Policy-Masked Private Experts: Auditable and Reversible Capability Access Control in Sparse MoE Models

A recent paper on arXiv (2608.06690) presents Policy-Masked Private Experts, a novel technique for managing capability access in sparse Mixture-of-Experts (MoE) models. This method involves freezing a pretrained MoE model, developing a separate expert branch, and choosing between public and private pools prior to top-k routing. This process allows for trusted authorization to dictate which newly trained parameters can be accessed during the forward pass, ensuring that unauthorized requests do not engage any private experts within the defined trusted computing base. The study evaluates this execution control versus task utility on Qwen3-30B-A3B and DeepSeek-V2-Lite. In testing, three Qwen BF16 seeds updated all 32 private experts, with the public fingerprint remaining constant. In 64 adversarial scenarios and 96 deny/fail-closed events, unauthorized execution of private experts was successfully blocked. This research tackles a systems-level inquiry: is it possible to enforce access control at the parameter level instead of merely controlling behavior? The full paper can be found on arXiv.