Rethinking Multi-User Semantic Communications with Deep Generative Models

TL;DR

This paper tackles multi-user semantic communication by enabling diffusion-model-based regeneration to fill in untransmitted content in OFDMA downlinks. It reframes the problem as inverse generation and introduces a null-space diffusion sampling mechanism that aligns generated content with partial observations, leveraging constructs such as $L=KN$, $oldsymbol{A}_k=\boldsymbol{H}_k\boldsymbol{B}_k$, and the projection into range and null spaces. Experiments on CelebA-HQ and ImageNet show robust reconstruction even when $N/M$ is as low as $0.4$, outperforming LDPC and DeepJSCC jointly in terms of SSIM, PSNR, LPIPS, and FID. The work demonstrates a viable path toward GenAI-based next-generation communications where semantic content is regenerated at receivers, reducing transmitted bits while preserving perceptual semantics.

Abstract

In recent years, novel communication strategies have emerged to face the challenges that the increased number of connected devices and the higher quality of transmitted information are posing. Among them, semantic communication obtained promising results especially when combined with state-of-the-art deep generative models, such as large language or diffusion models, able to regenerate content from extremely compressed semantic information. However, most of these approaches focus on single-user scenarios processing the received content at the receiver on top of conventional communication systems. In this paper, we propose to go beyond these methods by developing a novel generative semantic communication framework tailored for multi-user scenarios. This system assigns the channel to users knowing that the lost information can be filled in with a diffusion model at the receivers. Under this innovative perspective, OFDMA systems should not aim to transmit the largest part of information, but solely the bits necessary to the generative model to semantically regenerate the missing ones. The thorough experimental evaluation shows the capabilities of the novel diffusion model and the effectiveness of the proposed framework, leading towards a GenAI-based next generation of communications.

Figures (9)

• Figure 1: Sample results of our method on natural images for two-user scenario and number of subcarriers $N < M$, with $M = 256$ and $N = 180$, equivalent to an $N/M$ ratio equal to $0.7$. The proposed null-space diffusion model effectively inpaints the gaps in the received information.
• Figure 2: Schematic representation of the proposed generative semantic communication system with the proposed null-space diffusion model to regenerate missing portions of the bitstream (black dots in the figure) at the receivers.
• Figure 3: Illustration of diffusion model forward process, utilized in training, diffusion model standard reverse process for sampling, and ours null-space diffusion model sampling with the reparameterization of the mean and variance.
• Figure 4: Random samples from the ImageNet test set regenerated by our method. We consider two users and three different channel assignment scenarios, where the total number of subcarriers $M$ is set to $256$, and we experiment with $N < M$ and $N = \{ 205, 180, 150 \}$, equivalent to an $N/M$ ratio of $\{0.8, 0.7, 0.6 \}$, respectively. We also zoom four heavily affected regions for a better visual evaluation. Here, it is clear how our method excellently regenerated missing portions of data even though it receives extremely degraded information.
• Figure 5: Comparison for different $N/M$ subcarriers ratio, where $N < M$ and $M = 256$, for $K=2$ users during transmission, evaluated with different metrics, namely SSIM$\uparrow$, PSNR$\uparrow$, FID$\downarrow$, LPIPS$\downarrow$ on the CelebA-HQ dataset. The proposed method (in red line) far exceeds any other method according to all the four metrics in each scenario.