AI Paper Atlas

Research map / Adversarial robustness and security / Certified and empirical robustness

On Lp-norm Robustness of Ensemble Decision Stumps and Trees

Yihan Wang, Huan Zhang, Hongge Chen, Duane Boning, Cho-Jui Hsieh

ICML 2020 · Adversarial robustness and security › Certified and empirical robustness · approach: Neural network robustness verification

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Replace ℓ∞ certified training with ℓ1 certified training for boosted stumps (ℓ1 robust training)ensemble decision stumps · MNIST 2 vs. 6 and Fashion-MNIST shoes, ℓ1 verified errorℓ1 verified test error: e.g. 44.22% -> 33.46% (MNIST 2 vs. 6, ϵ1=1.0)Improved
Replace ℓ∞ certified training with ℓ1 certified training for boosted stumpsensemble decision stumps · breast-cancer and diabetes, ℓ1 verified errorℓ1 verified test error: breast-cancer 99.27% -> 35.77% (improves); diabetes 35.06% -> 31.82% (improves)Mixed
Replace ℓ∞ certified training with ℓ1 certified training for boosted treesensemble decision trees · Fashion-MNIST shoes, MNIST 1 vs. 5, MNIST 2 vs. 6, breast-cancer (ℓ1 verified error)ℓ1 verified test error: e.g. 89.54% -> 65.18% (Fashion-MNIST shoes); breast-cancer 99.63% -> 47.05%Improved
Use the DP-based approximate ℓ1 verification instead of MILP for stump ensemblesensemble decision stumps · ℓ1 verification on five datasetsverification time per sample: e.g. 0.11s MILP vs 0.0013s DP on MNIST 1 vs 5 (about 71X speedup)Improved
Use the DP-based approximate ℓ1 verification instead of MILP for stump ensemblesensemble decision stumps · Fashion-MNIST shoes ℓ1 verificationverified err.: 10.45% (MILP) vs 10.55% (DP)Hurt
Use approximate multi-level ℓ1 verification (K=3, L=2) instead of MILP for tree ensemblesensemble decision trees · breast-cancer, diabetes, Fashion-MNIST shoes, MNIST 1 vs. 5 ℓ1 verificationverified err. and avg. time: verified error matches MILP on 4 of 5 datasets (e.g. 8.03% vs 8.03%); MNIST 2 vs. 6 8.60% MILP vs 8.80% approx; speedupsMixed
Use ℓ0 complete verification for stump ensembles (linearithmic algorithm)ensemble decision stumps · five datasets, ϵ0=1verified err. and avg. time: avg time under 0.1 s per sample (e.g. 0.010s on MNIST 1 vs 5)Improved
Use an ε schedule (gradually increasing ε) during certified ℓp training instead of a large fixed ε from the startensemble stumps and trees ℓp training · ℓp certified trainingqualitative (argued)No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy