Experimental quantum adversarial learning with programmable superconducting qubits.

Area of Science:

• Quantum Computing
• Machine Learning
• Artificial Intelligence

Background:

• Quantum computing offers potential advancements for machine learning and artificial intelligence.
• Theoretical studies indicate quantum classifiers, similar to classical deep neural networks, are susceptible to adversarial perturbations.

Purpose of the Study:

• To experimentally demonstrate quantum adversarial learning using programmable superconducting qubits.
• To investigate the robustness of quantum classifiers against adversarial attacks and the effectiveness of adversarial training.

Main Methods:

• Trained quantum classifiers using variational quantum circuits with ten transmon qubits.
• Utilized superconducting qubits with high average lifetimes (150 μs) and gate fidelities (>99.94% for single- and >99.4% for two-qubit gates).
• Tested classifiers on real-life images (e.g., MRI scans) and quantum data.

Main Results:

• Demonstrated that well-trained quantum classifiers (up to 99% testing accuracy) can be deceived by small adversarial perturbations.
• Showed that adversarial training substantially enhances the robustness of quantum classifiers to these perturbations.

Conclusions:

• Quantum classifiers are vulnerable to adversarial perturbations, mirroring classical machine learning.
• Quantum adversarial learning, implemented with superconducting qubits, offers a viable method to improve classifier robustness.