Research map / Adversarial robustness and security / Attacks on text and graphs
A Provable Energy-Guided Test-Time Defense Boosting Adversarial Robustness of Large Vision-Language Models
Mujtaba Hussain Mirza, Antonio D'Orazio, Odelia Melamed, Iacopo Masi
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 |
|---|---|---|---|
| apply ET3 test-time energy minimization to robust CLIP zero-shot classifiers | TeCoA/FARE ViT-L/14 · 14-dataset zero-shot, AutoAttack eps 4/255 | robust accuracy: 32.85 -> 40.93 (TeCoA, eps_t 4/255) | Improved |
| apply ET3 to models adversarially trained with the weaker eps_t 2/255 | FARE ViT-L/14 · 14-dataset zero-shot avg | robust accuracy: 20.39 -> 31.75 | Improved |
| apply ET3 to robust CLIP | ViT-L/14 TeCoA/FARE · 14-dataset zero-shot avg | clean accuracy: 67.57 -> 67.42 (FARE eps_t 2/255) | No clear effect |
| add ET3 to the lightweight TTC-style transformation baseline | TeCoA CLIP-ViT-B/32 · 8 fine-grained datasets, PGD-100 | robust accuracy: 13.90 (TTC) -> 16.51 (ET3) | Improved |
| combine ET3 with augmentation ensemble | Ensemble · 8 fine-grained datasets, PGD-100 | robust accuracy: 19.4 -> 22.26 | Improved |
| combine ET3 with R-TPT | R-TPT · 8 fine-grained datasets, PGD-100 | robust accuracy: 21.41 -> 23.88 | Improved |
| apply ET3 to LLaVA vision encoder, standard CLIP | LLaVA 1.5-7B · captioning and VQA avg, eps 4/255 | robust score: 1.0 -> 38.8 | Improved |
| apply ET3 to LLaVA with standard CLIP | LLaVA 1.5-7B · captioning and VQA avg, clean | clean score: 76.2 -> 73.9 | Hurt |
| apply ET3 under defense-aware adaptive attack on ImageNet RN-50 | Salman et al. RN-50 · ImageNet worst-case robustness | robust accuracy: 34.96 -> 37.70 | Improved |
| apply ET3 to RN-50 on clean ImageNet | Salman et al. RN-50 · ImageNet clean | clean accuracy: 64.02 -> 63.12 | Hurt |
| apply prior TTT defense Singh et al. to RN-50 | Salman et al. RN-50 · ImageNet worst-case robustness | robust accuracy: 34.96 -> 34.68 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- R-TPT — R-TPT: Improving Adversarial Robustness of Vision-Language Models through Test-Time Prompt Tuning (CVPR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- R-TPT: Improving Adversarial Robustness of Vision-Language Models through Test-Time Prompt Tuning (CVPR 2025 · similarity 0.89)
- TAPT: Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models (CVPR 2025 · similarity 0.89)
- One Prompt Word is Enough to Boost Adversarial Robustness for Pre-trained Vision-Language Models (CVPR 2024 · similarity 0.87)
- Inducing High Energy-Latency of Large Vision-Language Models with Verbose Images (ICLR 2024 · similarity 0.86)
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.86)
- ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time (ICLR 2025 · similarity 0.86)
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.
