Fast Provably Robust Decision Trees and Boosting

The implementation of Fast Provably Robust Decision Trees and Boosting.

Simple example

FPRDT

To train and evaluate FPRDT on dataset mnist1v7, one can use the following command:

python main.py -m FPRDT --dataset mnist1v7 --epsilon 0.4

PRAdaBoost

To train and evaluate PRAdaBoost on dataset mnist1v7, one can use the following command:

python main.py -m PRAdaBoost --dataset mnist1v7 --epsilon 0.4

Other optional parameters

'''
--max_tree_num: int
  The number of maximum trees in PRAdaBoost.
--max_depth: int 
  The maximum depth of the FPRDT.
--min_samples_split: int 
  The minimum samples needed to split a leaf.
--min_samples_leaf: int
  The minimum samples needed to create a leaf.
--epsilon: float
  The maximum perturbation size.
--k_times: int
  The times to evaluate the model.
--k_folds: int
  The number of folds in each evaluating time.
--random_seed: int
  The random seed used to split the dataset.
--record_loss
  If record the curve of the loss in the training process.
--record_img
  If generate adversarial images after training process.
'''