AL on Edge: Client-Side Local-Global Disagreement for Active Split Learning

Split learning (SL) lets resource-constrained edge devices train deep networks by running only a shallow model locally and offloading the rest of the model to a server. However, every local sample still sends a cut-layer activation to the server and receives a gradient back, so communication and server computation grow with the amount of local data. Training on an informative subset would remove much of this cost, but existing selection criteria rely on predictions, losses, or gradients of the full model, which in SL become available only after the activation has been transmitted. We propose AL-on-Edge, a framework that makes server-informed selection decisions entirely on the client. AL-on-Edge attaches two lightweight heads to the cut layer: a local-view head trained only on hard labels, and a global surrogate distilled from server logits that are already computed for the selected samples. Each candidate is scored by the Jensen–Shannon disagreement between the two heads plus the surrogate's predictive entropy, and only the highest-scoring activations enter the split-learning pipeline. The same heads enable agreement-gated early exit at inference, answering confident and consistent inputs on the device without contacting the server. Across CIFAR-10, SVHN, and FMNIST under IID and non-IID client partitions, AL-on-Edge reduces server computation and client-to-server communication by up to 90% at the smallest budget. Using half of each client's data, it outperforms full-data SL, random selection, and server-side active selection on CIFAR-10 and SVHN. At inference, 73–83% of test inputs exit locally while matching server-only accuracy.