A plethora of excellent surveys have been published regarding the security of vehicular networks, which have covered the overview, requirements, characteristics, challenges, and solutions against attacks. The different research works provide their points and contributions (See Table 1). The survey of this research is fully based on the authentication mechanism of vehicular communication; therefore, the discussion of the related research is covered in this section.

Ali et al. [10] presented a survey on the privacy schemes and authentication in the VANET. Moreover, the security requirements, limitations, attacks, and efficiency of performance are shown in this research. In 2020, Farooq et al. [11] reviewed different authentication techniques in VANET. Additionally, they showed comparisons of several authentication protocols in the context of privacy preservation, batch verification, signature, attack mitigation, and communication. In 2021, Abbas et al. [12] proposed a complete review of authentication based on the blockchain in the Internet of Vehicles (IoV) and VANET. To emphasize clearly, the detailed discussion on blockchain, attacks, and mitigation to vehicular networks is discussed in this research. Moreover, the research shows a thorough comparative study of the technique used, authentication scheme, evaluation tools, network models, and attack prevention. In the same year, Al-Shareeda et al. [13] emphasized a survey on VANET’s different authentication and privacy schemes. Azam et al. [14] presented a detailed discussion on the taxonomy of authentication schemes on VANET. The scalability requirements, security, and privacy were compared with the existing research. Additionally, recent technologies such as 5G, blockchain, and 5 G-SDN were discussed to develop low-cost, low-overhead, and low-communication-powered authentication. In the same context, Muhammad et al. [15] reviewed a few DL-based authentications in autonomous vehicles. In 2022, Jenefa and Mary Anita [16] presented the broad classification of VANET authentication based on message signing and verification methods. Moreover, the security attacks, performance parameters, and requirements are compared with other research. In 2023, Dong et al. [17] proposed a survey on security challenges and properties with respect to attacks and builders. Furthermore, this research discussed the availability of systems, the authenticity of nodes, integrity, confidentiality, and non-repudiation of information. Sripathi Venkata Naga et al. [18] presented the classification of certificate authentication and features of the VANET in their research. The classifications were further extended based on the attacks addressed, security requirements, type of authentication, and the technique used. The review showed the complete performance analysis with the existing VANET authentication. In 2024, Sutradhar et al. [19] surveyed vehicular communication on the basis of privacy preservation. Alternatively, Shawky et al. [20] proposed a review of PHY-layer, cross-layer, and crypto-based authentication on VANET. Alternatively, Soujanya and Azam [21] discuss the problems in general authentication in vehicular networks in their studies. On the other hand, Aljehane [22] studied the roles and challenges of DL in autonomous vehicles; however, the research was not very comprehensive. In the same context, Zhang et al. [23] reviewed the application of ML and DL in ITS. However, the discussion on authentication was neglected in both of the studies. In 2025, Yang et al. [24] studied privacy concerns of Vehicular Cloud Computing (VCC) on the basis of ML-based approaches.

Traditional authentication mechanisms—such as certificate-based schemes and symmetric key cryptography—struggle to meet the real-time, scalable, and adaptive demands of highly dynamic vehicular environments. Deep Learning (DL), with its powerful pattern recognition capabilities, has recently gained attention as a promising tool to enhance the robustness and intelligence of authentication protocols. However, despite the broad interest in DL across ITS applications, its application in vehicular authentication remains underexplored. The research, as mentioned earlier, contributed to the knowledge of VANET information, which is respectable. Although there are few surveys on autonomous vehicle security [25], security issues in IoV [26], motion control security on road vehicles [27], and context-aware specified [28] cyber-physical security [29] in recent days, our research has been completely based on the recent advancements in deep learning for authentication in vehicular networks. This survey aims to bridge this critical research gap by offering a comprehensive overview of DL-based authentication techniques tailored for VANETs. While previous surveys have reviewed general security protocols or specific technologies such as blockchain or cryptography, a focused and up-to-date review on DL techniques for authentication is notably missing. Addressing this deficiency is vital, not only to consolidate current knowledge but also to guide future innovations towards building secure, scalable, and intelligent vehicular networks. To show the unique contribution of our research, the comparisons are shown with the existing related research in Table 1. The indicators ‘✔’ and ‘✖’ show if the specified factors are discussed in the mentioned review. Moreover, ‘-✖-’ represents if the specified factors are discussed by providing sufficient knowledge of content.

