Attack as Defense

This repository contains the code for the ISSTA'21 paper Attack as Defense: Characterizing Adversarial Examples using Robustness and its extended version.

Requirements & Setup

The experiment environment is based on the baseline, https://github.com/rfeinman/detecting-adversarial-samples ("Detecting Adversarial Samples from Artifacts", Feinman et al. 2017).

We implemented 'Attack as Defense' based on this repo.

Req:

Foolbox == 1.8.0

Code Structure

/attack_method, attack methods code, from Foolbox.

/data, samples and models

/results, output results

/scripts, scripts that users will run

Running the Code

Follow the baseline, all of the scripts for running the various parts of the code are located in the scripts/ subfolder.

In this repo, we will show how to run 'attack as defense' method on the MNIST dataset.

We provide a natural trained model in the data folder so user do not need to train a new model, the training steps are followed the baseline repo. We also provide the adversarial examples generate by the baseline, for C&W attack, as the baseline has not provide this part of the attack code yet, so we used the Foolbox tool with default parameters to generate these examples.

1. Attack adversarial examples

CUDA_VISIBLE_DEVICES=XX python attack_iter_generate.py -a BIM -i fgsm

where <-a> defines the attack methods used for defense, and <i> defines the input data.

After executing this code, we can generate the BIM attack costs of fgsm examples.

User can use the scripts bash attack_iter_generate.sh to generate adversarial and benign samples attack costs automatically.

2. Attack benign samples

CUDA_VISIBLE_DEVICES=XX python attack_iter_generate.py -a BIM -i benign

User can use the scripts bash attack_iter_generate.sh to generate adversarial and benign samples attack costs automatically.

3. Get the AUROC results

python calc_auc.py -a BIM

where <-a> defines the attack methods used for defense.

4. Detect

The code contains two detection methods, k-nn based detector and z-score based detector.

CUDA_VISIBLE_DEVICES=XX python main_adv_detector.py -d knn -a BIM --init

CUDA_VISIBLE_DEVICES=XX python main_adv_detector.py -d zscore -a BIM

For the first time to execute it, please add the --init parameter, which generates the training set for the detection.

Note: We extended the work to detect other types of abnormal samples, and the related code is located under the \extension folder.