Research map / Adversarial robustness and security / Attacks on text and graphs
GLEAM: Enhanced Transferable Adversarial Attacks for Vision-Language Pre-training Models via Global-Local Transformations
Yunqi Liu, Xue Ouyang, Xiaohui Cui
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 NURBS local deformation and global distribution expansion to the attack (ablation) | GLEAM · ALBEF -> CLIP-CNN, Flickr30K TR | attack success rate (%): to 80.7 (TR) and 81.3 (IR) with LFE+GDE | Improved |
| add cross-modal feature alignment using intermediate adversarial images (ablation) | GLEAM (LFE+GDE) · ALBEF -> CLIP-CNN, Flickr30K | attack success rate (%): +6.0 TR, +5.3 IR points (80.7 -> 86.7 TR) | Improved |
| use LFE alone instead of SGA/SA-AET augmentation (ablation) | GLEAM · ALBEF -> CLIP-CNN | attack success rate (%) | Improved |
| replace SA-AET with GLEAM for black-box transfer to MLLMs | GPT-4o · MLLM attack success rate | attack success rate (%): 6% -> 39% | Improved |
| replace SA-AET with GLEAM for cross-model transfer | ALBEF source, CLIP-ViT target · Flickr30K TR | attack success rate (%): 55.58 -> 86.66 | Improved |
| replace SA-AET with GLEAM in white-box setting | ALBEF/TCL source and target · Flickr30K/MSCOCO | attack success rate (%) | No clear effect |
| increase image perturbation budget epsilon_v (ablation) | GLEAM · TR attack success | attack success rate | Improved |
| apply defenses (NRP, JPEG, DiffPure, etc.) against GLEAM | GLEAM · ITR with defenses | TR R@1: ALBEF TR 73.04 under NRP+LT | Mixed |
| replace SA-AET with GLEAM for cross-task transfer to captioning | ITR adversarial examples · MSCOCO captioning | CIDEr: 64.0 -> 56.2 | Improved |
Closest papers by meaning
- Set-level Guidance Attack: Boosting Adversarial Transferability of Vision-Language Pre-training Models (ICCV 2023 · similarity 0.91)
- 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 0.89)
- Chain of Attack: On the Robustness of Vision-Language Models Against Transfer-Based Adversarial Attacks (CVPR 2025 · similarity 0.89)
- SLADE: Shielding against Dual Exploits in Large Vision-Language Models (CVPR 2025 · similarity 0.89)
- Pandora's Box: Towards Building Universal Attackers against Real-World Large Vision-Language Models (NeurIPS 2024 · similarity 0.89)
- Stage-wise Attention-Guided Region Sequencing for Adversarial Attacks on Large Vision-Language Models (NeurIPS 2026 · 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.
