SCLBD / iccv2021_meta_attack

Meta-Attack: Class-agnostic and Model-agnostic Physical Adversarial Attack

this paper is received by iccv 2021

Method

Dataset and pretrained model download

1. download the pre-trained classifier on GTSRD from https://drive.google.com/drive/folders/1l4s6XLx9PVRSUl1ZXUguK8Wa_1pGXkBH?usp=sharing
2. collect digital and physical images dataset of GTSRD or ImageNet , put them in image folder like this:

image
	|
	|--label1
	|       |--TrainA(digital)
	|       |--TrainB(physical)
	|
	|--label2
	|       |--TrainA(digital)
	|       |--TrainB(physical)
	|--label3
	|       |--TrainA(digital)
	|       |--TrainB(physical)
	.
	.
	.

4. run ./run.sh to train the GAN networks

5. run ./run_test.sh to finetune the GAN on unseen images and unseen classifiers

6. evaluate the generated adversarial images in checkpoints folder