Dual-Circulation Generative AI for Optimizing Resource Allocation in Multi-Granularity Heterogeneous Federated Learning

The rapid proliferation of Internet of Things (IoT) devices and the deployment of edge computing infrastructures have significantly advanced data processing and network communications. Federated Learning (FL), leveraging these developments, offers a decentralized approach to learning across edge devices while preserving local privacy. However, traditional FL paradigms encounter challenges related to non-IID data distributions, diverse computational capabilities of edge devices, and multimodal tasks. This paper proposes a novel Dual-Circulation Generative Artificial Intelligence (GAI) framework for Clustered Federated Learning (GAI-CFL), designed to address multi-granularity heterogeneity including data, resources, and task heterogeneity. The GAI-CFL framework integrates Generative Diffusion Models (GDMs) and Reinforcement Learning (RL) strategies to optimize data generation and resource allocation processes. GAI techniques are employed to reduce intra-cluster data heterogeneity by generating data, thereby enriching local datasets and enhancing model convergence. To address inter-cluster data and task heterogeneity, a dual-circulation optimization strategy is implemented, utilizing RL with GDMs to generate optimal strategies based on input data characteristics dynamically. Additionally, a Mixture of Experts (MoE) approach is incorporated within GDMs to select the best expert models for denoising steps, ensuring efficient data generation strategies. Comprehensive simulations clarify that the proposed GAI-CFL framework significantly presents benchmark schemes, achieving lower energy consumption and enhanced system performance under the same delay and performance constraints.