Efficient Split Learning with Client-Side Active Selection and Early Exiting

Split learning enables privacy-preserving distributed training by partitioning a neural network between resource-constrained clients and a server, but its efficiency is limited by repeated transmission of cut-layer activations and corresponding server computation. We introduce, a client-side data-selection framework that reduces these costs by identifying informative samples before server communication. Proposed method attaches two lightweight prediction heads at the cut layer: a locally supervised head and a global surrogate distilled from server predictions. Samples are ranked using a combination of local–global predictive disagreement and global-surrogate uncertainty, and only the highest-scoring subset is processed through the split-learning pipeline. This enables server-informed acquisition without querying the server over the full candidate pool. Experimental results shows that proposed method substantially reduces computation and communication while maintaining competitive predictive performance.