Research map / Adversarial robustness and security / Attacks on text and graphs
Identifying Robust Neural Pathways: Few-Shot Adversarial Mask Tuning for Vision-Language Models
Wonjeong Choi, Sejong Ryu, Jungmoon Lee, Dong-Jun Han, Jaekyun Moon
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 |
|---|---|---|---|
| use AdvMask mask tuning instead of FAP prompt tuning | CLIP ViT-B/32 · base-to-new, new classes | adversarial accuracy: 21.9 -> 26.1 | Improved |
| use AdvMask instead of FAP | CLIP ViT-B/32 · base-to-new, base classes | adversarial accuracy: 38.0 -> 43.6 | Improved |
| use AdvMask instead of FAP in zero-shot robustness (16-shot) | CLIP ViT-B/32 · TinyImageNet to unseen datasets | adversarial accuracy: 16.8 -> 19.4 | Improved |
| use AdvMask 16-shot vs TGA-ZSR 16-shot | CLIP ViT-B/32 · zero-shot robustness | adversarial accuracy: 13.0 -> 19.4 | Improved |
| apply masks to MHSA only instead of MLP only (ablation) | AdvMask · 16-shot, 5 datasets | adversarial accuracy: 45.95 -> 47.13 | Improved |
| add MLP masks to MHSA masks (ablation) | AdvMask · 16-shot, 5 datasets | clean accuracy: 67.34 -> 66.01 | Hurt |
| use LAFA loss in addition to adversarial CE (ablation) | AdvMask · 16-shot, 5 datasets | adversarial accuracy: 46.4 -> 47.1 | Improved |
| add KL divergence auxiliary loss to adversarial CE (ablation) | AdvMask · 16-shot, 5 datasets | adversarial accuracy: 46.4 -> 43.6 | Hurt |
| add JS divergence auxiliary loss to adversarial CE (ablation) | AdvMask · 16-shot, 5 datasets | adversarial accuracy: 46.4 -> 46.5 | No clear effect |
| add adaptive weighting to LAFA (ablation) | AdvMask · 16-shot, 5 datasets | adversarial accuracy: 46.8 -> 47.1 | Improved |
| add adaptive weighting to LAFA in 1-shot (ablation) | AdvMask · 1-shot, 5 datasets | adversarial accuracy: 17.8 -> 18.4 | Improved |
| combine AdvMask with CoOp prompts, then further train prompts | AdvMask + CoOp · 16-shot, 5 datasets | adversarial accuracy: 37.7 -> 44.7 | 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 0.91)
- Robust CLIP: Unsupervised Adversarial Fine-Tuning of Vision Embeddings for Robust Large Vision-Language Models (ICML 2024 · similarity 0.90)
- One Prompt Word is Enough to Boost Adversarial Robustness for Pre-trained Vision-Language Models (CVPR 2024 · similarity 0.89)
- R-TPT: Improving Adversarial Robustness of Vision-Language Models through Test-Time Prompt Tuning (CVPR 2025 · similarity 0.89)
- On the Zero-shot Adversarial Robustness of Vision-Language Models: A Truly Zero-shot and Training-free Approach (CVPR 2025 · similarity 0.88)
- LORE: Lagrangian-Optimized Robust Embeddings for Visual Encoders (NeurIPS 2025 · 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.
