Applications of Generative AI (GAI) for
Mobile and Wireless Networking: A Survey

AI techniques have become widespread in various research areas in mobile networking and the IoT, such as network management, wireless security, resource allocation, and edge services. In recent years, data-driven and AI-enabled approaches have established clear performance superiority over many conventional optimization-based designs, which are limited primarily by network scalabilities in handling massive connections and heterogeneous variations [4]. Additionally, AI techniques enable efficient approximation of computationally challenging algorithms, to effectively overcome nonlinear and unknown patterns in wireless networks, and rapidly accelerate existing model-based algorithms. On this basis, there have been several surveys of different AI paradigms and their applications for mobile networking, such as Deep Learning (DL) [5, 6], Deep Reinforcement Learning (DRL) [7, 8], Federated Learning (FL) [9, 10], and swarm intelligence [11]. However, current AI-based approaches for mobile and wireless networking face several issues. Specifically, conventional AI algorithms usually require a large volume of labeled training data to facilitate better performance. Unfortunately, collecting and labeling training data in complex environments like wireless networks are costly, time-consuming, and require intensive human interventions. In addition, with an enormous number of diverse users, future wireless systems are highly dynamic and uncertain. It can significantly degrade the performance of conventional AI models, which are usually trained in a specific environment/condition. Although the distributed learning paradigm can help to facilitate the learning process of AI-based approaches, it still suffers heavily from scarcity of data required for efficient training [12]. Moreover, due to the differences in local system characteristics, the distributed AI system suffers from the long-tailed distributed data phenomenon. This data imbalance results in a negative transfer toward the data cluster with a dominant size. To address all these practical challenges, GAI is considered an efficient solution thanks to its properties of domain adaptation, data generation, and abnormal detection. Aiming to address these challenges, there are several surveys in the literature focusing on the applications of GAI in various domains such as mobile networking, Semantic Communication (SemCom), and security and privacy, as summarized in Table I.

For example, GANs, a notable GAI technique, are reviewed in several recent works [13, 14, 15, 16, 17, 18], providing preliminary discussions about the use of this GAI-specific technique in emerging networks. More recently, GAI has been reviewed for several research issues in wireless networks. For instance, in [19], the authors provide a detailed survey of training solutions for GANs that deal with challenges of model collapsing and gradient vanishing, providing meaningful techniques and guidance for mobile network deployments. Similarly, in [20], the authors analyze GANs with various stability issues and survey several training solutions by assessing the obstacles. In [21], several critical GAN analyses of computer and communication networks are outlined, with a focus on network management, security, and mobile networks. Differently, the authors in [22] introduce the concept of on-device GAI, where multiple devices, functioning as distinct learning agents, deploy on-device GAI models for the collaborative processing of complex tasks. In [23], a direction on GAI services at the network edge has been outlined in various use cases, followed by guidance on building a hierarchical collaboration architecture of IoT devices, mobile edges, and centralized clouds. In [24], wireless network management efficiency can be improved by DL models in general and particular applications of DGM. In addition, the authors in [25] present a review of the applications of GAI for physical layer communications, focusing on emerging problems such as signal detection, physical layer security, and Joint Source-Channel Coding (JSCC). Considering a different perspective, the authors in [26] discuss the potential of common GAI techniques, e.g., GANs, VAEs, and normalizing flows, in addressing the design challenges of unmanned vehicle swarms.

On another front, the interplay of GAI and SemCom has opened up a new avenue for reducing resource consumption while maintaining the reliability of wireless communication networks [27]. Fundamentally, SemCom relies on two fundamental techniques [28, 29] of knowledge extraction and lossy data compression, both rooted in AI. As such, the current landscape of SemCom designs represents a paradigm shift, with these innovative concepts emerging across various disciplines. In the literature, several surveys aim to elucidate the intricate connections between the potential AI techniques within GAI and their application in the field of SemCom. From the perspective of Knowledge Graph (KG), the current SemCom is based on the KG embeddings from Natural Language Processing (NLP), which are comprehensively reviewed in [30, 31, 32, 33]. From the perspective of lossy compression, despite a huge range of research that has both integrated and without knowledge representation to improve data reconstruction performance, no studies have surveyed these techniques in general and lossy compression-aided SemCom in particular.

