Data Poisoning based on Adversarial Attacks using Non-Robust Features

Usage

python main.py [-h] [--gpu | -g GPU]  [--eps |-e EPSILON] [--pert | -p PERTURBATION_PERCENTAGE] [--loss_fn | -l LOSS_FUNCTION] [--layer_cuts | -c LAYER_CUTS] [--target_class | -t TARGET_CLASS] [--new_class | -n NEW_CLASS] [-v | --eva] [--dataset | -d DATASET] [--resnet | -m] [--transfer | -f] [--rand | -a] [--iters | -s ITERATIONS]

Arguments

Examples

python main.py --gpu 0 --eps 2 1 0.75 0.5 0.25 0.1 --pert 0.5 --loss_fn 2 --layer_cuts 1 2 --dataset 0 --target_class "deer" --new_class "horse"

Would use deer as the target class and horse as the new class to create 12 datasets. Six datasets with ​epsilon = [2, 1, 0.75, 0.5, 0.25, 0.1] and the activations from the last dense layer and six datasets with the same epsilon values but the activations from the penultimate dense layer. Both datasets contain 50% perturbed images and the generation as well as the training is performed on GPU:0. The model used is the standard CNN while the dataset is a unmodified CIFAR10 dataset.

python main.py --gpu 1 --dataset 1 --target_class "bee" --new_class "beetle" --resnet --transfer --rand --iters 10 --best

Would load the attack parameters from results/attack_results.pkl for the chosen class combination and would choose 10 times a random target image to test these parameters on.

Untargeted Attack Test-Calls

python3 main.py --gpu 0 --dataset 0 --eps 0.5 --pert 1.0 --loss_fn 2 --resnet --transfer --untargeted --rand --cluster 1 --iters 10

Download TinyImageNet

wget -nc http://cs231n.stanford.edu/tiny-imagenet-200.zip