This survey focuses on topics that are not covered comprehensively in any of the research, as far as we know. Most importantly, the main focus of this survey is to investigate various DL methods in vehicular authentication. The main contributions are discussed as follows:

To provide independent information, this survey discussed the overview of VANET in detail, along with the architecture and communication standards in VANET.

The survey highlights the potential of DL in enhancing the security and efficiency of authentication mechanisms in vehicular networks. It underscores the need for robust, adaptive models to address evolving threats and ensure reliable communication in dynamic vehicular environments.

The organization of the rest of the paper (see Fig. 3) are as follows. In Section 1, the introduction, and the contribution towards vehicular authentication are discussed. Section 2 shows the overview of the vehicle network, including the architecture and communication standards in VANET. A detailed discussion on DL-based authentication in vehicular communication is shown in Section 3. In addition, the challenges, open issues, and future directions on the DL method-based vehicular authentication are discussed in Section 4. Finally, the survey is concluded in Section 5.

VANETs are a specialized form of Mobile Ad-hoc Networks (MANETs) designed to facilitate communication among vehicles and between vehicles and roadside infrastructure. The architecture of VANETs typically involves several key components. The architecture of the vehicular network is shown in Fig. 4.

Vehicular networks rely on several communication standards to facilitate efficient and reliable communication among vehicles and between vehicles and roadside infrastructure. The primary communication standards in VANETs are discussed as follows.

Supervised deep learning has emerged as a powerful approach for enhancing authentication in vehicular communication systems, addressing the growing need for robust security in Vehicle-to-Everything (V2X) networks. By leveraging hierarchical neural architectures, such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), deep supervised learning models can analyze complex patterns in heterogeneous data sources, including sensor data, communication logs, and behavioral patterns, to distinguish between legitimate and malicious entities.

In the realm of deep learning for vehicular authentication, a Multi-Layer Perceptron (MLP) serves as a foundational neural network architecture that can be leveraged to address the challenges of secure and efficient vehicle identity verification. The architecture of authentication using MLP is shown in Fig. 6. Recently, Zhang and Li [46] proposed an automatic irrigation system with authentication features in VANET using a neural network. They used MLP to analyze sensor data to reduce water waste. Alternatively, Artificial Neural Networks (ANNs), inspired by the structure and function of biological neural networks, consist of interconnected layers of nodes that process input data to extract meaningful patterns and make decisions. ANN can learn to recognize unique signatures, such as cryptographic keys, vehicle-specific sensor data, or driving behavior, to authenticate vehicles in real-time scenarios [41]. Islam et al. [47] proposed a license plate authentication in a barrier access control tailoring with ANN to recognize characters in license plates. In addition, ANN can further use the maxout layer to extract features for authentication in a Deep Maxout Network (DMN). Recently, Kaur and Kakkar [34] used the Fractional Aquila Spider Monkey Optimization (FASMO) algorithm to train the bias and weights of DMN for attack detection. Alternatively, Convolutional Neural Networks (CNNs) have emerged as a highly effective architecture for processing structured and spatial data, such as images, sensor inputs, or communication patterns, to enhance security. CNNs leverage convolutional layers to automatically extract hierarchical features from input data, followed by pooling layers to reduce dimensionality and fully connected layers for decision-making. In vehicular authentication, CNNs can be employed to analyze visual data from cameras, such as license plate recognition or driver behavior monitoring, to verify vehicle identity and detect anomalies. In their research, Xun et al. [48] proposed an authentication scheme by leveraging the secure driver fingerprint using CNN and support vector domain description. On the other hand, Borra et al. [40] proposed biometric authentication for transport users using Multilayer CNN (ML-CNN) and Deep Hashing Component Analysis (DHCA) to extract high-level and low-level features. In addition, Qiu et al. [49] proposed a signal enhancement-based authentication using a deep convolutional generative adversarial network (DCGAN). On the other hand, Recurrent Neural Networks (RNNs) offer a powerful framework for addressing the dynamic and sequential nature of data in ITS. Unlike traditional feedforward networks, RNNs are designed to process sequential data by maintaining a hidden state that captures temporal dependencies, making them particularly suitable for tasks involving time-series data. Additionally, advanced recurrent architectures such as Long Short-Term Memory (LSTM), Bidirectional RNN/LSTM (Bi-LSTM), and Gated Recurrent Units (GRUs) have gained prominence in authentication due to their ability to model complex temporal dependencies and sequential data. LSTM networks address the limitations of traditional RNNs by incorporating memory cells and gating mechanisms that enable them to capture long-term dependencies and mitigate the vanishing gradient problem [50]. Alternatively, Bi-LSTM extends the capabilities of LSTMs by processing sequential data in both forward and backward directions, allowing the network to capture contextual information from past and future states simultaneously. GRUs [51], on the other hand, offer a simplified yet powerful alternative to LSTMs by combining the forget and input gates into a single update gate and reducing the number of parameters. In recent research, Shen et al. [52] proposed batch-based authentication using LSTM to predict workflow. Alternatively, Transformer-Based Deep Neural Networks (TDNNs) have emerged as a cutting-edge approach for addressing the challenges of secure and efficient identity verification in VANETs by leveraging the self-attention mechanism to capture global dependencies within sequential data, enabling them to process long-range interactions and complex patterns more effectively [53].

