Week Seven: Writing and Experiment Planning

Hello everyone! I spent this last week focusing on writing an introduction to my paper and on planning some experiments to do over the course of the next two weeks. Reading papers on saturated autoencoders and how they tend to be more overfitted to the data has piqued my interest in adversarial subspaces. As theory goes, adversarial subspaces are areas of the input space (in the case of my autoencoder, a 784-dimensional space because each pixel of the 28×28 image is considered one dimension) where there are no data points yet the classifier has a high certainty of a label for that space. This allows adversarial attacks to find adversarial examples close to correctly labeled data by finding the closest subspace rather than finding an area close to correctly labeled points of a different class to fool the classifier.

Some of the experiments I want to conduct are:

• Using latent dimension data of my autoencoder to create a visualization of the properties of the high dimensional space in 2D. Possible create a classifier and visualize how it labels the input space to see if we can see adversarial examples
• Using the distance of adversarial examples from the attacked image to the closest data point of the same label as the adversarial image to see if these adversarial attacks are really using these subspaces.
• See if it is possible to map out these subspaces using an algorithm that takes an adversarial example and finds the borders of the input group it lays on