UC Davis

The rapid proliferation of data- and AI-driven applications in wireless and edge networks has ushered in a paradigm shift, delivering substantial advances over conventional communication systems. Contemporary communication infrastructures increasingly exploit the vast amounts of user-generated data to enable learning directly from distributed sources. Nevertheless, this shift introduces several pressing challenges. Issues such as user privacy, communication and bandwidth efficiency, and heterogeneous data distributions continue to constrain the full potential of distributed learning. In parallel, the emergence of generative AI has driven an exponential increase in image and video traffic, which far outpaces the comparatively slow growth of available bandwidth resources. Motivated by these developments, this dissertation investigates both the theoretical underpinnings and practical applications of generative AI models to address the aforementioned challenges. First, we develop methods to reconstruct radio maps from sparse information collected from distributed users, employing generative adversarial networks (GANs). We then examine the problem of learning a unified AI model from distributed user data while preserving privacy. To this end, we introduce a GAN-splitting mechanism, establish its convergence properties, and demonstrate its advantages over existing GAN-based federated learning (FL) approaches. Furthermore, we extend this framework to accommodate computation- and resource-constrained user environments. Building upon these results, we broaden the application of generative AI models to goal-oriented semantic communication (GO-Com) systems. In this context, we investigate the integration of advanced generative models, such as diffusion models, and ultimately derive lightweight solutions that enable their practical deployment.