Research map / Adversarial robustness and security / Black-box and transfer attacks
Prompt-Driven Contrastive Learning for Transferable Adversarial Attacks
Hunmin Yang, Jongoh Jeong, Kuk-Jin Yoon
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 GAMA CLIP loss with prompt-driven contrastive loss using hand-crafted prompt (ablation) | generative attack with VGG-16 surrogate · Cross-domain avg | top-1 accuracy after attack (lower is better): 48.56 -> 46.69 | Improved |
| replace GAMA CLIP loss with L_PDCL (hand-crafted prompt) (ablation) | generative attack with VGG-16 surrogate · Cross-model avg | top-1 accuracy after attack (lower is better): 40.91 -> 40.35 | Improved |
| add learned prompt context (L_context, Phase 1) (ablation) | PDCL-Attack · Cross-domain avg | top-1 accuracy after attack (lower is better): 46.69 -> 43.91 | Improved |
| add learned prompt context (ablation) | PDCL-Attack · Cross-model avg | top-1 accuracy after attack (lower is better): 40.35 -> 38.60 | Improved |
| increase number of learnable context words from 4 to 16 (ablation) | PDCL-Attack · Cross-domain avg | top-1 accuracy after attack (lower is better): 45.44 -> 43.91 | Improved |
| use randomly initialized word vector prompt (domain randomization) instead of heuristic prompts (ablation) | PDCL-Attack · Cross-domain avg | top-1 accuracy after attack (lower is better): 47.81 vs 46.14-47.70 heuristic | No clear effect |
| use PDCL-Attack across test-time perturbation budgets 6 to 10 (ablation) | PDCL-Attack · Cross-domain avg | top-1 accuracy after attack (lower is better): 75.23 -> 72.07 at eps 6; 48.56 -> 43.91 at eps 10 (vs GAMA) | Improved |
| PDCL-Attack vs GAMA on FGVC Aircraft SE-Res101 and on Res-50/Dense-169/MNasNet victims | PDCL-Attack · Cross-domain and cross-model per-victim | top-1 accuracy after attack (lower is better): 47.25 vs 43.41 (Aircraft SE-Res101) | Mixed |
| use CLIP-based training | PDCL-Attack · Training cost | training time: ~33 vs. 12 hours (BIA) | Hurt |
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.
- GAMA — GAMA: Generative Adversarial Multi-Object Scene Attacks (NeurIPS 2022) · better on 5 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
- Towards Highly Transferable Vision-Language Attack via Semantic-Augmented Dynamic Contrastive Interaction (CVPR 2026 · similarity 0.91)
- X-Transfer Attacks: Towards Super Transferable Adversarial Attacks on CLIP (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)
- Chain of Attack: On the Robustness of Vision-Language Models Against Transfer-Based Adversarial Attacks (CVPR 2025 · similarity 0.88)
- AIM: Additional Image Guided Generation of Transferable Adversarial Attacks (AAAI 2025 · similarity 0.88)
- Mutual-Modality Adversarial Attack with Semantic Perturbation (AAAI 2024 · 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.
