Privacy-Preserving Split Learning for Federated LLM Fine-Tuning

To address the problem that the autoregressive nature of LLMs in federated split learning causes intermediate activations to leak inputs and renders existing perturbation defenses ineffective, the researchers propose a learnable obfuscation–recovery scheme that protects participants' private datasets while still enabling the training of an independently deployable model on the server side. Experiments show that the method achieves strong privacy protection while incurring only moderate utility loss and system overhead, making split-based federated LLM fine-tuning practically feasible.