] proposed a stacked sparse autoencoder with Softmax classifier neural network architecture for identifying DDoS assaults aimed at SDN-powered VANET. Their approach dimensionally reduced features using SSAE to extract the important features, which are then utilised as input for the Softmax categorizer. Evaluation outcomes demonstrated that their recommended approach attained 96.9% precision, surpassing other models in identifying DDoS intrusions in SDN-driven VANET. However, the model's effectiveness against new attack patterns needs further testing.

Wang et al. [24] suggested a lightweight and effective authentication system for safe VANET transmission (LESPP) that preserves privacy. The proposed technique only requires the construction of a fast MAC re-generation and a lightweight symmetric encryption and message authorization code (MAC) for message signing. To safeguard security and conditional tracking, such strategy employs a self-created phony identity. The suggested approach significantly reduces calculation costs.

Alfadhli et al. [25] demonstrated the application of genetic hashing function to resolve the issues of unsafe driving sequences. Furthermore, the vehicle authorization is performed exclusively one time by the VANET framework manager thereby increasing the authentication process. To mitigate the attacks, the framework offers the confidentiality to maintain safety of the vehicles. The framework offers the more superior performance in maintaining the privacy of the vehicles against the multiple attacks. But the framework needs improvisation in detecting the attacks which will be occurring in unknown occasions.

Fatemidokht et al. [26] introduced a cluster-based routing protocol termed QoS-based Monitoring of Malicious Activity (QMM-VANET) to improve network QoS. The protocol comprises of three components: CH identification, optimal neighbor identification, and gateway renewal method. The experiment is carried out using NS2 in the highway situation. Packet delivery ratio, latency, and network reliability are the key metrics focused on in the result performance analysis. However, characteristics such as detection ratio and high-end privacy are overlooked.

Guo et al. [27] presented the game-theoretic-relied incentive framework for collaborative recognition of multiple –threats in the VANET framework. These algorithms combine the different machine learning models. But fails to improve the safety measures and confidentiality breaches in the VANET framework.

Figure 2 illustrates the recommended system consist of four components including Dataset Collection, Data-Preprocessing unit (DPU), Centrality (Sway) Extraction (CFU) and Modified Gated Recurrent learning network (MGRLN). In the event of classification, proposed model detects the two important parameters such as type of attack (TA) and malicious node provide in the VANET framework. The comprehensive explanation of the recommended framework is provided in the previous section.

4.1 Dataset collection

For an efficient data collection, powerful integration of the SUMO [28], VEINS [29] and OMNET [30] are used in this research. To induce the attacks in the networks, python-based attack injection module has been introduced. Nearly 4,00,000 data are collected, with 70% allocated for training and 30% for testing. The complete description of data generation process is depicted in Algorithm-1.

4.2 Data pre-processing unit

The data gathered in the preceding phase may contain the misleading or null values, hence data prep-processing technique is needed before implementing to the proposed learning model. The pre-processing stage involves the data labelling and data normalization. Labeling the dataset is a crucial phase in data preprocessing. Based on the events gathered from the prior phase, all traffic was classified as normal (0-label) and attack (1-sybil, 2-wormhole, 0-Normal) according to the information like source address, destination address, time, and duration. After categorizing the data, a min-max normalization method was applied to scale the features to a uniform range, typically between 0 and 1. This guarantees that every attribute plays an equal role in the training model and prevents bias towards attributes with greater ranges. The min-max normalization formula is expressed as:

$Normalization=\frac{x-Min}{Max-Min}$                  (1)

where, Min denotes Minimum data, Max indicates maximum data and x is the collected raw data.

After normalizing the data, these data feed to the recommended DL model for the further identification of various threats.

2.png

Figure 2. Proposed framework for the deep sway networks and hybrid encryption process

4.3 Centrality feature extractor

The major purpose of this research is to design an effective feature database capable of classifying normal and influential nodes. Numerous centrality measure detection techniques have been introduced in existing research to quantify node importance. However, this paper highlights the application of an expanded set of centrality metrics to attain the precise categorization of significant nodes.

To capture both the structural and functional traits of the nodes, the subsequent centralities are evaluated, as detailed below.

