Research map / Adversarial robustness and security / Attacks on text and graphs
Adversarial Attacks Already Tell the Answer: Directional Bias-Guided Test-time Defense for Vision-Language Models
Liangsheng Liu, Si Chen, Jiamin Wu, Weiwei Feng, Zhixin Cheng, Xiaotian Yin, Wenfei Yang, Tianzhu Zhang
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 DBD test-time defense | CLIP-ViT-B/32 · 10 fine-grained datasets, PGD-100 eps=4/255 | robust accuracy: 0.0 -> 91.3 (R-TPT 38.1) | Improved |
| add DBD test-time defense | CLIP-ViT-B/16 · 5 ImageNet-OOD, PGD-100 eps=4/255 | robust accuracy vs clean: clean 63.7 vs robust 97.7 | Improved |
| add DBD test-time defense | CLIP-ViT-B/32 · ImageNet-OOD avg | clean accuracy: 52.7 -> 55.0 | Improved |
| add linear feature shift along Defense Direction (ablation) | DBD with 5 transformation types · 15 datasets, three attack settings | robust accuracy: 34.8 -> 92.1 | Improved |
| add linear feature shift along Defense Direction (ablation) | DBD with 5 transformation types · 15 datasets, three attack settings | clean accuracy: 61.5 -> 58.9 | Hurt |
| add DB-score threshold two-stream strategy (ablation) | DBD · 15 datasets, three attack settings | clean accuracy: 58.9 -> 61.2 | Improved |
| add DB-score threshold two-stream strategy (ablation) | DBD · 15 datasets, three attack settings | robust accuracy: 92.1 -> 91.7 | No clear effect |
| use multiple transformations instead of a single one (ablation) | DBD with feature shift · 15 datasets, three attack settings | clean accuracy: 53.0 -> 58.9 (crop-resize-flip only vs all) | Improved |
| set feature shift lambda from 1.0 to 2.5 (ablation) | DBD · 15 datasets, three attack settings | robust accuracy: 34.8 -> 92.1 | Improved |
| set DB-score threshold tau to 0.8 (ablation) | DBD · 15 datasets, three attack settings | overall mean accuracy | Improved |
| use pseudo-labels for attack generation | DBD · 6 fine-grained datasets, CLIP-ViT-B/16 | robust accuracy: 66.8 vs clean 67.5 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- R-TPT — R-TPT: Improving Adversarial Robustness of Vision-Language Models through Test-Time Prompt Tuning (CVPR 2025) · better on 8 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- TAPT: Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models (CVPR 2025 · similarity 0.90)
- R-TPT: Improving Adversarial Robustness of Vision-Language Models through Test-Time Prompt Tuning (CVPR 2025 · similarity 0.89)
- One Prompt Word is Enough to Boost Adversarial Robustness for Pre-trained Vision-Language Models (CVPR 2024 · similarity 0.89)
- CLIP is Strong Enough to Fight Back: Test-time Counterattacks towards Zero-shot Adversarial Robustness of CLIP (CVPR 2025 · similarity 0.89)
- Improving Zero-Shot Adversarial Robustness in Vision-Language Models by Closed-form Alignment of Adversarial Path Simplices (ICML 2025 · similarity 0.88)
- On the Zero-shot Adversarial Robustness of Vision-Language Models: A Truly Zero-shot and Training-free Approach (CVPR 2025 · similarity 0.87)
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.
