Faster Repeated Evasion Attacks in Tree Ensembles
Authors: Lorenzo Cascioli, Laurens Devos, Ondrej Kuzelka, Jesse Davis
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, we address three questions: (Q1) Is our approach able to improve the run time of generating adversarial examples? (Q2) How does ensemble complexity affect our approach s performance? (Q3) What is our empirical false negative rate? ... we present numerical experiments for ten binary classification tasks on high-dimensional datasets... |
| Researcher Affiliation | Academia | Lorenzo Cascioli Department of Computer Science KU Leuven Leuven, Belgium lorenzo.cascioli@kuleuven.be Laurens Devos Department of Computer Science KU Leuven Leuven, Belgium Ondrej Kuzelka Faculty of Electrical Engineering Czech Technical University in Prague Prague, Czech Republic Jesse Davis Department of Computer Science KU Leuven Leuven, Belgium |
| Pseudocode | Yes | Algorithm 1 Fast repeated adversarial example generation |
| Open Source Code | Yes | The source code for the presented algorithms and all the experiments is publicly available at https: //github.com/lorenzocascioli/faster-repeated-evasion-tree-ensembles. |
| Open Datasets | Yes | Table 3 gives specific reference to each of the considered datasets. covtype https://www.openml.org/d/1596 fmnist https://www.openml.org/d/40996 higgs https://www.openml.org/d/42769 miniboone https://www.openml.org/d/44128 mnist https://www.openml.org/d/554 prostate https://www.openml.org/d/45672 roadsafety https://www.openml.org/d/45038 sensorless https://archive.ics.uci.edu/dataset/325 vehicle https://www.openml.org/d/357 webspam https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary.html#webspam |
| Dataset Splits | Yes | We apply 5-fold cross validation for each dataset. We use four of the folds to train an XGBoost [9], random forest [3, 25] or GROOT forest (a robustified ensemble type [32]) ensemble T . |
| Hardware Specification | Yes | The experiments were run on an Intel(R) E3-1225 CPU with 32Gi B of memory. |
| Software Dependencies | No | The paper mentions software like XGBoost, random forest (implicitly scikit-learn), GROOT forest, and Gurobi (for Kantchelian), but does not provide specific version numbers for these software components or libraries. |
| Experiment Setup | Yes | Table 1 also reports the adopted values of maximum perturbation δ and the hyperparameters of the learned ensembles, which were selected via tuning using the grid search described in Appendix B. |