Research map / Adversarial robustness and security / Black-box and transfer attacks
CGBA: Curvature-aware Geometric Black-box Attack
Md Farhamdur Reza, Ali Rahmati, Tianfu Wu, Huaiyu Dai
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 |
|---|---|---|---|
| search along a semicircular path in the 2D plane of normal and boundary directions instead of 1-D normal-vector search | decision-based black-box attack · ImageNet ResNet50 non-targeted, 20000 queries | median l2 perturbation: 0.97 (SurFree) -> 0.59 (CGBA) | Improved |
| use CGBA-H (high-curvature variant) instead of CGBA for targeted attacks (ablation) | CGBA · ImageNet ResNet50 targeted, 1000 queries | median l2 perturbation: 78.99 -> 56.01 | Improved |
| use CGBA-H instead of CGBA for non-targeted attacks with small query budget (ablation) | CGBA · ImageNet ResNet50 non-targeted, 1000 queries | median l2 perturbation: 6.03 -> 5.78 | Improved |
| use CGBA-H instead of CGBA for non-targeted attacks with large budget (ablation) | CGBA · ImageNet ResNet50 non-targeted, 20000 queries | median l2 perturbation: 0.59 -> 0.62 | Hurt |
| use CGBA instead of CGBA-H for targeted attack with large budget (ablation) | CGBA-H · ImageNet ResNet50 targeted | median l2 perturbation: CGBA 78.99 at 1000 queries, 5.38 at 20000 vs CGBA-H 56.01, 4.03 | Mixed |
| use K=50 target-class samples for initial boundary point instead of random (K=1) (ablation) | targeted CGBA · ImageNet targeted | initial perturbation: about 85 -> reduction of more than 30, at about 110 extra queries | Improved |
| increase dimension reduction factor f for CGBA in frequency subspace (ablation) | CGBA · ImageNet non-targeted | median l2 perturbation: f=4 best at 10000 queries (0.88); f=8 best at 1000 queries (4.72) | Mixed |
| attack adversarially trained ResNet50 with CGBA methods | CGBA · adversarially-trained ResNet50 | perturbation | Improved |
| CGBA on low-dimensional CIFAR10 targeted attack with large budget (ablation) | CGBA-H · CIFAR10 targeted | median l2 perturbation: WRN40 20000 queries: 0.238 (CGBA-H) -> 0.231 (CGBA) | Mixed |
| CGBA vs SurFree on WRN40 non-targeted at 1000 queries | SurFree · CIFAR10 WRN40 non-targeted, 1000 queries | median l2 perturbation: 0.493 (SurFree) vs 0.498 (CGBA) | Hurt |
Later papers that compared with this method
3 later accepted papers report CGBA as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- SeRI: Gradient-Free Sensitive Region Identification in Decision-Based Black-Box Attacks (ICLR 2026 · better on 7 of 7 shared results)
- Bias in Zeroth-Order Normal Estimation for Decision-Based Attacks (ICML 2026 · better on 3 of 4 shared results)
- TtBA: Two-third Bridge Approach for Decision-Based Adversarial Attack (ICML 2025 · better on 4 of 4 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
- GeoDA: A Geometric Framework for Black-Box Adversarial Attacks (CVPR 2020 · similarity 0.91)
- Curvature-Invariant Adversarial Attacks for 3D Point Clouds (AAAI 2024 · similarity 0.87)
- ADBA: Approximation Decision Boundary Approach for Black-Box Adversarial Attacks (AAAI 2025 · similarity 0.86)
- Explicitly Perceiving and Preserving the Local Geometric Structures for 3D Point Cloud Attack (AAAI 2024 · similarity 0.86)
- GSBA$^K$: $top$-$K$ Geometric Score-based Black-box Attack (ICLR 2025 · similarity 0.85)
- BRUSLEATTACK: A QUERY-EFFICIENT SCORE- BASED BLACK-BOX SPARSE ADVERSARIAL ATTACK (ICLR 2024 · similarity 0.85)
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.
