Methods for inducing a covert misclassification

A non-transitory machine-readable storage medium having a machine-readable program stored therein, wherein the machine-readable program, when executed on a processing system, causes the processing system to perform a method of producing a set of generated data operable to induce a covert misclassification of a data input, wherein the method comprises:

• receiving a high dimensional space input data representing a spectrum from a generator of a generative adversarial network (GAN) from a data storage media;
• identifying a target dataset, wherein the target dataset represents a reduced dimensional representation of the high dimensional space input data with a perturbation;
• creating a spectral perturbation tensor, wherein the spectral perturbation tensor is configured to shift the high dimensional space input data to generate a low dimensional dataset;
• defining a function based on the spectral perturbation tensor, the low dimensional dataset, and the target dataset;
• adjusting the spectral perturbation tensor by determining a minimum of the defined function to optimize the perturbation, thereby generating an obtained low dimensional dataset approximating the target low dimensional dataset;
• obtaining an output high dimensional dataset based on a combination of the high dimensional space input data and the adjusted spectral perturbation tensor;
• outputting the output high dimensional dataset;
• and
• classifying the output high dimensional dataset with an established classifier to a misclassified classification.