PaperlayerView exact revision on arXiv

Source-linked AI summary

Unleashing the Power of Edge-Cloud Generative AI in Mobile Networks: A Survey of AIGC Services

Minrui Xu, Hongyang Du, Dusit Niyato, Jiawen Kang, Zehui Xiong, Shiwen Mao, Zhu Han, Abbas Jamalipour, Dong In Kim, Xuemin Shen, Victor C. M. Leung, H. Vincent Poor

arXiv:2303.16129v4cs.NI

TL;DR

Deploying generative AI at mobile edges promises personalized services but raises infrastructure, resource, security, and privacy challenges. This survey synthesizes mobile AIGC architectures, service lifecycles, applications, and open issues, concluding that deployment requires coordinated cloud-edge-mobile technologies and further research.

  • Problem

    Mobile AIGC networks face deployment challenges spanning resource constraints, infrastructure coordination, security, and privacy protection.

  • Method

    The paper surveys generative models, mobile AIGC service lifecycles, collaborative cloud-edge-mobile infrastructure, applications, use cases, and deployment challenges.

  • Results

    The survey identifies implementation, security, privacy, and resource-allocation challenges and organizes future research directions for realizing mobile AIGC networks.

  • Takeaways & Limitations

    Mobile AIGC research must address mobility, limited resources, privacy concerns, and cloud-edge collaboration across diverse applications and service stages.

Abstract

from arXiv · show

Artificial Intelligence-Generated Content (AIGC) is an automated method for generating, manipulating, and modifying valuable and diverse data using AI algorithms creatively. This survey paper focuses on the deployment of AIGC applications, e.g., ChatGPT and Dall-E, at mobile edge networks, namely mobile AIGC networks, that provide personalized and customized AIGC services in real time while maintaining user privacy. We begin by introducing the background and fundamentals of generative models and the lifecycle of AIGC services at mobile AIGC networks, which includes data collection, training, finetuning, inference, and product management. We then discuss the collaborative cloud-edge-mobile infrastructure and technologies required to support AIGC services and enable users to access AIGC at mobile edge networks. Furthermore, we explore AIGCdriven creative applications and use cases for mobile AIGC networks. Additionally, we discuss the implementation, security, and privacy challenges of deploying mobile AIGC networks. Finally, we highlight some future research directions and open issues for the full realization of mobile AIGC networks.

I. INTRODUCTION … C. Related Works and Contributions

AIGC automates diverse content creation, while mobile AIGC networks aim to provide real-time, personalized, localized, and privacy-preserving services despite edge-resource and implementation challenges. This survey covers AIGC fundamentals, collaborative mobile-edge-cloud infrastructure, applications, use cases, challenges, and future directions.

  • B. Motivation: Mobile AIGC networks address cloud services’ limitations by moving fine-tuning and inference toward edge servers and mobile devices for low-latency, interactive service access [1],,,.Cloud data centers can pre-train models such as GPT-3 and GPT-4, while edge deployment supports real-time access.
  • B. Motivation: Edge-based AIGC enables localization, mobility-aware provisioning, customization, personalization, and privacy preservation by adapting services to geographic contexts, user preferences, and local requirements [3],,.Users submit service requests to edge servers rather than sending preferences to cloud servers in the core network.
  • B. Motivation: Mobile AIGC deployment must balance accuracy, latency, and energy under limited edge resources, while addressing technical maturity, transparency, robustness, impartiality, insightfulness, and bias-related fairness risks.Computationally intensive tasks can be offloaded from mobile devices to edge servers, but biased raw data can produce unfair algorithmic results.
  • C. Related Works and Contributions: Distinct from existing surveys, this survey focuses on deploying mobile AIGC networks for real-time, privacy-preserving service provisioning through collaborative mobile-edge-cloud communication, computing, and storage infrastructure.It introduces AIGC definitions, lifecycles, models, metrics, and provisioning technologies.
  • C. Related Works and Contributions: The survey examines creative AIGC applications and use cases spanning text, images, video, and 3D generation, and summarizes the advantages of constructing mobile AIGC networks.The development roadmap traces AIGC from text and audio toward 3D content while computing shifts from cloud data centers toward mobile devices.
  • C. Related Works and Contributions: It identifies implementation challenges from dynamic channel conditions, meaningless or insecure content, and privacy leaks, then discusses open issues in networking and computing, ML, and practical implementation.The survey is organized around fundamentals, infrastructure, applications, use cases, implementation challenges, and future research directions.

II. BACKGROUND AND FUNDAMENTALS OF AIGC … 3) AIGC:

The section distinguishes PGC, UGC, and AIGC as three primary content forms, then characterizes AIGC as automatically generated, creative, multimodal, diverse, and socially valuable content produced from user inputs. It also frames the section around AIGC’s definition, classification, lifecycle, and ChatGPT as a use case.

