In short
A VQC can take one feature per qubit, and only a handful of qubits are available. The malware data has 262,144 features, so it has to be squeezed down to between 2 and 12 numbers first. What gets thrown away matters.
Variational quantum classifiers (VQCs) can only encode as many features as they have qubits, and current hardware only offers a low number of qubits. Because of this, high-dimensional datasets, like datasets containing malware features, have to be reduced before they can be encoded.
Datasets generally need to be “flattened” through some pipeline to make it feasible to train and run a simulated quantum neural network, especially in research where access to actual quantum computers is limited and researchers are required to simulate quantum computers.
In this research that meant going from 262,144 features to 2–12 values, one per qubit.
Why the way of reducing matters
Current quantum data reduction techniques rely on generic dimensionality reduction techniques such as Principal Component Analysis (PCA) to compress datasets. These techniques do not take into account the type of features being encoded and the importance that a feature can play on a classification task. This can result in classification-critical data being discarded indiscriminately, or data not needed for classification being included, which could negatively impact performance on real-world high-dimensional datasets such as malware feature sets. If it was possible to know which features mattered for classification before reducing dimensions and encoding the data, the information loss could be avoided.