In addition to the aforementioned dominant features of GAI, it also possesses significant merits in the cutting-edge area of security. For instance, in [34], Zhang et al. explore GANs in transportation, including autonomous driving, data completion, and traffic anomaly detection. In [35], Liu et al. explore how adversarial ML can be both a threat and a defense mechanism in mobile networks. In [36], Cai et al. provides a discussion of the potential benefits and challenges of GANs in the context of privacy and security in mobile networks. In [37], Deldjoo et al. present the role and application of GANs in improving data distribution learning of recommendation systems. In [38], Dunmore et al. survey and discuss the potential of GAN applications in enhancing cybersecurity in mobile networks, particularly in Intrusion Detection Systems (IDS). In [39], Lin et al. investigate and inspect the use of GANs in enhancing trajectory prediction, security, and anomaly detection of the wireless-based transportation domain. However, a recent investigation in [40] points out that the use of GAI exposes potential security vulnerabilities, and this will generate a double-edged sword for wireless security if GAI is not exploited properly.

A common idea in existing studies is that most focus is on clarifying either applications or challenges. The increasing emphasis on generative AI technologies is expanding into various applications and domains within mobile and wireless networking. However, current studies do not provide comprehensive GAI evaluation and its applications. To the best of the authors’ knowledge, the fundamentals and applications of GAI as outlined above have not been thoroughly examined in the literature. Consequently, we aim to bridge this gap by providing the necessary preliminaries to GAI and its key applications in the future of the IoT and mobile networking. This work centralizes all recent developments in GAI and complementary technologies and remarkable advancements in various aspects of mobile networks. At the same time, this work also seeks to gather the recent works in this trending research field and provide a unified repository. The notable contributions of our work are listed below.

• •

  We present an overview of GAI, covering its historical development, foundational structures (e.g., autoencoders, GANs, and diffusion models), data creation, and multimodal aspects (e.g., meta-learning and multi-tasking).

• •

  We summarize the contributions of GAI to network management, particularly in anomaly detection and security through its ability to learn and flag deviations.

• •

  We explore the applications of GAI in wireless security based on baseline models, detecting anomalies, fortifying against cyber threats, improving intrusion detection, and optimizing security policies.

• •

  We examine the applications of GAI in enhancing SemCom through NLP for human-like text, chatbots, virtual assistants, automated content creation, and improved language translation.

• •

  We reveal challenges of GAI in mobile networking, emphasizing technical complexities, legacy systems, ethical concerns, security, and privacy and necessitating comprehensive solutions.

In the study, Section II presents a preliminary to GAI, including the development history, data-processing and data-creation techniques, and representative models. In Section III, we comprehensively review the role of GAI in mobile network management. Section IV reviews the use of GAI in various wireless security issues, including data obscurity, intrusion detection, jamming attacks, and generative steganography, while Section V describes the applications of GAI for research advances in SemCom. Section VI discusses crucial challenges and explores potential research directions to facilitate applications of GAI techniques. We finally conclude the paper in Section VII.

AI has a long history and has developed a lot over many years. Due to big investments from governments and large companies, AI has reached important milestones and is used in almost every industry, including infotainment, finance, advertising, defense, agriculture, and telecommunications.

In general, AI techniques and their applications are designed by researchers in the field of mobile networking. AI-driven algorithms are used by researchers to analyze and optimize data traffic, improving network efficiency and performance, while automating network management tasks like configuration and troubleshooting [45]. These algorithms also facilitate the dynamic allocation of network resources based on real-time demand and usage patterns, optimizing resource utilization and enhancing network capacity and scalability [46]. To ensure network security, AI-powered IDS analyzes network traffic patterns in real time, identifying and mitigating potential threat, and supporting network resilience against cyber attacks and unauthorized access [47]. Additionally, AI-based encryption and authentication mechanisms ensure data confidentiality and integrity, protecting sensitive information transmitted over mobile networks from unauthorized access and tampering [48].

Despite showing great potential in various domains, conventional AI approaches may not be effective in highly dynamic and complex environments (e.g., future mobile and wireless networking), as discussed in Section I. To overcome this drawback, the potential of GAI has recently been demonstrated by advances in VAEs [59] and GANs [60]. These innovations were based on Deep Neural Network (DNN) architectures capable of learning generative models for complex data types like images. In 2017, Google introduced the Transformer architecture [61] that further enhanced the capability of GAI in learning generative models. This led to the development of the first Generative Pre-trained Transformer (GPT), as known as GPT-1, in 2018. Since then, GAI has received attention from both academia and industry, with various advantaged techniques and applications, e.g., DALL-E and GPT-4. The advantages and challenges of GAI use cases can be summarized in Table II. In the following, we present the fundamentals of common GAI techniques together with their potential in mobile and wireless networking areas.

One of the core objectives of generative adversarial learning-enabled trust management services for Beyond Fifth-Generation (B5G)/Sixth-Generation (6G) wireless networks is to enhance intelligence and security while ensuring reliable and real-time network security and service quality optimization [15]. The core elements inherent in the basic structure of GAI models empower them to provide robust and reliable services for both modern wireless service providers and users.

