Lethean Attack: An Online Data Poisoning Technique

Heavily based on the source code by Y. Sun, X. Wang, Z. Liu, J. Miller, A. A. Efros, and M. Hardt. Test-time training for out-of-distribution generalization, 2019

See https://arxiv.org/abs/2011.12355 for more details.

Prerequisites

• Python 3.5+
• PyTorch
• torchvision

Fetch Data

CIFAR-10 is fetched automatically using torchvision. To download CIFAR-10-C (used by test.py):

mkdir -p data && cd data
wget https://zenodo.org/record/2535967/files/CIFAR-10-C.tar
tar xvf CIFAR-10-C.tar

Train a Test-Time Training Model

python main.py --batch_size 128 --outf [MODEL_DIRECTORY] --epochs 150

Perform a Lethean Attack

python adversarial_lethean.py --resume [MODEL_DIRECTORY] --epochs 5000

Perform an FGSM Attack

python adversarial_fgsm.py --resume [MODEL_DIRECTORY] --epochs 5000

Test using CIFAR-10-C data

python test.py --resume [MODEL_DIRECTORY] --epochs 5000 --corruption [CORRUPTION] --level [1/2/3/4/5]

Possible corruption types:

• brightness
• defocus_blur
• fog
• gaussian_blur
• glass_blur
• jpeg_compression
• motion_blur
• saturate
• snow
• speckle_noise
• contrast
• elastic_transform
• frost
• gaussian_noise
• impulse_noise
• pixelate
• shot_noise
• spatter
• zoom_blur