Federated Quantum Learning With Adaptive Differential Privacy via Quantum Noise

This paper studies a novel quantum federated learning (QFL) framework that leverages inherent quantum noise to enforce differential privacy (DP) and protect model updates from information leakage in current near-term quantum devices. Specifically, we achieve targeted DP guarantees during model training by i) varying the number of measurement shots, and ii) leveraging unavoidable depolarizing channel noise. To adapt privacy levels based on varying training conditions, we further introduce an adaptive DP strategy using a multi-armed bandit (MAB) agent at each quantum device, enabling dynamic control over the trade-off between privacy and model performance. Moreover, we present a rigorous convergence analysis of our proposed QFL-DP model. Our theoretical analysis and numerical evaluation reveal the following key properties: i) the global QFL-DP model converges to a bounded error floor, which indicates that, despite leveraging inherent quantum noise for DP, the model’s performance remains stable, making our approach suitable for real-world QFL applications; ii) a clear trade-off exists between privacy guarantees and model accuracy, and the adaptive MAB strategy further enhances this trade-off by enabling adaptive privacy budgeting with efficient noise allocation based on training dynamics and client needs; iii) our theoretical results are consistent with simulations, thereby demonstrating the superiority of the proposed framework over baseline approaches in terms of training accuracy, adaptive privacy preservation, and robustness against adversarial examples. Overall, the results highlight a tunable trade-off between privacy and robustness, offering an efficient and secure QFL solution for current near-term quantum devices.