Efficient Training of Robust Decision Trees Against Adversarial Examples

Authors: Daniël Vos, Sicco Verwer

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our results on both single trees and ensembles on 14 structured datasets as well as on MNIST and Fashion-MNIST demonstrate that GROOT runs several orders of magnitude faster than the state-of-the-art works and also shows better performance in terms of adversarial accuracy on structured data.
Researcher Affiliation Academia 1Cyber Security Group, Delft University of Technology, Delft, The Netherlands. Correspondence to: Dani el Vos <d.a.vos@tudelft.nl>.
Pseudocode Yes Algorithm 1 Find Best Robust Split on Numerical feature; Algorithm 2 Fit Robust Tree on Numerical Data
Open Source Code Yes We implement and publish GROOT s source code in a Scikitlearn (Pedregosa et al., 2011) compatible classifier. To foster further research we implemented GROOT according to the Scikit-learn API and published the code on Git Hub1. 1https://github.com/tudelft-cda-lab/GROOT
Open Datasets Yes We present results on 14 structured datasets as well as MNIST (Le Cun et al., 2010) and Fashion-MNIST (Xiao et al., 2017). ... All datasets can be retrieved from Open ML2, their specific versions, size and corresponding ϵ values can be found in Table 2. 2https://www.openml.org/
Dataset Splits Yes We train each model on each dataset with 5 fold stratified cross validation. The datasets were randomly split in a 70%-30% stratified train-test split and were evaluated against an L radius of 0.4.
Hardware Specification Yes All experiments ran on a Linux machine with 16 Intel Xeon CPU cores and 72GB of RAM total. Each algorithm instance ran on a single core and therefore did not use any parallel optimizations. All models ran on a system with 8GB RAM and 4 Intel i7-4710MQ CPU cores (8 logical cores), the hyperparameters were the same as in the previous experiment.
Software Dependencies No The paper mentions 'Scikit-learn' and 'XGBoost' but does not specify their version numbers, nor does it list specific versions for other key software components.
Experiment Setup Yes The used hyperparameters are summarized in Table 5. All models required at least 10 samples to make a split and 5 samples to create a leaf. All ensembles were limited to training 100 trees.