Next Article in Journal
Evolutionary Optimisation of Runge–Kutta Methods for Oscillatory Problems
Previous Article in Journal
Prediction of Mental Fatigue for Control Room Operators: Innovative Data Processing and Multi-Model Evaluation
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Mathematics
Volume 13
Issue 17
10.3390/math13172795
mathematics-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Review

Generative AI for the Internet of Vehicles: A Review of Advances in Training, Decision-Making, and Security

by
Xiaoming Yuan
1,*,
Xinling Zhang
1,
Aiwen Wang
1,
Jiaxin Zhou
1,†,
Yingying Du
1,†,
Qingxu Deng
1 and
Lei Liu
2
1
Hebei Key Laboratory of Marine Perception Network and Data Processing, Northeastern University at Qinhuangdao, Qinhuangdao 066004, China
2
Xidian Guangzhou Institute of Technology, Xidian University, Guangzhou 510555, China
*
Author to whom correspondence should be addressed.
†
These authors contributed equally to this work.
Mathematics 2025, 13(17), 2795; https://doi.org/10.3390/math13172795
Submission received: 23 July 2025 / Revised: 13 August 2025 / Accepted: 26 August 2025 / Published: 31 August 2025
Browse Figures
Versions Notes

Abstract

Generative artificial intelligence (GAI) has emerged as a transformative force in the Internet of Vehicles (IoV), addressing limitations of traditional AI such as reliance on large labeled datasets and narrow task applicability. This paper aims to systematically review recent advances in applying GAI to the IoV, with a focus on training, decision-making, and security. We begin by introducing the fundamental concepts of vehicular networks and GAI, laying the groundwork for readers to better understand the subsequent sections. Methodologically, we adopt a structured literature review, covering developments in synthetic data generation, dynamic scene reconstruction, traffic flow prediction, anomaly detection, communication management, and resource allocation. In particular, we integrate multimodal GAI capabilities with 5G/6G-enabled edge computing to support low-latency, reliable, and adaptive vehicular network services. Our synthesis identifies key technical challenges, including lightweight model deployment, privacy preservation, and security assurance, and outlines promising future research directions. This review provides a comprehensive reference for advancing intelligent IoV systems through GAI.

1. Introduction

Intelligent Internet of Vehicles (IoV) is a frontier field that integrates artificial intelligence, communication technology, and transportation systems. Increasing academic interest has been observed in this area in recent studies [1]. This technological fusion architecture is driving the transportation industry to accelerate its evolution from “single-device interconnection” to “full-domain intelligent collaboration”. Furthermore, with the advancements in vehicle-mounted sensors, 5/6G communication technologies, and edge computing infrastructures, the IoV has transcended its traditional role of merely providing entertainment services and navigation. By leveraging the ultra-low latency and high bandwidth of 5G/6G together with the proximity processing capabilities of edge computing, GAI models can perform real-time data generation, predictive scene modeling, and dynamic resource scheduling, thereby enhancing IoV responsiveness in latency-critical scenarios such as cooperative driving and collision avoidance. The IoV has become a complex ecosystem covering multiple scenarios, such as vehicle–road collaborative decision-making, intelligent traffic flow prediction, and autonomous driving safety assurance [2,3]. Moreover, it can deeply integrate into traffic management, smart city services, and people’s livelihood application systems, bringing more convenience to people.
This transformation is closely linked to recent breakthroughs in artificial intelligence (AI), especially the emergence of generative AI (GAI). GAI acts as a foundational component within the broader framework of artificial intelligence-generated content (AIGC) [4]. GAI demonstrates the capability to learn complex data distributions and generate new data. This remarkable feature has enabled it to exhibit exceptional competence in generating coherent texts, images, and audio and creating semantically rich multimodal content. Different from traditional discriminative artificial intelligence, which focuses on classification or regression tasks, generative models [5] are capable not only of learning the underlying data distribution but also of producing novel and coherent outputs to reinforce the learning process.
The application of GAI in vehicular networks is becoming a transformative force in advancing modern transportation ecosystems. Table 1 organizes existing works into six thematic groups, arranged chronologically within each category, to better capture the evolution of research priorities. GAI overviews for IoV provide the conceptual and technical foundations for subsequent studies. Early efforts [6,7,8,9] outlined how AI supports autonomous perception, localization, mapping, traffic analysis, and secure VANET operations, offering comprehensive roadmaps for integrating AI into vehicular contexts. In terms of GAI for training and data synthesis, refs. [10,11,12] investigated the role of generative models such as GANs and GAI in enhancing training efficiency through synthetic data generation, traffic scene simulations, and multimodal user experience design, thereby reducing dependence on costly labeled datasets. GAI for vehicular communications addressed communication optimization problems, including handover, beam selection, and bandwidth allocation. Christopoulou et al. [13] extend this focus to V2X communications, exploring how AI/ML optimizes real-time operations like handover management and resource allocation, while addressing challenges like computational complexity. Similarly, in enhancing vehicular networks with GAI, Teef et al. [14] explored its applications in traffic optimization, communication enhancement, and security, noting challenges like scalability and privacy. GAI for traffic prediction and ITS optimization [15,16] highlighted the synergy between AI algorithms, big data analytics, and emerging GAI paradigms such as the mixture of experts (MoE) and multimodality, enabling more accurate traffic flow prediction, vehicle recognition, and route optimization in the ITS.
In terms of GAI-enabled edge intelligence and resource management, refs. [17,18,19,20,21] examined the convergence of GAI and edge intelligence in IoV systems. These works propose architectures and strategies for optimizing computation, communication, and resource allocation at the network edge. A representative case is the work by Zhang et al. [21], who introduced a multimodality semantic-aware GAI–IoV framework augmented by deep reinforcement learning (DRL) for resource allocation. In the simulation, their DDQN-based approach achieved an average reward 30% higher than random and 15% higher than greedy within the same number of training steps, converging 40% faster than DQN. In testing, when the image payload size reached 8 MB, the proposed method’s quality of experience (QoE) was 25% higher than the second-best benchmark. It demonstrating the tangible optimization benefits of GAI-assisted decision-making. In terms of GAI for IoV security and privacy, refs. [22,23,24,25] explored the application of AI and GAI in defending vehicular networks against cyber-attacks. These studies covered intrusion detection, threat simulations, blockchain-enhanced access control, and privacy preservation, underscoring the importance of proactive, adaptive defense strategies in future IoV deployments.

Motivations and Contributions

As the above analysis reveals, most existing reviews primarily focus on the specific applications of GAI in the IoV, often overlooking the underlying technologies themselves. Current studies typically approach the topic from application-driven, system-level, or transportation scenario perspectives. In practical deployments, GAI supports diverse functions including data augmentation, predictive maintenance, and anomaly detection, making it applicable across multiple use cases and environments. However, organizing these discussions solely by scenario may lead to repetitive or fragmented descriptions of the same technological principles. In contrast, this survey centers on the enabling role of GAI in supporting IoV systems, offering a unified perspective that connects diverse viewpoints within a coherent framework. To consolidate scattered research efforts, this work reviews recent advancements in GAI and synthesizes its technical and application-level progress across multiple domains of vehicular networking. The notable contributions of our work are as follows.
  • We provide the first systematic taxonomy of generative AI applications in IoV systems, categorizing them into three primary domains: training enhancement, decision-making support, and network management.
  • We analyzes how generative AI models such as GANs, diffusion models, and LLMs can be adapted to different IoV scenarios, bridging the gap between AI research and vehicular technologies.
  • We provide an in-depth analysis of security implications when integrating generative AI into safety-critical IoV applications, offering novel perspectives on trustworthy AI deployment.
  • A clear roadmap is presented to guide future exploration in lightweight model deployment, real-time adaptive generation, and secure integration with next-generation networks such as 6G.
The remainder of this paper is organized as follows: We start with an overview of GAI in Section 2, where we introduce core GAI technologies and IoV system architectures. Section 3 examines how GAI enhances IoV training through synthetic data generation and traffic scene simulation. Section 4 investigates GAI applications in decision-making processes, including traffic prediction and autonomous driving decisions. Section 5 addresses GAI’s role in IoV management and security, covering network protection, resource allocation, and others. Section 6 discusses current challenges and future research directions. Finally, Section 7 concludes the survey with key findings and research implications.

2. Foundations of Generative Artificial Intelligence and the Internet of Vehicles

This part outlines the historical evolution and core architectures of GAI, along with key techniques in multimodal learning and fine-tuning. The main concepts and representative models are summarized in Figure 1.

2.1. History of Generative AI Models

As shown in the lower part of Figure 1, early progress in machine learning and deep learning laid the groundwork for the emergence of GAI. The foundations of GAI were established in the 1950s with several pioneering contributions. In 1952, Arthur Samuel developed the first machine learning algorithm for playing checkers and coined the term “machine learning.” This was followed by Frank Rosenblatt’s creation of the perceptron in 1957, which represented the first trainable neural network capable of basic pattern recognition. The 1960s witnessed early applications of generative concepts with Joseph Weizenbaum’s introduction of ELIZA in 1961. This chatbot, designed to simulate a psychotherapist, is considered one of the earliest forms of GAI. The 1970s brought significant theoretical advances when Seppo Linnainmaa introduced backpropagation for training deep neural networks, and Kunihiko Fukushima developed the cognitron in 1975 for handwritten character recognition.
The field experienced its first AI winter from 1973 to 1979, which separated machine learning from AI due to unmet expectations and reduced funding. Despite this setback, machine learning continued to evolve for commercial applications. The 1980s faced another AI winter from 1984 to 1990, marked by slow progress, though key developments persisted. By 1989, Yann LeCun’s team made deep learning functional for handwritten postal code recognition. The 1990s marked a turning point with the emergence of 3D graphics cards, which served as precursors to modern GPUs and enhanced computational capabilities. In 1998, LeCun’s team developed LeNet-5, the first deep learning convolutional neural network designed for recognizing handwritten digits in checks.
The computational landscape was further transformed in 2007 when NVIDIA released CUDA, enabling GPU acceleration for general-purpose computing. The modern era of GAI began with several breakthrough developments. In 2014, Ian Goodfellow published his seminal paper on Generative Adversarial Networks [26], while Geoffrey Hinton’s team achieved significant breakthroughs in image classification, enabling realistic data generation. The introduction of Transformer models [27] in 2017 revolutionized natural language processing and generative modeling capabilities.
Recent years have witnessed rapid advancements in GAI applications. OpenAI released DALL-E in 2021, advancing text-to-image generation capabilities. The public release of ChatGPT in 2022 revolutionized text-based tasks and brought GAI into mainstream attention. GPT-4’s release in 2023 represented a significant leap in capabilities, followed by Anthropic’s release of Claude 3 and Claude 3.5 in 2024, further advancing the field’s sophistication and practical applications.

