摘要
Federated Learning (FL) is a promising framework for edge intelligence, allowing clients to preserve user privacy by collaboratively training a shared model. However, in real-world scenarios, heterogeneous resource challenges, such as the straggler effect among devices, degrade training efficiency. Even worse, existing approaches, such as model heterogeneity that rely heavily on contributions from strong devices, tend to trigger model bias issues. In contrast, resource-constrained devices are only able to train the initial few layers, eventually leading to uneven parameter training. In this paper, we propose a HeteroSplit framework that enables all devices to iteratively train the full depths of the model, thereby addressing the FL training challenges mentioned above. We design a Progressive Least Contribution First training strategy based on D2D-assisted split learning to dynamically balance each device's contribution and prevent model bias from strong devices. Additionally, we introduce an early transmission mechanism to accelerate overall training time. The experiments show that HeteroSplit improves accuracy by 15% and reduces training time by 41% compared to other state-of-the-art baselines under various device distributions and Non-IID data settings. © 2025 IEEE.