Research map / Adversarial robustness and security / Black-box and transfer attacks
Boosting Adversarial Transferability via Ensemble Non-Attention
Yipeng Zou, Qin Liu, Jie Wu, Yu Peng, Guo Chen, Hui Zhou, Guanghui Ye
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 non-attention meta-testing step to meta-gradient optimization (ablation) | NAMEA · ImageNet, avg ASR over 9 ViTs and 8 CNNs | average ASR: drop of 9.2% when removed | Improved |
| add attention meta-training step (ablation) | NAMEA · ImageNet, avg ASR over 9 ViTs and 8 CNNs | average ASR: drop of 7.4% when removed | Improved |
| add gradient scaling optimization (GSO) module (ablation) | NAMEA · ImageNet, avg ASR over 9 ViTs and 8 CNNs | average ASR: drop of 2% when removed | Improved |
| fill masked attention areas with random noise instead of 0s or 1s (ablation) | NAMEA NAE module · ImageNet, I-FGSM | average ASR: about +2.7% | Improved |
| set attention threshold eta to 0.6 rather than smaller or larger values (ablation) | NAMEA · ImageNet, I-FGSM and MI-FGSM | average ASR | Improved |
| extract non-attention areas instead of attention areas or random patches for meta-testing (ablation) | NAMEA · ImageNet | average ASR | Improved |
| choose meta-testing model by different-architecture, same-architecture or random strategy (ablation) | NAMEA · ImageNet | average ASR | No clear effect |
| replace NAE module with BSR random input transformation in meta-testing (ablation) | NAMEA · ImageNet | average ASR | Hurt |
| combine BSR input transformation with NAMEA under DI-MI-FGSM | NAMEA · ImageNet, DI-MI-FGSM | average ASR | Hurt |
| use NAMEA instead of SMER as ensemble attack | DI-MI-FGSM · ImageNet, 17 target models | average ASR: 78.5 -> 83.7 (CNNs); 72.1 -> 77.5 (ViTs) | Improved |
| use NAMEA instead of SMER against DiffPure defense | DI-MI-FGSM · ImageNet, DiffPure | attack success rate: 39.9 -> 50.3 | Improved |
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.
- AdaEA — AdaEA++: Enhanced Adaptive Model Ensemble Adversarial Attack for Boosting Adversarial Transferability (IJCV 2026) · better on 2 of 2 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
- An Adaptive Model Ensemble Adversarial Attack for Boosting Adversarial Transferability (ICCV 2023 · similarity 0.92)
- Rethinking Model Ensemble in Transfer-based Adversarial Attacks (ICLR 2024 · similarity 0.90)
- Understanding Model Ensemble in Transferable Adversarial Attack (ICML 2025 · similarity 0.89)
- Ensemble Diversity Facilitates Adversarial Transferability (CVPR 2024 · similarity 0.88)
- Enhancing Adversarial Transferability with Adversarial Weight Tuning (AAAI 2025 · similarity 0.87)
- Enhancing Adversarial Transferability with Checkpoints of a Single Model's Training (CVPR 2025 · 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.
