Robustness-Estimation-Experiments

This repository contains the code and neural networks from the paper:

Empirical Analysis of Upper Bounds for Robustness Distributions using Adversarial Attacks

Author(s): Aaron Berger, Nils Eberhardt, Annelot W. Bosman, Henning Duwe, Holger H. Hoos and Jan N. van Rijn
Published in: [Conference/Journal Name, Year]
citation key:

Overview

This repository contains:

• Pre-trained models in ONNX format.
• Experiment scripts and instructions for reproducing the results on MNIST.

Networks

• MNIST
  • mnist-net_256x2
  • mnist_relu_3_50
  • mnist_relu_4_1024
  • mnist_relu_9_100

Experiment Scripts

In the experiments folder, one can find scripts to compute robustness distributions using the following algorithms:

• alpha-beta-CROWN
• AutoAttack
• fab-attack
• Fast Gradient Sign Method (FGSM) Goodfellow et al., 2015
• Projected Gradient Descent (PGD) Madry et al., 2018

External Packages

This project uses the following external packages:

• VERONA: An open-source package for creating Robustness Distributions.