In contrast, the deep supervised learning models are trained using labeled datasets to minimize loss functions, such as cross-entropy, enabling accurate and real-time authentication while mitigating threats such as spoofing and replay attacks. However, challenges such as the need for large labeled datasets, computational overhead, and vulnerability to adversarial attacks remain, prompting research into techniques such as transfer learning and ensemble methods to improve robustness and scalability. Deep supervised learning thus offers a promising solution for securing vehicular communication systems, ensuring the integrity and reliability of intelligent transportation networks. A critical analysis of strengths, weaknesses, and limitations of supervised learning in the context of vehicular authentication has been shown in Table 2.

From the perspective of deep learning for vehicular authentication, unsupervised deep learning plays a pivotal role in addressing the challenges of secure and efficient identity verification in ITS. Unlike supervised methods that require labeled datasets, unsupervised learning techniques leverage unlabeled data to discover hidden patterns, structures, and anomalies, making them highly suitable for real-world vehicular environments where labeled data may be scarce or costly to obtain. Techniques such as autoencoders, Restricted Boltzmann Machines (RBMs), Deep Belief Networks (DBNs), and Self-Organizing Maps (SOMs) enable the extraction of meaningful features and the detection of anomalies by learning low-dimensional representations of high-dimensional vehicular data, such as sensor readings, communication logs, or driving patterns. By leveraging unsupervised deep learning, vehicular authentication systems can achieve greater adaptability, scalability, and robustness, ensuring secure and reliable operation in dynamic and evolving connected and autonomous vehicle ecosystems.

Self-Organizing Maps (SOMs), a type of unsupervised neural network, excel in clustering and visualizing high-dimensional data by mapping it onto a low-dimensional space while preserving topological relationships [54]. In vehicular authentication, SOMs are utilized to identify patterns and anomalies in vehicle behaviour, sensor data, or communication logs, enabling the detection of unauthorized access or spoofing attempts by clustering normal and abnormal activities. Alternatively, Generative Adversarial Networks (GANs), which consist of a generator and a discriminator network engaged in a competitive learning process, are particularly effective in generating synthetic data and enhancing anomaly detection capabilities [55]. Subsequently, GAN can be represented by a generator G(z) which maps a random noise vector z∼pz(z) to a generated sample G(z) and a discriminator D(x) which is a classifier that outputs the probability that an input x is real from the true data distribution data(x) rather than fake from G(z) as per (1).(1)G:z→G(z),D:x→[0,1]where D(x) should be close to 1 for real data and 0 for generated (fake) data, and G(z) aims to fool D into predicting 1 for fake data.

