Hierarchical Multimodal Federated Learning Over Cell-Free Massive MIMO Systems

Abstract

Cell-Free massive MIMO (CF-mMIMO) is a promising technology for enabling Federated Learning (FL) in the next generation of wireless networks due to its uniform service coverage. However, existing approaches that optimize FL over CF-mMIMO networks rely on a single Control Unit (CU), limiting scalability in terms of geographic coverage and user participation, while also overlooking multimodal data heterogeneity, which further increases latency. To address these challenges, we propose Hierarchical Multimodal Federated Learning (HMFL) over CF-mMIMO networks, which employs multiple CUs, managed by a Cloud Data Center (CDC). Instead of a single CU for global aggregation, HMFL uses a hierarchical approach where each CU aggregates local updates from the users before forwarding the edge models to the CDC for global aggregation. Moreover, we formulate an optimization problem for long-term decision-making in HMFL over CF-mMIMO networks, aiming to balance latency and user participation under a long-term energy budget. To solve this problem, we propose Long-Term Device-Modality Selection and Resource Allocation (LT-DeMoSRA) that employs optimization techniques to enable per-round decision-making with a long-term perspective over CUs without requiring future information. Additionally, our HMFL framework personalizes the fusion process based on the available modalities for each user, ensuring more adaptive and efficient multi-modal learning. Experimental results demonstrate that HMFL over multi-CU CF-mMIMO networks supports a larger number of users and outperforms existing alternatives by reducing training latency and improving user participation for both unimodal and multi-modal data.