This is code for "Margin Maximization for Robust Classification using Deep Learning". The idea is to maximize the decision margin which in turn improves the robustness of classifier. Unlike adversarial training, information about the dependency between model parameters and adversarial noise is not discarded which helps to improve robustness.

Links to the paper and slides.

• Theano
• Lasagne
• tensorboard

To reproduce margin_maximization results run train_margin.py --lmbd 0.1. For other method (standard, at, vat), you can run scripts train_standard.py, train_at.py, train_vat.py with standard parameters. Take a note that in the paper results for vat were reported using full dataset (60000 examples) while for other methods 50000 examples were used.

To test robustness of the model with DeepFool, use test.py --load_dir <load_path>.

If you find this code useful for your research, please consider citing it:

@InProceedings{	  matyasko2017margin,
  author	= {A. Matyasko and L. P. Chau},
  title		= {Margin maximization for robust classification using deep learning},
  booktitle	= {2017 International Joint Conference on Neural Networks (IJCNN)},
  year		= {2017},
  pages		= {300-307},
  month		= {May},
  doi		= {10.1109/IJCNN.2017.7965869}
}