Federated Generative Diffusion Model for Secure Communications

Generative artificial intelligence (AI) facilitates secure communications by modeling signal and channel characteristics. However, its nature of centralized training raises privacy and scalability challenges. Federated learning (FL) is a decentralized machine learning paradigm that enables collaborative model training across distributed devices while ensuring data privacy by keeping the training data locally. In this paper, we outline the limitations of existing standalone and federated generative models. Furthermore, we propose a federated diffusion model (FDM) that comprehensively considers both the training and sampling phase for IoT scenarios. Then, we explore the applications relying on the proposed framework across various tasks in wireless security. To demonstrate the effectiveness, we provide a case study under a multi-user physical layer authentication scenario. Experimental results show that the proposed FDM substantially matches the performance of centralized diffusion model, while also ensuring secure deployment in distributed IoT environments.