Identifying Attack-Specific Signatures in Adversarial Examples

This repository is the official implementation of the Identifying Attack-Specific Signatures in Adversarial Examples.

Getting Started

Clone this repository into a directory of your choosing with the following command:

cd path/to/your/directory/
git clone https://github.com/hsouri/REDRL

Next go into the cloned directory cd ./REDRL and create a directory for future checkpoints by the following command:

mkdir checkpoints

At this point the repository structure should be ready. In the next step make sure that the provided path to the pre-trained image classifiers models and the data is correct in the ./tools/path.py script. If not modify them accordingly.

To start trainig the pipeline go to ./experiment_scripts/ and run the following command:

sh $dataset_train.sh

Before start to train, make sure all the settings such as involved loss functions, their weights, and etc are the ones you would want.

Citations

Please cite our paper:

@inproceedings{souri2024identifying,
  title={Identifying Attack-Specific Signatures in Adversarial Examples},
  author={Souri, Hossein and Khorramshahi, Pirazh and Lau, Chun Pong and Goldblum, Micah and Chellappa, Rama},
  booktitle={ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={7050--7054},
  year={2024},
  organization={IEEE}
}