  • II. BACKGROUND AND FUNDAMENTALS OF AIGC: The section introduces AIGC’s definition, classification, technological lifecycle in mobile networks, and ChatGPT as a prominent use case.
  • A. Definitions of PGC, UGC, and AIGC: Web 3.0 and the Metaverse feature three primary content forms: professionally-generated content (PGC), user-generated content (UGC), and AIGC.
  • 1) Professionally-generated Content:: PGC is professional-generated digital content created by skilled individuals or organizations, offering automation and multimodality but potentially limited diversity and creativity.
  • 2) User-generated Content:: UGC is digital material created by users rather than experts or organizations, spanning text, photos, video, and audio as creation barriers decline.
  • 3) AIGC:: AIGC uses generative AI models to learn patterns from user inputs and produce varied content, with diffusion-based text-to-image systems [46] and transformer-based ChatGPT driving recent attention.
  • 3) AIGC:: AIGC is automatically produced from user inputs, supports creative prompt-driven generation, and handles multimodal inputs and outputs such as text, images, voice, and 3D models.Specific prompts can improve originality and relevance; examples include ChatGPT for text conversations and DALL-E 2 [48] for text-to-image generation.
  • 3) AIGC:: AIGC supports personalized and customized outputs, generates diverse results, and should provide extended value to society, economics, and humanity.Examples include representing global population diversity in DALL-E 2 outputs and supporting medical-report writing and medical-image interpretation.

B. Serving ChatGPT at Mobile Edge Networks … 1) Data Collection:

Serving ChatGPT at mobile edge networks follows four stages—pre-training, fine-tuning, inference, and product management—while AIGC data collection shapes model quality and diversity. The lifecycle uses diverse collection methods, including crowdsourcing, data markets, IoT sensing, and passive edge-network sensing, before model training.

  • B. Serving ChatGPT at Mobile Edge Networks: ChatGPT deployment comprises pre-training, fine-tuning, inference, and product management stages.These stages structure the development and operation of ChatGPT in mobile edge networks.
  • 1) Pre-training:: Pre-training uses large text corpora to train GPT-3, a 175-billion-parameter autoregressive Transformer language model that learns linguistic patterns and relationships.The corpus includes books, articles, and other information sources.
  • 2) Fine-tuning:: Fine-tuning adapts ChatGPT to specific tasks or domains through supervised dialogue learning and response-quality ranking for reinforcement learning.This process aims to improve task-specific accuracy and relevance and supports conversational AI.
  • 3) Inference:: Inference generates coherent, contextually relevant responses from prompts by leveraging pre-training and fine-tuning knowledge and analyzing dialogue history and user profiles.In-context learning analyzes the entire input context [54].
  • 4) Product Management:: Product management deploys ChatGPT in production and maintains efficient operation, with mobile-edge applications such as new Bing and Office 365 Copilot.These tools are described as providing personalized, contextually appropriate responses while conserving resources.
  • 4) Product Management:: New Bing provides detailed replies, summarized answers, personalized follow-ups, creative writing assistance, and consolidated web sources for mobile-edge users.Its summarized-answer capability may help users with limited resources.
  • 1) Data Collection:: Data collection determines the quality and diversity of AIGC outputs by shaping the patterns and relationships learned during model training.Methods include crowdsourcing, purchased datasets, IoT sensing of GPS and wireless data, and passive edge-network sensing for applications such as air-quality, traffic, and pedestrian insights; collected data then trains the generative model.

2) Pre-training: … 2) Frechet Inception Distance:

The mobile AIGC lifecycle spans centralized pre-training, edge-based fine-tuning and inference, and management of generated products, while model quality is assessed using metrics such as IS and FID.

  • 2) Pre-training:: Pre-training uses centrally collected data and powerful servers to learn data patterns and predict target outcomes through models including GANs, VAEs, flow-based models, and diffusion models.
  • 3) Fine-tuning:: Fine-tuning adapts a pre-trained model to new tasks or domains with modest additional data, and mobile edge networks can perform it using small datasets uploaded by users.
  • 4) Inference:: Inference generates requested content from inputs; deploying services at the edge is intended to reduce centralized request congestion and optimize latency.
  • 5) Product Management:: AIGC product management preserves and manages generated content’s ownership, copyright, and value, with producers such as users or companies obtaining products for distribution and trading.
  • III. TECHNOLOGIES AND COLLABORATIVE INFRASTRUCTURE OF MOBILE AIGC NETWORKS: Mobile AIGC infrastructure is designed around collaborative edge computing systems whose performance metrics indicate whether they maximize user satisfaction and utility.
  • A. Evaluation Metrics of Generative AI Models and Services: Generative AI model and service quality metrics help providers and users assess AIGC capabilities in mobile networks.
  • 1) Inception Score:: Inception Score evaluates generated-image accuracy by combining high class probabilities with low KL divergence from a reference class distribution.
  • 2) Frechet Inception Distance:: Frechet Inception Distance evaluates GAN image quality and diversity by measuring distances between real and synthetic image embeddings from a pre-trained Inception network.