Additionally, GANs can be employed to create realistic synthetic datasets for training robust authentication models [56], as well as to improve intrusion detection systems by learning the distribution of legitimate data and identifying deviations indicative of cyberattacks [42]. In their recent studies, Fei et al. [57] used a Deep Convolution Generative Adversarial Network (DCGAN) to utilize the pseudo-random number generator for entropy-stopping method-based training in vehicular networks. On the other hand, autoencoders (AEs) and their variants, such as Sparse Autoencoder (SAE), Denoising Autoencoder (DAE), Contractive Autoencoder (CAE), and Variational Autoencoder (VAE) provide powerful frameworks for feature extraction, anomaly detection, and data reconstruction, which are critical for ensuring secure and reliable vehicle identity verification in the domain of deep learning for vehicular authentication [39]. In this context, the architecture of AE-based vehicular authentication is shown in Fig. 7. Eventually, Sparse Autoencoders (SAEs) introduce sparsity constraints during training, encouraging the network to activate only a small subset of neurons, which enhances feature selection and improves the interpretability of learned representations. Hemavathi et al. [58], in their studies, used a deep stacked sparse autoencoder unsupervised algorithm for authentication in HetNet. On the other side, Denoising Autoencoders (DAEs) are trained to reconstruct clean data from corrupted or noisy inputs, making them robust to noise and perturbations in sensor data or communication logs. Likewise, Saponara et al. [59] utilized SAE for reconstructing fingerprints to use it for authentication. In 2024, Chen et al. [60] proposed a physical layer authentication by utilizing DAE to reduce noise and feature dimension from the vehicular data. On the other hand, Contractive Autoencoders (CAEs) add a regularization term to the loss function that penalizes sensitivity to small input variations, resulting in more robust and stable feature representations. Azri et al. [61] used CAE for capturing robust and effective features of user and item to build a temporal recommender system. On the other hand, Variational Autoencoders (VAEs) introduce a probabilistic approach by learning a distribution over the latent space, enabling the generation of new data samples and improving anomaly detection by modeling the likelihood of observed data. In 2023, Meng et al. [62] used the VAE model to improve the representational ability of Channel Impulse Responses (CIR) for an Industrial Internet of Things (IIoT) physical layer authentication. In 2024, Qiu et al. [63] proposed a hardware fingerprint authentication utilizing VAE on optical spectra and trained the model for feature extraction. In the same year, Li et al. [64] used VAE for data augmentation and data reduction for mobile user authentication. Recently, Wang et al. [65] proposed a CNN-based authentication for digital therapeutics and used VAE for data augmentation. In the other aspects, Restricted Boltzmann Machines (RBMs) excel in learning probabilistic representations of input data by modeling the joint distribution between visible and hidden layers as generative stochastic neural networks [66,67]. Furthermore, multiple RBMs stack and construct Deep Belief Networks (DBNs) [68], which further enhance feature extraction by learning hierarchical representations of data. Eventually, DBNs can capture complex, non-linear relationships in vehicular data, enabling more accurate detection of anomalies. With this note, Althubiti [69] proposed a trust-aware authentication scheme protocol for Wireless Sensor Networks (WSNs) using DBNs to select threshold trust values dynamically. A critical analysis of unsupervised deep learning in the context of vehicular authentication is shown in Table 3.

DRL, a novel and adaptive approach to enhancing security and decision-making in ITS. DRL combines the representational power of deep neural networks with the decision-making capabilities of reinforcement learning, enabling systems to learn optimal policies through interaction with their environment. In vehicular authentication, DRL can be employed to dynamically adapt authentication mechanisms based on real-time data, such as vehicle behavior, communication patterns, or environmental conditions. For example, a DRL agent can learn to detect and respond to evolving cyber threats, such as spoofing or intrusion attempts, by continuously optimizing its actions to maximize security while minimizing false positives.

