Generative Network Layer for Communication Systems with Artificial Intelligence

TL;DR

The paper proposes a generative network layer that leverages GenAI at edge/ intermediate nodes to produce approximate data replicas from prompts, enabling non-traditional use of network capacity. It formalizes a network model with nodes s, r, g, d, capacities $R_{ij}$, and flows $f_{ij}$, and analyzes end-to-end rate gains with a max-flow framework, where the gain is $G_{flow} = 1 + y_g / f'_sd$ and prompting rate $\lambda^* = min( c_{sg} / L_p, c_{gd} / L )$. A case study on image transmission uses latent representations as prompts, evaluates distortion with $\delta_D(L_p)$ and perceptual quality with $\delta_P(L_p)$ via FID, and compares to JPEG as a baseline, reporting that the latent-based scheme can achieve higher perceptual quality at small prompts and, overall, greater end-to-end flow under perceptual constraints. The results show convex relationships and that, under a low quality constraint, max flow gain can exceed 100% for the proposed scheme, while JPEG provides no such gain.

Abstract

The traditional role of the network layer is the transfer of packet replicas from source to destination through intermediate network nodes. We present a generative network layer that uses Generative AI (GenAI) at intermediate or edge network nodes and analyze its impact on the required data rates in the network. We conduct a case study where the GenAI-aided nodes generate images from prompts that consist of substantially compressed latent representations. The results from network flow analyses under image quality constraints show that the generative network layer can achieve an improvement of more than 100% in terms of the required data rate.

Figures (4)

• Figure 1: Exemplary multi-path network topology with a relay node $r$ that replicates any incoming data and a node $g$ that generates approximations of the data. $R_{ij}$ is the maximum achievable data rate from node $i$ to $j$.
• Figure 2: Rate-Perception functions for the Prompt Extension() and Pixel Swapping () approaches for the and JPEG compression schemes. The dots indicate the measured data points and the lines are the curves fitted to these data points.
• Figure 3: Rate-Distortion functions for the Pixel Swapping () approaches for the and JPEG compression schemes. The dots indicate the measured data points and the lines are the curves fitted to these data points.
• Figure 4: Results of the optimization problem with $c_{sg}=3.184$ : (a) The Optimal prompt size given rate-distortion and -perception functions for the two proposed approaches. (b) The $G_\text{flow}$ calculated from the optimal prompt size and the size of the content after node $g$.