Constructing Optimal Noise Channels for Enhanced Robustness in Quantum Machine Learning

Quantum Machine Learning (QML) is rapidly evolving. Recent studies suggest that quantum noise could improve classifier robustness. To verify this claim, we developed a semidefinite programming model that, given a dataset and a QML ansatz, constructs the most robust noise channel possible. In our smallscale experiments, we first explore the range of behaviors generated by this optimized noise channel. Secondly, we show that, despite recent claims, the direct impact of noise is modest compared to the substantial robustness improvements achieved through increasing the number of qubits. Aside from that, we utilize our framework to assess different noise channels in terms of their robustness and certifiability.