Multi-modal GAI encompasses a range of tasks, including but not limited to image recognition, semantic segmentation, NLP, speech synthesis, and emotion detection. Through the integration of diverse modalities like visual, auditory, and textual information, multi-modal GAI is capable of capturing a holistic representation of the world, mirroring the way humans perceive and interpret multiple sources of information concurrently [74]. In what follows, we explain how GAI deals with multi-tasking problems.

Across a diverse range of tasks and domains, several representative GAI models have demonstrated impressive capabilities. One such example is GPT-3, a renowned model developed by OpenAI. GPT-3 stands out for its exceptional language processing abilities, excelling in tasks like language generation, translation, summarization, and question-answering [88]. In addition, OpenAI’s DALL-E model is a specialized form of GAI that focuses on generating images from textual descriptions, effectively transforming natural language prompts into accurate visual representations [89]. It enables tasks such as image categorization, object detection, and text-based image synthesis. Contrastive Language-Image Pre-training, another creation from OpenAI, integrates language and image understanding in a cohesive manner [90]. These GAI models exemplify the potential for developing intelligent systems that excel across various activities and domains, showcasing cutting-edge progress in the field of AI.

Software-Defined Networking (SDN) represents a pivotal convergence point between GAI and network management. To be specific, SDN’s separation of the control plane and data plane facilitates centralized network management and programmability. Integrating GAI into SDN frameworks empowers network administrators with enhanced automation, intelligent resource allocation, and adaptable network optimization. Specifically, GAI is in charge of real-time decision-making for effective traffic routing, load balancing, and Quality-of-Service (QoS) management, enabling SDN controllers to dynamically learn from and adapt to changing network conditions. For example, DGMs are used in [24] to address the issues of data scarcity, limited flexibility, and dynamic network states using a diffusion model to generate effective contracts. In [91], NetGPT is proposed to employ LLMs at the edge and the cloud in conjunction with location-based information to speed up personalized procedures and cloud interaction, enhancing intelligent network management and orchestration.

Network slicing, a vital aspect of network management, has great potential to leverage GAI and opens a new chance to customize network instances within a shared infrastructure [92]. On this basis, GAI can exploit this property to perform network data analysis, pattern prediction, and generation of optimized network slices tailored to specific application requirements. In this way, mobile networks can dynamically allocate resources, balance workloads, and adapt to changing traffic demands, contributing to better user experiences, increased network efficiency, and greater flexibility in service provisioning [93].

In the context of dynamic resource allocation, GAI contributes to dynamically changing network conditions in real time through continuous monitoring and analysis of network data. This enables efficient utilization of resources, ensuring optimal performance and responsiveness to the evolving needs of various network factors such as network load, user demands, and QoS demands. It also helps systems use proper resource allocation strategies to ensure an optimal allocation that aligns with network objectives and user expectations.

For instance, the authors in [94] propose GAI-based resource allocation for the SemCom framework based on YOLOv7-X. This method helps reduce costs and improve information transfer in communication services by extracting the necessary semantic information from edge device images before deciding on resource allocation via confidence and AI-generated models. In [44], an AIGC-as-a-Service (AaaS) architecture is introduced to wireless edge networks, where GAI at edge servers is responsible for repairing damaged images and naturally creating high-resolution/augmented reality/virtual video content, reducing computational resources. Compared to DRLs, this architecture has not only a simple transmission protocol, but also an efficient and scalable method for mobile users to access intelligent services on demand without limited resource concerns.

To deal with the dynamics and uncertainty of network conditions, introducing GAI into network routing can be an innovative alternative to conventional AI approaches. Specifically, GAI models can assess the efficacy of various routing algorithms and optimize routing choices for certain network objectives by simulating and creating synthetic network scenarios. This guarantees effective resource utilization, low latency, and better overall network performance. Moreover, by analyzing network data to spot potential weaknesses and learn from historical security incidents, GAI helps navigate and enhance routing patterns to reduce the possibility of assaults or unauthorized access, proactively mitigating security concerns.

A typical example of this feature is the adoption of DGMs in network economics to improve the efficiency of wireless network management by exemplifying and generating effective contracts for incentivizing mobile AIGC services [95]. Another example of GAI in routing solutions for different network status distributions and topology is exploiting a transfer RL algorithm to improve training efficiency by rapidly transferring knowledge [96].

