Research map / Adversarial robustness and security / Certified and empirical robustness
Can Shape Structure Features Improve Model Robustness Under Diverse Adversarial Settings?
Mingjie Sun, Zichao Li, Chaowei Xiao, Haonan Qiu, Bhavya Kailkhura, Mingyan Liu, Bo Li
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 |
|---|---|---|---|
| replace Canny with Robust Canny (add noise masking by gradient-magnitude thresholding) (ablation) | EdgeNetRob · Fashion MNIST, PGD-40 | robust accuracy: 39.99 -> 76.75 | Improved |
| replace RCF edge detector with Canny (ablation) | EdgeNetRob · Fashion MNIST, PGD-40 | robust accuracy: 0.18 -> 39.99 | Improved |
| replace Canny with Robust Canny (ablation) | EdgeNetRob · Fashion MNIST | clean accuracy: 88.32 -> 87.00 | Hurt |
| add GAN texture refilling (EdgeGANRob vs EdgeNetRob) | EdgeNetRob · Fashion MNIST | clean accuracy: 87.00 -> 87.14 | No clear effect |
| add GAN texture refilling (EdgeGANRob vs EdgeNetRob) | EdgeNetRob · CelebA | clean accuracy: 94.51 -> 95.88 | Improved |
| add GAN texture refilling (EdgeGANRob vs EdgeNetRob) | EdgeNetRob · CIFAR-10 | clean accuracy: 79.21 -> 76.25 | Hurt |
| add GAN texture refilling (EdgeGANRob vs EdgeNetRob) | EdgeNetRob · Tiny ImageNet, PGD-40 | robust accuracy: 19.53 -> 13.55 | Hurt |
| add CNN-based edges to Robust Canny (Combined Edge) (ablation) | EdgeNetRob · CIFAR-10 | clean and PGD-40 accuracy: 67.85 -> 79.21 clean; 36.31 -> 33.08 robust | Mixed |
| use EdgeNetRob instead of PGD adversarial training | EdgeNetRob · Fashion MNIST, PGD-40 | robust accuracy: 72.62 -> 76.75 | Improved |
| use EdgeNetRob instead of PGD adversarial training | EdgeNetRob · CIFAR-10, PGD-40 | robust accuracy: 44.15 -> 33.08 | Hurt |
| use EdgeNetRob instead of a vanilla network | EdgeNetRob · CIFAR-10 backdoor (Pixel, 5% poison) | poison attack success rate: 99.50 -> 9.42 | Improved |
| use EdgeGANRob instead of Spectral Signature | EdgeGANRob · Fashion MNIST backdoor (Pixel, 20% poison) | poison attack success rate: 45.62 -> 0.07 | Improved |
Closest papers by meaning
- Harnessing Edge Information for Improved Robustness in Vision Transformers (AAAI 2024 · similarity 0.90)
- Edges to Shapes to Concepts: Adversarial Augmentation for Robust Vision (CVPR 2023 · similarity 0.88)
- Revisiting Residual Networks for Adversarial Robustness (CVPR 2023 · similarity 0.87)
- Does enhanced shape bias improve neural network robustness to common corruptions? (ICLR 2021 · similarity 0.87)
- RobustBench: a standardized adversarial robustness benchmark (NeurIPS 2021 · similarity 0.86)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.86)
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.
