Research map / Adversarial robustness and security / Attacks on text and graphs
Enhance the Visual Representation via Discrete Adversarial Training
Xiaofeng Mao, YueFeng Chen, Ranjie Duan, Yao Zhu, Gege Qi, Shaokai Ye, Xiaodan Li, Rong Zhang, Hui Xue'
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 DAT (discrete adversarial training) | ResNet50 · ImageNet FGSM | accuracy (%): 12.19 -> 30.66 | Improved |
| add DAT | AugReg-ViT · ImageNet val | top-1 accuracy (%): 79.91 -> 81.46 | Improved |
| add DAT | AugReg-ViT · ImageNet-A | top-1 accuracy (%): 19.03 -> 30.15 | Improved |
| add DAT | MAE-H finetuning · ImageNet val | top-1 accuracy (%): 86.90 -> 87.02 | Improved |
| add DAT in pre-training | MoCov3 · ImageNet linear eval | top-1 accuracy (%): 68.63 -> 69.60 | Improved |
| add DAT | SimSiam · ADE20K mIoU / Flowers | mIoU: 0.2212 -> 0.2203 | Mixed |
| add DAT instead of Det-AdvProp | EfficientDet-Lite0 · COCO-C | mAP: 16.11 -> 17.89 | Improved |
| remove adversarial component (alpha=0, VQGAN augmentation only) (ablation) | DAT ResNet50 · ImageNet FGSM | accuracy (%): 30.66 -> 23.94 | Hurt |
| increase perturbation magnitude alpha from 0.1 to 0.4 (ablation) | DAT ResNet50 · ImageNet val | top-1 accuracy (%): 76.52 -> 74.28 | Hurt |
| replace adversarial discrete perturbation with random discrete perturbation (ablation) | DAT ResNet50 · ImageNet FGSM | accuracy (%): 30.66 -> 29.01 | Hurt |
| keep discretizer Q at inference (ablation) | DAT ResNet50 · ImageNet | clean top-1 accuracy (%) and FGSM: 76.52 -> 74.8 clean; 30.66 -> 55.4 FGSM | Mixed |
| bound delta with l_inf eps=1/255 (ablation) | DAT ResNet50 · ImageNet val | top-1 accuracy (%): 76.52 -> 76.10 | Hurt |
Closest papers by meaning
- Large-Scale Adversarial Training for Vision-and-Language Representation Learning (NeurIPS 2020 · similarity 0.89)
- Revisiting Adversarial Training at Scale (CVPR 2024 · similarity 0.88)
- DAT: Improving Adversarial Robustness via Generative Amplitude Mix-up in Frequency Domain (NeurIPS 2024 · similarity 0.87)
- Few-Shot Adversarial Prompt Learning on Vision-Language Models (NeurIPS 2024 · similarity 0.86)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.86)
- LORE: Lagrangian-Optimized Robust Embeddings for Visual Encoders (NeurIPS 2025 · similarity 0.86)
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.