Furthermore, DRL can be categorized as model-based and model-free deep reinforcement learning. Model-based DRL relies on learning an explicit model of the environment, which simulates state transitions and rewards, enabling the agent to plan and optimize actions efficiently. In vehicular authentication, this approach can be used to predict potential cyber threats, such as spoofing or intrusion attempts, by modeling the behavior of malicious actors and proactively adapting authentication protocols [70]. With this note, the Markov Decision Process (MDP), a model-based DRL, serves as a fundamental framework in modeling decision-making problems under uncertainty, making it particularly relevant in the context of deep learning-based vehicular authentication. On the other hand, in intelligent transportation systems, vehicular authentication must dynamically adapt to evolving network conditions, adversarial threats, and varying authentication costs [71]. By formulating the authentication process as an MDP, the system can optimize security decisions based on states representing vehicle credentials, trust scores, and environmental factors [72]. Alternatively, Deep Transfer Learning (DTL) plays a crucial role in enhancing the efficiency and adaptability of deep learning-based vehicular authentication by leveraging knowledge learned from related domains to improve authentication performance in dynamic vehicular environments. DTL enables the reuse of pre-trained models, allowing authentication systems to transfer learned features and patterns from previously seen vehicular contexts to new but related authentication tasks [73]. This approach not only reduces training time and data dependency but also improves the model’s generalization ability across different vehicular scenarios. Given the high mobility and real-time constraints of vehicular networks, authentication mechanisms must adapt to continuously changing conditions while ensuring security and minimal latency [74]. Dynamic Programming facilitates optimal policy selection by breaking down the authentication process into subproblems, solving them recursively, and leveraging stored solutions to avoid redundant computations [75]. Dynamic Programming relies on defining a problem recursively [76]. If f(n) represents the optimal solution for a problem of size n, it can often be expressed in terms of smaller subproblems as per (2).(2)f(n)=a∈Amin{g(n,a)+f(T(n,a))}where V(s) is the value of being in a state s, R(s,a) is the immediate reward of taking action a, P(s′∣s,a) is the probability of transitioning to a state s′, γ is a discount factor (0≤γ≤1).

On the other hand, Value iteration leverages Bellman equations to iteratively update authentication value functions iteratively, converging toward an optimal security strategy [30]. In contrast, policy iteration alternates between policy evaluation and improvement to enhance authentication decisions [77]. The Bellman equation serves as the foundation for these methods, providing a recursive framework to evaluate authentication state transitions based on security risks, trust scores, and latency constraints [78]. By integrating these approaches with deep reinforcement learning, vehicular authentication systems can dynamically adapt to evolving cyber threats and optimize authentication policies, ensuring both security and efficiency in intelligent transportation networks.

On the other hand, model-free DRL directly learns optimal policies or value functions without explicitly modeling the environment, making it highly flexible and suitable for dynamic and complex vehicular networks. For instance, a model-free DRL agent can learn to detect anomalies in real-time communication patterns or driving behaviors by interacting with the environment and refining its decision-making process through trial and error. On the same note, Deep Q-learning (DQL), a model-free DRL, emerged as a powerful technique for vehicular authentication. DQL combines the strengths of Q-learning, a model-free reinforcement learning algorithm, with deep neural networks to approximate the Q-value function, which estimates the expected utility of actions in a given state [79]. In vehicular authentication, DQL can be employed to develop adaptive and intelligent systems capable of detecting and responding to cyber threats. Roy et al. [80] proposed a secure healthcare model utilizing DQL and DNN. On the other hand, Monte Carlo Control (MCC) plays a significant role in optimizing deep learning-based vehicular authentication by enabling model-free reinforcement learning in dynamic vehicular networks [81]. Given the unpredictability of vehicular environments, where authentication requests, network conditions, and security threats continuously evolve, MCC provides an effective approach to learning optimal authentication policies through experience. Unlike dynamic programming methods that require complete knowledge of the environment, MCC estimates value functions based on sampled authentication interactions, allowing the system to improve authentication strategies over time [82]. Alternatively, the Actor-criticism method combines the advantages of both policy-based and value-based approaches, where the actor learns an optimal authentication policy by interacting with the environment [83]. At the same time, the critic evaluates the policy using value functions. This dual-network structure accelerates learning and improves stability, allowing the authentication system to quickly adapt to changing vehicular behaviors, network conditions, and adversarial threats [83]. In the other context, Policy Gradient (PG) approaches parameterize the authentication policy and adjust it iteratively using gradients of expected rewards [84]. This allows the authentication mechanism to handle complex, high-dimensional vehicular environments where traditional rule-based or heuristic methods fail. On the same note, Jiu et al. [85] proposed an authentication scheme for an unknown network using a deep deterministic policy gradient.

