Research map / Adversarial robustness and security / Attacks on text and graphs
Self-Calibrated Consistency can Fight Back for Adversarial Robustness in Vision-Language Models
Jiaxiang Liu, Jiawei Du, Xiao Liu, Shangyang Li, Songchen Ma, Changshuo Wang, Prayag Tiwari, Mingkun Xu
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 semantic consistency (ablation) | SCC without modules · 16-dataset avg, PGD-10 eps=1/255 (ablation) | robust accuracy (%): 39.18 -> 39.76 | Improved |
| add spatial consistency (ablation) | SCC without modules · 16-dataset avg, PGD-10 eps=1/255 (ablation) | robust accuracy (%): 39.18 -> 48.01 | Improved |
| add spatial consistency on top of semantic consistency (ablation) | SCC semantic only · 16-dataset avg, PGD-10 eps=1/255 (ablation) | robust accuracy (%): 39.76 -> 51.68 | Improved |
| add semantic consistency on top of spatial consistency (ablation) | SCC spatial only · 16-dataset avg, PGD-10 eps=1/255 (ablation) | robust accuracy (%): 48.01 -> 51.68 | Improved |
| replace TTC with SCC | TTC · 16-dataset avg, PGD-10 eps=1/255 | robust accuracy (%): 39.17 -> 51.68 | Improved |
| replace TTC with SCC at larger attack budget | TTC · 16-dataset avg, PGD-10 eps=4/255 | robust accuracy (%): 20.63 -> 27.88 | Improved |
| replace TTC with SCC under CW attack | TTC · 16-dataset avg, CW-10 eps=1/255 | robust accuracy (%): 38.17 -> 49.42 | Improved |
| replace TTC with SCC on BioMedCLIP | TTC · 6 medical datasets avg, eps=1/255 | robust accuracy (%): 10.62 -> 34.00 | Improved |
| replace TTC with SCC | TTC · DTD, adversarial attack | robust accuracy (%): 27.4 -> 34.6 | Improved |
| apply SCC instead of vanilla CLIP | CLIP · 16-dataset avg, PGD-10 eps=1/255 | clean accuracy (%): 61.51 -> 60.21 | Hurt |
| use clean-image predictions for pseudo-labels instead of warm-up pseudo-labels | SCC · 16-dataset avg, PGD-10 eps=1/255 | robust accuracy (%): 51.68 -> 53.62 | Improved |
| increase number of views L from 1 to 2-4 (ablation) | SCC · 16-dataset avg (sensitivity) | robust accuracy (%) | Improved |
Closest papers by meaning
- Enhancing CLIP Robustness via Cross-Modality Alignment (NeurIPS 2025 · similarity 0.90)
- ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time (ICLR 2025 · 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 0.89)
- One Prompt Word is Enough to Boost Adversarial Robustness for Pre-trained Vision-Language Models (CVPR 2024 · similarity 0.89)
- TAPT: Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models (CVPR 2025 · 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.
