Research map / Adversarial robustness and security / Attacks on text and graphs
Exploring Geometry of Blind Spots in Vision models
Sriram Balasubramanian, Gaurang Sriramanan, Vinu Sankar Sadasivan, Soheil Feizi
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 |
|---|---|---|---|
| Traverse by projecting the target direction onto the gradient-orthogonal subspace | LST on vision classifiers · ImageNet | average path confidence: 0.96 ± 0.05 (ResNet-50 Normal); 0.94 ± 0.02 (DeiT-S Normal) | Improved |
| Use adversarial training rather than normal training | ResNet-50 with LST · ImageNet | average triangle confidence: 0.56 ± 0.10 -> 0.83 ± 0.09 | Improved |
| Use adversarial training rather than normal training | DeiT-S with LST · ImageNet | LPIPS distance to target: 0.024 ± 0.017 -> 0.219 ± 0.068 | Hurt |
| Increase LST iterations (ablation) | LST on ResNet-50 (Normal) · ImageNet | RMSE distance to target: 0.141 ± 0.029 (100) -> 0.008 ± 0.001 (400) | Improved |
| Increase LST iterations (ablation) | LST on ResNet-50 (Normal) · ImageNet | average triangle confidence: 0.91 ± 0.111 (100) -> 0.564 ± 0.119 (400) | Hurt |
| Increase parallel gradient step ε (ablation) | LST with m=100 on ResNet-50 (Normal) · ImageNet | average triangle confidence: 0.91 ± 0.111 (0.002) -> 0.92 ± 0.102 (0.012) | Improved |
| Increase parallel gradient step ε (ablation) | LST with m=100 on ResNet-50 (Normal) · ImageNet | RMSE distance to target: 0.141 ± 0.029 (0.002) -> 0.141 ± 0.029 (0.012) | No clear effect |
| Increase orthogonal step η at m=100 (ablation) | LST on ResNet-50 (Normal) · ImageNet | RMSE distance to target: 0.141 ± 0.029 (0.01) -> 0.008 ± 0.012 (0.04) | Improved |
| Increase orthogonal step η at m=100 (ablation) | LST on ResNet-50 (Normal) · ImageNet | average triangle confidence: 0.91 ± 0.111 (0.01) -> 0.622 ± 0.102 (0.04) | Hurt |
| Reduce orthogonal step η to 0.001 (ablation) | LST on ResNet-50 (Normal) · ImageNet | RMSE distance to target: 0.008 ± 0.001 (η=0.01) -> 0.257 ± 0.052 (η=0.001) | Hurt |
Closest papers by meaning
- On the Structural Sensitivity of Deep Convolutional Networks to the Directions of Fourier Basis Functions (CVPR 2019 · similarity 0.87)
- Stereoscopic Universal Perturbations Across Different Architectures and Datasets (CVPR 2022 · similarity 0.86)
- Empirical Study of the Topology and Geometry of Deep Networks (CVPR 2018 · similarity 0.86)
- Inductive Bias of Deep Convolutional Networks through Pooling Geometry (ICLR 2017 · similarity 0.86)
- Understanding Deep Networks via Extremal Perturbations and Smooth Masks (ICCV 2019 · similarity 0.86)
- Contrasting Adversarial Perturbations: The Space of Harmless Perturbations (AAAI 2025 · 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.
