search

psandovalsegura / autoregressive-poisoning

Code for the paper "Autoregressive Perturbations for Data Poisoning" (NeurIPS 2022)
MIT License
18 stars 5 forks source link

Attack on CIFAR-100 #3

Open EhanW opened 2 months ago

EhanW commented 2 months ago

Hi, sir! I found that the poisoned images generated for one class (class 9) in CIFAR-100 are completely black. What could be the reason for this?

psandovalsegura commented 2 months ago

Can you specify which file from the Google Drive? I'm assuming it's ar-cifar-100.zip but I want to make sure. I uploaded the AR processes for 100 classes in params-classes-100-mr-3.pt, so let me check if there was any issue with those.

EhanW commented 2 months ago

yes, ar-cifar-100.zip

psandovalsegura commented 2 months ago

Using Generate-AR-Perturbations-from-Coefficients.ipynb I checked params-classes-100-mr-3.pt and it does look like the AR process at index 9 returns nan values.

I've committed (424cc2e) a new copy of params-classes-100-mr-3.pt which you should be able to see generates correct AR perturbations using the notebook above. I also added a check to the param finder to prevent this for new AR process param searches.

You should now be able to regenerate the CIFAR-100 poison using python create_ar_poisons_3channel.py ar-cifar-100 CIFAR100 --epsilon 1.0 --p_norm 2 or if you want I will generate and reupload to the Google Drive. I'll update here when that's done.

psandovalsegura commented 2 months ago

@EhanW A fixed CIFAR-100 poison (ar-cifar-100.zip) is now in the Google Drive. I've also uploaded a notebook which checks that class 9 is now ok: Check-CIFAR100-Perturbations.ipynb.

Let me know if this helped!

EhanW commented 2 months ago

Thanks a lot!