Research map / Adversarial robustness and security / Attacks on text and graphs
AEMIM: Adversarial Examples Meet Masked Image Modeling
Wenzhao Xiang, Chang Liu, Hang Su, Hongyang Yu
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add adapters (norm layer) for adversarial/clean branches (ablation) | AEMIM ViT-S/16 · ImageNet-1k val (ablation, 300 ep) | top-1 accuracy: 81.2 -> 81.4 | Improved |
| use L2 distance as adversarial loss rather than reconstruction loss (ablation) | AEMIM ViT-S/16 · ImageNet-1k val (ablation, 300 ep) | top-1 accuracy: 81.1 -> 81.4 | Improved |
| use L2 distance rather than KLD as adversarial loss (ablation) | AEMIM ViT-S/16 · ImageNet-1k val (ablation, 300 ep) | top-1 accuracy: 81.3 -> 81.4 | Improved |
| use 2 attack steps rather than 1 (ablation) | AEMIM ViT-S/16 · ImageNet-1k val (ablation, 300 ep) | top-1 accuracy: 81.2 -> 81.4 | Improved |
| use 3 attack steps rather than 2 (ablation) | AEMIM ViT-S/16 · ImageNet-1k val (ablation, 300 ep) | top-1 accuracy: 81.4 -> 81.4 | No clear effect |
| perturbation budget eps=2 rather than eps=4 (ablation) | AEMIM ViT-S/16 · ImageNet-1k val (ablation, 300 ep) | top-1 accuracy: 81.2 -> 81.4 | Improved |
| perturbation budget eps=2 rather than eps=1 (ablation) | AEMIM ViT-S/16 · ImageNet-1k val (ablation, 300 ep) | top-1 accuracy: 81.3 -> 81.4 | Improved |
| add AEMIM auxiliary adversarial reconstruction to SimMIM | SimMIM Swin-B 100 ep · ImageNet-1k val | top-1 accuracy: 83.5 -> 83.7 | Improved |
| add AEMIM to SimMIM on OOD variants | SimMIM Swin-B 100 ep · ImageNet-A | accuracy: 34.9 -> 36.5 | Improved |
| add AEMIM to SimMIM on ImageNet-V2 | SimMIM Swin-B 100 ep · ImageNet-V2 | accuracy: 72.8 -> 72.7 | Hurt |
| add AEMIM to MAE pre-training (800 ep) | MAE ViT-B/16 · ImageNet-1k val | top-1 accuracy: 83.3 -> 83.8 | Improved |
| add AEMIM to MAE under large FGSM budget | MAE ViT-B/16 800 ep · FGSM eps=8 | accuracy: 32.1 vs 33.0 | Hurt |
Closest papers by meaning
- Adversarial Masking for Self-Supervised Learning (ICML 2022 · similarity 0.87)
- Defending Against Adversarial Examples Via Modeling Adversarial Noise (IJCV 2025 · similarity 0.85)
- Learning With Noisy Labels via Self-Supervised Adversarial Noisy Masking (CVPR 2023 · similarity 0.84)
- Architecture-Agnostic Masked Image Modeling -- From ViT back to CNN (ICML 2023 · similarity 0.84)
- Masked Generative Adversarial Networks are Data-Efficient Generation Learners (NeurIPS 2022 · similarity 0.84)
- Adversarial Examples Improve Image Recognition (CVPR 2020 · similarity 0.83)
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.
