AI Paper Atlas

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

Robust Learning against Relational Adversaries

Yizhen Wang, Mohannad Alhanahnah, Xiaozhu Meng, Ke Wang, Mihai Christodorescu, Somesh Jha

NeurIPS 2022 · Adversarial robustness and security › Certified and empirical robustness · approach: Attack evaluation and defense critique

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
add normalization (N&P) to RFVanilla-RF · Authorship, non-adaptive attackaccuracy (%): 13.2±3.3 -> 72.3±4.0Improved
add normalization (N&P) to LSTMVanilla-LSTM · Authorship, non-adaptive attackaccuracy (%): 21.1±2.8 -> 73.7±3.6Improved
add normalization to RF under adaptive attackVanilla-RF · Authorship, adaptive attackaccuracy (%): 13.2±3.3 -> 70.5±4.6Improved
add normalization to LSTM under Adaptive+ attack (7 unnormalized transformations)Vanilla-LSTM · Authorship, Adaptive+ attackaccuracy (%): 0.9±0.4 -> 49.9±7.4Improved
add normalization to RF under Adaptive+ attackVanilla-RF · Authorship, Adaptive+ attackaccuracy (%): 1.0±0.4 -> 37.3±13.3Improved
add normalization to RF on clean inputsVanilla-RF · Authorship, cleanaccuracy (%): 90.4±1.7 -> 76.1±3.8Hurt
add normalization to LSTM on clean inputsVanilla-LSTM · Authorship, cleanaccuracy (%): 88.4±3.7 -> 78.7±4.8Hurt
replace adversarial training with N&P-LSTMAdv-LSTM · Authorship, Adaptive+ attackaccuracy (%): 25.2±5.1 -> 49.9±7.4Improved
replace adversarial training with normalizationAdv-LSTM · Authorship, adaptive attackaccuracy (%): 30.8±4.0 -> 71.2±4.8Improved
replace adversarial training with normalization (training cost)Adv-LSTM · Authorship training timedays / hours: more than 20 days -> less than 12 hours overheadImproved
add normalization before adversarial training (Unified) vs adversarial training aloneAdv-Trained · Sleipnir adversarial inputsFNR (%): 27.9±8.2 -> 5.5±0.5Improved
add API substitution normalization on top of Al-Dujaili adversarial trainingAl-Dujaili et al. [2018] · Sleipnir adversarial inputsFNR (%): 89.9±7.8 -> 5.5±0.5Improved

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