On Lp-norm Robustness of Ensemble Decision Stumps and Trees
Authors: Yihan Wang, Huan Zhang, Hongge Chen, Duane Boning, Cho-Jui Hsieh
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the first certified defense method for training ensemble stumps and trees with respect to ℓp norm perturbations, and verify its effectiveness empirically on real datasets.In this section we empirically test the proposed algorithms for ℓp robustness verification and training. The code is implemented in Python and all the experiments are conducted on a machine with 2.7 GHz Intel Core i5 CPU with 8G RAM. Our code is publicly available at https://github.com/Yihan Wang617/On-ell_p-Robustnessof-Ensemble-Stumps-and-Trees |
| Researcher Affiliation | Academia | 1 Tsinghua University, Beijing, China 2UCLA, Los Angeles, USA 3MIT, Cambridge, USA. |
| Pseudocode | No | The paper describes algorithms (e.g., 'efficient dynamic programming based algorithm') and states that details are in an appendix ('We include the detail algorithm for enumerating the size-K cliques in Appendix 1'), but no pseudocode or algorithm blocks are present in the provided text. |
| Open Source Code | Yes | Our code is publicly available at https://github.com/Yihan Wang617/On-ell_p-Robustnessof-Ensemble-Stumps-and-Trees |
| Open Datasets | Yes | breast-cancer, diabetes, Fashion-MNIST shoes, MNIST 1 vs. 5, MNIST 2 vs. 6 |
| Dataset Splits | No | The paper mentions 'training set' and 'test error' but does not specify exact training/validation/test dataset splits (percentages or sample counts) needed for reproduction. |
| Hardware Specification | Yes | all the experiments are conducted on a machine with 2.7 GHz Intel Core i5 CPU with 8G RAM. |
| Software Dependencies | No | The paper states 'The code is implemented in Python' but does not provide specific version numbers for Python or any other software libraries or dependencies. |
| Experiment Setup | Yes | Our goal is to select the 4 parameters (b, j, wl, wr) robustly by minimizing the minimax loss: min j,b,wl,wr (x,y) S max δ p ϵ L(y F(x + δ))... When features are correlated in ℓp cases, we find that it is important to have an ϵ schedule during the training process the ϵ increases gradually from small to large, instead of using a fixed large ϵ in the beginning. We also include the choice of ϵ schedules in Appendix D.1. |