Research map / Adversarial robustness and security / Black-box and transfer attacks
PA-Attack: Guiding Gray-Box Attacks on LVLM Vision Encoders with Prototypes and Attention
Hefei Mei, Zirui Wang, Chang Xu, Jianyuan Guo, Minjing Dong
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 |
|---|---|---|---|
| PA-Attack vs VEAttack at eps=2/255 | gray-box LVLM attack · LLaVA-1.5-7B average SRR | average SRR %: 65.2 -> 77.1 | Improved |
| PA-Attack vs VEAttack on OpenFlamingo-9B at eps=4/255 | gray-box LVLM attack · OpenFlamingo-9B average SRR | average SRR %: 61.5 -> 66.4 | Improved |
| add prototype-anchored guidance (PG) (ablation) | vision encoder attack (100 steps) · LLaVA-1.5-7B TextVQA | SRR %: 72.8 -> 76.3 | Improved |
| add attention enhancement (AE) alone (ablation) | vision encoder attack (100 steps) · LLaVA-1.5-7B, COCO/Flickr30k/TextVQA/VQAv2 | SRR %: VQAv2 48.4 -> 50.9; TextVQA 72.8 -> 72.3 | Mixed |
| add AE on top of PG (ablation) | PA-Attack (100 steps) · LLaVA-1.5-7B VQAv2 | SRR %: 52.2 -> 54.2 | Improved |
| add two-stage attention refinement (TS) (ablation) | PA-Attack (PG+AE, 150 steps) · LLaVA-1.5-7B TextVQA | SRR %: 81.1 -> 86.3 | Improved |
| use nearest guidance sample instead of farthest prototype (ablation) | PA-Attack · LLaVA-1.5-7B guidance ablation | average SRR %: 82.9 -> 61.4 | Hurt |
| use farthest prototype instead of farthest sample (ablation) | PA-Attack · LLaVA-1.5-7B guidance ablation | average SRR %: 80.6 -> 82.9 | Improved |
| increase loss weight lambda from 1.0 to 2.0 (ablation) | PA-Attack · LLaVA-1.5-7B | average SRR %: 83.7 -> 82.5 | Hurt |
| decrease lambda from 1.0 to 0.5 (ablation) | PA-Attack · LLaVA-1.5-7B | average SRR %: 83.7 -> 81.7 | Hurt |
| increase PCA dimension from 1024 to 2048 (ablation) | PA-Attack prototypes · LLaVA-1.5-7B | average SRR % (Flickr30k, TextVQA, VQAv2): 74.4 -> 70.7 | Hurt |
| change number of prototypes K from 20 to 10 or 30 (ablation) | PA-Attack prototypes · LLaVA-1.5-7B | average SRR %: 74.4 vs 74.3 (K=10) vs 73.6 (K=30) | 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.
- VEAttack — VEAttack: Downstream-agnostic Vision Encoder Attack against Large Vision Language Models (ICLR 2026) · better on 6 of 6 shared results
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
- Pandora's Box: Towards Building Universal Attackers against Real-World Large Vision-Language Models (NeurIPS 2024 · similarity 0.90)
- A Frustratingly Simple Yet Highly Effective Attack Baseline: Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5/4o/o1 (NeurIPS 2025 · similarity 0.90)
- SLADE: Shielding against Dual Exploits in Large Vision-Language Models (CVPR 2025 · similarity 0.89)
- Stage-wise Attention-Guided Region Sequencing for Adversarial Attacks on Large Vision-Language Models (NeurIPS 2026 · similarity 0.89)
- Chain of Attack: On the Robustness of Vision-Language Models Against Transfer-Based Adversarial Attacks (CVPR 2025 · similarity 0.89)
- Black-Box Membership Inference Attack for LVLMs via Prior Knowledge-Calibrated Memory Probing (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.