2.2. Overview of Generative AI Models

Generative models, unlike traditional models that merely perform recognition or classification tasks, are designed to simulate, reconstruct, or create new data instances. In the subsequent development, many different variants evolved due to different applications. As shown in the top-left corner of Figure 1, the basic architectures of many evolved variants include a (1) VAE, a (2) GAN, a (3) Transformer, and a (4) diffusion model.

2.2.1. VAE

Variational Auto-Encoders: VAEs function as probabilistic models that encode data into a latent space structured by a predefined prior. The model is composed of two neural networks. One is an encoder, which maps input data to a distribution of latent variables. The other is a decoder, which reconstructs the original data from samples extracted from this latent distribution. VAEs optimize the evidence lower bound (ELBO) to balance the trade-off between reconstruction fidelity and the regularization of the latent space [28].
In IoV scenarios, VAEs have been applied to reconstruct missing or noisy sensor data [29,30], enabling more reliable perception in situations where certain inputs are temporarily unavailable. Additionally, VAEs can be employed for trajectory prediction tasks, thereby further enhancing driving safety [31,32].

2.2.2. GAN

Generative Adversarial Networks (GANs) are composed of two networks operating within a game-theoretic framework. One network, the generator, produces synthetic data, while the other, the discriminator, attempts to distinguish real data from the generated samples. These two networks are trained simultaneously, with the generator improving its ability to deceive the discriminator over time [33].
GANs are capable of generating high-fidelity samples, making them highly valuable in automotive vision applications [34,35]. Their ability to produce high-quality synthetic data proves particularly useful in scenarios where real-world data is difficult to collect. For example, GANs have been applied in hazardous driving conditions such as close-range collisions [36,37], nighttime obstacles [38], and extreme weather conditions [39]. These synthetic datasets are crucial for enhancing safety training databases for autonomous vehicles. Additionally, GANs are effective in image enhancement, allowing them to improve low-quality input data from vehicle-mounted sensors [40,41,42].

2.2.3. Transformer

Transformer-based models, especially those like GPT, employ self-attention to capture dependencies across distant elements in data sequences. These models adopt an autoregressive training strategy, where each token is estimated using the prior context within the sequence. This methodology facilitates the production of fluent and semantically consistent content [43,44].
In the context of the Internet of Vehicles (IoV), Transformer models are increasingly being explored for multimodal reasoning [45] and semantic understanding [46]. For example, they can translate raw sensor data into natural language descriptions of traffic scenes, assist in developing in-vehicle conversational agents, and support cross-agent communication in cooperative driving scenarios [47,48,49]. Their scalability and versatility make them ideal for integrating structured traffic knowledge, the environmental context, and user intent into decision-making processes.

2.2.4. Diffusion Model

Diffusion models represent a recent and powerful class of generative frameworks relying on iterative denoising processes [50]. These models learn to reverse a gradual noising process that transforms data into pure noise, effectively modeling the data distribution through a series of latent variables. By mastering this reverse process, diffusion models are capable of producing highly realistic and diverse samples, often outperforming other generative models in terms of quality and stability [51].
In Internet of Vehicles (IoV) systems, diffusion models are particularly well-suited for generating complex visual scenes with fine-grained details [52,53], such as urban traffic simulations or sensor fusion reconstructions [54]. Their robustness and mode coverage make them highly attractive for safety-critical applications, where diversity and realism in generated samples are essential. Additionally, diffusion models can be used to enhance degraded inputs, such as recovering visibility in low-light or foggy conditions, directly contributing to improved situational awareness in autonomous driving [55,56].
These generative architectures offer complementary strengths that align well with the heterogeneous and dynamic needs of intelligent vehicular systems. Variational autoencoders (VAEs) provide structured latent representations for uncertainty-aware modeling. GANs excel at visual realism and data augmentation. Transformers offer cross-modal flexibility and semantic control, and diffusion models deliver high-quality generation under complex data distributions. Within the broader IoV development lifecycle, architecture selection and customization must account for limited computing resources, stringent latency demands, and diverse deployment scenarios.

2.3. Multimodal Learning and Large Model Fine-Tuning

The integration of GAI into Internet of Vehicles systems necessitates sophisticated approaches to handle diverse data modalities and adapt pretrained models to specific vehicular applications [57].
Modern IoV systems generate heterogeneous data streams from various sensors, including visual cameras, LiDAR, radar, GPS, and communication transceivers. GAI models must effectively process and synthesize information across these multiple modalities to provide comprehensive scene understanding and decision support [48,58]. Contemporary multimodal generative models have demonstrated remarkable capabilities in cross-modal understanding and generation. Vision-language models such as CLIP and BLIP have established strong foundations for aligning visual and textual representations, enabling models to understand relationships between images and natural language descriptions. As shown in the top-right corner of Figure 1, CLIP’s architecture exemplifies how visual and textual encoders are jointly trained to achieve this alignment [59]. Recent advances in multimodal Transformers have further enhanced the integration capabilities across different data types. Models like GPT-4V and Flamingo demonstrate the ability to process and generate content across multiple modalities simultaneously, providing more comprehensive understanding of complex vehicular environments.
In practical IoV applications, such multimodal capabilities significantly enhance decision-making processes. For example, in autonomous driving, a GAI model can fuse high-resolution camera images with LiDAR point clouds and vehicle telemetry (e.g., speed, acceleration, and steering angle) to generate a unified environmental representation. This representation can be further enriched by integrating in-cabin audio signals (such as driver vocal commands or abnormal mechanical sounds) to identify potential hazards or adjust driving strategies in real time. Such multimodal synthesis not only improves object detection and trajectory prediction but also supports context-aware decision-making under complex traffic conditions [60,61].
While large pretrained generative models possess impressive general capabilities, their deployment in IoV systems requires careful adaptation to specific vehicular tasks and constraints [62,63]. Fine-tuning strategies play a crucial role in customizing these models for optimal performance in automotive environments. Parameter-efficient fine-tuning methods have gained prominence due to their ability to adapt large models without extensive computational overhead. Techniques such as Low-Rank Adaptation (LoRA) and prefix-tuning enable targeted modification of model behavior while preserving the majority of pre-trained parameters. These approaches are particularly valuable in resource-constrained vehicular environments where full model retraining is impractical.

3. Generative AI for Training in Vehicle Network Systems

The deployment of reliable and robust vehicle networking systems requires extensive training on diverse and comprehensive datasets. However, real-world data collection in automotive environments presents significant challenges, including safety risks, cost constraints, and the rarity of critical scenarios. GAI has emerged as a transformative solution for addressing these challenges by creating synthetic training data, reconstructing complex scenarios, and augmenting existing datasets. This section examines how GAI technologies enhance the training processes for vehicle networking systems across multiple dimensions.

3.1. Synthetic Data Generation for Traffic and Sensor Inputs

The generation of synthetic data for vehicle networking systems addresses fundamental challenges in automotive AI development. Traditional data collection methods face significant limitations in capturing the full spectrum of driving scenarios, particularly those involving dangerous or rare situations that are crucial for comprehensive system safety validation.
Generative AI models excel at creating realistic traffic scenarios that encompass diverse driving conditions and vehicle interactions [64]. Generative Adversarial Networks have demonstrated remarkable capabilities in synthesizing traffic flow patterns that maintain statistical consistency with real-world observations. These models can generate traffic scenarios across various densities while preserving the complex behavioral patterns exhibited by human drivers. Beyond traffic scenario generation, the synthesis of sensor data represents another critical application of GAI in vehicle networking systems [65]. Different sensor modalities require specialized generation approaches to maintain both physical realism and measurement accuracy. For visual sensor applications, diffusion models serve as effective generative frameworks capable of simulating diverse environmental image conditions with high fidelity.
For example, Tamayo-Urgilés et al. [66] generated synthetic vehicle driving event data via TimeGAN and RTSGAN, evaluated it with PCA/T-SNE, and used it for accident risk classification, proving its training efficacy comparable to real data. Similarly, Cao & Ramezani [67] utilized the CARLA simulator and cGAN to create realistic synthetic data, proposing a multi-level learning framework to enhance autonomous driving perception in harsh conditions. Song et al. [68] systematically reviewed autonomous driving synthetic datasets, categorized them into single/multi-task types, discussed their roles in evaluation and trustworthiness, and outlined future directions.
GAN-based frameworks (e.g., cGAN, TimeGAN, and RTSGAN) efficiently generate statistically consistent traffic and temporal event data but risk mode collapse and often miss rare edge cases. Diffusion models produce more diverse, high-fidelity sensor imagery under varied conditions, yet their heavy computational demands limit real-time or edge use. In practical IoV settings, GANs are better suited for fast, iterative traffic pattern generation in centralized training pipelines, while diffusion models are preferable for safety-critical perception training where fidelity outweighs generation speed. A combined strategy that employs GANs for large-scale data synthesis and leverages diffusion models to refine rare or safety-critical scenarios offers a promising balance between scalability and quality.

3.2. Scene Reconstruction for Edge Cases and Rare Events

The reconstruction and generation of rare events represents one of the most critical applications of GAI in vehicle networking systems. These scenarios, while infrequent in normal operation, are essential for comprehensive system validation and safety assurance.
GAI models can synthesize safety-critical scenarios [69] that are too dangerous or impractical to collect from real-world driving. These include near-collision events, sudden obstacle appearances, and emergency braking situations. The models learn from limited real-world examples and expert knowledge to generate plausible variations of these critical scenarios. Advanced generative models can create scenarios with controlled risk levels, allowing for systematic testing of vehicle safety systems [70]. By varying parameters such as vehicle speeds, reaction times, and environmental conditions, these models MDPIcan generate comprehensive test suites that cover the full range of possible critical situations. Additionally, the integration of digital twin technology with GAI introduces a powerful approach for high-fidelity scenario simulation in IoV environments. A digital twin serves as a real-time virtual replica of physical vehicular systems, road infrastructure, and dynamic environments [71]. When coupled with generative models, this enables the simulation of realistic and adaptive scenarios that respond to system behaviors and contextual feedback.
Similarly, weather-related scenarios pose significant challenges for vehicle networking systems, yet collecting comprehensive data under all weather conditions is often impractical. GAI models can synthesize various weather conditions and their effects on sensor performance and vehicle dynamics [39]. For example, Singh et al. [72] proposed a lightweight visibility restoration network (LVR-Net) that employs modified adaptive distributed differential evolution (MADE) to optimize the Deep Multi-Patch Hierarchical Network (DMPHN), enhancing image restoration in adverse weather for autonomous vehicles. Similarly, in article [55], Li et al. introduced LightDiff, a multi-condition diffusion framework that leverages depth maps, text captions, and reinforcement learning to enhance low-light images for autonomous driving.

