Performance of quantum reservoir processors under adversarial attacks

Abstract

This study investigated the performance of a selected quantum neural network, a quantum polariton reservoir, under adversarial attacks. First, adversarial examples were generated using a white-box Fast Gradient Sign Method from two classical neural networks: a multilayer perceptron and a convolutional neural network. These examples were then tested for their transferability to the quantum polariton reservoir. Next, a similar technique that involved estimation of the gradients was devised to be used on the quantum polariton reservoir itself. Finally, a Generative Adversarial Network-based black-box method was utilized to test all three networks at once. Overall, the quantum polariton reservoir showed some level of robustness in all tests despite the adversarial attacks still being effective. There are multiple possible explanations for this, such as the networks' respective initial classification accuracies, the quantum polariton reservoir's method of data preparation, and how the quantum nature of the exciton-polaritons was harnessed. However, the definitive reasons remain unclear, and this project can only serve as a starting point for further testing in order to prove this robustness and how it may be applied to other quantum reservoir processors.