Research map / Adversarial robustness and security / Black-box and transfer attacks
Boosting Adversarial Transferability by Achieving Flat Local Maxima
Zhijin Ge, Hongying Liu, Xiaosen Wang, Fanhua Shang, Yuanyuan Liu
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 gradient-norm regularizer to the attack loss (ablation) | I-FGSM · Inc-v3 to 7 black-box models | avg attack success rate: +5.3% | Improved |
| add gradient-norm regularizer to the attack loss (ablation) | MI-FGSM · Inc-v3 to 7 black-box models | avg attack success rate: +7.2% | Improved |
| approximate Hessian-vector product with finite differences instead of exact Hessian (ablation) | I-FGSM with gradient penalty · Inc-v3 -> Inc-v4 | attack success rate / time: 39.2 -> 37.9 success; 469.54 s -> 96.42 s | Mixed |
| use smaller finite-difference step size (1e-4) (ablation) | PGN · Inc-v3 -> Inc-v4 | attack success rate: 91.6 -> 86.4 | Hurt |
| increase number of sampled examples N from 1 (ablation) | PGN · Inc-v3 to black-box models | attack success rate | Improved |
| raise balance coefficient delta up to 0.5 (ablation) | PGN · Inc-v3 to black-box models | attack success rate | Improved |
| enlarge zeta ball beyond 4.5 x epsilon (ablation) | PGN · Inc-v3 to seven black-box models | attack success rate | Hurt |
| combine PGN with DIM | DIM · Inc-v3 source, 7 models avg | attack success rate: 54.56 -> 84.06 | Improved |
| combine PGN with SSA | SSA · Inc-v3 source, 7 models avg | attack success rate: 71.39 -> 78.03 | Improved |
| combine PGN with RAP | RAP · Inc-v3 source, 7 models avg | attack success rate: 68.80 -> 73.91 | Improved |
| combine PGN with SGM | SGM · Res-18 source, 7 models avg | attack success rate: 54.77 -> 68.64 | Improved |
| use PGN instead of RAP on transformer targets | Inc-v3 source · ViT/PiT/Visformer/Swin avg | attack success rate: 44.45 -> 56.98 | 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.
- Admix — Admix: Enhancing the Transferability of Adversarial Attacks (ICCV 2021) · better on 1 of 1 shared result
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
- Transferability Bound Theory: Exploring Relationship between Adversarial Transferability and Flatness (NeurIPS 2024 · similarity 0.91)
- Enhancing Adversarial Transferability with Adversarial Weight Tuning (AAAI 2025 · similarity 0.91)
- Boosting Adversarial Transferability through Augmentation in Hypothesis Space (CVPR 2025 · similarity 0.90)
- Towards Transferable Adversarial Attacks with Centralized Perturbation (AAAI 2024 · similarity 0.89)
- Rethinking Model Ensemble in Transfer-based Adversarial Attacks (ICLR 2024 · similarity 0.89)
- LRS: Enhancing Adversarial Transferability through Lipschitz Regularized Surrogate (AAAI 2024 · similarity 0.89)
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.