3.3. Data Augmentation for Perception and Learning Models

Data augmentation through GAI extends beyond simple transformations to create semantically meaningful variations that improve model robustness and generalization capabilities [73]. A key limitation in automotive applications is the lack of diverse training data, which this method effectively mitigates. Traditional data augmentation techniques such as rotation, scaling, and color adjustment provide limited semantic diversity. GAI enables semantic augmentation that creates meaningful variations in scene content while preserving label consistency.
For example, Baresi et al. [74] proposed an efficient domain augmentation method for autonomous driving testing using diffusion models, evaluating strategies like instruction editing and inpainting with refinement to expand operational design domain (ODD) coverage. Similarly, in article [75], Zheng et al. addressed robust perception under adverse conditions by integrating unpaired image-to-image translation for data augmentation and designing a two-branch architecture to fuse the original and enhanced images. Yu et al. [76] utilized the MUNIT framework for labeled image augmentation, generating synthetic data for drivable area detection and object recognition. They also developed an auto-labeling tool to create a closed-loop pipeline, demonstrating enhanced CNN performance with augmented datasets.

3.4. Summary of Generative AI for Training

Generative AI methods for IoV training each serve distinct roles. GANs offer fast, large-scale scenario generation but may miss rare edge cases. Diffusion models excel in high-fidelity, diverse outputs for safety-critical perception, though their heavy computation limits deployment. Digital twin-based approaches provide controllable, adaptive simulations but require complex infrastructure. In practice, GANs suit bulk traffic synthesis, while diffusion models target precision-critical vision tasks, and digital twins enable dynamic system-level testing. Hybrid strategies combining scalability and fidelity, alongside improved rare-event modeling and reduced sim-to-real gaps, present promising directions for future research.

4. Generative AI for Enhancing Decision-Making in Vehicle Network Systems

The complexity of modern vehicle network systems demands sophisticated decision-making capabilities that can process vast amounts of real-time data, predict future scenarios, and generate optimal responses under dynamic conditions. GAI is increasingly recognized for its capacity to support decision-making through predictive modeling, policy generation, anomaly identification, and reasoning in complex traffic environments. This section examines how GAI technologies contribute to more intelligent and adaptive decision-making in vehicle networking systems.

4.1. Traffic Flow Prediction

Traffic flow prediction represents a fundamental challenge in vehicle network systems, requiring accurate forecasting of vehicle movements, congestion patterns, and system-wide traffic dynamics. GAI models offer sophisticated approaches for modeling and predicting these complex spatiotemporal phenomena [77,78]. Generative models excel at capturing the intricate spatiotemporal dependencies inherent in traffic flow patterns. Graph neural networks combined with generative architectures can model traffic flow across road networks, where nodes represent intersections or road segments and edges capture the flow relationships between them. These models learn to generate realistic traffic patterns that preserve both spatial correlations between connected road segments and temporal dependencies across different time periods. Additionally, vehicle network systems require real-time traffic predictions that can adapt quickly to changing conditions.
For example, Katariya et al. [79] proposed DeepTrack, a lightweight deep learning model using temporal convolutional networks (TCNs) for real-time vehicle trajectory prediction in highways. In article [80], Sun et al. developed an online traffic flow prediction model by combining BiLSTM and a CNN within a Generative Adversarial Network (GAN) framework. Using multi-source data and a rolling time-domain scheme, it outperforms ARIMA and BiLSTM in peak-hour predictions, especially with signal timing integration. Additionally, Meng et al. [81] introduced NetGPT, the first generative pretrained Transformer for network traffic. It addresses heterogeneous headers and payloads via multi-pattern modeling, header field shuffling, and packet segmentation, excelling in traffic understanding and generation tasks across encrypted, DNS, and industrial protocol datasets.

4.2. Decision Policy Generation for Autonomous Vehicles

The generation of decision policies for autonomous vehicles represents one of the most critical applications of GAI in vehicle networking systems [82,83]. These policies must balance safety, efficiency, and passenger comfort while adapting to diverse and dynamic driving environments [84].
Generative models can synthesize driving policies that exhibit human-like behaviors while maintaining safety and efficiency objectives [85]. Generative adversarial imitation learning combines the power of GANs with imitation learning to generate driving policies that mimic expert human drivers. Additionally, vehicle networking systems require coordination between multiple autonomous vehicles to optimize traffic flow and safety [86,87]. Generative models can create coordination policies that enable effective multi-agent interactions in complex traffic scenarios. Moreover, safety represents the paramount concern in autonomous vehicle decision-making [88]. Generative models must incorporate safety constraints throughout the policy generation process to ensure that all generated decisions maintain acceptable risk levels.
Multimodal GAI-based decision-making frameworks extend beyond unimodal perception by synthesizing heterogeneous inputs from vehicular sensors [89]. For instance, when approaching an intersection with obstructed views, the system can combine camera feeds, radar reflections, V2X communication data, and ambient audio (e.g., sirens from emergency vehicles) to generate a predictive scene model. This model enables the decision-making module to preemptively adjust speed or trajectory, thereby reducing reaction time and enhancing safety. For example, with respect to multimodal fusion for robust semantic perception in driving scenes, Brödermann et al. [90] proposed CAFuser, a condition-aware framework that uses RGB input to generate a condition token guiding sensor fusion. It employs a shared backbone with modality-specific feature adapters to align diverse sensor inputs, dynamically adjusting fusion based on environmental conditions.
Moreover, combining generative models with deep reinforcement learning [91,92] allows for the synthesis of adaptive driving policies that can continuously improve through interaction with dynamic environments, enabling more robust decision-making under uncertainty. For example, Huang et al. [93] proposed an efficient deep reinforcement learning (DRL) framework that integrates imitative expert priors for autonomous driving. They used behavioral cloning and uncertainty estimation to derive an expert policy, then guided a DRL agent [94], and developed VWP, an autonomous driving model combining VAE-WGAN for feature extraction and enhanced PPOE for decision-making.

4.3. Anomaly Detection Through Generative Modeling

Anomaly detection in vehicle network systems requires the capability to identify unusual patterns, behaviors, or events that may indicate safety threats, system failures, or security breaches [95,96]. Generative models offer powerful frameworks for learning normal operational patterns and detecting deviations from expected behaviors [97]. These systems must continuously monitor their own performance to identify system-level anomalies that could signal cyber-attacks, hardware failures, or software malfunctions [23]. Furthermore, given the diverse and dynamic nature of vehicular environments, these systems require adaptive anomaly detection mechanisms with adjustable thresholds that can accommodate varying operating conditions and environmental contexts [98].
For example, Khan et al. [99] presented EDL-CMSO, a secure intrusion detection framework for the IoV integrating AES-256 encryption, secure multi-party computation, and homomorphic encryption. Oucheikh et al. [100] proposed a deep learning framework using LSTM auto-encoders and a CNN for real-time anomaly detection in connected autonomous vehicles.

4.4. Summary of Generative AI in Vehicular Decision-Making

Generative AI techniques applied to decision-making in vehicle networks exhibit varied strengths and trade-offs across key tasks. For traffic flow prediction, probabilistic generative models effectively capture complex spatiotemporal dependencies but may struggle with scalability in large, dynamic networks. Decision policy generation benefits from generative models’ ability to simulate diverse driving behaviors and rare events, enhancing robustness; however, integrating these models with real-time control systems remains challenging due to computational demands and interpretability concerns. Anomaly detection through generative modeling provides powerful unsupervised methods for identifying novel or malicious behaviors, though sensitivity to false positives and adaptability to evolving threats require further improvements. Overall, the suitability of each approach depends on the task’s real-time requirements, data availability, and complexity. Future work should focus on improving model efficiency, interpretability, and adaptive capabilities.

5. Generative AI for Communication and Resource Management in Vehicle Network Systems

The management and security of vehicle network systems present complex challenges that require sophisticated approaches to communication optimization and resource allocation. As vehicle networks continue to evolve toward more interconnected and autonomous operations, the need for intelligent management systems becomes increasingly critical. This section examines how GAI technologies contribute to more efficient and reliable vehicle network operations.

5.1. Generative AI for Communication Efficiency

Traditional communication systems focus primarily on syntactic accuracy, transmitting exact bit sequences regardless of semantic meaning [101]. GAI enables semantic communication approaches that prioritize the preservation of meaning over syntactic precision, resulting in significant efficiency improvements for vehicle network communications [102,103]. Semantic compression utilizes generative models to identify and communicate task-relevant semantic content tailored to vehicular application needs. For traffic scenario sharing between vehicles, a generative model can compress complex sensor observations into compact semantic representations that capture the critical elements needed for decision-making.
For example, Feng et al. [104] proposed a scalable AIGC system for vehicular network semantic communication and used an encoder–decoder architecture to convert images into textual representations for optimized transmission. They integrated reinforcement learning to enhance content reliability, demonstrating improved blind-spot vehicle perception and data compression compared to baselines. Similarly, Du et al. [105] introduced a task-oriented semantic communication framework for large multimodal models (LMMs) in vehicle networks, leveraging a large language and vision assistant (LLaVA) to optimize image slicing and resource allocation. Cheng et al. [106] introduced TCC-SemCom, an architecture that integrates Transformer and CNN modules in a complementary block design tailored for semantic-level image communication. The parallel TCC block combined CNNs for local feature extraction and Transformers for global semantic modeling.

5.2. Resource Allocation

Resource allocation in vehicle networks refers to the systematic distribution and management of computational, communication, and energy resources across distributed vehicular systems to optimize performance while meeting diverse service requirements [107]. This process involves coordinating heterogeneous resources including processing power, memory, bandwidth, storage, and energy across vehicles, edge computing nodes, and cloud infrastructure to support various applications ranging from safety-critical autonomous driving functions to infotainment services. GAI optimizes resource allocation by analyzing the environment through intelligent scheduling and real-time monitoring to respond to real-time vehicle network conditions and various application requirements. The integration of generative AI with advanced models such as DRL demonstrates significant potential for enhancing resource management in vehicle-to-everything (V2X) communications [108,109].
For example, Jahan et al. [110] proposed a game-theoretic–GAI approach for computation offloading and resource management in mobile edge collaborative vehicular networks. They integrated a Stackelberg game framework with GAI-driven simulations, designed a two-stage algorithm to optimize offloading ratios and resource allocation, and used Lagrangian optimization with Karush–Kuhn–Tucker conditions to balance latency, energy consumption, and computational efficiency. Zheng et al. [111] proposed a framework for SIoT-based image retrieval services in automotive market analysis. They introduced adversarial attack schemes during image transmission and proposed a defense strategy using Generative Diffusion Models (GDMs). Considering mobile devices’ resource constraints, they used GDMs to design resource allocation strategies, optimizing energy use and balancing image transmission and defense energy consumption. Wu et al. [112] focused on GAI-based resource management in RIS-aided next-generation networks by designing a channel distribution learning (CDL) method for cascade channel estimation and integrating GAI with distributional reinforcement learning (DBRL) to optimize system utility for energy efficiency and QoS satisfaction rates (QoSSRs).

