Research map / Adversarial robustness and security / Black-box and transfer attacks
Improving Generalizability and Undetectability for Targeted Adversarial Attacks on Multimodal Pre-trained Models
Zhifang Zhang, Jiahan Zhang, Shengjie Zhou, Qi Wei, Shuo He, Feng Liu, Lei Feng
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 |
|---|---|---|---|
| replace single-target alignment with proxy targets plus source-modal proxies | Illusion Attack · Classification (XmediaNet+ImageNet), no defense | Cls ASR (%) avg: 93.18 -> 98.31 | Improved |
| add source-modal proxy loss and target proxies under anomaly detection | Illusion Attack · Classification (XmediaNet+ImageNet), with anomaly detection | Cls ASRD (%) avg: 20.04 -> 84.17 | Improved |
| proxy-based targets | Illusion Attack · Retrieval (XmediaNet+MSCOCO), no defense | R@1 ASR (%) avg: 46.49 -> 77.72 | Improved |
| proxy-based targets plus source proxies | Illusion Attack · Retrieval (XmediaNet+MSCOCO), with anomaly detection | R@1 ASRD (%) avg: 14.84 -> 61.80 | Improved |
| use PTA instead of Illusion Attack for poisoning a retrieval gallery | Illusion Attack · MSCOCO retrieval poisoning, ImageBind | R@1 (%) after injection: 37.42 -> 20.04 (PTA 1% vs Illusion 1%) | Improved |
| use PTA at 0.1% injection vs Illusion Attack at 1% | Illusion Attack · MSCOCO retrieval poisoning, One-PEACE | R@1 (%) after injection: 36.55 vs 37.26 | No clear effect |
| add more proxies via higher injection ratio of PTA AEs (ablation) | PTA · MSCOCO retrieval poisoning, LanguageBind | R@1 (%) after injection: 36.23 -> 15.47 -> 12.26 (0.1%, 0.5%, 1%) | Improved |
| PTA instead of Illusion Attack under query-based black-box attack | Illusion Attack · MSCOCO retrieval, ImageBind, 2x10^4 queries | R@1 ASR (%): 11.56 -> 41.81 | Improved |
| PTA instead of Illusion Attack for transfer black-box attack | Illusion Attack · ImageNet zero-shot classification, ImageBind -> OpenCLIP-ViT | Cls ASR (%): 28.65 -> 39.19 | Improved |
| PTA instead of Illusion Attack for transfer attack OpenCLIP-ViT -> ImageBind | Illusion Attack · ImageNet zero-shot classification | Cls ASR (%): 30.27 -> 30.81 | No clear effect |
| PTA for textual AEs | Illusion Attack · Textual AEs, classification, no defense | Cls ASR (%): 26.01 -> 37.32 | Improved |
| PTA for audio AEs under defense | Illusion Attack · Audio AEs, classification, with defense | Cls ASRD (%): 21.45 -> 91.03 | Improved |
Closest papers by meaning
- Adversarial Attacks against Closed-Source MLLMs via Feature Optimal Alignment (NeurIPS 2025 · similarity 0.86)
- Dynamic Generative Targeted Attacks With Pattern Injection (CVPR 2023 · similarity 0.85)
- Towards Transferable Targeted Adversarial Examples (CVPR 2023 · similarity 0.85)
- Perturbation Towards Easy Samples Improves Targeted Adversarial Transferability (NeurIPS 2023 · similarity 0.85)
- 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 0.84)
- Improved Generation of Adversarial Examples Against Safety-aligned LLMs (NeurIPS 2024 · similarity 0.84)
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.
