Generative diffusion models for wireless communication networks: detection, recovery, and decision-making
Year: 2026
Type: dissertation
Abstract: Modern wireless communication systems are increasingly required to operate under uncertain and heterogeneous conditions, spanning signal inference under unknown noise distributions, reliable transmission in the presence of adversarial interference, and adaptive spectrum management in dynamic environments. While classical approaches have been developed for each of these settings individually, they typically rely on simplifying statistical assumptions that become harder to satisfy as deployment environments grow more complex. This thesis proposes a unified perspective: these three challenges, despite arising at different layers of the communication stack, are all instances of the same fundamental problem of conditional generation, namely, learning and sampling from the posterior distribution p(x \ y) of a wireless variable given an observation. Under this view, a generative model captures prior knowledge about the signal or environment, the wireless system specifies the likelihood, and their combination through the posterior score drives inference, recovery, and control within a single probabilistic framework. Three contributions instantiate this perspective at progressively larger operational scales, each showing how conditional generation addresses limitations that are difficult to handle using purely model-based or discriminative approaches. In Chapter 3, the conditional generation perspective is applied at the signal level, addressing regression, classification, and multiple-input multiple-output (MIMO) detection tasks where the noise distribution is unknown or non-Gaussian. This chapter replaces direct prediction with posterior-based inference, enabling the diffusion mechanism to act as a universal scoring tool rather than a sampler. The Supervised Score-Based Model (SSM) reinterprets score-based inference as a noise-free gradient descent process, establishing a formal equivalence with gradient boosting and enabling deterministic, stable predictions without the stochastic overhead of generative models. The Maximum-Likelihood Estimation Inference (MLEI) framework repurposes the diffusion evidence lower bound as a distribution-free likelihood proxy, enabling maximum-likelihood-equivalent MIMO detection without any parametric assumption on the noise. Together, these contributions demonstrate that the diffusion mechanism extends naturally beyond generation to serve as a principled inference engine. In Chapter 4, the scale advances to the physical layer, addressing wireless image transmission under jamming and spoofing attacks. This chapter reformulates physical-layer security as a posterior sampling problem under a corrupted observation model, where the attack structure is encoded directly into the likelihood term. The SecDiff framework decomposes adversarial decoding into two attack-specific likelihoods: a masked inpainting constraint for subcarrier jamming, and an expectation-maximisation (EM) channel refinement loop for pilot spoofing. This enables plug-and-play adversarial recovery without retraining the underlying codec and without prior knowledge of the attack, establishing that the conditional generation viewpoint remains effective under structured, non-stochastic perturbations. In Chapter 5, the conditional generation perspective reaches the network level, addressing unmanned aerial vehicle (UAV)-based spectrum sensing and trajectory planning in dynamic, partially observable environments. This chapter reformulates spectrum cartography and multi-agent trajectory optimisation as two coupled conditional generation problems over complementary target distributions: the full radio power map given sparse sensor observations, and the optimal UAV action given the inferred spectrum state. A masked autoencoder reconstructor and a conditional diffusion policy are jointly optimised through the reconstruction loss, enabling perception and control to reinforce each other within a single generative formalism. This establishes that the conditional generation principle can scale from point-wise signal inference to multi-agent policy generation. In summary, the three chapters support a unified perspective that connects task-specific model-based and discriminative designs with posterior-driven generative intelligence. By reformulating bespoke estimators, classifiers, and controllers as task-specific conditional generation problems, this thesis provides a scalable and principled foundation for wireless intelligence. Building on this foundation, Chapter 6 identifies future research directions in which the same conditional generation principle can be extended further, including richer environment modelling for cognitive wireless agents, more energy-efficient posterior sampling for latency-constrained deployments, and tighter integration of spectrum perception and decision-making in next-generation low-altitude networks.
Author: Changyuan Zhao
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7212335836