5.3. Summary of Communication and Resource Management

Generative AI techniques in managing and securing vehicle networks address critical challenges in communication efficiency and resource allocation, each with distinct benefits and limitations. For communication efficiency, generative models enable data compression, semantic encoding, and synthetic data generation to reduce bandwidth and latency, but they face challenges in maintaining robustness under diverse network conditions and adversarial environments. Resource allocation methods leveraging generative AI optimize network load balancing and power management by predicting traffic patterns and user demands; however, their effectiveness depends heavily on accurate modeling of highly dynamic vehicular contexts and real-time adaptability. While these approaches demonstrate promising potential to enhance system scalability and security, practical deployment requires advances in model generalization, resilience to network fluctuations, and integration with existing vehicular protocols. Future research should focus on developing lightweight, adaptive generative models tailored for heterogeneous IoV environments. Figure 2 presents a central–branch framework outlining four generative AI-driven domains for enhancing resource management in vehicular networks.

6. Challenge and Future Directions

GAI has demonstrated its effectiveness in various aspects of the IoV, including training support, decision-making assistance, and communication enhancement. However, as GAI continues to evolve and exhibit remarkable capabilities within intelligent vehicular networks, several critical challenges must be addressed to fully realize its potential in real-world deployments. In the following sections, we explore key technical barriers and outline possible solutions related to model lightweighting and edge deployment, continual learning and adaptability, as well as emerging security concerns.

6.1. Model Lightweighting and Edge Deployment Challenges

The integration of generative AI into the IoV can significantly enhance training and decision-making processes. However, the high demand for data, memory, and computation in generative architectures makes them difficult to implement efficiently on edge devices with limited resources. As noted in [113], training generative AI models typically requires high-performance hardware such as GPUs (e.g., Tesla V100 16 GB, RTX 2080Ti, or NVIDIA RTX 3090 24 GB) or TPUs. For smaller-scale models, a GTX 1060 with 6 GB of DDR5 memory may be sufficient, while basic sample generation can be achieved with more modest configurations like an i7 3.4 GHz CPU and a GTX 970 GPU. However, these increasing hardware demands significantly extend the training time and limit the feasibility of deploying generative models on resource-constrained edge environments. In addition, the real-time performance requirements of in-vehicle applications demand ultra-low-latency responses, which traditional cloud-based generative AI architectures are unable to deliver. Safety-critical tasks such as collision avoidance and emergency responses require decision-making within a few milliseconds. This makes the deployment of generative models at the edge essential for ensuring both reliability and timeliness [114,115].
To address these constraints, several lightweight model optimization techniques can be employed in conjunction with edge intelligence:
  • Model pruning removes redundant weights and neurons to reduce model size and computation while preserving task-specific accuracy.
  • Quantization compresses model parameters from a 32-bit floating point to lower-precision formats (e.g., INT8 or FP16), significantly reducing memory usage and inference latency.
  • Knowledge distillation transfers knowledge from a large “teacher” generative model to a smaller “student” model, enabling comparable performance with reduced complexity.
  • Parameter-efficient fine-tuning methods such as Low-Rank Adaptation (LoRA) or prefix-tuning allow adaptation to specific vehicular tasks without retraining the full model.
Additionally, to overcome the constraints of in-vehicle systems in running GAI, edge intelligence [18] has been proposed as a viable solution, as it relocates computation to the network edge to minimize delays and enhance responsiveness. In this work, Zeng et al. [116] proposed a generative AI-aided multimodal parallel offloading framework for AIGC metaverse services. They partitioned multimodal content into ROIs, offloaded tasks to multiple edge servers, and used generative AI to adapt to dynamic resources, minimizing delay via Beckmann transformations. Similarly, Yang et al. [117] proposed a multi-user offloading framework for personalized diffusion models, splitting inference into edge-based low-level semantic processing and local personalized refinement. The team formulated the joint optimization of offloading and split points as an extended generalized quadratic assignment problem (GQAP), transforming it into a Markov decision process (MDP) and solving it with a DRL-convex hybrid method. These works emphasize task partitioning and resource coordination, with less focus on internal model compression. Therefore, integrating model-lightweighting methods (e.g., pruning, quantization, knowledge distillation, and LoRA) with intelligent offloading and split-inference strategies offers a practical, complementary pathway: lightweighting reduces local resource demands, while offloading handles computational peaks and heavy-modal subtasks, jointly enabling low-latency and reliable GAI deployment in IoV environments.

6.2. Privacy and Security Challenges

As GAI enables a growing range of applications and services within the IoV, it also introduces new challenges related to privacy and security. In particular, the increasing transparency of data exchange through wireless communication and collaborative sharing frameworks underscores the urgency of addressing privacy protection [118,119]. As generative models continue to grow in complexity and scale, stronger safeguards are required to mitigate risks such as data leakage, adversarial exploitation, and model misuse.
Recent studies have made notable efforts to enhance privacy protection and prevent data leakage in vehicular networks. For example, Wang et al. [120] proposed puncturable registered ABE (PR-ABE) for vehicular social networks, addressing three key limitations in existing ABE solutions: data leakage upon key compromise, reliance on fully trusted authorities, and privacy breaches from transparent policies. PR-ABE enables flexible access control, allows precise data deletion via key updating, and eliminates trusted authorities by letting vehicles generate keys independently. Zhou et al. [121] developed an efficient multilevel threshold changeable homomorphic data encapsulation (MCTh-HDEM) strategy for privacy-preserving vehicle positioning. MCTh-HDEM supports batch encryption, lightweight matrix computations, and dynamic threshold decryption for different security levels.
In addition, further attention has been directed toward securing collaborative operations within networked environments. Zheng et al. [111] proposed a framework for Semantic IoT-based image retrieval, using Generative Diffusion Models (GDMs) to defend against adversarial attacks. Furthermore, in the field of collaborative misbehavior detection for untrusted vehicular environments, Sedar et al. [122] introduced a deep reinforcement learning approach with transfer learning. They performed selective knowledge transfer between roadside units (RSUs) to counter label-flipping and policy induction attacks, reducing training time at target RSUs.
Beyond traditional cryptographic and access control techniques, GAI can generate privacy-preserving solutions in IoV environments by synthesizing realistic yet anonymized datasets, such as camera images, LiDAR point clouds, or V2X communication traces. By integrating differential privacy or federated learning, these synthetic datasets can support tasks like perception and anomaly detection while safeguarding privacy. For example, Khan et al. [123] developed a GenAI-driven intrusion detection system that generates synthetic data to enhance threat detection and protect privacy. Similarly, Li et al. [124] proposed a trajectory privacy-preserving (TPP) scheme using WGAN with differential privacy, effectively balancing privacy and data utility in VANETs. These approaches highlight how GAI can complement existing security frameworks to enable privacy-preserving collaborative intelligence in vehicular networks.
In IoV-specific deployments, such privacy-preserving data synthesis can enable safe cross-fleet learning among autonomous taxis operating in different cities, collaborative traffic flow optimization without exposing proprietary fleet behaviors, and realistic simulations of rare driving scenarios without compromising real-world identities or locations. However, deploying these methods in the IoV faces unique challenges: Edge devices may have limited computational resources for real-time synthetic data generation. Model updates must remain consistent across geographically distributed vehicles, and long-term performance can degrade due to domain shifts in traffic patterns and environmental conditions. Addressing these IoV-specific constraints is essential for ensuring that GAI-based privacy solutions are not only theoretically secure but also practically deployable in safety-critical vehicular networks.

6.3. Latency and Adaptability Challenges

In addition to the aforementioned challenges, time complexity and latency remain critical barriers to the practical deployment of generative AI in Internet of Vehicles (IoV) systems. Safety-critical applications such as autonomous driving require real-time or near-real-time inference, where any delay in model processing or data transmission could jeopardize operational safety. The computationally intensive nature of many generative models makes meeting stringent latency requirements difficult, highlighting the need for model compression, efficient inference acceleration, and optimized deployment strategies.
Furthermore, IoV environments are highly dynamic, with frequent changes in traffic patterns, road conditions, and environmental factors. Generative models must exhibit strong adaptability to new and evolving scenarios without requiring costly and time-consuming retraining. Techniques such as online learning, transfer learning, and continual learning could enhance robustness and responsiveness [125].
Lastly, the performance of generative AI in IoV often depends on stable, high-bandwidth network connectivity. In real-world deployments, fluctuating latency, intermittent coverage, and limited bandwidth can significantly impair model performance. Reducing reliance on centralized cloud infrastructures through edge computing, distributed inference, and hybrid architectures can mitigate these issues. Addressing latency, adaptability, and network dependence in a holistic and coordinated manner will be crucial for ensuring the safe, efficient, and scalable deployment of generative AI in IoV systems.

6.4. Future Research Directions

Advancing generative AI in the Internet of Vehicles (IoV) will require tackling both the efficiency constraints of large-scale models and the long-term adaptability challenges inherent in dynamic mobility environments. Two complementary research avenues, namely the adoption of mixture-of-experts (MoE) architectures and the mitigation of catastrophic forgetting, emerge as particularly promising.
(1) Mixture-of-Experts for Context-Aware and Resource-Efficient Generation: MoE architectures provide a scalable and computationally efficient mechanism for deploying large generative models in IoV systems [16]. By conditionally activating only the most relevant experts based on the current operational context, MoE architectures enable targeted specialization without incurring the cost of full-model inference. For instance, weather-adapted experts could be engaged during adverse conditions, while high-speed experts could be employed on expressways. This selective computation not only reduces latency but also allows for the coexistence of multiple domain-specific generative capabilities, such as simulating rare traffic hazards, modeling seasonal variations in road conditions, or predicting diverse driver behaviors. Future work should investigate adaptive expert routing strategies, potentially leveraging reinforcement learning for real-time selection, as well as lightweight expert module designs suitable for deployment on edge-grade vehicular hardware. The integration of MoE architectures with federated learning further offers a pathway for maintaining global adaptability while respecting the privacy constraints of distributed vehicular datasets.
(2) Continual Learning to Overcome Catastrophic Forgetting: While MoE architectures address specialization and efficiency, long-term deployment in IoV environments demands models that can adapt to evolving data without erasing prior knowledge. This challenge, known as catastrophic forgetting, is particularly severe in highly variable vehicular contexts where traffic patterns, infrastructure layouts, and environmental factors can shift dramatically across regions and seasons [126]. A model optimized for low-density suburban roads may underperform when exposed to high-congestion metropolitan grids, just as a model trained under summer conditions may fail in icy winter scenarios. Addressing this requires continual learning strategies that maintain a stable knowledge base while integrating novel information. Replay-based systems can retain representative historical samples for incremental retraining, while parameter isolation methods such as Elastic Weight Consolidation (EWC) safeguard task-critical weights. Modular expansion architectures offer the ability to incorporate new capabilities without disturbing existing ones, and hybrid approaches that blend continual learning with domain adaptation could enable seamless expertise transfer across deployment contexts. Promising directions include generative replay, in which models rehearse by synthesizing prior scenarios, and meta-learning frameworks that improve rapid adaptation without sacrificing retention.
In concert, MoE architectures and advanced continual learning strategies can form a robust foundation for generative AI in the IoV, achieving both context-aware specialization and resilience against knowledge degradation. Progress in these areas will depend on coordinated advances in architectural design, learning algorithms, and deployment frameworks, ultimately enabling generative systems that evolve in step with the diverse and dynamic landscapes of future transportation networks.