3) R-Precision: … 4) Flow-based Generative Models:

The section presents evaluation metrics for mobile AIGC and surveys foundational generative models, emphasizing how these metrics assess text-image alignment, user satisfaction, and model behavior. It also outlines the architectures, benefits, and limitations of GANs, energy-based models, VAEs, and flow-based models.

  • 3) R-Precision:: R-Precision measures the proportion of relevant text items among the top-R retrieved from 100 candidates using an AI-generated image as the query, typically with R=1 and DAMSM.It evaluates alignment between generated images and text inputs.
  • 4) CLIP-R-Precision:: CLIP-R-Precision replaces DAMSM with CLIP to provide a more objective, less model-specific text-image retrieval assessment.CLIP uses contrastive learning on large-scale image-caption pairs to align visual and linguistic embeddings.
  • 5) Quality of Experience:: QoE evaluates mobile AIGC user satisfaction through visual quality, relevancy, and utility, using surveys, interactions, and behavioral data.Its definition can vary with the mobile network designer’s objectives and user needs.
  • B. Generative AI Models: The survey introduces five fundamental generative model families: GANs, energy-based models, VAEs, flow-based models, and diffusion models.These models aim to learn input-data distributions through iterative training and generate novel data aligned with those distributions.
  • 1) Generative Adversarial Networks:: GANs use competing generative and discriminative networks to produce realistic data, supporting image synthesis and text-to-image translation.Adversarial training can improve both networks, but GANs are difficult to train because competition makes training unstable and slow, and they mainly augment existing data rather than create entirely new multimodal content.
  • 2) Energy-based Generative Models:: Energy-based models represent configurations with energy values, optimizing them so correct latent values receive lower energy than incorrect values [72].They are described as comprehensible, flexible, and capable of generalization for AIGC services.
  • 3) Variational Autoencoder:: VAEs comprise encoder and decoder networks: the encoder maps inputs to latent-space mean and variance parameters, while the decoder generates data from sampled latent variables.The passage contrasts their training method with the supervised approach attributed to GANs.
  • 4) Flow-based Generative Models:: Flow-based generative models use probabilistic flow formulations and backpropagation-based gradients, enabling efficient learning and direct probability-density computation during generation.The passage identifies computational efficiency as a benefit in mobile edge networks.

5) Generative Diffusion Models: … 2) Edge-Mobile Collaborative Inference for AIGC Services:

The paper surveys diffusion models, large language models, and the collaborative cloud-edge-mobile infrastructure for delivering personalized, low-latency AIGC services while addressing resource, communication, mobility, and privacy constraints.

  • 6) Large Language Models:: Large language models trained on billions of parameters and large-scale datasets [75] understand prompts, generate human-like text, and can support multimodal reasoning and embodied tasks, as demonstrated by PaLM-E [76].PaLM-E transfers knowledge across domains for robot planning and embodied question answering.
  • C. Collaborative Infrastructure for Mobile AIGC Networks: Mobile AIGC networks combine AI-generated content with mobile edge networks to enable rapid content creation, delivery, and processing at the network edge for improved user experience and reduced latency.The architecture supports pre-training, fine-tuning, and inference across collaborative infrastructure.
  • 1) Cloud Computing:: Cloud computing centrally supplies servers, storage, and databases for AIGC data collection, training, fine-tuning, and inference, with IaaS, PaaS, and SaaS providing infrastructure, development platforms, and applications.Users access these services through the core network rather than maintaining physical infrastructure.
  • 2) Edge Computing:: Edge computing places limited computing and storage near the RAN, supporting real-time fine-tuning and inference rather than generative-model training, while reducing latency, bandwidth use, and privacy exposure.Location-aware and customizable edge delivery can improve user experience compared with centralized cloud delivery.
  • 3) Mobile Computing:: Mobile computing enables devices to run generative models locally or offload AIGC services to edge and cloud servers, trading device computation and energy for lower latency and stronger privacy when processing locally,.This device-level execution forms the mobile layer of the collaborative infrastructure.
  • 1) Cloud-Edge Collaborative Training and Fine-tuning for Generative AI Models:: Cloud-edge collaborative training assigns resource-intensive pre-training to cloud data centers and uses edge data for more customized, personalized fine-tuning, balancing real-time interaction and privacy against edge computing and storage limits.The approach also faces substantial communication and bandwidth requirements.
  • 2) Edge-Mobile Collaborative Inference for AIGC Services:: Because user location and mobility change over time, dynamic edge-mobile collaboration is required for inference, with federated learning and distributed training among techniques for handling changing service-forwarding requirements.Wireless communication, computing, storage, and infrastructure compatibility also constrain LLM deployment in wireless networks [77], [78].

