Research map / Adversarial robustness and security / Attacks on text and graphs
Complementary Text-Guided Attention for Zero-Shot Adversarial Robustness
Lu Yu, Haiyang Zhang, Chunqiang Xu
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 Global Attention Constraint Module (L_GACM) to L_CE + L_LARM (ablation) | Comp-TGA · 16 datasets avg, PGD eps 1/255 | A_Robust: 30.69 -> 44.46 | Improved |
| add L_GACM to L_CE + L_LARM, clean accuracy (ablation) | Comp-TGA · 16 datasets avg, clean | A_Clean: 46.68 -> 55.44 | Improved |
| add Local Attention Refinement (L_LARM) to adversarial fine-tuning with L_CE (ablation) | Comp-TGA · 16 datasets avg, PGD eps 1/255 | A_Robust: 29.45 -> 30.69 | Improved |
| use text-guided attention instead of Grad-CAM vision-based attention (ablation) | TGA-ZSR · 16 datasets avg, PGD eps 1/255 | A_Robust: 29.47 -> 42.09 | Improved |
| add non-class prompt attention to class-prompt attention (ablation) | Comp-TGA · 16 datasets avg, PGD eps 1/255 | A_Robust: 41.45 -> 44.46 | Improved |
| use non-class prompt instead of other prompt (background of photo) as second prompt (ablation) | Comp-TGA · 16 datasets avg, PGD eps 1/255 | A_Robust: 43.24 -> 44.46 | Improved |
| add other prompt attention to class prompt only (ablation) | Comp-TGA · 16 datasets avg, PGD eps 1/255 | A_Robust: 41.45 -> 43.24 | Improved |
| add complementary non-class attention on top of TGA-ZSR, robustness | TGA-ZSR · 16 datasets avg, PGD eps 1/255 | A_Robust: 42.09 -> 44.46 | Improved |
| add complementary non-class attention on top of TGA-ZSR, clean accuracy | TGA-ZSR · 16 datasets avg, clean | A_Clean: 56.44 -> 55.44 | Hurt |
| use Comp-TGA instead of PMG-AFT under PGD | Comp-TGA · 16 datasets avg, PGD eps 1/255 | A_Robust: 32.51 -> 44.46 | Improved |
| use Comp-TGA instead of PMG-AFT under adaptive attack | Comp-TGA · 16 datasets avg, adaptive attack | A_Robust: 29.82 -> 42.40 | Improved |
| use Comp-TGA instead of test-time defense TTC under AutoAttack | Comp-TGA · 16 datasets avg, AutoAttack | A_Robust: 6.97 -> 31.41 | Improved |
Closest papers by meaning
- Text-Guided Attention is All You Need for Zero-Shot Robustness in Vision-Language Models (NeurIPS 2024 · similarity 0.98)
- Few-Shot Adversarial Prompt Learning on Vision-Language Models (NeurIPS 2024 · similarity 0.90)
- On the Zero-shot Adversarial Robustness of Vision-Language Models: A Truly Zero-shot and Training-free Approach (CVPR 2025 · similarity 0.89)
- Enhancing CLIP Robustness via Cross-Modality Alignment (NeurIPS 2025 · similarity 0.88)
- Pre-trained Model Guided Fine-Tuning for Zero-Shot Adversarial Robustness (CVPR 2024 · similarity 0.88)
- One Prompt Word is Enough to Boost Adversarial Robustness for Pre-trained Vision-Language Models (CVPR 2024 · similarity 0.88)
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.
