AI Paper Atlas

Research map / Adversarial robustness and security / Black-box and transfer attacks

LGV: Boosting Adversarial Example Transferability from Large Geometric Vicinity

Martin Gubri, Maxime Cordy, Mike Papadakis, Yves Le Traon, Koushik Sen

ECCV 2022 · Adversarial robustness and security › Black-box and transfer attacks · approach: Transferable adversarial examples · 40 citations (OpenAlex)

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 single DNN surrogate with LGV (collect weights along high-lr SGD)1 DNN surrogate · ImageNet L-inf, target RN50success rate: 45.3 -> 95.4Improved
use SGD-collected LGV weights instead of random weight directions (RD) (ablation)LGV · ImageNet L-inf, target RN50success rate: 60.6 -> 95.4Improved
use ensemble of 40 LGV models instead of their average (LGV-SWA) (ablation)LGV · ImageNet L-inf, target RN50success rate: 84.9 -> 95.4Improved
add random directions around LGV-SWA (ablation)LGV-SWA · ImageNet L-inf, 8 targetssuccess rate: +1.32 to +10.18 points (average 6.90)Improved
add input diversity and momentum to LGVLGV · ImageNet L-inf, target IncV3success rate: 25.3 -> 45.4Improved
add Ghost Networks to LGVLGV · ImageNet L-inf, target RN50success rate: 95.4 -> 94.2Hurt
add SGM to LGVLGV · ImageNet L-inf, target RN50success rate: 95.4 -> 86.9Hurt
add random Gaussian noise in weight space to single DNN (ablation)1 DNN · ImageNet L-inf, target RN50success rate: 45.3 -> 60.6Improved
shift LGV deviations to another solution (LGV-SWA + (LGV' - LGV-SWA')) (ablation)LGV-SWA+RD · ImageNet L-infsuccess rate: +3.27 to +12.32 points over LGV-SWA + RD (mean 8.61)Improved
use unscaled instead of scaled (gamma 0.5) LGV deviations on regular DNN (ablation)1 DNN + gamma(LGV'-LGV-SWA') · ImageNet L-infsuccess rate: -32.8 points on averageHurt

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