/adversarial-patch

PyTorch implementation of adversarial patch

Primary LanguagePython

adversarial-patch

PyTorch implementation of adversarial patch

This is an implementation of the Adversarial Patch paper. Not official and likely to have bugs/errors.

How to run:

Data set-up:

Run attack:

  • python make_patch.py --cuda --netClassifier inceptionv3 --max_count 500 --image_size 299 --patch_type circle --outf log

Results:

Using patch shapes of both circles and squares gave good results (both achieved 100% success on the training set and eventually > 90% success on test set)

I managed to recreate the toaster example in the original paper. It looks slightly different but it is evidently a toaster.

Alt text This is a toaster

Square patches are a little more homogenous due to that I only rotate by multiples of 90 degrees.

Alt text This is also a toaster

Issues:

  • Cannot make a perfect circle with numpy/pytorch. The hack I came up with makes the boundary slightly hexagonal.

  • Rather slow if max_count and conf_target are large.

  • Probably lots of redundant calls and variables.