IV. HOW TO DEPLOY AIGC AT MOBILE EDGE NETWORKS: APPLICATIONS AND ADVANTAGES OF AIGC … 5) AI-generated 3D:

The section surveys mobile-edge AIGC applications across text, audio, images, video, and 3D, emphasizing resource-efficient models, network benefits, creative capabilities, and associated ethical or legal challenges.

  • 1) AI-generated Texts:: ALBERT and MobileBERT reduce computational or memory demands while retaining strong language-processing performance, making them suitable for smartphones, IoT devices, and other edge deployments,.ALBERT performs comparably to BERT on question answering and sentiment analysis, while MobileBERT uses a compact design and quantization.
  • 2) AI-generated Audio:: Audio AIGC improves mobile-network call quality, transmission efficiency, automation, personalization, security, and accessibility, while Audio Albert reduces average inference time by 20%,.Applications include speech synthesis, enhancement, recognition, compression, and speech-to-text transcription.
  • 3) AI-generated Images:: AI-generated images support enhancement, compression, recognition, and text-to-image creation, while Make-a-Scene and SPADE enable realistic generation, translation, inpainting, and attribute-controlled editing,.These capabilities can support visual storytelling, advertising, road-map representation, object detection, facial recognition, and image search.
  • 3) AI-generated Images:: AI-generated images also create deep-fake risks by depicting nonexistent events or individuals, potentially disrupting mobile-user tasks and raising ethical and legal concerns requiring further study and legislation.The passage identifies deep fakes as a limitation of image-generation technology rather than a demonstrated quantitative result.
  • 4) AI-generated Videos:: AI-generated videos support compression, enhancement, summarization, and synthesis, offering more immersive experiences and customizable style, resolution, and frame rate [95],.Imagen Video generates high-definition videos from text using cascaded spatial and temporal video super-resolution models [13].
  • 5) AI-generated 3D:: AI-generated 3D content supports AR/VR applications and latency reduction through optimal base-station placement, using complementary techniques such as Latent-NeRF, LPD, and Diffusion-SDF –[135].Latent-NeRF supports reconstruction, scene understanding, and shape editing; LPD provides diverse shapes and fine details; Diffusion-SDF generates detailed shapes from natural-language descriptions.

B. Advantages of Mobile AIGC … A. AI-Generated Incentive Mechanism

Mobile AIGC improves mobile networks through efficiency, reconfigurability, accuracy, scalability, sustainability, and potential security and privacy benefits, while also introducing associated risks. A diffusion-model-based AI-generated contract algorithm addresses utility maximization in mixed-reality semantic communication and achieves higher utility than SAC and PPO through improved sampling quality and long-term dependence processing [143].

  • 1) Efficiency:: Generative AI improves mobile-network efficiency by automating content creation, reducing human labor, boosting productivity, and supporting edge deployment of generated audio models such as Audio Albert,,, –.The cited applications include AI-generated text models ALBERT and MobileBERT and edge-implementable audio generation.
  • 2) Reconfigurability:: AIGC is reconfigurable because it generates diverse content that can be adjusted to changing network demands and user preferences, including through image and audio-generative models such as Make-a-Scene and SPADE, [140].ChatGPT illustrates adaptable content generation for evolving requirements.
  • 3) Accuracy:: Generative AI enhances prediction, decision-making, and the quality and accuracy of network-provided content, enabling more accurate and efficient services tailored to diverse mobile users [47],.The applications span AI-generated visuals and audio for domains including advertising and entertainment.
  • 4) Scalability and Sustainability:: AIGC supports scalability and sustainability by generating broad content ranges while reducing reliance on human resources and streamlining production from initial capture through retouching [13],.
  • 5) Security and Privacy:: Embedding sensitive information in generated content can provide steganographic security and privacy benefits, but AIGC also faces risks including adversarial attacks and malicious misuse.
  • A. AI-Generated Incentive Mechanism: The case study uses full-duplex device-to-device semantic communication to share generated content in mixed-reality environments, reducing redundant in-view-image generation despite HMD processing limits [143],.The approach targets utility maximization through optimal contract designs for constrained mixed-reality systems, [163].
  • A. AI-Generated Incentive Mechanism: The diffusion-based contract algorithm improves incentive design through iterative denoising rather than neural-network backpropagation or DRL parameter optimization [1] [143].Its performance is attributed to higher-quality sampling with diffusion step 10 and multiple refinement steps, plus enhanced long-term dependence processing across additional time steps,.
  • A. AI-Generated Incentive Mechanism: Utility reached 189.1 with the AI-generated contract algorithm, exceeding SAC’s 185.9 and PPO’s 184.3 in a specific environmental state [143].The evaluation is shown in Fig. 10 and demonstrates superior performance over traditional DRL incentive-design schemes.

