v-wangg/orthogonalpgd

Evading Adversarial Example Detection Defenses with Orthogonal Projected Gradient Descent

This repository contains the code used in the paper

The main attack class can be found in attack.py. We then give a few examples of the case studies we talk about in the paper.

We include defense scripts for the trapdoor and DLA defenses because we re-implemented these in pytorch ourself, and so we need no author permission to distribute these files. We do not include pre-trained models because the original defense authors trained models for us---we will obtain their permission to release those pretrained models in a full code release. Similarly we will release other defense with author permission on code reuse.

The script "run_experiment.py" contains the necessary boilerplate to run an attack given pretrained model weights.

orthogonalpgd's Issues

Could you provide cifar_model.h5, torch_cifar_res.p, torch_cifar_model.h5, and cifar_alarm.h5 in the repo?

Thanks a lot.

Detector model

I want to run the code, but there is no trained weights file. Could you please release the trained model file or the detector training code?

pretrain models

Hi, I need cifar_model.h5 and cifar_alarm.h5.
thanks

Pretrained models

Hey, I was wondering when you will upload the pretrained models of the detectors.

Thank you in advance.

Could you add a LICENSE in this repo?

Thanks a lot.