Split Federated Learning Under Joint System and Data Heterogeneity

Split Federated Learning integrates model partitioning with federated aggregation, making the paradigm suitable for collaborative learning in edge environments. Heterogeneity is a characteristic of these environments, where client devices vary in computation power and network bandwidth, as well as the distribution of the data they own. Many existing SFL architectures assume the same cut layer for all clients and are tested under homogeneous data settings. This assumption limits the applicability of the models in edge environments where devices differ widely in data distribution, computation, and communication capacity. This paper introduces a generalized SFL architecture, CC-SFL, with a client specific cut layer setting that can adapt to the resource heterogeneity of clients and allow them select a cut layer suitable for their resources.We study the effectiveness of learning with this architecture under various cut layer settings using two benchmark datasets: CIFAR-10 and MNIST. We employ a Dirichlet distribution for data partitioning, resulting in various degrees of data heterogeneity. Our results provide insights on howa CC-SFL model reacts to joint heterogeneity in data and cut layer settings, establishing a foundation for scalable and resource aware CC-SFL deployments in real world edge environments.