In contrast, DRL can facilitate the development of adaptive authentication protocols that adjust to the dynamic nature of vehicular networks, ensuring robust performance in diverse and unpredictable scenarios. By leveraging its ability to learn from experience and improve over time, deep reinforcement learning provides a powerful framework for enhancing the security, resilience, and efficiency of vehicular authentication systems, contributing to the safety and reliability of connected and autonomous vehicle ecosystems. The critical analysis of deep reinforcement learning in the context of vehicular authentication is shown in Table 4.

Hybrid learning, which integrates multiple learning paradigms, plays a vital role in enhancing deep learning-based vehicular authentication by improving adaptability, efficiency, and security in dynamic vehicular networks. Given the challenges of high mobility, evolving cyber threats, and latency constraints, a hybrid learning approach combines supervised, unsupervised, and reinforcement learning techniques to optimize authentication strategies. Supervised learning helps recognize known authentication patterns, while unsupervised learning detects anomalies and potential threats in real time. Subsequently, reinforcement learning enables adaptive decision-making by continuously refining authentication policies based on environmental interactions. By leveraging hybrid learning, vehicular authentication systems can achieve robust, context-aware security, minimizing authentication delays and improving resistance against adversarial attacks, making them well-suited for intelligent transportation systems. Hybrid learning, such as deep ensemble learning and Adaptive Deep Neural Networks (ADNN), play a crucial role in securing vehicular communications by providing authentication facilities. Recently, Pan et al. [86] proposed a vehicle license plate detection and recognition model using hybrid DL algorithms. The model is further extended to combine the CNN-based You Only Look Once (YOLO) algorithm and Convolutional Recurrent Neural Network (CRNN) for license-plate character recognition. On the other hand, SSD-MobileNet is a popular deep learning architecture for real-time object detection, combining the efficiency of the MobileNet convolutional neural network with the Single Shot MultiBox Detector (SSD) algorithm, allowing for fast and accurate identification of multiple objects within an image, making it ideal for applications where low latency and high throughput are crucial, like mobile devices and surveillance systems [87].

On the other hand, Deep ensemble learning [32] has gained significant attention as a robust methodology for addressing the challenges of vehicular authentication in dynamic and security-sensitive environments. By aggregating the predictions of multiple deep learning models, this approach mitigates the limitations of single-model systems, which are often prone to overfitting, sensitivity to noisy data, and inadequate generalization in heterogeneous vehicular networks. The ensemble framework typically incorporates diverse architectures, such as CNNs for spatial feature extraction, RNNs for capturing temporal dependencies, and transformers for handling sequential data to enhance authentication accuracy and resilience collectively. Additionally, deep ensemble learning provides uncertainty estimates, enabling risk-aware decision-making in real-time authentication scenarios, which is critical for mitigating sophisticated threats like spoofing and replay attacks.

In deep ensemble learning [88], authentication can be framed as a binary classification problem, where the goal is to predict whether an input x belongs to the legitimate class (y=1) or the malicious class y=0. Let f1,f2,f3,…,fm be M deep learning models, each trained to predict the probability of legitimacy as per (3). (3)Pi(y=1∣x)=fi(x),where Pi(y=1∣x) is the probability of x being legitimate, as predicted by the model fi.

In the same context, Song et al. [89] proposed two-layer security on the authentication layer and ensemble learning-based monitoring layer.