Numerous benefits can be achieved by incorporating GAI approaches into mobile networking channel estimation. First, GAI enhances the ability to detect faulty signals and provide appropriate fixes to improve channel estimation procedures. Second, GAI can streamline the resources used for channel estimation, which lowers the computational complexity and energy usage of conventional estimation techniques. Third, by learning the fundamental patterns and properties of wireless channels, GAI makes channel estimation more reliable and precise, particularly in difficult situations with little training data or time-varying channel circumstances.

To verify these promising traits, the work in [97] designs a high dimensional channel estimation approach based on a deep generative network combined with a compressive sensing strategy. The experiments from [97] indicate that GAI not only uses fewer channel pilots than conventional techniques, but also improves the execution time of channel estimation procedures for one-bit quantized pilot measurements. Similarly, another study for wideband channel estimation [98] has demonstrated the efficacy of GANs in dealing with channel estimation in the conditions of limited channel pilots and low transmit SNR. Specifically, it significantly reduces the number of pilots by about 70%, even at mmWave and THz frequencies, without affecting the estimation errors. On another front, the work in [99] displays the potential of deep score-based generative models to multi-input multi-output systems, with high fidelity channel estimation for channel communication of sizes 64×256 with pilot density up to 25%.

Modern networks have undergone a revolution because of the use of GAI in network management, which has made it possible to provide accurate and high-quality data. Among GAI techniques, GANs have become popular in many applications, such as picture synthesis, text generation, and data augmentation, providing innovative and varied training data for AI models. In wireless networks, GAI is crucial for allocating resources effectively, optimizing input data to cut transmission costs and increase accuracy. Additionally, generative models have solved practical issues in channel estimation by identifying network risks with greater precision and eliminating data imbalance problems.

The landscape of network routing presents challenges, encompassing aspects like security, user verification, and data integrity. Recent research indicates that the GAI is instrumental in addressing network routing challenges. It excels in enhancing security, conducting user verification, ensuring secure communication, and fortifying against advanced attacks. Future investigations should not be confined to refining GAI algorithms for optimized mobile network operations. Instead, they should focus on fortifying data security and continually enhancing user experiences. Moreover, the dynamic potential of GAI in mobile networking requires ongoing exploration for improved creativity and security in communication networks.

In summary, the key takeaways from this section are as follows.

• •

  GAI enhances network management by employing predictive models to optimize traffic flow and minimize downtime.

• •

  By dynamically adjusting computing resources based on real-time demand and ensuring optimal efficiency, the resource allocation capabilities are improved.

• •

  Channel estimation and security benefit are achieved from generative models with improved estimation and detection accuracy.

• •

  GAI can strengthen network security utilizing advanced anomaly detection algorithms to identify and mitigate potential threats.

• •

  GAI can contribute toward data confidentiality and integrity by implementing encryption protocols and monitoring data for unauthorized access or alterations.

Data sharing has become a popular trend in intelligent systems for training machine learning models. However, this activity poses numerous risks related to user privacy, in that sensitive information may be disclosed by third parties without unauthorized access or interception [100]. In this context, GAI has recently demonstrated its promising capabilities in obfuscating sensitive data by creating synthetic data that closely mimics real datasets [101].

There are multiple methods for generating synthetic data using GAI models, such as using GAN, VAE, or GPT models. In particular, data perturbation is one of GAI’s main techniques in safely hiding sensitive data [102]. This technique uses synthetic data to complicate extraction of useful information, by applying controlled perturbations to sensitive data [103]. In addition, distributing and transforming only the sensitive elements in datasets can also increase the security of data sharing while preserving the features and statistical utility of the original data for model training activities [104]. Moreover, GAI models can apply other techniques like differential privacy and data anonymization.

GAI techniques also aid in encrypted communication. Indistinguishable noise can be created from encrypted data using generative models, complicating decryption of the original data without the proper decryption keys [105]. With this strategy, sensitive data are sent more securely over wireless networks while maintaining their confidentiality and integrity. However, despite the apparent advantages, there are issues when using GAI to obfuscate sensitive data, such as the trade-off between data value and obscurity, the scalability of generative models of huge datasets, as well as the criteria and requirements for the reliability and efficacy evaluation of data obscurity strategies.

Intrusion detection in wireless security is essential for identifying and combating threats or intrusions on mobile networks. Applying GAI techniques opens new avenues to strengthen IDSs and enhance wireless network security as a whole [106]. A valuable tool for discovering anomalies and distinguishing malicious activity in wireless network traffic is provided by GAI models like GANs. These models are capable of learning the typical patterns and behaviors of network traffic through training on large-scale datasets, allowing them to detect unforeseen attack patterns from expected norms with high accuracy [107].

