Self-enhanced multi-task and split federated learning framework for RIS-aided cell-free systems

Collaborative learning-based beamforming schemes have been exploited to improve spectral efficiency (SE) with low privacy risks in reconfigurable intelligent surface (RIS)-aided cell-free (CF) systems. However, a single-task-driven federated learning (FL) scheme needs to run a large model on local devices with limited computing capacity due to its imbalanced computing resources. Although a single-task-driven split learning (SL) can split a large model into multiple smaller portions, it concerns the training time overhead due to its relay-based training. Meanwhile, annotation for well-labeled channel state information (CSI) still affects beamforming performance with high labeling costs. In this paper, we first propose a collaborative learning framework, named multi-task and split federated learning (M-SFL), for joint channel semantic reconstruction and beamforming for RIS-aided CF systems. The proposed M-SFL framework simultaneously tackles channel semantic reconstruction and beamforming with shared knowledge to distinguish the inherent information of user equipments (UEs). The proposed M-SFL splits large model into multiple lightweight parts friendly with the limited computing local devices and trains local and global models parallelly with the Federated server. Then, we expand the proposed M-SFL framework into a self-enhanced multi-task and split federated learning (SM-SFL) framework by integrating the contrastive learning technique. The SM-SFL framework pre-trains by predicting and distinguishing the target CSI and others without annotation, and then we fine-tune the local and global models with limited labeled CSI. Simulation results show that the proposed framework can jointly achieve better channel semantic reconstruction and higher SE with balanced computing resources, faster beamforming, and low labeling costs.