Semantic Communications Via Denoising Diffusion Autoencoder Models

Existing semantic communication (SemCom) frameworks typically emphasize on channel-adaptive neural encoding-decoding schemes, lacking full exploration of signal distribution. Moreover, having autoencoder architectures as backbone, their encoding is tightly coupled to a matched decoder, causing scalability issues in practice. To solve these issues, diffusion autoencoder models are proposed. A neural encoder extracts the high-level semantics, and then a conditional denoising diffusion model probabilistic model (C-DDPM) at the decoder learns the source distribution through signal-space denoising, while the noisy semantic latents are incorporated as the conditioning input, to "steer" the decoding process towards the semantics intended by the transmitter. Simulations over CIFAR-10 highlights 50% and 33% improvement in terms of the learned perceptual image patch similarity (LPIPS) metric compared to autoencoders with matched decoder architecture, and variational autoencoders (VAE) benchmarks, respectively.