Recent research related to authentication in vehicular communication has been successful in using face detection by DCNN, which has proven to be a significant result. However, Du et al. [90] outperformed DCNN-based user authentication by incorporating the PelFace model in parallel deep ensemble learning. On the other hand, ADNN has the power to adapt the new features by adjusting the parameters based on new data. A standard DNN with L layers is represented as (4).(4)h(l)=σ(W(l)h(l−1)+b(l))where h(l) is the activation at layer l, W(l) is the weight matrix, b(l) is the bias vector, and σ(⋅) is the activation function. However, instead of using all layers, an ADNN selects only a subset S of layers to compute based on an adaptive gating function G(x) as per (5).(5)S={l|G(h(l))>τwhere, G(h(l)) determines layer importance, τ is a threshold controlling adaptivity.

Instead of fixed weights, ADNNs update weights based on the input dynamically as per (6).(6)W(l)(x)=f(W(l),x)where f(⋅) is an adaptive function such as an attention mechanism or meta-learning.

Recently, Jia et al. [91] used ADNN to perform node authenticity and analyzed the trust score to minimize the attack in the VANET system. This approach not only enhances security but also ensures adaptability to the dynamic and evolving nature of vehicular networks, making it a promising direction for future research and deployment. On the other hand, Zhang et al. [92] proposed a user identification method by extracting the user’s gait information using a convolution kernel and applying ANN to authenticate.

On the other hand, merging two or three DL models together to create individual safeguards is potentially beneficial for VANET authentication, especially for offering enhanced adaptability, robustness, and accuracy by combining the strengths of multiple learning paradigms. For instance, Inzillo et al. [93] combined CNNs for spatial feature extraction with LSTM networks for temporal pattern recognition in vehicle movement data. Alternatively, Chougule et al. [94] proposed CNN-LSTM to bolster the in-vehicle network security. In the first stage of the proposed model, LSTM is used to detect weather a communication is an attack or not and in the second stage the category of the attaches are judged by using CNN. On the other hand, Khan et al. [95] proposed a hybrid intrusion detection system combining CNN, LSTM networks, and DBN with feature selection techniques such as Random Projection (RP) and Principal Component Analysis (PCA). This framework achieved a detection accuracy of 99.4% for DoS and DDoS attacks, surpassing traditional machine learning models. Recently, Minu et al. [96] proposed an authentication framework for vehicular network using hybrid approaches. ADBN is used to enhance the reliability of the network messages and Hybrid Attribute-Based Advanced Encryption Standard (HABAES) encryption techniques used for secure communication. In another research, Eman et al. [97] combined deep-learning-based mask detection, landmark and oval face detection for key features, and Robust Principal Component Analysis (RPCA) to separate occluded and non-occluded image parts. Particle Swarm Optimization (PSO) is used to optimize k-nearest neighbors (KNN) features and the number of ‘k’ for improved performance. Experimental results show the proposed method achieves a 97% recognition rate, significantly outperforming existing methods in accuracy and robustness to occlusion.

In a nutshell, deep learning has revolutionized vehicular authentication by providing adaptive, efficient, and highly secure mechanisms for identity verification in intelligent transportation systems. Unlike traditional authentication methods, deep learning enables real-time decision-making, anomaly detection, and dynamic adaptation to evolving cyber threats. Techniques such as deep reinforcement learning, transfer learning, and hybrid learning enhance authentication resilience by leveraging past experiences and optimizing authentication strategies under varying network conditions [22]. A comprehensive analysis of deep learning is shown in Table 5.

Deep learning has emerged as a transformative technology in various domains, including vehicular authentication. Its ability to learn complex patterns from large datasets makes it a promising solution for enhancing the security and efficiency of vehicular systems. However, the deployment of deep learning in vehicular authentication is not without significant challenges and open issues. This section delves into the key challenges and unresolved problems that must be addressed to ensure the reliable and secure implementation of deep learning-based authentication systems in vehicles.

As the automotive industry continues to evolve toward connected and autonomous vehicles, the role of deep learning in vehicular authentication is expected to grow significantly. While current research has demonstrated the potential of deep learning for enhancing security and user experience, several future directions can further advance the field. These directions aim to address existing challenges, leverage emerging technologies, and explore novel applications of deep learning in vehicular authentication. This section outlines key areas of focus for future research and development.