Layered Uploading for Quantum Convolutional Neural Networks

TL;DR

The paper addresses the challenge of limited data and qubit resources in quantum convolutional neural networks (QCNNs) by proposing layered uploading, a data reuploading strategy that injects $2(n-1)$ features into an $n$-qubit QCNN without increasing qubits. By reencoding PCA-derived features before each layer, the approach leverages more information from the input while keeping circuit size fixed. Across MNIST tasks (e.g., 0 vs 1, 0 vs 8) and malware image classification, layered uploading yields higher accuracy and F1-scores than standard QCNNs, though benefits depend on model depth and parameter count. The results suggest layered uploading as a practical method to improve QCNN performance on small datasets and resource-constrained quantum hardware, with potential applications in malware detection and beyond.

Abstract

Continuing our analysis of quantum machine learning applied to our use-case of malware detection, we investigate the potential of quantum convolutional neural networks. More precisely, we propose a new architecture where data is uploaded all along the quantum circuit. This allows us to use more features from the data, hence giving to the algorithm more information, without having to increase the number of qubits that we use for the quantum circuit. This approach is motivated by the fact that we do not always have great amounts of data, and that quantum computers are currently restricted in their number of logical qubits.

Figures (6)

• Figure 1: Examples of PE files transformation into grayscale images. One can see that there is no distinguishable pattern that allows to visually classify malicious and benign files.
• Figure 2: Standard architecture of a quantum convolutional neural network, where convolutional and pooling layers alternate up to the measurement by an observable.
• Figure 3: Description of the convolutional and pooling layers used in our algorithm. In the pooling layer, the symbol $\bullet$ means that the gate $R_Z(\theta_1)$ is activated only if the first qubit is in the state $|1\rangle$. Conversely, the symbol $\circ$ means that the gate $R_X(\theta_2)$ is activated only if the first qubit is in the state $|0\rangle$. At the end of the pooling subcircuit we trace out the control qubit in order to reduce the dimension.
• Figure 4: Description of our architecture for the layered uploading QCNN. We add an encoding layer before each convolutional layer, almost doubling the number of features injected in the quantum circuit.
• Figure 5: Comparison of 5 random initializations of parameters for models with uploading and 2 layers, without uploading and 2 layers, and without uploading and 3 layers, on the MNIST dataset. For the models with 2 layers we use the same initialization parameters. We also plot the average of the 5 initializations, showing that the model with layered uploading is indeed better than the models without layered uploading.