B. AIGC Service Provider Selection

This section frames ASP selection as a resource-constrained task-assignment problem in edge-based AIGC-as-a-service and presents SAC-based deep reinforcement learning to maximize user utility while avoiding provider overload. It also identifies practical and future challenges involving adaptation, privacy, security, and multi-objective optimization [138].

  • AIGC Service Provider Selection: Edge-deployed AIGC-as-a-service reduces latency and resource consumption by enabling users to access AI models through wireless networks [138].The architecture places ASP models on edge servers for instantaneous service delivery.
  • AIGC Service Provider Selection: ASP selection assigns sequential user tasks to available providers with distinct utility functions under resource constraints, where violations can crash providers and restart running tasks [138].The objective is to improve mobile-user QoE by matching task requirements with providers’ model capabilities and computational resources.
  • AIGC Service Provider Selection: SAC-based DRL outperforms overloading-avoidance, random, and round-robin policies and approximates the optimal ASP-selection policy [138].The comparison is shown through cumulative rewards under different ASP selection algorithms.
  • AIGC Service Provider Selection: The DRL simulation models 20 ASPs and 1,000 edge users with randomly varying provider capacities and asynchronously arriving AIGC requests.ASP capacity ranges from 600 to 1,500 diffusion timesteps per time frame.
  • AIGC Service Provider Selection: Future work includes federated or distributed training, adaptive DRL and meta-learning, real-world latency, privacy and security constraints, and multi-objective optimization of quality, energy, and cost.These directions target lower communication overhead and better adaptation to changing network conditions and user requirements.

C. Generative AI-empowered Traffic and Driving Simulation · D. Blockchain-Powered Lifecycle Management for AI-Generated Content Products

Generative AI reduces the cost of traffic-data collection and labeling while enabling resource-efficient vehicular Metaverse simulation. A blockchain-based AIGC lifecycle framework coordinates stakeholders, protects ownership and exchanges, and selects service providers through reputation.

  • C. Generative AI-empowered Traffic and Driving Simulation: Generative AI synthesizes traffic and driving data, reducing the costly data-collection and labeling burden that hinders fully automated transportation.The approach opens a vehicular Metaverse paradigm in which data and resources are used more efficiently.
  • C. Generative AI-empowered Traffic and Driving Simulation: MTEPViSA prices roadside-unit resources through online and offline submarkets for heterogeneous digital-twin tasks with differing resource demands and deadlines.Connected autonomous vehicles offload digital-twin tasks to roadside units for real-time remote execution.
  • C. Generative AI-empowered Traffic and Driving Simulation: The AIGC-empowered mechanism doubles total surplus versus PViSA and EPViSA across different numbers of autonomous vehicles.The mechanism coordinates supply and demand in vehicular Metaverse markets to improve market efficiency.
  • D. Blockchain-Powered Lifecycle Management for AI-Generated Content Products: The blockchain framework assigns producers, ESPs, consumers, and attackers distinct roles across AIGC generation, ownership, trading, and disruption attempts.Producers propose prompts and retain ownership, ESPs generate content for fees, consumers trade products, and attackers target ownership or plagiarism vulnerabilities.
  • D. Blockchain-Powered Lifecycle Management for AI-Generated Content Products: The blockchain provides a traceable, immutable ledger validated by consensus, with ESPs acting as full nodes and producers and consumers as clients.The platform combines ledger functionality with on-chain mechanisms to secure and trace AIGC transactions.
  • D. Blockchain-Powered Lifecycle Management for AI-Generated Content Products: Proof-of-AIGC records products and supports challenges against suspected plagiarism, while HTLC-based incentives protect exchanges of funds and ownership.Challenges require a pledged deposit and involve proof retrieval, identity verification, similarity measurement, and result checking.
  • D. Blockchain-Powered Lifecycle Management for AI-Generated Content Products: Reputation-based ESP selection ranks providers and assigns tasks to trustworthy services, while negative reputation can reduce earnings and motivate prompt, honest execution.The scheme uses local and recommended producer opinions to calculate reputation and accommodate heterogeneous ESPs.
  • D. Blockchain-Powered Lifecycle Management for AI-Generated Content Products: Traditional ESP-selection methods produce uneven workloads and extended service latencies, motivating reputation-based allocation in the demonstrated AIGC lifecycle framework.The demonstration uses three ESPs and three producers with Draw Things, and reports reputation trends and assigned-task workloads in Fig. 15 and Fig. 16.

