bravolu / ssae

the source code of ACM MM'21 paper "Discriminator-free Generative Adversarial Attack"

[ACM MM'21] Discriminator-Free Generative Adversarial Attack

Installation

> git clone https://github.com/BravoLu/SSAE.git
> cd SSAE
> pip -r install requirements.txt

Get Started

Adversarial Attack for Image Classification

Adversarial Attack for Image Retrieval

Demo

• We deploy the demo on http://124.156.185.206:9988/, you can use the images in directory ./images/original_examples/ for testing.

Citation

@inproceedings{lu2021s,
  title={Discriminator-Free Generative Adversarial Attack},
  author={Shaohao Lu, Yuqiao Xian, Ke Yan, Yi Hu, Xing Sun, Xiaowei Guo, FeiyueHuang, and Wei-Shi Zheng},
  booktitle={In Proceedings of the 29th ACM International Conference onMultimedia (MM ’21)},
  year={2021}
}

ssae's Issues

A question about SSAE structure

sir, I wonder why use deepcopy rather than copy, copy use the same parameters in the three decode layers but deepcopy need relearn parameters which maybe consume more time?

    self.mask_decoder_1 = deepcopy(self.perturb_decoder_1)
    self.mask_decoder_2 = deepcopy(self.perturb_decoder_2)
    self.mask_decoder_3 = nn.Sequential(*end_layers2)

a question about saving pictures after the attack

I saved the pictures after the attack, read them back into CV2, and then calculate the cosine similarity. The effect almost disappeared. Is it because the RGB value is rounded and the information is lost when saving?

the results of an experiment are as follows:
cosine_direct: 0.6383863687515259
cosine_readback: 1.8832021951675415

Cannot connect to the demo site

There seems to be an issue connecting to the demo website?
I cannot access to the demo properly.

Hello，Based on the code you provided, AdvGAN could not produce the results of the paper. May I ask what is the cause.

Hello，Based on the code you provided, AdvGAN could not produce the results of the paper. May I ask what is the cause.

Originally posted by @zzx16045223 in #3 (comment)

Your work is the same as adding the random noises

Hello, I found that adding the random noises to the images get the same attack performances as your work. It seems that your work doesn't work at all.

AdvGAN