This is an easy approach for the competition "Facial Adversary Examples" in TIANCHI, which can get 3.5 in score based the evaluation criterion of the competition.

1. Download the dataset from TIANCHI. Suppose the directory is $DATA_DIR.

2. Download the pretrained Face-Recognition models from Baidu (Extraction code: sjqs).

3. Download the feature files from Baidu (Extraction code: jf2z). Or you can use the script attack/preprocess_eval.py to generate these files.

4. Init attack mask directory:

   mkdir attack/masks
   

   Your directory tree should look like this:

   ${PROJECT_HOME}
   ├── attack
       ├── log
       ├── masks
       ├── state
       └── *.py
   ├── model
       └── downloaded models
   ├── result
       └── downloaded features
   ├── ...
   └── ...
   

• python 3.6
• PyTorch 1.0.1
• CUDA 9.0
• CUDNN 7.1.2
• opencv 3.4.2
• numpy 1.15
• scipy 1.2.0

• The code is developed using python 3.6 on Ubuntu 18.04. NVIDIA GPUs are needed.
• The code is tested using 1 NVIDIA 1080Ti GPU card. Other platforms or GPU cards are not fully tested.
• OpenCV is installed through anaconda, which is a little different with installed through pip.

cd $PROJECT_HOME/attack

python attack.py \
    --root $DATA_DIR/securityAI_round1_images \
    --dev_path $DATA_DIR/securityAI_round1_dev.csv \
    --output_path $OUTPUT_PATH

We develop our attack codes based wujiyang's Face_Pytorch.