Preprint cuts encryption depth for shared quantum model training

Researchers report a way to train quantum neural networks across separate clients while keeping their updates encrypted, without the costly encryption resets previously associated with rotation weights. Their preprint represents each rotation as a unit quaternion, making the operation a low-degree calculation that uses one multiplicative level of homomorphic encryption; averaging updates uses none. The authors tested two cryptographic backends and studied training with up to 20 clients. On a 156-qubit processor, the encrypted approach achieved 0.9918 fidelity against 0.99957 for an unencrypted control. The results remain preprint findings, and the measured hardware fidelity is lower than the control.