Research map / Adversarial robustness and security / Black-box and transfer attacks
On Success and Simplicity: A Second Look at Transferable Targeted Attacks
Zhengyu Zhao, Zhuoran Liu, Martha Larson
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 cross-entropy loss with logit loss | TI-MI-DI simple attack · Res50->Dense121, 300 iterations | targeted success rate: 42.6 -> 72.5 | Improved |
| replace Po+Trip loss with logit loss | TI-MI-DI simple attack · Res50->Dense121, 300 iterations | targeted success rate: 54.7 -> 72.5 | Improved |
| increase attack iterations from 20 to 300 (ablation) | Logit attack · Res50->Dense121 | targeted success rate: 29.3 -> 72.5 | Improved |
| increase iterations from 20 to 100 for Po+Trip vs CE (ablation) | Po+Trip · easy ensemble transfer average | targeted success rate: 56.8 -> 84.4 | Hurt |
| replace CE with logit loss (ablation) | simple attack · easy ensemble transfer, 100 iter | targeted success rate: 87.7 -> 88.1 | No clear effect |
| replace CE with logit loss for lowest-ranked target class | simple attack · 1000th ranked target | targeted success rate: 25.1 -> 52.8 | Improved |
| replace CE with logit loss for 2nd-ranked target (ablation) | simple attack · 2nd ranked target | targeted success rate: 89.9 -> 83.8 | Hurt |
| use logit loss attack instead of TTP at eps=16 single-model | Logit · Res50->D121 | targeted success rate: 79.6 -> 75.9 | Hurt |
| use logit loss attack instead of TTP at eps=8 ensemble | Logit · D121-ens | targeted success rate: 63.2 -> 92.6 | Improved |
| use logit loss rather than C&W loss (ablation) | Logit attack · single-model transfer | targeted success rate | Improved |
| vary step size (ablation) | Logit/CE/Po+Trip attacks · single-model transfer | targeted success rate | No clear effect |
| use only DI vs only TI or MI (ablation) | simple attack · targeted transfer | targeted success rate | Improved |
Closest papers by meaning
- Everywhere Attack: Attacking Locally and Globally to Boost Targeted Transferability (AAAI 2025 · similarity 0.89)
- Improving Transferable Targeted Adversarial Attacks with Model Self-Enhancement (CVPR 2024 · similarity 0.88)
- Rethinking Model Ensemble in Transfer-based Adversarial Attacks (ICLR 2024 · similarity 0.88)
- Enhancing Adversarial Transferability with Adversarial Weight Tuning (AAAI 2025 · similarity 0.88)
- Understanding Model Ensemble in Transferable Adversarial Attack (ICML 2025 · similarity 0.87)
- PubDef: Defending Against Transfer Attacks From Public Models (ICLR 2024 · similarity 0.87)
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.