Generating synthetic data samples that closely resemble actual network traffic is one of the main benefits of utilizing GAI in intrusion detection. IDS is capable of accomplishing this by enhancing its training data and producing a variety of testing and evaluation situations. IDS can increase its robustness and adaptability to various sorts of threats by creating synthetic attack patterns to improve its detection capabilities. The early discovery of zero-day attacks, vulnerabilities or exploits of which security experts are unaware, can also be helped by GAI [108]. IDS can detect suspicious actions that depart from usual behavior by continually monitoring network traffic and analyzing patterns using generative models, acting as an early warning system for potential zero-day attacks.

Furthermore, the effectiveness and precision of IDS can be increased using GAI approaches. IDS can improve their detection accuracy and minimize the impact of false alarms by lowering false positives and false negatives [109]. Additionally, the use of generative models can lower the computing burden of IDS, enabling instantaneous network traffic monitoring and analysis [110]. However, integrating GAI into IDSs can be difficult. The necessity for extensive wide-ranging training datasets, the interpretability of generative models, and the adversarial robustness of the models against complex attacks are a few complications.

The integrity and availability of mobile networks are seriously threatened by jamming assaults in the area of wireless security. These assaults encompass deliberate disruption or interference of wireless signals, which results in a reduction in service or prevention of communication [111]. Innovative methods for countering and lessening the effects of jamming attacks in mobile networking are possible using GAI techniques. GANs can be useful in recognizing and preventing jamming attacks. These models can recognize irregularities brought on by jamming activities because they can learn and comprehend the typical patterns and properties of wireless signals [112]. GAI models can identify the presence of jamming signals by continuously observing and analyzing network traffic and can then activate the necessary countermeasures.

To create synthetic samples that closely match real signals, one method is to train GAI models on massive datasets of typical wireless signal patterns. Any appreciable differences between the generated samples and the received signals can be used to identify potential jamming assaults [40]. This improves the responsiveness of security systems by enabling the quick and precise detection of jamming actions. Intelligent jamming mitigation solutions can be developed with the help of GAI. Models can learn to identify and categorize many sorts of jamming attacks by being trained on a wide variety of signals [14]. Using this knowledge, adaptive and robust defenses can be created to dynamically modify their tactics to overcome particular jamming approaches.

Furthermore, jamming attack prediction and prevention can be accomplished through GAI approaches. These algorithms can identify potential weaknesses and forecast the risk of upcoming jamming occurrences by examining past data and patterns [40]. This proactive strategy enables network operators to implement preventive actions to lessen the effects of prospective jamming assaults, such as modifying frequencies or adding security measures. The use of GAI for jamming attack detection and mitigation is not without challenges. These include the need for substantial diverse training datasets, the ability of models to adapt to evolving jamming strategies, and the need for real-time processing for quick response [112]. These issues can be resolved and the defense against jamming assaults in mobile networking can be strengthened with additional studies and developments in GAI.

Data privacy and sensitive information security are of the highest priority for modern wireless networks. To improve data security in mobile networking, generative steganography combines the ideas of GAI and steganography to conceal sensitive information within apparently harmless wireless signals [113]. With the help of steganography, we can hide information such that undesirable recipients cannot detect it. GANs can be used to produce cover signals with secret information. These models study the statistical properties of typical wireless signals and employ this information to include encrypted data within them [114].

In terms of mobile networking security, generative steganography algorithms offer several benefits. They provide a covert way to send sensitive data without prompting suspicion [101]. Generative steganography overcomes usual detection techniques by hiding data within wireless signals that appear to be harmless. Additionally, generative steganography improves mobile networks’ resistance to unauthorized access and data breaches. Since the secret data are integrated into the cover signals, it is safe and impractical to decrypt without proper tools even if attackers intercept the transmission [115].

Furthermore, generative steganography offers a strong defense against assaults meant to compromise the secrecy and authenticity of wireless communications. The process of embedding and extracting concealed information can be extremely efficient and resilient to detection by unauthorized parties using GAI models [116]. However, some issues must be resolved for generative steganography to be used in real-world settings. The challenges include maximizing the capacity of concealed information while minimizing the influence on the cover signals through optimization of embedding algorithms and the technique’s resilience to various sorts of attacks.

Numerous applications of GAI in wireless security have demonstrated promise for bolstering mobile networking defenses against a variety of threats and weaknesses. Differential privacy and safe multi-party computing are two methods that have been successful in obscuring critical information and preserving user privacy. Network security is strengthened against DoS attacks and zero-day vulnerabilities due to the integration of GAI models into network IDS, which enables quick detection of unauthorized access attempts and security breaches. To guarantee network integrity, GAI’s proactive and dynamic intrusion detection algorithms are essential.

