Research map / Adversarial robustness and security / Certified and empirical robustness
Scaling Adversarial Training to Large Perturbation Bounds
Sravanti Addepalli, Samyak Jain, Gaurang Sriramanan, R. Venkatesh Babu
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 LPIPS-based Oracle-Invariant attack (ablation) | OA-AT · CIFAR-10 ResNet-18 | clean accuracy: 78.47 -> 80.24 | Improved |
| add LPIPS-based Oracle-Invariant attack (ablation) | OA-AT · CIFAR-10 ResNet-18 | GAMA 16/255 accuracy: 24.05 -> 22.73 | Hurt |
| add convex combination target (alpha below 1) (ablation) | OA-AT · CIFAR-10 ResNet-18 | clean accuracy: 79.29 -> 80.24 | Improved |
| add both LPIPS and convex combination (ablation) | OA-AT · CIFAR-100 ResNet-18 | clean accuracy: 57.77 -> 60.27 | Improved |
| compute LPIPS with weight-averaged model (ablation) | OA-AT · CIFAR-10 ResNet-18 | Square 16/255 accuracy: 30.76 -> 31.16 | Improved |
| use 2x eps perturbation in AWP weight step (ablation) | OA-AT · CIFAR-10 ResNet-18 | clean accuracy: 79.96 -> 80.24 | Improved |
| maximize CE instead of KL in attack generation (ablation) | OA-AT · CIFAR-100 ResNet-18 | GAMA 16/255 accuracy: 7.93 -> 10.47 | Improved |
| use LPIPS instead of discriminator for OI attack (ablation) | OA-AT · CIFAR-10 ResNet-18 | GAMA 8/255 accuracy: 50.75 -> 51.40 | Improved |
| start eps at eps_max/4 instead of increasing from beginning (ablation) | OA-AT · CIFAR-10 ResNet-18 | GAMA 8/255 accuracy: 50.77 -> 51.40 | Improved |
| apply AutoAugment with p=1 (ablation) | OA-AT · CIFAR-10 ResNet-18 | GAMA 8/255 accuracy: 51.40 -> 48.15 | Hurt |
| add AutoAugment (p=0.5) (ablation) | OA-AT · CIFAR-100 ResNet-18 | clean accuracy: 58.08 -> 60.27 | Improved |
| train standard AT directly at eps 16/255 instead of OA-AT | PGD-AT · CIFAR-10 | clean accuracy: 81.12 -> 64.93 | Hurt |
Closest papers by meaning
- Certified Adversarial Robustness with Additive Noise (NeurIPS 2019 · similarity 0.90)
- Towards Stable and Efficient Adversarial Training against $l_1$ Bounded Adversarial Attacks (ICML 2023 · similarity 0.90)
- Adversarial Training With Anti-Adversaries (TPAMI 2024 · similarity 0.90)
- Scaling provable adversarial defenses (NeurIPS 2018 · similarity 0.90)
- Adversarial Training and Robustness for Multiple Perturbations (NeurIPS 2019 · similarity 0.90)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.88)
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.
