shikiw / si-adv Goto Github PK

Thanks for your great work.
I wonder how to visualize the sensitive maps as in your paper.
Thank you in advance!

i ran this command
python main.py --transfer_attack_method ifgm_ours

I would like visualize the before and after the results of the point cloud attack,
inside the class function

def run(self, points, target):
# i added a code to save the input, after the adversarial attack, save the output

but i am noticing the number of points of the output reduced to 1/2 of the input

how can i get it the number of output points to be the same as the number of input points?

this picture show the input on the left and output on the right

Thanks for your great work! I'm a little bit confused about the attack pipeline.

I notice that you put the defense_head(e.g., sor) inside the attack iterations by:

if not self.defense_method is None:
    logits = self.classifier(self.pre_head(inputs.detach()))
else:
    logits = self.classifier(inputs.detach()) # [1, num_class]

However, in the implementation of IF-Defense, they do not put the sor in the attack iterations to generate adversarial points. I think your setting simulates where the attacker gets full information of the model AND the defense methods while the IF-Defense assume the attack only knows the model but does not know the defense methods.

I also tried the 2nd setting (e.g., do not use defense_head in the attack iterations) and found the ASR of SI-Adv largely decreased. I'm wondering which setting is more valid.

Just as the author of IF-Defense mentioned in Wuziyi616/IF-Defense#12, I think the sor defense is also non-differentiable, so can we plug it into the attack iterations?

Hello, I have completed the construction of the conda environment according to the environment.yml you gave, but when I use "python main.py --transfer_attack_method ifgm_ours" to train, the error "RuntimeError: CUDA error: no kernel image is available for execution on the device", how do I solve this problem? (Is my pytorch version too high?)

Why does replacing the dataset in command
‘python main.py
--dataset ModelNet40
--data_path /your/path/to/dataset/
--transfer_attack_method ifgm_ours
--surrogate_model pointnet_cls
--target_model pointnet_cls
--step_size 0.007
--max_steps 50
--eps 0.16’
with ShapeNetPart report an error as follows：
‘UnboundLocalError: local variable 'batch_id' referenced before assignment’？Looking forward to your response, thanks!