Research map / Adversarial robustness and security / Certified and empirical robustness
Randomized Adversarial Training via Taylor Expansion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 first- and second-order Taylor terms to Wu et al. (2020) adversarial training (ablation) | WRN-34-10 adversarial training · CIFAR-10 ℓ∞ | Auto Attack accuracy: 56.2 -> 57.4 | Helped |
| Add first- and second-order Taylor terms to Wu et al. (2020) adversarial training (ablation) | WRN-34-10 adversarial training · CIFAR-100 ℓ∞ | Auto Attack accuracy: 28.6 -> 30.2 | Helped |
| Add first- and second-order Taylor terms to Wu et al. (2020) adversarial training (ablation) | PreAct ResNet-18 adversarial training · CIFAR-10, CIFAR-100, and SVHN ℓ2 | Auto Attack accuracy: 71.6 -> 72.0; 41.8 -> 42.4; 63.8 -> 64.5 | Helped |
| Add first- and second-order Taylor terms to Zhang et al. (2019) training (ablation) | VGG16 adversarial training · CIFAR-10 ℓ∞ | Auto Attack accuracy: 44.3 -> 44.4 | Helped |
| Add first- and second-order Taylor terms to Wu et al. (2020) training (ablation) | MobileNetV2 adversarial training · CIFAR-10 ℓ∞ | Auto Attack accuracy: 47.7 -> 49.4 | Helped |
| Add the second-order Taylor term after the first-order term, η=0.2 (ablation) | RAT on PreAct ResNet-18 · CIFAR-10 ℓ∞ | Auto Attack accuracy: 48.8 -> 50.3 | Helped |
| Use η=0.3 instead of η=0.2 for first-order-only optimization (ablation) | RAT first-order optimization · CIFAR-10 ℓ∞ | Auto Attack accuracy: 48.8 -> 49.7 | Helped |
| Increase η from 0.2 to 0.3 with first- and second-order terms (ablation) | RAT first- and second-order optimization · CIFAR-10 ℓ∞ | Auto Attack accuracy: 50.3 -> 48.9 | Hurt |
| Add the first-order Taylor term to Zhang et al. (2019) training (ablation) | WRN-34-10 adversarial training · CIFAR-100 ℓ∞ | Auto Attack accuracy: 26.9 -> 28.1 | Helped |
| Add the second-order term after first-order optimization to Zhang et al. (2019) training (ablation) | WRN-34-10 adversarial training · CIFAR-100 ℓ∞ | Auto Attack accuracy: 28.1 -> 27.9 | Hurt |
| Add first- and second-order Taylor terms to AT (ablation) | ResNet-18 adversarial training · CIFAR-10 ℓ∞ | Auto Attack accuracy: 47.1 -> 48.7 | Helped |
| Add randomized-weight first- and second-order Taylor terms to TRADES (ablation) | ResNet-18 adversarial training · CIFAR-10 ℓ∞ | RayS accuracy: 56.06 -> 57.03 | Helped |
Closest papers by meaning
- Adversarial Robustness through Random Weight Sampling (NeurIPS 2023 · similarity 0.90)
- Adversarial Robustness via Deformable Convolution with Stochasticity (ICML 2025 · similarity 0.90)
- Mitigating Adversarial Effects Through Randomization (ICLR 2018 · similarity 0.90)
- Efficient Robust Training via Backward Smoothing (AAAI 2022 · similarity 0.90)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.88)
- Treatment of Statistical Estimation Problems in Randomized Smoothing for Adversarial Robustness (NeurIPS 2024 · 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.