LENS-SFL: Learning-Driven Contracts Framework for Personalized Split Federated Learning under Strong Uncertainty

Split Federated Learning (SFL) enables collaborative training of deep models by partitioning computation across distributed system components while preserving data locality at clients. Existing SFL systems typically rely on static split configurations or assume obedient clients with known capability distributions, which limits their effectiveness in heterogeneous environments. In practice, however, clients exhibit heterogeneous and privately held computational, communication, and data characteristics, and behave as rational and self-interested agents that strategically respond to server policies, rendering such assumptions invalid. In this paper, we propose LENS-SFL, a learning-driven framework for personalized SFL under strong incomplete information. LENS-SFL formulates personalized model splitting as an incentive-aware online learning problem, where the server jointly designs model split points and rewards as contract items. Through this contract-based abstraction, unobservable client heterogeneity is mapped to observable contract-selection behaviors, enabling the server to learn the latent client-type distributions purely from data, despite strategic client responses and the absence of prior knowledge. To support effective personalization, LENS-SFL integrates an online learning mechanism that adaptively optimizes contracts based on observed behavioral feedback. We show that the resulting learning framework guarantees behavior-consistent participation and voluntary client engagement regardless of the underlying client distribution, ensuring safe learning throughout the process. Furthermore, we establish a sublinear regret bound with respect to the complete-information oracle, demonstrating that the performance loss induced by learning vanishes asymptotically. Extensive experiments on benchmark datasets validate the effectiveness of LENS-SFL, showing that it achieves near-oracle model performance, exhibits stable learning dynamics under heterogeneous settings, and significantly outperforms static and non-adaptive baselines. These results show that transforming strategic client behaviors into learnable signals provides a principled and effective pathway toward robust and personalized SFL.