7. Conclusions

GAI is emerging as a transformative force in the IoV, offering powerful capabilities for synthesizing multimodal data, enhancing decision-making processes, and reinforcing system security. By leveraging models such as VAEs, GANs, Transformers, and diffusion models, GAI enables the creation of high-fidelity synthetic datasets, the reconstruction of rare traffic scenarios, the augmentation of perception models, and the generation of optimized decision policies. These advances hold significant potential for improving traffic safety, optimizing network resource allocation, and enabling privacy-preserving collaborative intelligence across distributed vehicular systems. The integration of GAI into IoV ecosystems also contributes to bridging the gap between simulated and real-world environments, accelerating model training, and enabling robust performance in edge-case conditions that are critical for autonomous driving and intelligent traffic management. Moreover, through privacy-preserving synthetic data generation and advanced threat detection, GAI offers innovative solutions to longstanding security challenges in distributed vehicular networks while providing novel perspectives on deploying trustworthy AI in safety-critical IoV contexts.
However, the real-world deployment of GAI in the IoV faces substantial challenges. Model lightweighting for resource-constrained edge devices, latency-sensitive inference, and reliable multimodal fusion remain key technical barriers. In addition, issues regarding data privacy, trustworthiness of generated outputs, and robustness against adversarial attacks demand careful attention. Beyond technological concerns, large-scale integration will require cross-industry collaboration, standardized evaluation benchmarks, and secure infrastructure for data sharing and model updates.
Looking forward, research should prioritize the development of parameter-efficient architectures tailored for vehicular environments, the integration of GAI with edge-cloud collaborative intelligence frameworks, and the creation of verifiable and interpretable generative models to enhance transparency and trust. Addressing these challenges will be essential to fully realize the promise of GAI in reshaping the IoV, thereby enabling safer, more efficient, and more secure transportation systems in the coming era of connected mobility.

Author Contributions

Conceptualization, X.Y.; Data curation, X.Z.; Investigation, X.Z.; Writing—original draft, X.Z., J.Z., Y.D.; Writing—review & editing, X.Y., A.W., Q.D., L.L.; Resources, X.Y., Q.D.; Supervision, Q.D., L.L.; Project administration, X.Y. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the National Natural Science Foundation of China (No. 62371116) and Innovation Support Project for Postgraduates in Hebei Province (CXZZSS2025166).

Conflicts of Interest

The authors declare no conflicts of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.

