Learning-Based Attack

Introduction

This repository contains all the code for the paper.

Learning-Based Attack is a black box attack method that uses a model to generate adversarial examples. The model is trained based on the the existing adversarial examples. It can generate sparse adversarial examples that are similar to the existing adversarial examples with fewer queries and more efficient.

The original models are already trained and saved in the saved_models folder.

Usage

Requirements

Run the requirements.txt to install the required packages.

pip install -r requirements.txt

Begin Attack

LBA needs an original attack, thus we provide the original attack methods in the attack folder.

Original Attack

In this paper, we use NVITA to generate the original adversarial examples.

FGSM and BIM are also supported.

Example: NVITA

python ./exp.py 
  --dataset Electricity
  --model CNN
  --attack NVITA
  --n 1
  --epsilon 0.1
  --save_csv
  --print_info
  --plot

Learning-Based Attack

When the original adversarial examples are generated, we can use the LBA to generate the adversarial examples.

Example: LBA

python ./exp.py 
  --dataset Electricity
  --model CNN
  --attack LBA
  --epsilon 0.1
  --save_csv
  --LBA_ori_atk_load_from_csv
  --print_info
  --plot

Note that all the results(*.csv) of the original attack are saved in the results folder.

By setting the LBA_ori_atk_load_from_csv parameter(Default: False), the LBA can directly use the csv file to load the original adversarial examples. If the parameter is set to False, the LBA will generate the original adversarial examples first.

The descriptions of all the parameters can be found in the exp.py file.