Research map / Adversarial robustness and security / Attacks on text and graphs
An Inverse-Adversarial and Difficulty-Adaptive Robust Vision-Language Model
Ruobing Xu, Junhao Dong, Xiaohua Xie
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 visual prompt tuning and text adapter (ablation) | CLIP head-only fine-tuning · Caltech101 novel | robust accuracy: 42.78 -> 52.52 | Improved |
| add visual prompt tuning and text adapter (ablation) | CLIP head-only fine-tuning · EuroSAT novel | robust accuracy: 9.36 -> 13.76 | Improved |
| use text adapter instead of text prompt tuning (ablation) | CLIP · Caltech101 novel | robust accuracy: 44.27 -> 48.92 | Improved |
| use visual prompt tuning instead of visual adapter (ablation) | CLIP · Caltech101 novel | robust accuracy: 43.30 -> 45.24 | Improved |
| combine text prompt tuning with visual adapter (ablation) | CLIP · EuroSAT novel | robust accuracy: 10.28 -> 10.32 | No clear effect |
| add inverse-adversarial learning, expert adaptation and difficulty-adaptive fusion (ablation) | CLIP · Caltech101 novel | robust accuracy: 39.52 -> 52.52 | Improved |
| add inverse-adversarial learning, expert adaptation and difficulty-adaptive fusion (ablation) | CLIP · EuroSAT novel | robust accuracy: 8.25 -> 13.76 | Improved |
| use IADA-RVLM instead of FAP (ablation) | robust few-shot CLIP · DTD novel | harmonic mean: 23.06 -> 26.02 | Improved |
| use IADA-RVLM instead of FAP | robust few-shot CLIP · EuroSAT novel | harmonic mean: 16.94 -> 19.02 | Improved |
| use IADA-RVLM instead of AdvMaPLe under AutoAttack | robust few-shot CLIP · OxfordPets novel (AutoAttack) | robust accuracy: 26.39 -> 26.02 | Hurt |
| increase attack budget from eps=2/255 to 4/255 | IADA-RVLM · Caltech101 novel (PGD) | robust accuracy: 52.52 -> 51.01 | Hurt |
| increase scaling factor beta (ablation) | IADA-RVLM · low vs high difficulty datasets | natural and robust accuracy | Mixed |
Closest papers by meaning
- Hierarchically Robust Zero-shot Vision-language Models (CVPR 2026 · similarity 0.92)
- Robustifying Vision-Language Models via Test-Time Prompt Adaptation (ICML 2026 · similarity 0.92)
- Improving Zero-Shot Adversarial Robustness in Vision-Language Models by Closed-form Alignment of Adversarial Path Simplices (ICML 2025 · similarity 0.91)
- On the Zero-shot Adversarial Robustness of Vision-Language Models: A Truly Zero-shot and Training-free Approach (CVPR 2025 · similarity 0.90)
- One Prompt Word is Enough to Boost Adversarial Robustness for Pre-trained Vision-Language Models (CVPR 2024 · similarity 0.90)
- Few-Shot Adversarial Prompt Learning on Vision-Language Models (NeurIPS 2024 · similarity 0.90)
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.
