Adversarial Parameter Attack

This is the official repository of "Adversarial Parameter Attack on Deep Neural Networks" (ICML 2023).

Attack

This repository supports training adversarial parameter attacks of pretrained robust networks on CIFAR-10.

Pretrained robust networks

A pretrained ResNet-18 using PGD-AT is saved as 'logs/resnet18/pgd.pth'.

If you want to train a network from scratch using PGD-AT or TRADES, run

python main.py \
    --model-name resnet18 \
    --log-path logs/resnet18 \
    --train [pgd / trades] \
    --device cuda 

$L_0$-norm attacks

To apply $L_0$-norm attacks on the pretrained ResNet-18, run:

python main.py \
    --pert-mode zero \
    --model-name resnet18 \
    --log-path logs/resnet18 \
    --ckpt-path logs/resnet18/pgd.pth \
    --dataset cifar10 \
    --device cuda

$L_\infty$-norm attacks

To apply $L_\infty$-norm attacks on the pretrained ResNet-18, run:

python main.py \
    --pert-mode inf \
    --model-name resnet18 \
    --log-path logs/resnet18 \
    --ckpt-path logs/resnet18/pgd.pth \
    --dataset cifar10 \
    --device cuda

Attacking using a small trainig set

To apply attacks with a small training set in CIFAR-10, run:

python main.py \
    --pert-mode [inf / zero] \
    --dataset cifar10-small