References

  1. Duan, W.; Gu, J.; Wen, M.; Zhang, G.; Ji, Y.; Mumtaz, S. Emerging Technologies for 5G-IoV Networks: Applications, Trends and Opportunities. IEEE Netw. 2020, 34, 283–289. [Google Scholar] [CrossRef]
  2. Yan, H.; Li, Y. A survey of generative ai for intelligent transportation systems. arXiv 2023, arXiv:2312.08248. [Google Scholar] [CrossRef]
  3. Yan, H.; Li, Y. Generative AI for Intelligent Transportation Systems: Road Transportation Perspective. ACM Comput. Surv. 2025, 57, 315. [Google Scholar] [CrossRef]
  4. Xu, M.; Du, H.; Niyato, D.; Kang, J.; Xiong, Z.; Mao, S.; Han, Z.; Jamalipour, A.; Kim, D.I.; Shen, X.; et al. Unleashing the Power of Edge-Cloud Generative AI in Mobile Networks: A Survey of AIGC Services. IEEE Commun. Surv. Tutor. 2024, 26, 1127–1170. [Google Scholar] [CrossRef]
  5. Sun, G.; Xie, W.; Niyato, D.; Mei, F.; Kang, J.; Du, H.; Mao, S. Generative AI for Deep Reinforcement Learning: Framework, Analysis, and Use Cases. IEEE Wirel. Commun. 2025, 32, 186–195. [Google Scholar] [CrossRef]
  6. Li, J.; Cheng, H.; Guo, H.; Qiu, S. Survey on artificial intelligence for vehicles. Automot. Innov. 2018, 1, 2–14. [Google Scholar] [CrossRef]
  7. Ma, Y.; Wang, Z.; Yang, H.; Yang, L. Artificial intelligence applications in the development of autonomous vehicles: A survey. IEEE/CAA J. Autom. Sin. 2020, 7, 315–329. [Google Scholar] [CrossRef]
  8. Mchergui, A.; Moulahi, T.; Zeadally, S. Survey on artificial intelligence (AI) techniques for vehicular ad-hoc networks (VANETs). Veh. Commun. 2022, 34, 100403. [Google Scholar] [CrossRef]
  9. Shoaib, M.R.; Emara, H.M.; Zhao, J. A Survey on the Applications of Frontier AI, Foundation Models, and Large Language Models to Intelligent Transportation Systems. In Proceedings of the 2023 International Conference on Computer and Applications (ICCA), Cairo, Egypt, 28–30 November 2023. [Google Scholar]
  10. Lin, H.; Liu, Y.; Li, S.; Qu, X. How Generative Adversarial Networks Promote the Development of Intelligent Transportation Systems: A Survey. IEEE/CAA J. Autom. Sin. 2023, 10, 1781–1796. [Google Scholar] [CrossRef]
  11. Stappen, L.; Dillmann, J.; Striegel, S.; Vögel, H.-J.; Flores-Herr, N.; Schuller, B.W. Integrating Generative Artificial Intelligence in Intelligent Vehicle Systems. In Proceedings of the 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC), Bilbao, Spain, 24–28 September 2023; pp. 5790–5797. [Google Scholar]
  12. Shu, D.; Zhu, Z. Generative models and connected and automated vehicles: A survey in exploring the intersection of transportation and ai. arXiv 2024, arXiv:2403.10559. [Google Scholar] [CrossRef]
  13. Christopoulou, M.; Barmpounakis, S.; Koumaras, H.; Kaloxylos, A. Artificial Intelligence and Machine Learning as key enablers for V2X communications: A comprehensive survey. Veh. Commun. 2023, 39, 100569. [Google Scholar] [CrossRef]
  14. Teef, D.; Muhammad, K.; Nassisid, K.; Farus, B. Enhancing Vehicular Networks with Generative AI: Opportunities and Challenges. arXiv 2024, arXiv:2407.11020. [Google Scholar]
  15. Abirami, S.; Pethuraj, M.; Uthayakumar, M.; Chitra, P. A systematic survey on big data and artificial intelligence algorithms for intelligent transportation system. Case Stud. Transp. Policy 2024, 17, 101247. [Google Scholar] [CrossRef]
  16. Xu, M.; Niyato, D.; Kang, J.; Xiong, Z.; Jamalipour, A.; Fang, Y.; Kim, D.I. Integration of mixture of experts and multimodal generative ai in internet of vehicles: A survey. arXiv 2024, arXiv:2404.16356. [Google Scholar] [CrossRef]
  17. Xie, G.; Xiong, Z.; Zhang, X.; Xie, R.; Guo, S.; Guizani, M.; Poor, H.V. GAI-IoV: Bridging Generative AI and Vehicular Networks for Ubiquitous Edge Intelligence. IEEE Trans. Wirel. Commun. 2024, 23, 12799–12814. [Google Scholar] [CrossRef]
  18. Yan, G.; Liu, K.; Liu, C.; Zhang, J. Edge Intelligence for Internet of Vehicles: A Survey. IEEE Trans. Consum. Electron. 2024, 70, 4858–4877. [Google Scholar] [CrossRef]
  19. Lu, S.; Shi, W. Vehicle as a Mobile Computing Platform: Opportunities and Challenges. IEEE Netw. 2024, 38, 493–500. [Google Scholar] [CrossRef]
  20. Shah, S.A.A.; Fernando, X.; Kashef, R. A survey on artificial-intelligence-based Internet of Vehicles utilizing unmanned aerial vehicles. Drones 2024, 8, 353. [Google Scholar] [CrossRef]
  21. Zhang, R.; Xiong, K.; Du, H.; Niyato, D.; Kang, J.; Shen, X.; Poor, H.V. Generative AI-Enabled Vehicular Networks: Fundamentals, Framework, and Case Study. IEEE Netw. 2024, 38, 259–267. [Google Scholar] [CrossRef]
  22. Haddaji, A.; Ayed, S.; Fourati, L.C. Artificial Intelligence techniques to mitigate cyber-attacks within vehicular networks: Survey. Comput. Electr. Eng. 2022, 104, 108460. [Google Scholar] [CrossRef]
  23. Rajapaksha, S.; Kalutarage, H.; Al-Kadri, M.O.; Petrovski, A.; Madzudzo, G.; Cheah, M. AI-based intrusion detection systems for in-vehicle networks: A survey. ACM Comput. Surv. 2023, 55, 1–40. [Google Scholar] [CrossRef]
  24. Delgado, L.; Luis, J.; Ramos, J.A.L. A Comprehensive Survey on Generative AI Solutions in IoT Security. Electronics 2024, 13, 4965. [Google Scholar] [CrossRef]
  25. Aouedi, O.; Vu, T.H.; Sacco, A.; Nguyen, D.C.; Piamrat, K.; Marchetto, G.; Pham, Q.V. A Survey on Intelligent Internet of Things: Applications, Security, Privacy, and Future Directions. IEEE Commun. Surv. Tutor. 2025, 27, 1238–1292. [Google Scholar] [CrossRef]
  26. Goodfellow, I.J.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; Bengio, Y. Generative adversarial nets. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, Montreal, QC, Canada, 8–13 December 2014; MIT Press: Cambridge, MA, USA, 2014; pp. 1–9. [Google Scholar]
  27. Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A.N.; Kaiser, Ł.; Polosukhin, I. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, Long Beach, CA, USA, 4–9 December 2017; pp. 1–11. [Google Scholar]
  28. Li, X.; Xiao, B.; Chen, X.; Jiang, P.; Wang, X.V.; Zheng, P.; Xia, L.; Yin, C. MSDF-VAE: A Cloud–Edge Collaborative Method for Fault Diagnosis Based on Transfer Learning. IEEE Internet Things J. 2025, 12, 22393–22403. [Google Scholar] [CrossRef]
  29. Liang, J.E. Cross-Modal Information Recovery and Enhancement Using Multiple-Input–Multiple-Output Variational Autoencoder. IEEE Internet Things J. 2024, 11, 26470–26480. [Google Scholar] [CrossRef]
  30. Chen, L.; Xu, Y.; Zhu, Q.-X.; He, Y.-L. Adaptive Multi-Head Self-Attention Based Supervised VAE for Industrial Soft Sensing With Missing Data. IEEE Trans. Autom. Sci. Eng. 2024, 21, 3564–3575. [Google Scholar] [CrossRef]
  31. Neumeier, M.; Botsch, M.; Tollkühn, A.; Berberich, T. Variational Autoencoder-Based Vehicle Trajectory Prediction with an Interpretable Latent Space. In Proceedings of the 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), Indianapolis, IN, USA, 19–22 September 2021; pp. 820–827. [Google Scholar]
  32. Xu, P.; Hayet, J.-B.; Karamouzas, I. Context-Aware Timewise VAEs for Real-Time Vehicle Trajectory Prediction. IEEE Robot. Autom. Lett. 2023, 8, 5440–5447. [Google Scholar] [CrossRef]
  33. Kammoun, A.; Slama, R.; Tabia, H.; Ouni, T.; Abid, M. Generative adversarial networks for face generation: A survey. ACM Comput. Surv. 2022, 55, 1–37. [Google Scholar] [CrossRef]
  34. Kozłowski, M.; Racewicz, S.; Wierzbicki, S. Image Analysis in Autonomous Vehicles: A Review of the Latest AI Solutions and Their Comparison. Appl. Sci. 2024, 14, 8150. [Google Scholar] [CrossRef]
  35. Liao, K.; Lin, C.; Zhao, Y.; Gabbouj, M. DR-GAN: Automatic Radial Distortion Rectification Using Conditional GAN in Real-Time. IEEE Trans. Circuits Syst. Video Technol. 2020, 30, 725–733. [Google Scholar] [CrossRef]
  36. Mo, T.; Zheng, S.; Chan, W.-Y.; Yang, R. Review of AI Image Enhancement Techniques for In-Vehicle Vision Systems Under Adverse Weather Conditions. World Electr. Veh. J. 2025, 16, 72. [Google Scholar] [CrossRef]
  37. Ding, F.; Yu, K.; Gu, Z.; Li, X.; Shi, Y. Perceptual Enhancement for Autonomous Vehicles: Restoring Visually Degraded Images for Context Prediction via Adversarial Training. IEEE Trans. Intell. Transp. Syst. 2022, 23, 9430–9441. [Google Scholar] [CrossRef]
  38. Zhou, W.; Wang, C.; Ge, Y.; Wen, L.; Zhan, Y. All-Day Vehicle Detection From Surveillance Videos Based on Illumination-Adjustable Generative Adversarial Network. IEEE Trans. Intell. Transp. Syst. 2024, 25, 3326–3340. [Google Scholar] [CrossRef]
  39. Wang, M.; Su, T.; Chen, S.; Yang, W.; Liu, J.; Wang, Z. Automatic Model-Based Dataset Generation for High-Level Vision Tasks of Autonomous Driving in Haze Weather. IEEE Trans. Ind. Inform. 2023, 19, 9071–9081. [Google Scholar] [CrossRef]
  40. Li, K.; Dai, Z.; Wang, X.; Song, Y.; Jeon, G. GAN-Based Controllable Image Data Augmentation in Low-Visibility Conditions for Improved Roadside Traffic Perception. IEEE Trans. Consum. Electron. 2024, 70, 6174–6188. [Google Scholar] [CrossRef]
  41. Chen, T.; Ren, J. Integrating GAN and Texture Synthesis for Enhanced Road Damage Detection. IEEE Trans. Intell. Transp. Syst. 2024, 25, 12361–12371. [Google Scholar] [CrossRef]
  42. Yurtsever, E.; Yang, D.; Koc, I.M.; Redmill, K.A. Photorealism in Driving Simulations: Blending Generative Adversarial Image Synthesis With Rendering. IEEE Trans. Intell. Transp. Syst. 2022, 23, 23114–23123. [Google Scholar] [CrossRef]
  43. Xu, P.; Zhu, X.; Clifton, D.A. Multimodal Learning With Transformers: A Survey. IEEE Trans. Pattern Anal. Mach. Intell. 2023, 45, 12113–12132. [Google Scholar] [CrossRef]
  44. Papa, L.; Russo, P.; Amerini, I.; Zhou, L. A Survey on Efficient Vision Transformers: Algorithms, Techniques, and Performance Benchmarking. IEEE Trans. Pattern Anal. Mach. Intell. 2024, 46, 7682–7700. [Google Scholar] [CrossRef]
  45. Li, H.; Li, C.; Zheng, A.; Tang, J.; Luo, B. MsKAT: Multi-Scale Knowledge-Aware Transformer for Vehicle Re-Identification. IEEE Trans. Intell. Transp. Syst. 2022, 23, 19557–19568. [Google Scholar] [CrossRef]
  46. Dong, S.; Xie, W.; Yang, D.; Tian, J.; Li, Y.; Zhang, J.; Lei, J. SeaDATE: Remedy Dual-Attention Transformer With Semantic Alignment via Contrast Learning for Multimodal Object Detection. IEEE Trans. Circuits Syst. Video Technol. 2025, 35, 4713–4726. [Google Scholar] [CrossRef]
  47. Zhou, J.; Yang, J.; Wu, X.; Zhou, W.; Wang, Y. TrVLR: A Transformer-Based Vehicle Light Recognition Method in Vehicle Inspection. IEEE Trans. Intell. Transp. Syst. 2024, 25, 19995–20005. [Google Scholar] [CrossRef]
  48. Zhang, X.; Ling, Y.; Li, K.; Shi, W.; Zhou, Z. Multimodality Adaptive Transformer and Mutual Learning for Unsupervised Domain Adaptation Vehicle Re-Identification. IEEE Trans. Intell. Transp. Syst. 2024, 25, 20215–20226. [Google Scholar] [CrossRef]
  49. Yin, H.; Tian, D.; Lin, C.; Duan, X.; Zhou, J.; Zhao, D.; Cao, D. V2VFormer++: Multi-Modal Vehicle-to-Vehicle Cooperative Perception via Global-Local Transformer. IEEE Trans. Intell. Transp. Syst. 2024, 25, 2153–2166. [Google Scholar] [CrossRef]