VI. IMPLEMENTATION CHALLENGES IN MOBILE AIGC NETWORKS · A. Edge Resource Allocation

Mobile AIGC networks face substantial implementation challenges because generative services are computation- and storage-intensive, requiring cloud-edge-mobile collaboration and careful allocation of constrained resources. Edge resource allocation must balance model accuracy, bandwidth utilization, resource consumption, dynamic user demands, and distributed model execution.

  • VI. IMPLEMENTATION CHALLENGES IN MOBILE AIGC NETWORKS: Generative AIGC services require substantial computation and storage, creating implementation challenges for existing mobile edge infrastructure and motivating cloud-edge-mobile collaborative architectures.
  • A. Edge Resource Allocation: AIGC provisioning is computationally and communication-intensive for resource-constrained edge servers and mobile devices, with model accuracy and resource consumption as common evaluation metrics,.Users send allocation requests to edge services, which execute AIGC tasks and return outputs.
  • A. Edge Resource Allocation: Resource allocation must optimize personalized and customized model accuracy while fully utilizing network resources, making evaluation more complex than traditional recognition and classification optimization.
  • A. Edge Resource Allocation: Edge servers must maximize bandwidth utilization by controlling channel access, reducing interference among user requests, and maintaining AIGC service quality in dense networks.
  • A. Edge Resource Allocation: AIGC training, inference, and continuous iteration consume substantial heterogeneous edge resources, making it difficult to balance unstable generation quality against resource consumption.User-dependent accuracy requirements can change the preferred trade-off between service quality and resource use.
  • A. Edge Resource Allocation: Transfer learning and model compression can improve accuracy with fewer resources or reduce model size, but personalization and customization make their applicability and accuracy evaluation unpredictable.
  • A. Edge Resource Allocation: Dynamic networks and user requirements require jointly considering model accuracy, networking, communication, and computation resources; a threshold-based collaborative-learning method reduced traffic effectively across system settings and data distributions.
  • A. Edge Resource Allocation: Frequent fine-tuning and retraining, limited edge storage, and heterogeneous customization demands motivate service-placement optimization, while partitioning large models enables local or partially offloaded collaborative execution,,,,.The placement and allocation problem can be formulated as an MINLP to minimize total time and energy consumption; distributed execution also requires effective model distribution and result aggregation.

B. Task and Computation Offloading

Task and computation offloading enables mobile AIGC networks to overcome mobile devices’ limited computation and battery resources by using nearby edge servers, while balancing latency, reliability, energy, task drops, privacy, and service quality. Effective cloud-edge collaboration requires adaptive offloading, model partitioning, distributed training, and task assignment tailored to service, user, resource, and network conditions.

  • Motivation and KPIs: Offloading generative-AI training, fine-tuning, and inference from mobile devices to nearby edge servers addresses resource constraints and can reduce service latency, despite transmission overhead.Mobile devices face limited processing power and battery life, while edge execution introduces additional transmission latency compared with local execution.
  • Cloud-edge collaboration: Cloud-edge collaboration uses federated learning and distributed training to aggregate locally updated model weights, while MINLP formulations can minimize provisioning delay during inference.These approaches support collaborative training and fine-tuning across cloud and edge servers, [192]–.
  • Offloading optimization: Adaptive translation and predictive offloading with random-forest regression extend deterministic designs toward dynamic intelligent-IoT environments, while optimization can formulate task execution as MINLP.The cited work automatically and dynamically offloads applications before making predictive decisions; provisioning-delay minimization is formulated as MINLP.
  • System factors: Offloading decisions must jointly account for service type, user characteristics, model complexity, computational resources, and network conditions because real-time and offline AIGC services have different requirements, –.Cloud-edge intelligence combines edge servers’ low latency with cloud servers’ high-quality services, and Neurosurgeon selects partition points using architecture, hardware, network, and server-load information.
  • Reliability and multi-user coordination: Reliable multi-task offloading requires efficient task assignment and dependable transmission, including redundant pathways that mitigate congestion or failures and reinforcement-learning policies optimizing latency, energy, task drops, and privacy.The cooperative offloading problem is modeled as an MDP whose state includes current tasks, local loads, and edge loads; agents select processing locations to maximize multiuser QoE.