In addition, recent developments in GAI have proven crucial in the fight against jamming attempts that interfere with wireless communication. GAI improves malware detection and stops the propagation of threats by analyzing network data and device behavior to quickly identify and isolate infected devices. Also, by encrypting data and disguising it within seemingly innocent material, steganography’s novel application of GAI enables secure data transmission. Concurrently, biometric-based authentication that uses GAI strengthens identity verification processes by recognizing distinctive biometric patterns like fingerprints or facial features. These uses serve as a reminder of the important of GAI for improving data security and facilitating safe and secure user verification.

Mobile networking offers more robust security measures due to the adaptability of GAI in varied security concerns. GAI approaches hold promise to increase the state-of-the-art in mobile network security and communication and data transmission integrity by strengthening defenses against sophisticated attacks. In addition, our review uncovered numerous innovative research methods that further demonstrate the potential of GAI to offer a thorough defence against changing threats and improve the mobile networking security environment, as summarized in Table IV. The future of mobile networking security is bright due to ongoing developments in GAI, which will ensure a safer and more secure digital environment.

In summary, the key takeaways from this section are as follows.

• •

  GAI strengthens wireless security against anomalies and intrusions in real time.

• •

  GAI combats jamming attacks and ensures secure data transmission with privacy protection.

• •

  Biometric-based authentication and malware detection are reinforced using GAI techniques.

• •

  Ongoing research in this area provides more resilient security measures for mobile networking.

Knowledge abstraction is crucial for representing abstract knowledge among users and acquiring shared knowledge from distributed data across device domains. KG are the most prominent candidate among contemporary techniques for this purpose. However, manually constructing KGs is labor-intensive and impractical for scalability [130, 131]. Toward automatic KG construction, three FL-enabled designs for GAI deployment in wireless networks have been recently proposed [43]. The success of these design lines is discussed using three key knowledge techniques.

Data compression is vital in SemCom, particularly for communication with limited resources. The use of lossy compression is believed to be beneficial in improving data compression performance but at the expense of data transmission. With the emergence of SemCom and Goal-Oriented Semantic Communication (GOSC), the shortcomings of lossy compression have been overcome. To achieve this, two main objectives were addressed: improving data representation structure and implementing knowledge-aided techniques. Fig. 7 illustrates the classification of data compression techniques using GAI, potentially strengthening mobile networking robustness.

GAI has demonstrated remarkable robustness in enhancing performance in distributed learning. To be specific, GAI techniques produce generative data representations that extract generalized knowledge or meaningful features, supporting training on distributed local devices without requiring access to unseen data domains.

In FedGen [181], the server based on local model performance cooperatively trains a GAI model as a lightweight generator to amass user information and that is capable of generating pseudo data representation without accessing their true data (see Fig. 8). Thus, each local user can have the additional knowledge of other distributed clients, without sharing data among distributed devices. This approach is robust as this innovative approach can significantly enhance the performance and communication efficiency of FL, all while ensuring that user data remain secure and private. In [44], an AaaS framework based on AIGC models is designed to address the challenges in training and deployment overheads for edge networks.

Similar to FedGen, which leverages GAI to train a generative AI model, DENSE [182] designs a GAN-based data generation network that enables the generation of on-server data. In contrast to FedGen, where GAI is employed for on-server data generation, DENSE employs a multi-teacher knowledge distillation technique (see Fig. 9). This technique ensure that the generated data accurately captures the entire system’s data distribution and characteristics. The local models collaboratively instruct a student, denoted as the global model, inheriting knowledge from all teachers. Furthermore, this approach can be enhanced by incorporating Label-Driven Knowledge Distillation (LKD) [183]. This methodology enables the student to selectively learn the most meaningful knowledge from each teacher, significantly improving distillation efficiency compared to other conventional knowledge distillation approaches. Consequently, it is feasible to apply the ensemble learning on the server, boosting the FL efficiency while retaining the communication efficiency.

Different from other GAN approaches, the work in [184] adopts a distinct approach by treating model parameters as data. Specifically, the authors introduce HCFL, a technique capable of compressing and reconstructing model parameters. It enables the FL system to circumvent substantial communication overheads. In particular, applying HCFL to FL systems can maintain stable accuracy when the distortion rate of the aggregated model decreases proportionally with network size.