  50. He, C.; Shen, Y.; Fang, C.; Xiao, F.; Tang, L.; Zhang, Y.; Zuo, W.; Guo, Z.; Li, X. Diffusion Models in Low-Level Vision: A Survey. IEEE Trans. Pattern Anal. Mach. Intell. 2025, 47, 4630–4651. [Google Scholar] [CrossRef]
  51. Huang, Y.; Huang, J.; Liu, Y.; Yan, M.; Lv, J.; Liu, J.; Xiong, W.; Zhang, H.; Cao, L.; Chen, S. Diffusion Model-Based Image Editing: A Survey. IEEE Trans. Pattern Anal. Mach. Intell. 2025, 47, 4409–4437. [Google Scholar] [CrossRef]
  52. Scarpellini, G.; Fiorini, S.; Giuliari, F.; Morerio, P.; Bue, A.D. DiffAssemble: A Unified Graph-Diffusion Model for 2D and 3D Reassembly. In Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 16–22 June 2024; pp. 28098–28108. [Google Scholar]
  53. Xu, H.; Lei, Y.; Chen, Z.; Zhang, X.; Zhao, Y.; Wang, Y.; Tu, Z. Bayesian Diffusion Models for 3D Shape Reconstruction. In Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 16–22 June 2024; pp. 10628–10638. [Google Scholar]
  54. Alexandru, R.; Blu, T.; Dragotti, P.L. Diffusion SLAM: Localizing Diffusion Sources From Samples Taken by Location-Unaware Mobile Sensors. IEEE Trans. Signal Process. 2021, 69, 5539–5554. [Google Scholar] [CrossRef]
  55. Li, J.; Li, B.; Tu, Z.; Liu, X.; Guo, Q.; Juefei-Xu, F.; Xu, R.; Yu, H. Light the Night: A Multi-Condition Diffusion Framework for Unpaired Low-Light Enhancement in Autonomous Driving. In Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 16–22 June 2024; pp. 15205–15215. [Google Scholar]
  56. Yang, B.; Su, H.; Gkanatsios, N.; Ke, T.W.; Jain, A.; Schneider, J.; Fragkiadaki, K. Diffusion-ES: Gradient-Free Planning with Diffusion for Autonomous and Instruction-Guided Driving. In Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 16–22 June 2024; pp. 15342–15353. [Google Scholar]
  57. Gupta, A.; Anpalagan, A.; Guan, L.; Khwaja, A.S. Deep learning for object detection and scene perception in self-driving cars: Survey, challenges, and open issues. Array 2021, 10, 100057. [Google Scholar] [CrossRef]
  58. Azadani, M.N.; Boukerche, A. A Novel Multimodal Vehicle Path Prediction Method Based on Temporal Convolutional Networks. IEEE Trans. Intell. Transp. Syst. 2022, 23, 25384–25395. [Google Scholar] [CrossRef]
  59. Xiao, Y.; Codevilla, F.; Gurram, A.; Urfalioglu, O.; López, A.M. Multimodal End-to-End Autonomous Driving. IEEE Trans. Intell. Transp. Syst. 2022, 23, 537–547. [Google Scholar] [CrossRef]
  60. Zhang, C.; Wang, H.; Cai, Y.; Chen, L.; Li, Y.; Sotelo, M.A.; Li, Z. Robust-FusionNet: Deep Multimodal Sensor Fusion for 3-D Object Detection Under Severe Weather Conditions. IEEE Trans. Instrum. Meas. 2022, 71, 2513713. [Google Scholar] [CrossRef]
  61. Prakash, A.; Chitta, K.; Geiger, A. Multi-modal fusion transformer for end-to-end autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, TN, USA, 20–25 June 2021; pp. 7077–7087. [Google Scholar]
  62. Wu, J.; Gao, B.; Gao, J.; Yu, J.; Chu, H.; Yu, Q.; Gong, X.; Chang, Y.; Tseng, H.E.; Chen, H.; et al. Prospective role of foundation models in advancing autonomous vehicles. Research 2024, 7, 0399. [Google Scholar] [CrossRef]
  63. Elallid, B.B.; Benamar, N.; Hafid, A.S.; Rachidi, T.; Mrani, N. A comprehensive survey on the application of deep and reinforcement learning approaches in autonomous driving. J. King Saud Univ.-Comput. Inf. Sci. 2022, 34, 7366–7390. [Google Scholar] [CrossRef]
  64. Feng, L.; Li, Q.; Peng, Z.; Tan, S.; Zhou, B. TrafficGen: Learning to Generate Diverse and Realistic Traffic Scenarios. In Proceedings of the 2023 IEEE International Conference on Robotics and Automation (ICRA), London, UK, 29 May–2 June 2023; pp. 3567–3575. [Google Scholar]
  65. Hui, S.; Wang, H.; Wang, Z.; Yang, X.; Liu, Z.; Jin, D.; Li, Y. Knowledge enhanced GAN for IoT traffic generation. In Proceedings of the ACM Web Conference, Lyon, France, 25–29 April 2022. [Google Scholar]
  66. Tamayo-Urgilés, D.; Sanchez-Gordon, S.; Valdivieso Caraguay, Á.L.; Hernández-Álvarez, M. GAN-Based Generation of Synthetic Data for Vehicle Driving Events. Appl. Sci. 2024, 14, 9269. [Google Scholar] [CrossRef]
  67. Cao, M.; Ramezani, R. Data generation using simulation technology to improve perception mechanism of autonomous vehicles. J. Phys. Conf. Ser. 2023, 2547, 012006. [Google Scholar] [CrossRef]
  68. Song, Z.; He, Z.; Li, X.; Ma, Q.; Ming, R.; Mao, Z.; Pei, H.; Peng, L.; Hu, J.; Yao, D.; et al. Synthetic Datasets for Autonomous Driving: A Survey. IEEE Trans. Intell. Veh. 2024, 9, 1847–1864. [Google Scholar] [CrossRef]
  69. Winter, K.; Vivekanandan, A.; Polley, R.; Shen, Y.; Schlauch, C.; Bouzidi, M.K.; Derajic, B.; Grabowsky, N.; Mariani, A.; Rochau, D.; et al. Generative AI for Autonomous Driving: A Review. arXiv 2025, arXiv:2505.15863. [Google Scholar] [CrossRef]
  70. Karunakaran, D.; Perez, J.S.B.; Worrall, S. Generating edge cases for testing autonomous vehicles using real-world data. Sensors 2023, 24, 108. [Google Scholar] [CrossRef]
  71. Hu, Z.; Lou, S.; Xing, Y.; Wang, X.; Cao, D.; Lv, C. Review and Perspectives on Driver Digital Twin and Its Enabling Technologies for Intelligent Vehicles. IEEE Trans. Intell. Veh. 2022, 7, 417–440. [Google Scholar] [CrossRef]
  72. Singh, D.; AlZubi, A.A.; Kaur, M.; Kumar, V.; Lee, H.-N. Deep Multi-Patch Hierarchical Network-Based Visibility Restoration Model for Autonomous Vehicles. IEEE Trans. Veh. Technol. 2025, 74, 7061–7071. [Google Scholar] [CrossRef]
  73. Liu, S.; Yang, H.; Zheng, M.; Xiao, L. Multi-UAV-Assisted MEC in IoV with Combined Multi-Modal Semantic Communication under Jamming Attacks. IEEE Trans. Mob. Comput. 2025, 24, 7600–7614. [Google Scholar] [CrossRef]
  74. Baresi, L.; Hu, D.Y.X.; Stocco, A.; Tonella, P. Efficient domain augmentation for autonomous driving testing using diffusion models. arXiv 2024, arXiv:2409.13661. [Google Scholar]
  75. Zheng, Z.; Cheng, Y.; Xin, Z.; Yu, Z.; Zheng, B. Robust Perception Under Adverse Conditions for Autonomous Driving Based on Data Augmentation. IEEE Trans. Intell. Transp. Syst. 2023, 24, 13916–13929. [Google Scholar] [CrossRef]
  76. Yu, W.; Sun, Y.; Zhou, R.; Liu, X. GAN Based Method for Labeled Image Augmentation in Autonomous Driving. In Proceedings of the 2019 IEEE International Conference on Connected Vehicles and Expo (ICCVE), Graz, Austria, 4–8 November 2019; pp. 1–5. [Google Scholar]
  77. Zhang, K.; Feng, X.; Jia, N.; Zhao, L.; He, Z. TSR-GAN: Generative adversarial networks for traffic state reconstruction with time space diagrams. Phys. A Stat. Mech. Its Appl. 2022, 591, 126788. [Google Scholar] [CrossRef]
  78. Lee, J.-M.; Kim, J.-D. A generative model for traffic demand with heterogeneous and spatiotemporal characteristics in massive wi-fi systems. Electronics 2022, 11, 1848. [Google Scholar] [CrossRef]
  79. Katariya, V.; Baharani, M.; Morris, N.; Shoghli, O.; Tabkhi, H. DeepTrack: Lightweight Deep Learning for Vehicle Trajectory Prediction in Highways. IEEE Trans. Intell. Transp. Syst. 2022, 23, 18927–18936. [Google Scholar] [CrossRef]
  80. Sun, T.; Sun, B.; Jiang, Z.; Hao, R.; Xie, J. Traffic flow online prediction based on a generative adversarial network with multi-source data. Sustainability 2021, 13, 12188. [Google Scholar] [CrossRef]
  81. Meng, X.; Lin, C.; Wang, Y.; Zhang, Y. Netgpt: Generative pretrained transformer for network traffic. arXiv 2023, arXiv:2304.09513. [Google Scholar] [CrossRef]
  82. Rasouli, A.; Tsotsos, J.K. Autonomous Vehicles That Interact With Pedestrians: A Survey of Theory and Practice. IEEE Trans. Intell. Transp. Syst. 2020, 21, 900–918. [Google Scholar] [CrossRef]
  83. Garikapati, D.; Shetiya, S.S. Autonomous vehicles: Evolution of artificial intelligence and the current industry landscape. Big Data Cogn. Comput. 2024, 8, 42. [Google Scholar] [CrossRef]
  84. Atakishiyev, S.; Salameh, M.; Yao, H.; Goebel, R. Explainable artificial intelligence for autonomous driving: A comprehensive overview and field guide for future research directions. IEEE Access 2024, 12, 101603–101625. [Google Scholar] [CrossRef]
  85. Padmaja, B.; Moorthy, C.V.; Venkateswarulu, N.; Bala, M.M. Exploration of issues, challenges and latest developments in autonomous cars. J. Big Data 2023, 10, 61. [Google Scholar] [CrossRef]
  86. Zhao, J.; Zhao, W.; Deng, B.; Wang, Z.; Zhang, F.; Zheng, W.; Cao, W.; Nan, J.; Lian, Y.; Burke, A.F. Autonomous driving system: A comprehensive survey. Expert Syst. Appl. 2024, 242, 122836. [Google Scholar] [CrossRef]
  87. Wang, Y.; Xing, S.; Can, C.; Li, R.; Hua, H.; Tian, K.; Mo, Z.; Gao, X.; Wu, K.; Zhou, S.; et al. Generative ai for autonomous driving: Frontiers and opportunities. arXiv 2025, arXiv:2505.08854. [Google Scholar] [CrossRef]
  88. Muhammad, K.; Ullah, A.; Lloret, J.; Ser, J.D.; de Albuquerque, V.H.C. Deep Learning for Safe Autonomous Driving: Current Challenges and Future Directions. IEEE Trans. Intell. Transp. Syst. 2021, 22, 4316–4336. [Google Scholar] [CrossRef]
  89. Zhang, L.; Wang, B.; Zhao, Y.; Yuan, Y.; Zhou, T.; Li, Z. Collaborative Multimodal Fusion Network for Multiagent Perception. IEEE Trans. Cybern. 2025, 55, 486–498. [Google Scholar] [CrossRef]
  90. Brödermann, T.; Sakaridis, C.; Fu, Y.; Van Gool, L. CAFuser: Condition-Aware Multimodal Fusion for Robust Semantic Perception of Driving Scenes. IEEE Robot. Autom. Lett. 2025, 10, 3134–3141. [Google Scholar] [CrossRef]
  91. Zhao, J.; Wu, Y.; Deng, R.; Xu, S.; Gao, J.; Burke, A. A Survey of Autonomous Driving from a Deep Learning Perspective. ACM Comput. Surv. 2025, 57, 1–60. [Google Scholar] [CrossRef]
  92. Zhao, R.; Li, Y.; Fan, Y.; Gao, F.; Tsukada, M.; Gao, Z. A Survey on Recent Advancements in Autonomous Driving Using Deep Reinforcement Learning: Applications, Challenges, and Solutions. IEEE Trans. Intell. Transp. Syst. 2024, 25, 19365–19398. [Google Scholar] [CrossRef]
  93. Huang, Z.; Wu, J.; Lv, C. Efficient Deep Reinforcement Learning With Imitative Expert Priors for Autonomous Driving. IEEE Trans. Neural Netw. Learn. Syst. 2023, 34, 7391–7403. [Google Scholar] [CrossRef] [PubMed]
  94. Jin, Y.-L.; Ji, Z.-Y.; Zeng, D.; Zhang, X.-P. VWP: An Efficient DRL-Based Autonomous Driving Model. IEEE Trans. Multimed. 2024, 26, 2096–2108. [Google Scholar] [CrossRef]
  95. Solaas, J.R.V.; Tuptuk, N.; Mariconti, E. Systematic Review: Anomaly Detection in Connected and Autonomous Vehicles. arXiv 2024, arXiv:2405.02731. [Google Scholar] [CrossRef]
  96. Ahmad, H.; Gulzar, M.M.; Aziz, S.; Habib, S.; Ahmed, I. AI-based anomaly identification techniques for vehicles communication protocol systems: Comprehensive investigation, research opportunities and challenges. Internet of Things 2024, 27, 101245. [Google Scholar] [CrossRef]
  97. Özdemir, Ö.; İşyapar, M.T.; Karagöz, P.; Schmidt, K.W.; Demir, D.; Karagöz, N.A. A Survey of Anomaly Detection in In-Vehicle Networks. arXiv 2024, arXiv:2409.07505. [Google Scholar] [CrossRef]
  98. Edozie, E.; Shuaibu, A.N.; Sadiq, B.O.; John, U.K. Artificial intelligence advances in anomaly detection for telecom networks. Artif. Intell. Rev. 2025, 58, 100. [Google Scholar] [CrossRef]
