Image classification and adversarial robustness analysis based on hybrid quantum–classical convolutional neural network

Abstract

Hybrid quantum and classical classification algorithms have provided a new solution to the classification problem with machine learning methods under a hybrid computing environment. Enlightened by the potential powerful quantum computing and the benefits of convolutional neural network, a quantum analog of the convolutional kernel of the classical convolutional neural network, i.e., quantum convolutional filter, is designed to enhance the feature extraction ability. Meanwhile, quantum convolutional layers stacked by quantum convolutional filters combine variational quantum circuits with tensor network architecture and convolution operations. In addition, a hybrid quantum–classical convolutional neural network model containing quantum convolution layers and classical networks is devised. The feasibility of the proposed hybrid model are tested on the classical MNIST dataset. Finally, the adversarial robustness of the presented hybrid network is compared with that of the classical convolutional neural network and the quanvolutional one under classical adversarial examples. It is demonstrated the presented hybrid quantum–classical convolutional neural network model outperforms the original convolutional neural network and the quanvolutional neural network in some adversarial cases.