4.3.1 Degree centralities

It reflects the count of connections associated with the nodes. It consists of two variations: indegree and outdegree centrality. These measures can be computed utilising the subsequent formulas.

i) Indegree centrality

$D_{i n}\left(P_i\right)=\left|P_{j i} \in P\right|, j \neq i$           (2)

$P_{j i}$ represents the connection extending from $P_i$ node to the assessed node P.

ii) Outdegree centrality

$D_{o t}\left(P_i\right)=\left|P_{i j} \in P\right|, i \neq j$            (3)

$P_{ij}$ represents the connection strength (i.e., edge) from the assessed unit $P_i$ to all another units $P_j$ in the system.

4.3.2 Betweenness centralities

It signifies the proportion of all shortest routes traversing the nodes. The numerical representation for this measure is presented by:

$D_B\left(P_i\right)=\sum_{P_m \neq P_i \neq P_n} \frac{\mu_{P m, P_n\left(P_i\right)}}{\mu_{P m, P_n}}$           (4)

where, $\mu_{P_m, P_n}\left(P_i\right)$ represents the count of minimal routes among nodes $P_m$ and $P_n$ that traverse through $P_i$ and $\mu_{P_m, P_n}$ denotes the count of all shortest paths among $P_m$ and $P_n$.

4.3.3 Closeness centralities

It represents the interval of nodes within the systems and its mathematical formulations provided below:

$D_c\left(P_i\right)=\frac{N}{\sum_{P y} d\left(P_y, P_i\right)}$                      (5)

where, N represents the count of vertices in the network and d (P_y, P_i) denotes the interval among Py and Pi nodes.

4.3.4 Eigen vector centralities

It is utilized for computing the centrality of other nodes in the network. The equation to calculate this is presented below:

$E_v\left(P_i\right)=1 / \alpha \sum_k \gamma_{P_k, P_i} * E_v\left(P_k\right)$                      (6)

where, $\mathrm{A}=\alpha(k, i)$ represents the adjacent matrix of a graph and γ is a constant.

4.3.5 PageRank centralities

This calculates the node ranking according to their centrality within the systems. Its mathematical formulation is determined as:

$R_p\left(P_i\right)=\rho \sum_k \frac{A_{P_k, P_i}}{d_k} * R_p\left(P_k\right)+\beta$                    (7)

where, ρ and β represent constants, and d_k signifies the out-degree of P_k, where this degree is positive, or d_k is 1 if the out-degree of P_k is zero. Furthermore, A = (ai,j) denotes the adjacent graph matrix, where A = α(k,i) is the adjacency matrix.

4.3.6 Position centrality

It is regarded as the major significant metric, representing the placement of the nodes in relation to the key nodes, that are computed using the Pagerank algorithm.

$\mathrm{H}_{\mathrm{c}}\left(\mathrm{P}_{\mathrm{i}}\right)=\beta \sum_{\mathrm{k}} \gamma_{\mathrm{P}_{\mathrm{i}}, \mathrm{P}_{\mathrm{k}}} * R_p\left(P_i\right)$                     (8)

where, A = (ai,j) denotes the adjacent graph matrix, and $R_p\left(P_i\right)$ represents the node PageRank, with β being a constant.

4.3.7 Clustering co-efficient

It signifies the proportion of triangles which are available within the total possible triangles in the neighborhood of the nodes. The numerical formula to calculate the clustering coefficient is expressed as:

$\mathrm{C}_{\mathrm{c}}=2 \mathrm{M}_{\mathrm{P}, i} / \mathrm{K}_{\mathrm{i}}\left(\mathrm{K}_{\mathrm{i}}-1\right)$                   (9)

where, Mp,i is the count of neighbor sets related to the hub pi. Within the expression, it is integrated to the count of potential neighbor sets of hub pi, where kpi = (kpi-1)/2, with kpi being the degree of hub pi. Figure 3 illustrates the representation of centrality measures used to analyze node importance within the VANET communication network.

3.png

Figure 3. Representation of centrality measures in a VANET topology

Table 1 provides the summary of feature vectors employed for classification.

Table 1. Overview of features utilised for the suggested classification

It is regarded as the most captivating form of Long Short-Term Memory (LSTM). This concept was introduced by Chung et al. [31], that seeks to integrate the forget gate and input array into a unified vector. This architecture accommodates extended sequences and prolonged memory. The intricacy is significantly minimized in contrast to the LSTM network.

