Leveraging Noise in Quantum Machine Learning to Improve Model Robustness

Description

This thesis explores the potential of utilizing quantum noise in Quantum Machine
Learning models to enhance robustness against adversarial attacks. Quantum noise,
typically seen as a challenge in quantum computing, is applied here with the hypothesis
that it may improve model resilience. This thesis employs a Variational Quantum
Algorithm to evaluate the effects of different types and strengths of quantum noise
on the robustness of Quantum Machine Learning (QML) models under adversarial
conditions created by Fast Gradient Sign Method and Projected Gradient Descent
attacks. The work examines both noisy and noiseless models, revealing four main
scenarios of adversarial performance across various noise combinations.
Interestingly, results indicate that while some noisy models achieve higher adversarial
accuracy at specific attack strengths, no direct linear relationship exists between
noise probability and model robustness. This lack of a straightforward correlation
suggests that certain noise configurations might indeed offer resilience advantages,
although the outcomes vary based on attack type and dataset. Additionally, quantum
noise utilized during the training phase does not alter the model’s final state, as the
weights are adapted to noise conditions in training yet evaluated noiselessly. This
distinction offers insights into how noise affects parameter adjustments rather than the
final classification states, proposing a nuanced understanding of robustness in noisy
QML environments.
This thesis contributes to the field of quantum adversarial machine learning by
demonstrating that quantum noise may serve as a resource for improving Quantum
Machine Learning model resilience under adversarial attacks. Potentially, this thesis
enables the development of QML architectures that inherently integrate quantum noise
for robustness purposes without sacrificing model performance on non-adversarial
samples.