Aside from the aforementioned methods, invariant learning is a unique approach to enhancing learning on local devices in distributed learning. In invariant learning, users collect their local invariant representations or cross-domain invariant representations by sharing the most consistent format of data (invariant features) across devices. In this way, it minimizes the need for extensive communication resources, a representative form of knowledge-aided SemCom in distributed learning. For instance, Federated Cross-Correlation and Continual Learning (FCCL) [185] and FCCL+ [186] are innovative methods inspired by Barlow Twins principles, which help identify invariants across domains in the FL system (See Fig. 10). As an alternative method to discover invariant features, the authors in [187] focus on class-wise invariance by introducing cluster prototypes of contrastive learning. Unlike [185, 186, 187], FedPAC employs a structured causal model to acquire goal-oriented invariant representations, demonstrating robustness in learning personalized representations for each client without extensive knowledge sharing among devices [188].

GAI plays a crucial role in the development of SemCom, especially in generating data suitable for various subject-specific requirements. Consequently, applying GAI to various subdomains brings several benefits to SemComs, including creating more compact and structured data representations, significantly improving data compression efficiency in SemCom, and enhancing communication efficiency.

Furthermore, the robust development of GAI, particularly in summarization and knowledge extraction, enables the system to extract common knowledge from data. In other words, identifying elements that are domain-invariant or label-invariant (consistent regardless of changes in data) allows the system to leverage such invariant information multiple times without passing it through the physical communication channel. This leads to a substantial reduction in communication overhead without diminishing the amount of transmitted information.

In summary, the key takeaways from this section are as follows.

• •

  GAI markedly enhances compression performance through the creation of concise and organized data representations.

• •

  GAI empowers the SemCom capability to abstract knowledge effectively, facilitating high-performance data compression without sacrificing information.

• •

  GAI facilitates the generation of pseudo data on distributed devices, allowing these devices to adapt to the domains of other devices in distributed learning scenarios.

• •

  GAI can assist the encoder in generating invariant features from data, comprising the most meaningful features that contribute significantly to AI performance. This improvement leads to enhanced learning while concurrently achieving a substantial reduction in communication costs.

• •

  While current knowledge abstraction heavily relies on NLP models, there exists significant research potential in SemCom. There is a pressing need for the introduction of a generalized, scalable approach to abstract knowledge across types, eliminating the necessity of intermediate processes like NLP.

To enable new services and use cases for future mobile networks (i.e., SemCom, mobile edge computing, user behavior analysis, wireless signal processing, channel rendering, 3D mapping, and integrated sensing), GAI faces some technical challenges that need to be addressed. The details of each challenge and the corresponding solution are discussed in the following.

Handling technical complexity: GAI models often require powerful computing ability and data to train and fine-tune, necessitating mobile network components to sustain the trade-off between performance and cost. For instance, ChatGPT-3, a text generation tool, requires 175 billion parameters and hundreds of billions of words from the internet. In this context, exploiting cloud or edge computing services is crucial to reducing computation and storage processing tasks from mobile devices or networks. Also, integrating model compression approaches into GAI systems like pruning, quantization, or distillation are also needed to reduce model size and complexity without affecting performance.

Tackling the legacy system: The incompatibility of GAI models with existing network infrastructure and protocols can limit their integration and scalability. In particular, current mobile networks, developed based on centralized servers and fixed bandwidth resources, may not work well with the GAI model and can even generate risks involving network security, privacy, and reliability. To cope with these issues, there is a need for new network paradigms such as SDN and network slicing to provide more flexibility, efficiency, and customization for GAI applications. In addition to this, it is also necessary to develop new standards and regulations to ensure the interoperability of GAI models on mobile networks.

Avoiding technical debt: The risk of falling behind the latest developments and best practices in GAI can affect the quality and reliability of services. Mobile network operators must keep up with the latest research and innovation in the field and update the deployed AI models to avoid errors, bugs, or vulnerabilities that may arise over time. Following that, two possible solutions should be considered: (1) employing a continuous learning and improvement approach to monitor the performance and feedback of GAI models and to make adjustments accordingly and (2) collaborating with other stakeholders such as researchers, developers, or customers to share general AI knowledge and best practices.

Potential misuse and AI hallucinations: GAI models without control limitations can compromise the quality and reliability of content, potentially producing fake news, spam, phishing emails, and malware. Attackers can exploit these to manipulate users or compromise their security. Hence, feedback and evaluation mechanisms, such as digital watermarking or blockchain, are needed to verify the authenticity of GAI content, along with moderation or filtering techniques to detect harmful content. To further prevent security threats related to the generation of adversarial examples, GAI-based systems must have strong and flexible security measures, such as encryption, authentication, anomaly detection, and prevention. In particular, GAI models may require access to personal data without proper consent or protection, resulting in sensitive information leaks. To ensure data privacy, GAI-based mobile systems need to have mechanisms to encrypt, anonymize, or delete data after use.

