Research map / Adversarial robustness and security / Attacks on text and graphs
Robustifying Zero-Shot Vision Language Models by Subspaces Alignment
Junhao Dong, Piotr Koniusz, Liaoyuan Feng, Yifei Zhang, Hao Zhu, Weiming Liu, Xinghua Qu, Yew-Soon Ong
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 intermediate adversaries, Omega alignment and antonymous prompts (ablation) | AoS · 15 datasets average | PGD robust accuracy (%): 41.20 -> 43.88 | Improved |
| use MaxExp subspace distance instead of Frobenius norm (ablation) | AoS · 15 datasets average | PGD robust accuracy (%): 40.86 -> 43.88 | Improved |
| use MaxExp instead of exact SVD for subspaces (ablation) | AoS · 15 datasets average | PGD robust accuracy (%): 42.35 -> 43.88 | Improved |
| use universal set perturbation instead of original-sample perturbation (ablation) | AoS · 15 datasets average | PGD robust accuracy (%): 41.43 -> 43.88 | Improved |
| use universal perturbation instead of instance-wise perturbation (ablation) | AoS · 15 datasets average | PGD robust accuracy (%): 44.12 -> 43.88 | No clear effect |
| increase augmentations n=q from 5 to 20 (ablation) | AoS · 15 datasets average | PGD robust accuracy (%): 42.95 -> 43.88 | Improved |
| increase augmentations n=q from 20 to 50 (ablation) | AoS · 15 datasets average | PGD robust accuracy (%): 43.88 -> 44.07 | No clear effect |
| use subspace alignment AoS instead of sample-wise TeCoA | AoS · 15 datasets average | PGD robust accuracy (%): 38.91 -> 43.88 | Improved |
| use AoS instead of FARE under larger radius 4/255 | AoS · 15 datasets average | PGD robust accuracy (%): 7.74 -> 9.51 | Improved |
| use AoS with visual prompt tuning instead of FARE (eps=1/255) | AoS · 15 datasets average (VPT) | PGD robust accuracy (%): 32.69 -> 34.38 | Improved |
| use AoS with BLIP instead of FARE for captioning | AoS · Nocaps captioning | robust CIDEr: 58.7 -> 62.1 | Improved |
Closest papers by meaning
- Improving Zero-Shot Adversarial Robustness in Vision-Language Models by Closed-form Alignment of Adversarial Path Simplices (ICML 2025 · similarity 0.94)
- On the Zero-shot Adversarial Robustness of Vision-Language Models: A Truly Zero-shot and Training-free Approach (CVPR 2025 · similarity 0.93)
- One Prompt Word is Enough to Boost Adversarial Robustness for Pre-trained Vision-Language Models (CVPR 2024 · similarity 0.91)
- R-TPT: Improving Adversarial Robustness of Vision-Language Models through Test-Time Prompt Tuning (CVPR 2025 · similarity 0.90)
- Few-Shot Adversarial Prompt Learning on Vision-Language Models (NeurIPS 2024 · similarity 0.90)
- Pre-trained Model Guided Fine-Tuning for Zero-Shot Adversarial Robustness (CVPR 2024 · similarity 0.89)
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.