C. Edge Caching · D. Mobility Management

Mobile AIGC networks use edge caching to reduce access latency and backhaul traffic while managing limited memory, model misses, and multiple cooperating models. Mobility management extends service coverage through vehicles and UAVs but must satisfy task completion, QoS, communication, computation, and privacy requirements in dynamic environments.

  • C. Edge Caching: Caching generative AI models at edge servers and mobile devices enables users to access AIGC without cloud data centers, but caches also require execution computing resources, [215].Unlike traditional content caches, generative-model caches must support model execution.
  • C. Edge Caching: Edge caching is evaluated by model access latency, backhaul traffic load, and model hit rate, with device caching providing the lowest model access latency.Latency also includes wireless-network delay for edge servers and core-network delay for cloud services; edge caching avoids core-network transfer of requests and results.
  • C. Edge Caching: Limited edge GPU memory makes caching all models infeasible, while cache misses require cloud downloads and functionally equivalent models complicate preload and eviction decisions,,, [218].Multiple cooperating base models are increasingly needed for classification, recognition, and multimodal generation.
  • C. Edge Caching: Model-aware eviction reduces model load delay by 1/3 versus a non-penalty-aware policy while managing heterogeneous requests and unpopular models [209].The policy uses model utility based on cache-miss penalty and request volume.
  • D. Mobility Management: Mobility management uses vehicles and UAVs to extend mobile AIGC coverage, leveraging UAV line-of-sight links and reconfigurable infrastructure for flexible deployment –,.Vehicles and UAVs can provide edge intelligence, generative models, content, and computing or caching services.
  • D. Mobility Management: Mobility-aware AIGC provisioning must complete tasks before users leave base-station coverage and optimize task accomplishment, coverage, latency, energy, and QoS under resource constraints,,.A joint vehicle-edge inference framework reduces DNN execution latency and energy consumption, while distributed scheduling addresses dynamic vehicle-network topologies.
  • D. Mobility Management: Highly dynamic IoV environments require joint communication-computation decisions, using Markov decision processes, quantum-inspired reinforcement learning, and spatiotemporal traffic modeling,, –.UAV-enabled federated learning further supports privacy-preserving edge intelligence, while adaptive switching between UAV compute and cache services remains future work.

E. Incentive Mechanism · F. Security and Privacy · 1) Privacy-preserving AIGC Service Provisioning:

The section presents incentives for increasing participation, computational capacity, service quality, and secure operations in mobile AIGC networks, while emphasizing privacy-preserving distributed learning against threats to sensitive data. It highlights economic and quality-aware mechanisms alongside federated-learning protections such as secure aggregation and differential privacy.

  • E. Incentive Mechanism: Suitable incentives attract more edge-node participation, increase computational capacity, improve AIGC service quality, and encourage secure operations through blockchain-recorded resource transactions [146], –.
  • E. Incentive Mechanism: Incentive mechanisms are evaluated through social welfare, provider revenue, and economic properties including individual rationality, incentive compatibility, and budget balance.
  • E. Incentive Mechanism: Edge-learning incentives must address prolonged training for satisfactory performance and provide appropriate monetary rewards for resource providers,,.
  • E. Incentive Mechanism: Quality-aware federated learning uses estimated device learning quality and reverse auctions under edge-server budgets to motivate high-quality contributions.
  • F. Security and Privacy: Cloud-edge collaborative computing and heterogeneous data enable beneficial AIGC use but also let malicious users produce destructive content, including phishing emails –[266].
  • 1) Privacy-preserving AIGC Service Provisioning:: Privacy must be protected throughout AIGC service provision because training data and user requests are generated and stored on edge servers and mobile devices with limited defense capacity [268]–.
  • 1) Privacy-preserving AIGC Service Provisioning:: Federated learning protects local updates through secure aggregation with authenticated encryption and secret sharing, or differential privacy that prevents servers from identifying update owners.
  • 1) Privacy-preserving AIGC Service Provisioning:: Privacy-preserving federated generative methods synthesize private-data examples, improve GAN efficiency and robustness under skewed data, and cluster heterogeneous participants, while centralized designs risk single-point failure [275]–.

2) Secure AIGC Service Provisioning: … 1) Decentralized Mobile AIGC Networks:

The paper frames mobile AIGC networks as collaborative, multi-tier systems requiring secure provisioning, multi-objective resource management, personalization, incentives, and decentralized governance to deliver reliable services while protecting privacy and data security.

