GainNet: Coordinates the Odd Couple of Generative AI and 6G Networks

TL;DR

GainNet proposes a GAI-native network framework that tightly couples generative AI with 6G networks to enable sustainable, knowledge-driven cloud-edge-end collaboration. It introduces GaiRom-ISCC, a unified sensing, communication, and computing resource orchestration mechanism, featuring a Z-shaped temporal structure, decoupled universal resource pools, graph-based resource-model representations, and DRL-based graph matching. Two case studies in medical text summarization and intelligent vehicular networks demonstrate improved QoS and robustness, while highlighting cost-performance trade-offs and adaptability under resource constraints. The paper also discusses privacy, incentive design, and joint optimization challenges critical for deploying GAI at the edge of next-generation networks.

Abstract

The rapid expansion of AI-generated content (AIGC) reflects the iteration from assistive AI towards generative AI (GAI) with creativity. Meanwhile, the 6G networks will also evolve from the Internet-of-everything to the Internet-of-intelligence with hybrid heterogeneous network architectures. In the future, the interplay between GAI and the 6G will lead to new opportunities, where GAI can learn the knowledge of personalized data from the massive connected 6G end devices, while GAI's powerful generation ability can provide advanced network solutions for 6G network and provide 6G end devices with various AIGC services. However, they seem to be an odd couple, due to the contradiction of data and resources. To achieve a better-coordinated interplay between GAI and 6G, the GAI-native networks (GainNet), a GAI-oriented collaborative cloud-edge-end intelligence framework, is proposed in this paper. By deeply integrating GAI with 6G network design, GainNet realizes the positive closed-loop knowledge flow and sustainable-evolution GAI model optimization. On this basis, the GAI-oriented generic resource orchestration mechanism with integrated sensing, communication, and computing (GaiRom-ISCC) is proposed to guarantee the efficient operation of GainNet. Two simple case studies demonstrate the effectiveness and robustness of the proposed schemes. Finally, we envision the key challenges and future directions concerning the interplay between GAI models and 6G networks.

Figures (5)

• Figure 1: The architecture of the proposed GainNet. Since the GAI model cannot be fully deployed on 6G terminal devices, to obtain personalized knowledge of local data of 6G terminals, the FL framework of distributed EI is adopted, and there are two solutions. The first is the direct fine-tuning based on FSL, which performs split learning on the GAI model, keeps the modules with high computing power demand (e.g. the Transformer layer of ChatGPT) on the edge server, and puts the lightweight modules with low computing power demand (e.g. the Embedding layer) on the 6G end devices. The other is FPL to perform indirect fine-tuning, that is, locking the original GAI model and adding additional modules to guide the optimization of the GAI models, such as adding external Adapters.
• Figure 2: Optimization of GAI edge-end subnetworks with integrated sensing, communication, and computing. Since the deployment locations of cloud servers and edge servers are relatively fixed and there are sufficient resources, the optimization of GainNet’s cloud-edge subnetworks is stable. Due to the limited resources, high mobility, and changes in data distribution of 6G terminals, the distributed optimization of edge-end subnetworks is full of uncertainty. Thus, this paper focuses on the optimization of the GAI model on edge-end subnetworks and the corresponding resource provisioning. In particular, the edge domain-specific model and end lightweight model are collectively referred to as the edge-end model.
• Figure 3: Overview of the proposed GaiRom-ISCC. First, based on Zeros, a time-efficient temporal structure, partial parallel execution of sensing, communication, and computing between neighboring GRs can be achieved, which reduces the resource waste caused by the resource idle due to the CSTCs between the intra-GR sensing, communication, and computing and improves the resource utilization rate. Then, based on the decoupled URPs, the multi-domain resources can be invoked without distinction among sensing, communication, and computing, creating the resource objects for the mapping between resources and GAI models. Furthermore, the graph-based representation of the resource-model relationship is used to fully explore the relationship characteristics between the clients’ resources and GAI models. Finally, the GAI-oriented generic resource orchestration is implemented with DRL-based graph matching.
• Figure 4: Implementation and results of the case study in the medical domain. In the summary generation task, the user’s input consists of the instruction that expresses the intention and the text file, which is GAI’s executing object. Three methods in the implementation process correspond to the traditional method a which only uses the cloud foundation model (i.e., GAI without GainNet), the method b, where the edge-end model only gives partial prompt (template labels) (i.e., GAI with GainNet (Part)), and the method c, where the edge-end model offers the full prompt (including the template labels and corresponding text) (i.e., GAI with GainNet (Full)).
• Figure 5: Implementation and results of the case study in the traffic domain. Consider a square area of $500m\times 500m$, where connected intelligent vehicles act as clients and other vehicles act as the targets. The client performs FL-based distributed model training based on the target-related sensing data samples. CR is taken as the period for sampling, and Fig \ref{['fig_5']} (a) shows the sampling results of 50 vehicles and 100 targets in 5 consecutive CRs represented by specific colors and demonstrates the effectiveness of the proposed GaiRom-ISCC. Furthermore, Fig. \ref{['fig_5']} (b) gives the graphical explanation of the proposed GaiRom-ISCC, and proves the robustness of the resource orchestration based on GaiRom-ISCC.