Multi-user semantic communications
File(s)
Author(s)
Yilmaz, Selim Firat
Type
Thesis
Abstract
This thesis presents novel frameworks for multi-user semantic communications that integrate machine learning with wireless communications to enable efficient, private, and semantically aware information exchange at the wireless edge. Traditional communication systems, built on Shannon's separation principle, face growing limitations when dense networks of edge devices must share limited spectral resources to exchange high-dimensional data under strict latency constraints. In such settings, separate source and channel coding is known to be suboptimal in the finite blocklength regime, and conventional orthogonal access schemes fail to exploit the correlation structure across users' signals. We address these limitations by developing semantic communication systems for scalable multi-user communications and privacy-preserving edge inference.
We first address perceptual quality in wireless image transmission by integrating denoising diffusion probabilistic models into the deep joint source-channel coding (DeepJSCC) framework, achieving enhanced perceptual quality under severe channel conditions while establishing principled perceptual-distortion trade-offs.
We then extend distributed source coding theory to modern deep learning frameworks, focusing on low-latency image transmission with decoder-only side information. Our architectures integrate side information at multiple stages, delivering superior performance across all channel conditions, with particular gains at low signal-to-noise ratios.
Next, we develop non-orthogonal approaches to multi-user semantic communications over multiple access channels using DeepJSCC\@. Our scheme enables devices to simultaneously transmit compressed image representations, achieving substantial improvements over orthogonal transmission. Furthermore, we introduce a method that unifies compression and channel coding through multi-view autoencoders, enabling non-orthogonal multiple access (NOMA). Our algorithm scales to 64 users or more, surpassing state-of-the-art NOMA methods.
Finally, we introduce a privacy-preserving collaborative inference framework at the wireless edge, where clients' independently trained models participate in ensemble inference. Leveraging over-the-air computation, we develop schemes that exploit channel superposition for bandwidth-efficient transmission. Our multi-view classification framework achieves statistically significant improvements over orthogonal approaches while ensuring robust privacy protection.
We first address perceptual quality in wireless image transmission by integrating denoising diffusion probabilistic models into the deep joint source-channel coding (DeepJSCC) framework, achieving enhanced perceptual quality under severe channel conditions while establishing principled perceptual-distortion trade-offs.
We then extend distributed source coding theory to modern deep learning frameworks, focusing on low-latency image transmission with decoder-only side information. Our architectures integrate side information at multiple stages, delivering superior performance across all channel conditions, with particular gains at low signal-to-noise ratios.
Next, we develop non-orthogonal approaches to multi-user semantic communications over multiple access channels using DeepJSCC\@. Our scheme enables devices to simultaneously transmit compressed image representations, achieving substantial improvements over orthogonal transmission. Furthermore, we introduce a method that unifies compression and channel coding through multi-view autoencoders, enabling non-orthogonal multiple access (NOMA). Our algorithm scales to 64 users or more, surpassing state-of-the-art NOMA methods.
Finally, we introduce a privacy-preserving collaborative inference framework at the wireless edge, where clients' independently trained models participate in ensemble inference. Leveraging over-the-air computation, we develop schemes that exploit channel superposition for bandwidth-efficient transmission. Our multi-view classification framework achieves statistically significant improvements over orthogonal approaches while ensuring robust privacy protection.
Version
Open Access
Date Issued
2025-09-23
Date Awarded
2026-04-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Gündüz, Deniz
Sponsor
European Commission
Grant Number
953775
Publisher Department
Department of Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)