Human-AI interaction: The integration of GAI into mobile networks to perform/replace some tasks can impact the skills and roles of the workforce, as well as the ethical and social aspects of employment, education, training, and motivation. This raises questions about GAI models’ accountability, transparency, fairness, and trustworthiness. To address this, human-in-the-loop or human-on-the-loop approaches are needed to create a tight connection between human workers and GAI in terms of design, operation, and evaluation. At the same time, building ethical principles and guidelines is critical to ensure the responsible and beneficial use of GAI in mobile networks as well as the shared values and norms of users and providers.

Lack of interpretability, another challenge of GAI for mobile networks, refers to lack of explaining and understanding the reasons behind GAI-based system outputs, particularly important in critical or sensitive tasks like network security, performance optimization, or service personalization. Interpretability acts as a bridge between mobile users and providers, fostering trust and confidence. For example, errors, failures, or unexpected outputs produced by GAI-based systems (i.e., generating harmful or biased content), interpretability becomes a powerful tool to help users and providers debug and improve the quality of GAI models. When there are growing concerns regarding ethical or legal issues (e.g., data violations, copyright infringement, or misinformation), interpretability will be the basis to assist users and providers in solving accountability and responsibility problems. It is not straightforward to obtain these benefits due to the following reasons.

Complexity: GAI-based systems are often developed based on complex and non-linear algorithms. For example, DNN contains a lot of hidden layers, parameters, and interactions that are difficult to interpret or visualize. Hence, GAI-based systems should adopt simpler or more transparent algorithms, such as linear models, decision trees, or rule-based methods.

Diversity: As the inputs of GAI models are diverse and heterogeneous (i.e., text, audio, video, or code), it is hard to perform comparison and integration due to their different formats, structures, and even semantics. A possible solution to overcome such challenges is to exploit visualization features to enhance the diversity of GAI-based systems due to graphical and interactive tools (e.g., charts, diagrams, or dashboards), which allow analysis and visualization of the data or outputs of GAI models more intuitively or understandably.

Uncertainty: Most GAI models rely on probabilistic and stochastic models to produce estimated outputs, such as GANs. Therefore, the randomness, variability, and noise during the training process complicate the measurement, processing, and control. The use of quantification features, in this case, is a reasonable choice to mitigate the uncertainty of GAI-based systems, e.g., numerical or statistical approaches such as metrics, scores, or confidence intervals, as well as in estimating the accuracy and reliability of the outputs.

Ethical and legal concerns arise when applying GAI to mobile networks as they relate to the ethical and legal implications of creating, using, and sharing new content. These may affect the rights, interests, values, and responsibilities of network stakeholders, such as developers, users, operators, and regulators; the law; and society in general.

Authenticity: In mobile networking, GAI-generated content introduces concerns regarding the authenticity and reliability of the information. This requires measures like digital watermarking to establish source identity and origin identification as well as digital verification techniques (i.e., cryptography or statistics) to ensure content integrity. More importantly, mobile network users need to be further educated with digital literacy-oriented programs to have a basis for evaluating and verifying GAI-generated content and its sources.

Ownership: The emergence of creative content from GAI systems raises concerns about ownership, authorship, and intellectual property rights in mobile networks. Digital rights management solutions become essential to protect and manage these rights, ensuring fair use and preventing unauthorized distribution. GAI-based systems should incorporate digital attribution mechanisms to acknowledge the contributions of original data sources from wireless networks and mobile devices. Furthermore, the development of digital licensing systems is crucial for specifying and enforcing terms and conditions governing the use, modification, and distribution of GAI-generated content.

Fairness: GAI systems raise concerns about bias in the contents generated from data sets, algorithms, or even with creators. Ensuring fairness characteristics of diversity and inclusion requires long-term investments during the deployment of GAI systems for mobile networks. In that, implementing fairness mechanisms, including auditing, debiasing, and calibration, becomes crucial to detect and rectify biases or errors. In addition to this, establishing and enforcing fairness policies are vital steps in defining rules and obligations for GAI systems in active mobile networking environments.

Responsibility: Impactful information generated by GAI-based systems raises concerns about responsibility and accountability. Determining who is liable for the outcomes of GAI systems and how to ensure adherence to ethical and legal standards becomes paramount, particularly for mobile networking applications. Addressing these concerns is not limited to responsibility mechanisms (e.g., feedback, reporting, and review processes) to monitor and assess the impacts of GAI systems but also involves cooperation between developers and users in establishing policies to uphold ethical norms and regulations in the mobile networking context.

Although GAI gains significant enhancements in the performance and efficiency of mobile networks, there are various challenges and risks associated with integrating GAI into mobile networks, especially in terms of security, privacy, and trust. In what follows, we will discuss the challenges and the respective solutions for these issues.