NIPS 2017: Targeted Adversarial Attack submission

This project is final submission on kaggle competition NIPS 2017: Targeted Adversarial Attack by team "erko" (Yerkebulan Berdibekov)

To read description of this project click here

Requirements:

Configure on host OS CUDA & cuDNN 6, docker-ce, nvidia-docker.

Instructions:

• Download dev_toolkit - https://www.kaggle.com/c/nips-2017-defense-against-adversarial-attack/data & extract to dev_toolkit folder;
• Download developement set - https://www.kaggle.com/google-brain/nips-2017-adversarial-learning-development-set and extract into dev_toolkit/dataset/;
• Inside of developement toolkit prepare sample_attacks, sample_targeted_attacks, sample_defenses by downloading checkpoints. Run sh download_checkpoints.sh in this folders;
• Put this project in new folder in sample_targeted_attacks;
• In this new folder, download checkpoint files: run sh download_checkpoints.sh;
• Edit file run_attacks_and_defenses.sh: append argument --gpu in line executing python script;
• Run sh run_attacks_and_defenses.sh;
• Compare this targeted-attack to other targeted-attacks in accuracy_on_targeted_attacks.csv file in resulting output folder.

Note: batch_size differs from actual submission, from 10 reduced to 4 to be able to run in 8GB GPU (GTX 1080)

Possible problems:

• Facing errors: "****.sh: Permission denied!" - may be need to make .sh file runnable.