Adaptive Split Federated Learning for IoT: Personalized Clustering and Efficient Communication

The growing number of Internet of Things (IoT) devices has led to massive datasets spread across different locations. While Federated Learning (FL) enables decentralized training without transmitting raw data, it often requires resource-limited IoT devices to train full models locally, causing high energy and computation costs. Split Federated Learning (SFL) addresses some of these issues by splitting the model between IoT devices and a central server. However, standard SFL does not fully accommodate the varied challenges of IoT environments, including uneven data distributions, devices joining or leaving over time, and frequent communication demands. To tackle these concerns, we introduce an enhanced SFL framework. Our approach groups devices based on partial feature representations from the early layers of the model, leading to a personalized training process that boosts performance under uneven data conditions. We further use hierarchical clustering to manage groups as devices join or leave, maintaining both stability and scalability. To reduce communication overhead, we compress the partial feature data and implement model migration, allowing devices that move between clusters to adapt with fewer training rounds. Through extensive experiments, we show that our method consistently surpasses existing approaches under various data imbalances, device participation rates, and communication constraints. These results confirm that our framework is a robust, efficient, and adaptive solution for IoT-based SFL, offering superior accuracy, reliability, and resource management in real-world applications.