Research map / Vision-language and multimodal / CLIP and zero-shot
TTP: Test-Time Padding for Adversarial Detection and Robust Adaptation on Vision-Language Models
Zhiwei Li, Yitian Pang, Weining Wang, Zhenan Sun, Qi Li
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 test-time padding defense (TTP) over R-TPT | CLIP ViT-B/32 · 8 fine-grained datasets, PGD eps=4.0 | robust accuracy (%): 35.3 -> 39.7 | Improved |
| TTP vs R-TPT on larger backbone | CLIP ViT-L/14 · 8 fine-grained datasets, PGD | robust accuracy (%): 49.6 -> 51.6 | Improved |
| TTP applied under CW, DeepFool and FGSM attacks | CLIP ViT-B/32 · Flowers and DTD | robust accuracy (%): 51.8 -> 54.1 (Flowers avg); 34.2 -> 38.7 (DTD avg) | Improved |
| detect-then-adapt keeps clean inputs unchanged | TTP · avg clean accuracy ViT-B/32 | clean accuracy (%): 57.4 -> 57.1 | No clear effect |
| add TPT on inputs detected as clean | TTP · avg clean accuracy ViT-B/32 | clean accuracy (%): 57.1 -> 57.9 | Improved |
| use all-black or all-white padding instead of random padding for detection (ablation) | TTP detector · detection accuracy ViT-B/32 | detection accuracy (%): 95.8 -> 98.5 (0), 98.7 (255) | Improved |
| add fixed spatial padding to CLIP (ablation) | TTP · 8 fine-grained datasets, PGD | robust accuracy (%): 0.0 -> 37.5 (B/32); 0.0 -> 38.0 (B/16) | Improved |
| add entropy minimization training of padding (ablation) | TTP · 8 fine-grained datasets, PGD | robust accuracy (%): 38.3 -> 39.7 (B/32); 39.5 -> 42.9 (B/16) | Improved |
| add similarity-aware ensemble (ablation) | TTP · 8 fine-grained datasets, PGD | robust accuracy (%): 39.0 -> 39.7 (B/32); 40.8 -> 42.9 (B/16) | Improved |
| increase padding size from 8 to 64 then to 256 (ablation) | TTP · DTD robust accuracy ViT-B/32 | robust accuracy (%): 19.6 -> 33.9 (size 64) -> 21.9 (size 256) | Mixed |
| use TTC detect-then-defend instead of TTP detection under strong attack | TTC · 8 fine-grained datasets, PGD eps=4.0 ViT-B/32 | robust accuracy (%): 39.7 -> 6.8 | Hurt |
Closest papers by meaning
- TAPT: Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models (CVPR 2025 · similarity 0.89)
- R-TPT: Improving Adversarial Robustness of Vision-Language Models through Test-Time Prompt Tuning (CVPR 2025 · similarity 0.88)
- The Illusion of Progress? A Critical Look at Test-Time Adaptation for Vision-Language Models (NeurIPS 2025 · similarity 0.88)
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.88)
- One Prompt Word is Enough to Boost Adversarial Robustness for Pre-trained Vision-Language Models (CVPR 2024 · similarity 0.87)
- Noisy Test-Time Adaptation in Vision-Language Models (ICLR 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.
