Bayesian Adversarial Learning

Introduction

We propose a novel framework for Bayesian adversarial learning that can be applied to various applications such as adversarial defense.

Prerequisites

Our code is based on Python3 (>=3.5). There are a few dependencies to run the code. Please stick to the versions listed. The major libraries are listed as follows:

• PyTorch (= 0.4.0)
• Cuda Toolkit (= 9)

Dataset

Dataset need to download before running

Traffic Sign Dataset Note that we only use the training set from the original dataset and split it into training and test in experiments. We provide a processed dataset which can be obtained from https://drive.google.com/open?id=1gPreM_0RWMCA0qwzZpoyWWy3g5FMGGYj Note that please cite the original dataset and meet the requirements.

How to quickly reproduce the results

cd bayesian_adversarial_learning_release/experiments/trafficsign

To test on FGSM Attack:

python run_advattackFGSM.py

python plot_resultFGSM.py

To test on CW Attack:

python run_advattackCW.py