Research map / Adversarial robustness and security / Certified and empirical robustness
Improving Viewpoint Robustness for Visual Recognition via Adversarial Training
Shouwei Ruan, Yinpeng Dong, Hang Su, Ning Chen, Haibin Duan, Xingxing Wei
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 |
|---|---|---|---|
| use GMVFool instead of ViewFool as VIAT inner maximization (ablation) | VIAT ResNet-50 · IM3D (ViewFool adversarial viewpoints) | accuracy: 55.12% -> 59.84% | Improved |
| use GMVFool instead of ViewFool as VIAT inner maximization (ablation) | VIAT ViT-B/16 · IM3D (ViewFool adversarial viewpoints) | accuracy: 79.53% -> 82.81% | Improved |
| use GMVFool instead of ViewFool as VIAT inner maximization (ablation) | VIAT ResNet-50 · IM3D (GMVFool adversarial viewpoints) | accuracy: 58.75% -> 59.61% | Improved |
| train with VIAT instead of standard training | ResNet-50 · IM3D (GMVFool adversarial viewpoints) | accuracy: 8.98% -> 59.61% | Improved |
| train with VIAT instead of standard training | ViT-B/16 · IM3D (ViewFool adversarial viewpoints) | accuracy: 25.70% -> 82.81% | Improved |
| replace random viewpoint augmentation with VIAT | ResNet-50 · IM3D (ViewFool adversarial viewpoints) | accuracy: 34.80% -> 59.84% | Improved |
| add VIAT training; clean ImageNet subset accuracy | ResNet-50 · ImageNet subset val | accuracy: 85.60% -> 85.70% | No clear effect |
| increase the number of Gaussian components K and sharing probability pi to excessive values (ablation) | VIAT · IM3D (GMVFool) | accuracy | Mixed |
| add VIAT training to ResNet-50 on ImageNet-V | ResNet-50 · ImageNet-V | accuracy: 28.15% -> 38.96% | Improved |
| add VIAT training to ResNet-50 on MIRO | ResNet-50 · MIRO | accuracy: 57.41% -> 65.86% | Improved |
| add VIAT training to ResNet-50 on ImageNet-A | ResNet-50 · ImageNet-A | accuracy: 19.03% -> 20.32% | Improved |
| use Gaussian mixture K=5 instead of unimodal ViewFool in attack | GMVFool · ResNet-50 attack | entropy: -10.14 -> -0.69 | Improved |
Closest papers by meaning
- Towards Viewpoint-Invariant Visual Recognition via Adversarial Training (ICCV 2023 · similarity 0.98)
- Vision Transformers Beat WideResNets on Small Scale Datasets Adversarial Robustness (AAAI 2025 · similarity 0.84)
- Gain from Neighbors: Boosting Model Robustness in the Wild via Adversarial Perturbations Toward Neighboring Classes (CVPR 2025 · similarity 0.84)
- Revisiting Adversarial Training at Scale (CVPR 2024 · similarity 0.84)
- View Transformation Robustness for Multi-View 3D Object Reconstruction with Reconstruction Error-Guided View Selection (AAAI 2025 · similarity 0.83)
- Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off (ICLR 2025 · similarity 0.83)
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.
