Quantum Machine Learning Defense via Steering-Based State Preparation
A recent preprint on arXiv introduces a defense strategy for quantum machine learning (QML) aimed at countering adversarial attacks. This innovative method substitutes the traditional quantum encoding phase with a controlled state preparation based on passive steering, directing the encoded state toward a specific intermediate state. By adjusting the strength of the steering and the number of iterations, this technique effectively mitigates adversarial disturbances while preserving accuracy and enhancing resilience against adversarial threats. The proposed approach has been validated through experimental results, demonstrating its efficacy.
Key facts
- arXiv:2605.10954v1 proposes a defense for QML against adversarial perturbations.
- The defense uses passive steering-based controlled state preparation.
- It replaces the conventional quantum encoding stage.
- Steering strength and number of iterations are tunable parameters.
- The method suppresses adversarial perturbations.
- It maintains high clean accuracy.
- It improves adversarial accuracy.
- Experimental results demonstrate effectiveness.
Entities
Institutions
- arXiv