Overview

This is pytorch code for NIPS adversarial attack/defence.

Stack: Python3.5, Pytorch, foolbox

DockerImage based on def. pytorch, contains: miniconda 3.5, pytorch, foolbox

Structure

• defence/ - code for the defence competition. For now simple inceptionv3 model on 6 TTA.
  • metadata.json uses my docker image with pytorch & foolbox (fork of pytorch default Dockerfile)
    • model_name point to a model in torch_defence/models.py
  • models/ contains pretrained CNN models

Defence

To download default model execute: cd defence/models; wget https://download.pytorch.org/models/inception_v3_google-1a9a5a14.pth or specify path inside run-defense.sh

Example of running the defence using Docker image (to replace prefixes):

OUTPUT_DATA=/home/tyantov/workspace/kaggle-nips-adversarial-attacks/output_torch
SUBMISSION_DIRECTORY=/home/tyantov/workspace/kaggle-nips-adversarial-attacks/defence
INPUT_IMAGES=/home/tyantov/workspace/kaggle-nips-adversarial-attacks/images_tf
DOCKER_CONTAINER_NAME=tyantov/nips-adv-pytorch

time nvidia-docker run --shm-size="1024m"\
  -v ${INPUT_IMAGES}:/input_images \
  -v ${OUTPUT_DATA}:/output_data \
  -v ${SUBMISSION_DIRECTORY}:/code \
  -w /code \
  ${DOCKER_CONTAINER_NAME} \
  ./run_defense.sh \
  /input_images \
  /output_data/result.csv

Docker installation

Some notes:

• install Docker CE
  • to run docker without sudo: sudo gpasswd -a <USERNAME docker
• install Nvidia-docker (nvidia drivers should be installes on the host)