This is the submissions of the UCNesl team during the Competition on Adversarial Examples and Defenses which is held as a part of NIPS'17 conference.
- Moustafa Alzantot (malzantot) - Leader for targeted attack.
- Yash Sharma (ysharma1126) - Leader for non-targeted attack.
- Supriyo Charkaborty (supriyogit)
- Tianwei Xing (TianweiXing)
- Sikai Yin (Sikai Yin)
- Prof. Mani Srivastava
The attacks in this repo were used as submissions for the final round of the competition.
Round 2 submission by team UCNesl can be found in this repo.
The following software is required to run this package:
- Python 2.7 with Numpy and Pillow Packages.
- Docker or NVIDIA Docker for GPU support.
- Download and extract the competition toolkit.
- Under the competition toolkit run the script
./download_data.sh
to download the development dataset. - Copy the attacks from this repo (e.g. ucnesl_targeted, ucnesl_nontargeted to the appropriate folders under the competition toolkit directory (e.g. sample_targeted, and sample_attacks for targeted and non-targeted attacks respectively).
- Under each attack folder, run the
./download_checkpoints.sh
script to download the models checkpoints. - Finally, you can run and evaluate the attacks by using
run_attacks_and_defenses.sh
script in the competition toolkit directory.
Note For GPU-support, you should edit the the line #43 of the run_attacks_and_defenses.sh
to add the following argument --gpu
to the python command.