BESplit: Bias-Compensated Split Federated Learning with Evidential Aggregation

Split Federated Learning (SFL) enables privacy-preserving collaborative training by partitioning models between clients and a server. However, under non-IID data distributions, SFL often suffers from biased optimization and unstable convergence, while existing solutions largely adapt techniques from conventional federated learning. In this work, we observe that the split architecture of SFL inherently alters how client information is represented and coordinated, opening opportunities for bias compensation beyond parameter-level aggregation. Based on this insight, we propose BESplit, an architecture-aware framework that exploits the intrinsic structure of SFL to mitigate non-IID effects. First, to prevent biased local data from dominating global updates, we introduce Evidential Aggregation (EA) to perform fine-grained reweighting of client contributions based on evidential uncertainty. Second, to further reduce distributional skew, we develop Bias-Compensated Collaboration (BCC) to align split-layer representations by pairing complementary clients. Finally, Dual-Teacher Distillation (DTD) is incorporated to synchronize knowledge between decoupled client and server models, enabling independent local inference. Extensive experiments on five benchmark datasets demonstrate that BESplit consistently outperforms state-of-the-art methods in accuracy, convergence stability, and computational efficiency under diverse non-IID settings.

Split Federated Learning allows organizations to train shared AI models without exposing private raw data, while also reducing the heavy computing demands placed on individual participants. However, in practice, different participants often hold highly uneven or biased local datasets, which can destabilize training, slow convergence, and significantly reduce model quality. Existing solutions mainly address hardware or communication differences, but they do not adequately solve this data imbalance problem. We developed BESplit, a new framework that specifically tackles biased data in Split Federated Learning by making better use of the system’s split design. BESplit first measures how reliable each participant’s data is, then adjusts global learning accordingly. It also identifies participants with complementary data patterns and strategically aligns their intermediate knowledge to offset local biases. Finally, it uses a dual-teacher learning strategy to better synchronize local and global models. Across multiple benchmark datasets, BESplit consistently improves accuracy, efficiency, and training stability compared with existing federated and split learning methods. This work shows that Split Federated Learning can become more practical, scalable, and trustworthy for real-world privacy-sensitive applications such as healthcare, finance, and distributed AI systems.