Growing Robust Trees (GROOT) is an algorithm that fits binary classification decision trees such that they are robust against user-specified adversarial examples. The algorithm closely resembles algorithms used for fitting normal decision trees (i.e. CART) but changes the splitting criterion and the way samples propagate when creating a split.
This repository contains the module groot that implements GROOT as a Scikit-learn compatible classifier, an adversary for model evaluation and easy functions to import datasets. For documentation see https://groot.cyber-analytics.nl
To train and evaluate GROOT on a toy dataset against an attacker that can move samples by 0.5 in each direction one can use the following code:
from groot.adversary import DecisionTreeAdversary
from groot.model import GrootTree
from sklearn.datasets import make_moons
X, y = make_moons(noise=0.3, random_state=0)
X_test, y_test = make_moons(noise=0.3, random_state=1)
attack_model = [0.5, 0.5]
is_numerical = [True, True]
tree = GrootTree(attack_model=attack_model, is_numerical=is_numerical, random_state=0)
tree.fit(X, y)
accuracy = tree.score(X_test, y_test)
adversarial_accuracy = DecisionTreeAdversary(tree, "groot").adversarial_accuracy(X_test, y_test)
print("Accuracy:", accuracy)
print("Adversarial Accuracy:", adversarial_accuracy)The experiment code runs on python version 3.6 and up. After installing python and pip one can install all required libraries using:
pip3 install -r requirements.txt
Or on windows:
pip install -r requirements.txt
We recommend using virtual environments.
To reproduce the results from the paper we provide generate_k_fold_results.py, a script that takes the trained models (from JSON format) and generates tables and figures. The resulting figures generate under /out/.
To not only generate the results but to also retrain all models we include the scripts train_kfold_models.py and fit_chen_xgboost.py. The first script runs the algorithms in parallel for each dataset then outputs to /out/trees/ and /out/forests/. Warning: the script can take a long time to run (about a day given 16 cores). The second script train specifically the Chen et al. boosting ensembles. /out/results.zip contains all results from when we ran the scripts.
To experiment on image datasets we have a script image_experiments.py that fits and output the results. In this script, one can change the dataset variable to 'mnist' or 'fmnist' to switch between the two.
The scripts summarize_datasets.py and visualize_threat_models.py output some figures we used in the text.
The TREANT implementation (groot.treant.py) is copied almost completely from the authors of TREANT at https://github.com/gtolomei/treant with small modifications to better interface with the experiments. The heuristic by Chen et al. runs in the GROOT code, only with a different score function. This score function can be enabled by setting chen_heuristic=True on a GrootTree before calling .fit(X, y). The provably robust boosting implementation comes almost completely from their code at https://github.com/max-andr/provably-robust-boosting and we use a small wrapper around their code (groot.provably_robust_boosting.wrapper.py) to use it. When we recorded the runtimes we turned off all parallel options in the @jit annotations from the code. The implementation of Chen et al. boosting can be found in their own repo https://github.com/chenhongge/RobustTrees, from whic we need to compile and copy the binary xgboost to the current directory. The script fit_chen_xgboost.py then calls this binary and uses the command line interface to fit all models.
To encode L-infinity norms correctly we had to modify TREANT to NOT apply rules recursively. This means we added a single break statement in the treant.Attacker.__compute_attack() method. If you are planning on using TREANT with recursive attacker rules then you should remove this statement or use TREANT's unmodified code at https://github.com/gtolomei/treant .
For any questions or comments please create an issue or contact me directly.