Universal adversarial examples and perturbations for quantum classifiers.

Area of Science:

• Quantum Computing
• Machine Learning
• Quantum Physics

Background:

• Quantum machine learning (QML) leverages quantum computation for machine learning tasks, potentially surpassing classical capabilities.
• QML systems are vulnerable to adversarial attacks, where small input perturbations cause confident misclassifications.

Purpose of the Study:

• Investigate the universality of adversarial examples and perturbations in quantum classifiers.
• Assess the robustness of quantum machine learning models against sophisticated attacks.

Main Methods:

• Developed concrete examples using real-life image and quantum phase classification.
• Analyzed the mathematical properties of adversarial perturbations for quantum classifiers.
• Quantified the relationship between perturbation strength, number of classifiers, and adversarial risk.

Main Results:

• Demonstrated the existence of universal adversarial examples that fool multiple quantum classifiers.
• Proved that a specific increase in perturbation strength can mitigate universal adversarial risk for multiple classifiers.
• Showed that universal adversarial perturbations can be generated for individual quantum classifiers.

Conclusions:

• Adversarial attacks exhibit universality in quantum machine learning, posing a significant security challenge.
• Understanding these universal attacks is crucial for the secure development and deployment of quantum technologies in machine learning.
• Findings inform the design of more robust quantum machine learning algorithms and systems.