Speaker
Description
Group-equivariant quantum models impose symmetry constraints that can improve trainability, but their effect on adversarial robustness remains unclear. We address this question through a feature-level analysis of equivariant quantum models under adversarial transfer attacks. For models with invariant readout, predictions depend only on the group-twirled input, separating accessible invariant information from an invisible subspace. For a rotationally equivariant quantum model, we show that the accessible feature space can be characterized as rotation-invariant image statistics resolved across symmetry sectors. Using targeted input transformations, we identify which of these statistics the model relies on across several datasets. We find that equivariance does not by itself provide transfer robustness, because the invariant feature space can still contain brittle features. We then show that suppressing the symmetry sector associated with the brittle feature improves robustness while preserving accuracy. Together, these results provide a way to diagnose brittle invariant features and improve robustness in equivariant QML models.
| Topic | Quantum-enhanced robustness in machine learning models |
|---|