Conditional Denoising Diffusion Probabilistic Models for Data Reconstruction Enhancement in Wireless Communications

TL;DR

Conditional denoising diffusion probabilistic models (CDiffs) are proposed to enhance the data transmission and reconstruction over wireless channels and it is shown that more than 10 dB improvement in the reconstruction could be achieved in low SNR regimes, without the need to reduce the information rate for error correction.

Abstract

In this paper, conditional denoising diffusion probabilistic models (DDPMs) are proposed to enhance the data transmission and reconstruction over wireless channels. The underlying mechanism of DDPM is to decompose the data generation process over the so-called "denoising" steps. Inspired by this, the key idea is to leverage the generative prior of diffusion models in learning a "noisy-to-clean" transformation of the information signal to help enhance data reconstruction. The proposed scheme could be beneficial for communication scenarios in which a prior knowledge of the information content is available, e.g., in multimedia transmission. Hence, instead of employing complicated channel codes that reduce the information rate, one can exploit diffusion priors for reliable data reconstruction, especially under extreme channel conditions due to low signal-to-noise ratio (SNR), or hardware-impaired communications. The proposed DDPM-assisted receiver is tailored for the scenario of wireless image transmission using MNIST dataset. Our numerical results highlight the reconstruction performance of our scheme compared to the conventional digital communication, as well as the deep neural network (DNN)-based benchmark. It is also shown that more than 10 dB improvement in the reconstruction could be achieved in low SNR regimes, without the need to reduce the information rate for error correction.

Figures (11)

• Figure 1: Data pipeline of our experimental communication setup for CDiff-based image reconstruction.
• Figure 2: Diffusion models: A general overview. A Markov chain is defined to mimic the forward diffusion steps, during which random perturbation noise is purposefully added to the original data. Then in a reverse process, the model learns to construct the desired data samples out of noise.
• Figure 3: Schematic of the Conditional DDPM Framework.
• Figure 4: General description of training the diffusion model and the notion of time embedding. Each time-step is mapped (via positional embedding) to an embedding vector, which is then incorporated into the hidden layers of the neural network. The figure shows time-step $t=21$ as an example.
• Figure 5: Data visualization for the performance evaluation of the proposed scheme for image reconstruction enhancement. From left to right, the ground-truth information signals $\boldsymbol{x}_0$, the degraded version $\widehat{\bm x}$ of the signals decoded at the output of the Sionna's simulator (before DDPM-based reconstruction), and the final reconstructed signals via our method are shown, respectively.

Theorems & Definitions (3)

• Remark 1
• Remark 2
• Remark 3