Awesome

Code for Stochastic Security: Adversarial Defense Using Long-Run Dynamics of Energy-Based Models

This repository contains PyTorch implementations of the training and evaluation experiments from our paper:

Stochastic Security: Adversarial Defense Using Long-Run Dynamics of Energy-Based Models<br/>Mitch Hill*, Jonathan Mitchell*, and Song-Chun Zhu (*equal contributions)<br/>https://arxiv.org/abs/2005.13525.

The file bpda_eot_attack.py is an implementation of the BPDA+EOT attack to evaluate EBM defense. The files train_ebm.py and train_clf.py will train an EBM and classifier network respectively.

NOTE: All config files use the pixel range [0, 255] for adversarial perturbation adv_eps and attack step size adv_eta. However, all experiments scale images so that the pixels range is [-1, 1]. Adversarial parameters are scaled accordingly during execution. The Langevin step size langevin_eps in the config files uses the pixel range [-1, 1].

Attack Code

The BPDA+EOT attack in Algorithm 1 of our paper is implemented by bpda_eot_attack.py. The folder config_attack has three files for three different attacks. The config bpda_eot_attack.json is the main attack used against EBM defense in our paper (see Section 4.2). The config clf_nat_attack.json is set up evaluate the base classifier without purification. The config clf_transfer_attack.json will create adversarial samples from the base classifier before a final evaluation using EBM purification (see Section 4.1).

EBM Training Code

The EBM training procedure in Algorithm 2 of our paper is implemented by train_ebm.py. The implementation is heavily based on this repository . The main difference is the introduction of an initial phase of non-convergent learning with the Adam optimizer followed by a final phase of convergent learning using SGD (see Section 2). The folder config_train_ebm has training files for Cifar-10, SVHN, and Cifar-100.

Classifier Training Code

The file train_clf.py is a minimal implementation of natural and adversarial training for classifier networks. The folder config_train_clf has config files for natural and adversarial training for Cifar-10, SVHN, and Cifar-100. Setting the config parameter adv_steps_train to 0 will lead to natural training, while a positive value will lead to adversarial training.

Running an Experiment

To run an experiment with bpda_eot_attack.py, train_ebm.py, or train_clf.py, just specify the JSON config file:

# json file with experiment config
CONFIG_FILE = './path_to/config.json'

before execution.

Other Files

Network structures are located in nets.py. Experiment setup a dataset import function are in utils.py.

Pre-trained Networks

A pre-trained natural classifier clf.pth and EBM ebm.pth for Cifar-10 are provided in the release section of the repository.

Contact

Please contact Mitch Hill (mitchell.hill@ucf.edu) or Jonathan Mitchell (jcmitchell@ucla.edu) for any inquiries.