timroith / adversarialflows

Code for the Paper "Adversarial Flows"

Adversarial Flows

This repository provides the code for the experiments in the paper "Adversarial Flows: A gradient flow characterization of adversarial attacks" by Lukas Weigand, Tim Roith, and Martin Burger.

Overview

The following Notebooks are available:

• TrainClassifier.ipynb: This notebook trains a simple fully connected neural network, defined in model.py, on data points sampled from the two-moons dataset. The concrete dataset used in the paper is provided in the folder data-weights as two_moons.npz. Furthermore the weights of the trained networks are also given as two_moons_ReLU.pt and two_moons_GeLU.pt

• Flows.ipynb: This notebook computes and visualizes the iterates of IFGSM and the minizing movement scheme. The optimizers are defined in flows.py

Furthermore, the file compute_diffs.py provides an executable script, computing the difference between IFGSM and the minimizing movement scheme for different initial vlaues.

Citation