AI Paper Atlas

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

Are Vision Transformers Robust to Patch Perturbations?

Jindong Gu, Volker Tresp, Yao Qin

ECCV 2022 · Adversarial robustness and security › Certified and empirical robustness · approach: Adversarial training with feature consistency · 82 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 ResNet50 with DeiT-small (same training recipe), 196 naturally corrupted patchesViT vs CNN · ImageNet-1k valfooling rate (%): 49.8 -> 38.9Improved
replace ResNet50 with DeiT-small, 1 adversarial patchViT vs CNN · ImageNet-1k valfooling rate (%): 30.6 -> 61.5Hurt
replace ResNet18 with DeiT-tiny, 1 adversarial patchViT vs CNN · ImageNet-1k valfooling rate (%): 39.4 -> 63.3Hurt
add Smoothed Attention (softmax temperature T>1) (ablation)DeiT · ImageNet-1k val adversarial patchfooling rateImproved
add Smoothed Attention to Swin TransformerSwin · ImageNet-1k val adversarial patchfooling rateNo clear effect
add Weight Standardization (ablation)ResNet18 · CIFAR10 single patch attackfooling rate (%): 35.9 -> 51.3Hurt
add Group Normalization (ablation)ResNet18 · CIFAR10 single patch attackfooling rate (%): 35.9 -> 52.6Hurt
add Weight Standardization and Group Normalization (ablation)ResNet18 · CIFAR10 single patch attackfooling rate (%): 35.9 -> 71.1Hurt
increase input size from 32 to 48 (ablation)ResNet18 · CIFAR10 single patch attackfooling rate (%): 35.9 -> 42.2No clear effect
align adversarial patch shift to patch grid (16 px vs 1 px) (ablation)DeiT-small · ImageNet-1k val patch transferabilitytransfer fooling rate: 0.27 -> 8.43Improved
use ViT-agnostic (unaligned) adversarial patch instead of aligned (ablation)DeiT-small · ImageNet-1k val adversarial patchfooling rate (%): 61.5 -> 47.9Hurt
replace self-attention with MLP-Mixer token mixingViT · ImageNet-1k adversarial patchfooling rate (%): 31.36Improved

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