Generative Diffusion Models for Wireless Networks: Fundamental, Architecture, and State-of-the-Art

TL;DR

This comprehensive review of Generative Diffusion Models for wireless networks is organized through a structured taxonomy that categorizes GDM-based schemes into the sensing, transmission, and Applications, complemented by a security plane.

Abstract

With the rapid development of Generative Artificial Intelligence (GAI) technology, Generative Diffusion Models (GDMs) have shown significant empowerment potential in the field of wireless networks due to advantages, such as noise resistance, training stability, controllability, and multimodal generation. Although there have been multiple studies focusing on GDMs for wireless networks, there is still a lack of comprehensive reviews on their technological evolution. Motivated by this, we systematically explore the application of GDMs in wireless networks. Firstly, we identify the core challenges of wireless networks and argue why GDMs are uniquely suited to address them. We then introduce the mathematical principles of GDMs and representative models. Furthermore, we organize our comprehensive review through a structured taxonomy that categorizes GDM-based schemes into the sensing, transmission, and Applications, complemented by a security plane. For each representative scheme, we analyze its innovative points, the role of GDMs, strengths, and weaknesses. Ultimately, we extract key challenges and provide potential solutions, with the aim of providing directional guidance for future research in this field.

Figures (10)

• Figure 1: The structure of this paper.
• Figure 2: Evolution of GDMs, where DDPM employs a fixed forward diffusion and a reverse denoising chain ho2020denoising; SGM reframes denoising as direct gradient estimation over noise scales song2019generativesong2020score; SDE and ODE supply a unifying limit that enables principled solver design and deterministic sampling song2020score; FM offers a more stable training paradigm for these continuous flows; DDIM introduces a non Markovian deterministic shortcut that skips many intermediate noise levels, yielding substantial speed gains song2020denoising; CMs further evolve this by enabling single-step generation; CDM augments the noise regression with external guidance signals nichol2021glideho2022classifier; and LDM relocates the objective into a compressed space for scalability rombach2022high.
• Figure 3: Illustration of the GDM-aided multi-layer network taxonomy. This framework organizes the application of GDMs into three functional layers and a cross-cutting security plane. The GDM module on the left serves as the foundational engine that provides high-quality generation and noise resilience to the entire system. At the bottom level, the sensing layer utilizes these generative priors to accurately model the physical environment through tasks such as channel estimation and radio map construction. Above this lies the transmission Layer which leverages the iterative denoising capabilities of GDMs to optimize end-to-end data delivery and enhance SemCom. The top level depicts vertical applications and services where the GDM functions as a world model for network digital twins or a solver for policy optimization. Finally, the security plane on the right interacts with all layers to defend against eavesdropping and adversarial attacks while simultaneously employing privacy-preserving techniques to secure the deployment of GDMs themselves.
• Figure 4: The architecture of the RadioDiff frameworkwang2024radiodiff utilizing a Latent Diffusion Model for efficient dynamic radio map construction. The system begins by employing a pre-trained VAE to compress high-dimensional radio map data into a low-dimensional latent space to significantly reduce the computational resources required for training. Within this latent representation, the model executes a decoupled diffusion process where a U-Net backbone predicts and removes noise conditioned on environmental prompts such as static building layouts and dynamic vehicle positions. A critical innovation embedded in the U-Net is the Adaptive Fast Fourier Transform module which dynamically filters features in the frequency domain to effectively capture high-frequency edge details that are typically missed by standard convolutional layers. By performing the diffusion process in the compressed latent space rather than the pixel space, the LDM achieves a superior balance between training efficiency and generation fidelity to synthesize accurate radio maps for complex dynamic environments.
• Figure 5: Illustration of Latent-Diff DNSC scheme xu2023latent. This architecture separates the semantic feature learning from the channel noise modeling to achieve robust SemCom. In Phase 1 shown at the top, a joint encoder-decoder is trained alongside a discriminator to compress images into a latent space while optimizing a composite objective function that includes reconstruction loss for fidelity, adversarial loss for local consistency, and regularization loss to constrain the latent space distribution. In Phase 2 shown at the bottom, a semantic de-noiser operates within this fixed latent space by first gradually adding Gaussian noise to the semantic vectors through a forward diffusion process to simulate physical channel impairments and then employing a U-Net to predict and eliminate these noises via a reverse diffusion process. This design allows the system to effectively model diverse channel noise characteristics during training and enables the adaptive elimination of interference during online transmission without requiring explicit CSI.