Research map / Adversarial robustness and security / Black-box and transfer attacks
Learning with Bilevel-Minimax Optimization for Efficient and Reliable Transfer Attacks
Yaohua Liu, Yifan Guo, Jiaxin Gao
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 BMAT to 9 base attackers (24 combinations) | PGD/MI/VMI/SI/DI/TI + SGM/Ghost/MBA · ImageNet (ResNet-50 surrogate) | average ASR: average ASR gain of 23.28% | Improved |
| use BMAT instead of RAP at equal backward passes | RAP · ImageNet BP-normalized | average ASR: 5.44 -> 15.03 | Improved |
| give RAP 10x more backward passes (BP=400) | RAP · ImageNet BP-normalized | average ASR: 9.68 vs BMAT 15.03 | Mixed |
| use BMAT instead of BETAK initialization learning | BETAK · ImageNet BP-normalized | average ASR and memory: 12.06 -> 15.03 ASR; 22.69 -> 7.89 GB | Improved |
| add IGA to MI (ablation) | MI · Cityscapes mIoU, FCN surrogate, UPerNet victim | mIoU: 3.57 -> 2.40 | Improved |
| add SWM on top of IGA (ablation) | MI + IGA · Cityscapes mIoU, DLV3-R50 surrogate, Setr victim | mIoU: 44.42 -> 42.61 | Improved |
| add SWM on top of IGA (ablation) | MI + IGA · Cityscapes mIoU, DLV3-R50 surrogate, UPerNet victim | mIoU: 1.56 -> 1.96 | Hurt |
| increase PGD iterations from K=10 to K=40 (ablation) | PGD · ImageNet, IncRes-V2 victim | ASR: 13.64 -> 13.22 | Hurt |
| add BMAT with (T,K~,K)=(3,10,10) against PGD with more iterations (ablation) | PGD · ImageNet attack iterations | average ASR gain: +25.87% at K=20, +41.45% at K=40 | Improved |
| keep SWM in Phase-II (Full BMAT instead of Fast) (ablation) | BMAT Single-Surrogate on DI-MI · ImageNet | average ASR: 46.53 -> 49.25 | Improved |
| use auxiliary Inc-v3 pseudo-surrogate instead of Bayesian single surrogate (ablation) | BMAT Fast on DI-MI · ImageNet | average ASR: 46.53 -> 49.37 | Improved |
| use single-surrogate Bayesian pseudo-surrogate (no extra victim access) (ablation) | BMAT · ImageNet | average ASR: 30.17% and 58.34% average ASR improvement over baselines | Improved |
Closest papers by meaning
- LRS: Enhancing Adversarial Transferability through Lipschitz Regularized Surrogate (AAAI 2024 · similarity 0.90)
- Enhancing Adversarial Transferability with Adversarial Weight Tuning (AAAI 2025 · similarity 0.88)
- Anchoring Adversarial Trajectories to Data Manifolds: A Bilevel Transfer Optimization Framework (NeurIPS 2026 · similarity 0.88)
- Towards Transferable Adversarial Attacks with Centralized Perturbation (AAAI 2024 · similarity 0.87)
- Enhancing Adversarial Transferability with Checkpoints of a Single Model's Training (CVPR 2025 · similarity 0.87)
- Boosting Adversarial Transferability through Augmentation in Hypothesis Space (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.
