Initialization and Rate-Quality Functions for Generative Network Layer Protocols

Abstract

Generative AI (GenAI) creates full content based on compact prompts. While GenAI has been used for applications where the generated content is returned to the prompt sender, it can play a vital role in extending the capacity of communication networks by transmitting compact prompts through links with limited capacity and, then, generating and forwarding approximations from the GenAI to the destination. This poses the challenge of evaluating the quality of those approximations as a function of the rate between the source and the GenAI node, while accounting for the communication overhead of learning. We present a method- and data-agnostic initialization protocol for learning rate-quality functions in GenAI-aided networks, defining three variants: (1) source-oriented, (2) node-oriented, and (3) destination-oriented. Each of them has different messaging flows based on where quality measurements are performed. The protocol augments node discovery protocols (e.g., MCP, A2A) when sources lack confidence in advertised model performance. We illustrate operation via statistical determination of required learning data, and validate using two prompting approaches. Results show successful rate-quality estimation with as few as 2 images, and positive gains over JPEG after just 1-18 post-learning transmissions, providing a practical, compression-agnostic foundation for GenAI-based network compression.

Figures (11)

• Figure 1: Illustration of the communication flow using generative compression with the decoder model placed at an intermediate network node. Alice uses a local encoder to compress data and chooses the compression ratio by selecting a prompt size $L_P$ based on an estimated rate-quality function.
• Figure 2: Example of image quality using the pixel swapping prompting method proposed in thorsager2024generative using the HiFiC HiFiC generative compression model. The generated image represents using the compression model alone and PixelSwapped 25% and 50% augments the generated image with 25 and 50 % of pixels from the original image chosen at random. Generated, pixel swapped 25%, and pixel swapped 50% correspond to the small, medium, and large prompt sizes from Figure \ref{['fig:overview']}.
• Figure 3: Message sequence diagram for the node discovery following the MCP and A2A protocols. The reference to the learning protocol indicates the beginning of one of the learning protocols presented in Section \ref{['subsec:Learning']} (Figures \ref{['fig:InitProtc_source-oriented']}, \ref{['fig:InitProtc_node-oriented']}, and \ref{['fig:InitProtc_destination-oriented']}).
• Figure 4: Message sequence diagram for the source-oriented learning process. For each of $N_L$ data points, $N_p = |\mathcal{P}_{N_p}|$ loops are required to measure the quality for the selected prompt sizes.
• Figure 5: Message sequence diagram for the node-oriented learning process. With augmented generation, the loop represents the $N_L$ data points; The generative node $g$ creates all the necessary prompt sizes locally in each loop. Without augmented generation, the loop represents $N_p$ prompt sizes per data point.