The subsequent equations were defined by Chung to describe the features of GRU.

$h_t=\left(1-x_t\right) \odot h_{t-1}+x_t \odot h_t$                 (10)

$\tilde{h_t}=g\left(W_h x_t+U_h\left(r_t \odot h_{t-1}\right)+b_h\right.$                (11)

Two gates of GRU are presented as

$z_t=\sigma\left(W_h x_t+U_z h_{t-1}+b_z\right)$                (12)

$r_t=\sigma\left(W_h x_t+U_r h_{t-1}+b_r\right)$               (13)

The complete GRU defining formula is expressed by:

$P=G R U\left(\sum_{t=1}^n\left[x_t, h_t, Z_t, r_t(W(t), B(t), \eta(\operatorname{tannh}))\right]\right.$             (14)

where, $x_t$ is the input attribute at the present state, $r_t$ is the resultant state, and $h_t$ is the output of the element at the current time step. $z_t$ and $r_t$ are the update and reset gates, while $W(t)$ and $B(t)$ represent the parameters and bias coefficients at the current point in time. To minimize the intricacy, each gate in the GRU is calculated using only the prior latent state and offset, thus reducing the overall count of variables by 2 times nm, compared with the established GRU model. Based on this modification, Eqn. (12) and (13) is modified as

$z_t=\sigma\left(U_z h_{t-1}+b_z\right)$                 (15)

$r_t=\sigma\left(U_r h_{t-1}+b_r\right)$                (16)

Again, the overall GRU characteristics is modified and expressed mathematically in Eq. (17)

$P=G R U\left(\sum_{t=1}^n\left[x_t, h_t, z_t, r_t(B(t), \eta(\operatorname{tannh}))\right]\right.$              (17)

4.5 Modified AES encryption and decryption

The proposed uses the same operations of the original AES with some modifications. DNA encoding is adopted instead of the traditional permutation and shifting technique. This alteration aims in minimizing the duration for encryption and decryption procedure while maintaining commands with strong defense properties. Dual-level Henon chaotic maps are employed in the creation of robust encryption. To begin with, the VANET data is divided into two separate units depends on the byte positioning. Firstly, Henon maps are utilized to generate the S1 box. Using the outcomes from the first phase, the initial conditions of the Henon maps are set and utilised to create the hybrid S2 box. These two S-boxes are then encrypted with DNA processing to produce the combined S3 box. At last, the information is enciphered utilising the recently generated S3 box. All such process aims to make AES lightweight by reducing its encryption/decryption time simultaneously, still strength to avoid VANET attacks. The detailed description of henon chaotic maps and DNA encoding process is provided below

4.5.1 Key generation process

To eliminate the intricacy involved in using matrices within the encryption method, the first positions of the sensor input bytes are considered. Initially, Henon maps are generated at random as described in Algorithm-2. These created logistic maps are utilized to construct the intermediate S1 box. The intermediate S1 box is developed by combining the Henon maps (H) and the input data (K). Instead of traditional permutations and diffusions, DNA addition encoding is employed to produce a highly secure intermediate S1-Box sequence. The process for generating S1 is illustrated in Algorithm-2.

$\mathrm{H}= Heon maps (K)$ For $\mathrm{K}=$ Input data Bytes             (18)

$S 1=\bmod ($ byte $\{(H) D N A K($ input $))$                  (19)

In the subsequent phase, Henon maps are once again generated, utilizing the outcome series from S1-box. The intermediate S2-box is formed utilising VANET inputs (O) and Henon maps (H). During this process, all permutations are substituted with DNA-based addition encrypting for developing a lightweight and easily deployable system, which still retains its robust defense capabilities resisting any threats.

$\mathrm{H}= Heon maps (K)$ For $\mathrm{K}=$ Input data Bytes              (20)

$S 2=\bmod ($ byte $\{(H) D N A K(input))$           (21)

4.5.2 Encryption process

Ultimately, the intermediate variables (S1 and S2) are merged to generate the new hybrid S-boxes. Upon being processed repeatedly, the input data, along with the hybrid S-box keys, undergo the DNA XOR operation, as outlined in Algorithm-3. Consequently, it produces robustly encrypted bytes that vary separately with every iteration. The entire encryption process involving the S-box is depicted in Algorithm-3.

$S=S 1 D N A-X o R-S 2$               (22)