Hardening Deep Neural Networks via Adversarial Model Cascades

Official code repository for our publication Hardening Deep Neural Networks via Adversarial Model Cascades, accepted at IJCNN'19. A short blog summarizing the paper can be found here.

Working

• Adversary gets hold of unlabelled data (not overlapping with target)
• Make queries to black-box model, use results to create dataset.
• Using this collected data, train a proxy network.
• Generate adversarial examples for the trained proxy network.
• Calculate error rate on the target network, for images generated by the proxy adversarial network.

Attacks present

• FGSM
• VAP
• PGM
• EAP

Setting it up

• bash prepare.sh to download required data and models
• python Code/test_accuracy.py --model_path <target_model> --dataset <dataset> to get test accuracy ,from Code/ folder
• bash test*.sh <dataset> <target_model> <proxy_model>, where * denotes anoy of the 7 attacks given in the repo ,from the Scripts/ folder
• For the basic bagging setup, run bash genericBagging.sh <dataset> <path_to_seed_model> <new_folder_for_bag> <path_to_file_containing_order_of_attacks> <transfer_parameters_per_bag?> ,from the Scripts/ folder

For example, bash genericBagging.sh mnist PlainModel MYBAG/ ORDER no

• For the adaptive bagging setup, run bash adaptiveBagging.sh <dataset> <path_to_seed_model> <new_folder_for_bag> <path_to_file_containing_order_of_attacks> <transfer_parameters_per_bag?> <path_to_proxy_model>

For example, bash adaptiveBagging.sh mnist PlainModel MYBAG/ ORDER no ProxyNormal

• For testing bagging on your own attack data, run python ../Code/bagging.py --mode test --dataset <dataset> --model_dir <model_bag_directory> --data_x <data_X> --data_y <data_Y> --predict_mode <voting/weighted>

Cite this work

You are encouraged to cite the following paper if you use AMC for academic research.

@INPROCEEDINGS{8851970,
  author={D. {Vijaykeerthy} and A. {Suri} and S. {Mehta} and P. {Kumaraguru}},
  booktitle={2019 International Joint Conference on Neural Networks (IJCNN)},
  title={Hardening Deep Neural Networks via Adversarial Model Cascades},
  year={2019},
  pages={1-8},
  doi={10.1109/IJCNN.2019.8851970},
  ISSN={},
  month={July}
}