  • 2) Secure AIGC Service Provisioning:: Blockchain-supported collaboration among cloud, edge, and mobile stakeholders is positioned as a basis for trustworthy, reliable, and secure AIGC service provisioning, –[281].Blockchain can organize application, blockchain, and computing-power network layers in mobile AIGC networks [254].
  • 1) Multi-Objective Quality of AIGC Services:: AIGC service quality depends on accuracy, latency, energy consumption, and revenue, requiring optimal edge-resource allocation and online decisions under dynamic network conditions, [153].Task and computation migration can improve service reliability and efficiency while supporting load balancing.
  • 2) Edge Caching for Efficient Delivery of AIGC Services:: Efficient edge caching must address constrained memory, model-missing costs, and functionally equivalent models through model-aware, preference-driven, and principled cache designs.These approaches aim to reduce AIGC service latency and energy consumption.
  • 3) Preference-aware AIGC Service Provisioning:: Preference-aware AIGC delivery uses historical user data, personalized recommendations, and feedback-driven strategy adjustment to improve satisfaction while reducing latency and resource consumption.Providers must collect and analyze user data before adapting service delivery to preferences.
  • 4) Life-cycle Incentive Mechanism throughout AIGC Services:: AIGC service lifecycles require incentives that distribute benefits according to stakeholder contributions across data collection, pre-training, fine-tuning, and inference.User evaluations based on transaction history can assess provider reputation and promote service optimization.
  • 5) Blockchain-based System Management of Mobile AIGC Networks:: Because heterogeneous devices create uncertain demand and security risks, mobile AIGC networks need secure device management and auditing for dynamic access, departure, and identification.Traditional centralized management remains vulnerable to central-node failure.
  • VII. FUTURE RESEARCH DIRECTIONS AND OPEN ISSUES: Future research addresses networking and computing, machine learning, and practical implementation issues for realizing mobile AIGC networks.These perspectives define the paper’s stated open-issue agenda.
  • 1) Decentralized Mobile AIGC Networks:: Blockchain-enabled decentralized mobile AIGC networks can combine distributed storage, computing-network convergence, digital identities, and smart contracts to protect user privacy and data security, [287].This structure also supports decentralized management across the AI service lifecycle.

2) Sustainability in Mobile AIGC Networks: … VIII. CONCLUSIONS

The paper identifies future directions for sustainable, robust, efficient, privacy-preserving, and human-centered mobile AIGC networks, while concluding that their deployment spans generative-model services, applications, and implementation, security, and privacy challenges.

  • 2) Sustainability in Mobile AIGC Networks:: Mobile AIGC networks require green operation strategies that reduce the substantial energy and carbon costs of model pre-training, fine-tuning, inference, and networking under dynamic configurations.[30], [293] motivate algorithms and frameworks for energy- and carbon-efficient operation.
  • 3) Wireless Communications in Mobile AIGC Networks:: Wireless conditions and unreliable synthesized data remain key challenges because transmit power, fading, mobility, and training-data quality can affect distributed AIGC performance and model reliability.Future work includes robustness to wireless communications [143], effects on distributed diffusion computing, and improving data quality through multimodal fusion [144], [298].
  • 1) Generative AI Model Compression:: Model compression can reduce AIGC service latency and resource consumption as models grow more complex, using pruning, quantization, and knowledge distillation,.Pruning removes unimportant weights, quantization lowers weight precision, and distillation trains a smaller model to mimic a larger one.
  • 2) AI-generated Network Design:: Generative models can support mobile-network design, analysis, control, monitoring, and traffic prediction by creating architectures, modeling behavior, detecting anomalies, and forecasting loads and demands [1], [301].This motivates research on improving machine-learning efficiency for mobile AIGC networks.
  • 3) Privacy-preserving AIGC Services:: Privacy-preserving AIGC requires privacy computing during both model training and inference, including differential privacy, secure multi-party computation, and homomorphic encryption, [142].These techniques protect sensitive data and prevent unauthorized access through noise addition, joint computation, and encrypted processing.
  • 1) Integrating AIGC and Digital Twins:: Integrating AIGC with digital twins can optimize mobile-network latency and quality and accelerate simulation-entity creation, but efficient and secure synchronization remains necessary.Digital twins maintain representations for monitoring, analysis, and prediction of physical entities.
  • 2) Immersive Streaming:: AIGC-enabled immersive AR and VR streaming can support real-time interaction in education, entertainment, and social media, while requiring research on risks including biased content.Immersive streaming creates content that transports viewers to virtual worlds and enhances AIGC delivery.
  • 3) Alignment; VIII. CONCLUSIONS: Human-oriented digital humans and avatars require aligned generative models for safety and ethnicity, with future work spanning personalization, ethics, transparency, emotion, culture, and adversarial robustness, [304],.The conclusion frames these directions within a broader survey of mobile-edge generative-model fundamentals, service lifecycles, creative applications, and implementation, security, and privacy challenges.
Loading 2303.16129v4…