  99. Khan, H.; Tejani, G.G.; AlGhamdi, R.; Alasmari, S.; Sharma, N.K.; Sharma, S.K. A secure and efficient deep learning-based intrusion detection framework for the internet of vehicles. Sci. Rep. 2025, 15, 12236. [Google Scholar] [CrossRef]
  100. Oucheikh, R.; Fri, M.; Fedouaki, F.; Hain, M. Deep real-time anomaly detection for connected autonomous vehicles. Procedia Comput. Sci. 2020, 177, 456–461. [Google Scholar] [CrossRef]
  101. Ye, S.; Wu, Q.; Fan, P.; Fan, Q. A survey on semantic communications in internet of vehicles. Entropy 2025, 27, 445. [Google Scholar] [CrossRef]
  102. Liang, C.; Du, H.; Sun, Y.; Niyato, D.; Kang, J.; Zhao, D.; Imran, M.A. Generative AI-Driven Semantic Communication Networks: Architecture, Technologies, and Applications. IEEE Trans. Cogn. Commun. Netw. 2025, 11, 27–47. [Google Scholar] [CrossRef]
  103. Xin, G.; Fan, P.; Letaief, K.B. Semantic communication: A survey of its theoretical development. Entropy 2024, 26, 102. [Google Scholar] [CrossRef]
  104. Feng, H.; Yang, Y.; Han, Z. Scalable ai generative content for vehicular network semantic communication. arXiv 2023, arXiv:2311.13782. [Google Scholar] [CrossRef]
  105. Du, B.; Du, H.; Niyato, D.; Li, R. Task-Oriented Semantic Communication in Large Multimodal Models-based Vehicle Networks. IEEE Trans. Mob. Comput. 2025, 1–15. [Google Scholar] [CrossRef]
  106. Cheng, G.; Chong, B.; Lu, H. TCC-SemCom: A Transformer-CNN Complementary Block-Based Image Semantic Communication. IEEE Commun. Lett. 2025, 29, 625–629. [Google Scholar] [CrossRef]
  107. Zhao, J.; Ren, R.; Zou, D.; Zhang, Q.; Xu, W. IoV-Oriented Integrated Sensing, Computation, and Communication: System Design and Resource Allocation. IEEE Trans. Veh. Technol. 2024, 73, 16283–16294. [Google Scholar] [CrossRef]
  108. Mekrache, A.; Bradai, A.; Moulay, E.; Dawaliby, S. Deep reinforcement learning techniques for vehicular networks: Recent advances and future trends towards 6G. Veh. Commun. 2022, 33, 100398. [Google Scholar] [CrossRef]
  109. Liang, H.; Zhang, X.; Hong, X.; Zhang, Z.; Li, M.; Hu, G.; Hou, F. Reinforcement Learning Enabled Dynamic Resource Allocation in the Internet of Vehicles. IEEE Trans. Ind. Inform. 2021, 17, 4957–4967. [Google Scholar] [CrossRef]
  110. Jahan, N.; Hasan, M.K.; Islam, S.; Nazri, M.Z.A.; Ariffin, K.A.Z.; Abbas, H.S.; Alqahtani, A.; Gohel, H. Game-Theoretic-GAI Approach for Computation Offloading and Resource Management for Mobile Edge Collaborative Vehicular Networks. IEEE Trans. Intell. Transp. Syst. 2025, 1–12. [Google Scholar] [CrossRef]
  111. Zheng, J.; Du, B.; Du, H.; Kang, J.; Niyato, D.; Zhang, H. Energy-Efficient Resource Allocation in Generative AI-Aided Secure Semantic Mobile Networks. IEEE Trans. Mob. Comput. 2024, 23, 11422–11435. [Google Scholar] [CrossRef]
  112. Wu, Z.; Zhang, H.; Li, L.; Lu, Y.; Yang, J. GAI-Based Resource Management in RIS-Aided Next-Generation Network and Communication. IEEE Trans. Cogn. Commun. Netw. 2025, 11, 847–857. [Google Scholar] [CrossRef]
  113. Bandi, A.; Adapa, P.V.S.R.; Kuchi, Y.E.V.P.K. The power of generative ai: A review of requirements, models, input–output formats, evaluation metrics, and challenges. Future Internet 2023, 15, 260. [Google Scholar] [CrossRef]
  114. Biswas, A.; Wang, H.-C. Autonomous vehicles enabled by the integration of IoT, edge intelligence, 5G, and blockchain. Sensors 2023, 23, 1963. [Google Scholar] [CrossRef]
  115. Hua, K.; Su, S.; Wang, Y. Intelligent service migration for the internet of vehicles in edge computing: A mobility-aware deep reinforcement learning framework. Comput. Netw. 2025, 257, 111021. [Google Scholar] [CrossRef]
  116. Zeng, W.; Zheng, J.; Gao, L.; Niu, J.; Ren, J.; Wang, H.; Cao, R.; Ji, S. Generative AI-Aided Multimodal Parallel Offloading for AIGC Metaverse Service in IoT Networks. IEEE Internet Things J. 2025, 12, 13273–13285. [Google Scholar] [CrossRef]
  117. Yang, W.; Xiong, Z.; Guo, S.; Mao, S.; Kim, D.I.; Debbah, M. Efficient Multi-user Offloading of Personalized Diffusion Models: A DRL-Convex Hybrid Solution. IEEE Trans. Mob. Comput. 2025, 24, 9092–9109. [Google Scholar] [CrossRef]
  118. Ibrahum, A.D.M.; Hussain, M.; Hong, J.E. Deep learning adversarial attacks and defenses in autonomous vehicles: A systematic literature review from a safety perspective. Artif. Intell. Rev. 2025, 58, 28. [Google Scholar] [CrossRef]
  119. Arifin, M.M.; Ahmed, M.S.; Ghosh, T.K.; Udoy, I.A.; Zhuang, J.; Yeh, J.H. A survey on the application of generative adversarial networks in cybersecurity: Prospective, direction and open research scopes. arXiv 2024, arXiv:2407.08839. [Google Scholar] [CrossRef]
  120. Wang, C.; Ming, Y.; Liu, H.; Deng, Y.; Yang, M.; Feng, J. Puncturable Registered ABE for Vehicular Social Networks: Enhancing Security and Practicality. IEEE Trans. Dependable Secur. Comput. 2025. [Google Scholar] [CrossRef]
  121. Zhou, T.; Zhou, J.; Cao, Z.; Dong, X.; Choo, K.-K.R. Efficient Multilevel Threshold Changeable Homomorphic Data Encapsulation With Application to Privacy-Preserving Vehicle Positioning. IEEE Trans. Intell. Transp. Syst. 2025, 26, 5494–5508. [Google Scholar] [CrossRef]
  122. Sedar, R.; Kalalas, C.; Dini, P.; Vázquez-Gallego, F.; Alonso-Zarate, J.; Alonso, L. Knowledge Transfer for Collaborative Misbehavior Detection in Untrusted Vehicular Environments. IEEE Trans. Veh. Technol. 2025, 74, 425–440. [Google Scholar] [CrossRef]
  123. Khan, M.A.; Alasiry, A.; Marzougui, M.; Bayhan, I.; Kuna, S.S.; Rao, G.S.N.; Algamdi, S.A.; Aldossary, H. Securing Intelligent Transportation Systems: A Dual-Framework Approach for Privacy Protection and Cybersecurity Using Generative AI. IEEE Trans. Intell. Transp. Syst. 2025, 1–12. [Google Scholar] [CrossRef]
  124. Li, T.; Xie, S.; Zeng, Z.; Dong, M.; Liu, A. ATPS: An AI Based Trust-Aware and Privacy-Preserving System for Vehicle Managements in Sustainable VANETs. IEEE Trans. Intell. Transp. Syst. 2022, 23, 19837–19851. [Google Scholar] [CrossRef]
  125. Song, Z.; Tao, Y.; Hua, Z.; Wang, S.; Pan, G.; An, J. Generative Artificial Intelligence-Empowered Multidomain Internet of Vehicles Systems: Scalability, Efficiency, and Suitability. IEEE Veh. Technol. Mag. 2025, 20, 53–62. [Google Scholar] [CrossRef]
  126. Yang, L.; Luo, Z.; Zhang, S.; Teng, F.; Li, T. Continual Learning for Smart City: A Survey. IEEE Trans. Knowl. Data Eng. 2024, 36, 7805–7824. [Google Scholar] [CrossRef]
Mathematics 13 02795 g001
Figure 1. Overview of the evolution, architectures, and key techniques of generative AI.
Figure 1. Overview of the evolution, architectures, and key techniques of generative AI.
Mathematics 13 02795 g001
Mathematics 13 02795 g002
Figure 2. Generative AI-driven resource management for vehicular networks.
Figure 2. Generative AI-driven resource management for vehicular networks.
Mathematics 13 02795 g002
Table 1. Summery of the current articles.
Table 1. Summery of the current articles.
RefYearDescriptionFocus
[6]2018An overview of AI applications in vehicles, focusing on deep learning and RL for autonomous perception and traffic analysisGeneral AI overviews for the IoV
[7]2020An overview of AI applications in AVs for perception, localization, mapping, decision-making, and future opportunities
[8]2022A survey of AI techniques (ML, DL, and swarm intelligence) in VANETs for routing, resource allocation, and security with future directions
[9]2023Reviews frontier AI, foundation models, and LLMs in an ITS for traffic management and autonomous driving
[10]2023A survey of GAN applications in the ITS, covering autonomous driving, traffic flow, and anomaly detection with future directionsTraining and data synthesis
[11]2023Investigates GAI integration in intelligent vehicles for multimodal interaction and user experience enhancement
[12]2024A survey on integrating generative models with CAVs to enhance prediction, simulations, and decision-making in transportation
[13]2023A comprehensive review of AI/ML in V2X communications for handover, beam selection, and resource managementVehicular communications
[14]2024Explores GAI applications in vehicular networks for communication optimization and security with framework proposals
[15]2024A systematic survey of big data and AI algorithms for traffic prediction, vehicle recognition, and route optimization in an ITSTraffic prediction and ITS optimization
[16]2024Surveys MoE and multimodal GAI integration in the IoV for advancing artificial general intelligence
[17]2024Proposes a GAI-IoV framework for edge intelligence, optimizing resource allocation and inference strategyEdge intelligence and resource management
[18]2024A comprehensive survey of edge intelligence in the IoV, covering architecture, inference, training, sensing, and future trends
[19]2024Explores vehicles as mobile computing platforms with five functions, discussing business models and challenges
[20]2024A survey of AI applications in UAV-assisted Internet of Vehicles, focusing on resource management, routing, and trajectory optimization
[21]2024An exploration of generative AI in vehicular networks with a multi-modality framework and a DRL-based resource allocation case study
[22]2022A survey of AI techniques for mitigating cyber-attacks in vehicular networks, exploring intrusion detection and defense strategiesIoV security and privacy
[23]2023A survey of AI techniques for intrusion detection in in-vehicle networks, discussing methods, datasets, and future directions
[24]2024A comprehensive review of GAI applications in IoT security, covering access control, blockchain, and cyber threat detection
[25]2025A comprehensive survey of intelligent IoT applications, security, privacy challenges, and future research directions
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Yuan, X.; Zhang, X.; Wang, A.; Zhou, J.; Du, Y.; Deng, Q.; Liu, L. Generative AI for the Internet of Vehicles: A Review of Advances in Training, Decision-Making, and Security. Mathematics 2025, 13, 2795. https://doi.org/10.3390/math13172795

AMA Style

Yuan X, Zhang X, Wang A, Zhou J, Du Y, Deng Q, Liu L. Generative AI for the Internet of Vehicles: A Review of Advances in Training, Decision-Making, and Security. Mathematics. 2025; 13(17):2795. https://doi.org/10.3390/math13172795

Chicago/Turabian Style

Yuan, Xiaoming, Xinling Zhang, Aiwen Wang, Jiaxin Zhou, Yingying Du, Qingxu Deng, and Lei Liu. 2025. "Generative AI for the Internet of Vehicles: A Review of Advances in Training, Decision-Making, and Security" Mathematics 13, no. 17: 2795. https://doi.org/10.3390/math13172795

APA Style

Yuan, X., Zhang, X., Wang, A., Zhou, J., Du, Y., Deng, Q., & Liu, L. (2025). Generative AI for the Internet of Vehicles: A Review of Advances in Training, Decision-Making, and Security. Mathematics, 13(17), 2795. https://doi.org/10.3390/math13172795

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop