GenAttack: Practical Black-box Attacks with Gradient-Free Optimization.

This repo has an implemntation for our paper GenAttack: Practical Black-box Attacks with Gradient-Free Optimization

Instructions

Setup

Install the required libraries:

pip install -r requirements.txt

ImageNet Experiment

Download Inception-v3 model checkpoint

python setup_inception.py

You can download test images from ImageNet test set.

To run the attack without dimensionality reduction and adaptive parameter scaling

 python main.py --input_dir=./images/ --test_size=1 \
    --eps=0.05 --alpha=0.15 --mutation_rate=0.005  \
    --max_steps=500000 --output_dir=attack_outputs  \
    --pop_size=6 --target=704 --adaptive=False

Original class: Squirrl, Adversarial class: Parking Meter, Number of queries=74,171

For more query efficiency

Run attack with dimensionality reduction and adaptive parameter scaling

python main.py --input_dir=./images/ --test_size=1 \
    --eps=0.05 --alpha=0.15 --mutation_rate=0.10  \
    --max_steps=100000 --output_dir=attack_outputs \
    --pop_size=6 --target=704 --adaptive=True --resize_dim=96

Original class: Squirrl, Adversarial class: Parking Meter, Number of queries=11,696

More options:

If you want to test on a single image, add the FLAG: --test_example=xx.
To specify a target class, instead of using a random target, add the flag --target=xx.

MNIST and CIFAR-10 Experiments

First, you need to train the classification models on MNIST and CIFAR-10 datasets.

python train_models.py

Attacking MNIST Model

 python main.py --model=mnist --test_size=1000 --mutation_rate=0.30 --alpha=0.5 --adaptive=False --max_steps=10000 --eps=0.30  --output_dir=mnist_output --pop_size=4  --temp=0.1

Attacking CIFAR-10 Model:

python main.py --model=cifar10 --test_size=1000 --mutation_rate=0.05 --alpha=0.25 --adaptive=False --max_steps=10000 --eps=0.05  --output_dir=cifar10_output --pop_size=4  --temp=0.1

Maintainer:

This project is maintained by: Moustafa Alzantot (malzantot)

nesl/adversarial_genattack