AI Paper Atlas

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

Pre-trained Adversarial Perturbations

Yuanhao Ban, Yinpeng Dong

NeurIPS 2022 · Adversarial robustness and security › Certified and empirical robustness · approach: Adversarial training with feature consistency · 5 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
add uniform Gaussian sampling (UGS) loss (ablation)L4A_base · SimCLRv2 ResNet101attack success rate AVG (%): 66.95 -> 72.20Improved
add uniform Gaussian sampling (UGS) loss (ablation)L4A_base · SimCLRv2 ResNet50attack success rate AVG (%): 71.16 -> 77.80Improved
add uniform Gaussian sampling (UGS) loss (ablation)L4A_base · MAE ViT16attack success rate AVG (%): 94.00 -> 95.30Improved
use fixed ImageNet mean/std noise instead of uniform sampling (ablation)L4A_ugs · ResNet101 / ResNet50 SimCLRv2attack success rate AVG (%): 68.39 vs 69.50Mixed
tune fusing weight lambda to 10^-0.5 (ablation)L4A_ugs · ResNet101attack success rate AVG (%): 1.52%Improved
attack lower layers of the network (ablation)L4A_base · 10 downstream tasksattack success rate AVG (%)Improved
L4A_ugs versus best baseline UAPEPGDL4A_ugs · SimCLRv2 ResNet101attack success rate AVG (%): 59.34 -> 72.20Improved
L4A_ugs versus best baseline FFF_noL4A_ugs · MAE ViT16attack success rate AVG (%): 77.69 -> 95.30Improved
add fused loss on top of base loss (ablation)L4A_base · CLIP ResNet50attack success rate AVG (%): 81.00 -> 91.41Improved
add UGS to fused lossL4A_fuse · MOCO ResNet50attack success rate AVG (%): 54.72 -> 59.74Improved
add UGS to L4A_base under adversarial fine-tuning (ablation)L4A_base · Adversarial fine-tuned ResNet50attack success rate AVG (%): 30.46 -> 30.04No clear effect
use L4A_ugs on adversarial-pretrained standard-finetuned modelL4A_ugs · Adversarial-pretrained ResNet50attack success rate AVG (%): 59.45 -> 68.91Improved

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