Research map / Adversarial robustness and security / Certified and empirical robustness
Squeeze Training for Adversarial Robustness
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 |
|---|---|---|---|
| Jointly regularize adversarial and collaborative examples within each perturbation neighborhood | ResNet-18 adversarial training · CIFAR-10 | AA accuracy: 49.37% -> 50.50% | Helped |
| Jointly regularize adversarial and collaborative examples within each perturbation neighborhood | ResNet-18 adversarial training · CIFAR-100 | AA accuracy: 25.13% -> 25.61% | Helped |
| Jointly regularize adversarial and collaborative examples within each perturbation neighborhood | ResNet-18 adversarial training · SVHN | AA accuracy: 49.51% -> 50.54% | Helped |
| Combine ST with AWP | ResNet-18 adversarial training · CIFAR-10 | AA accuracy: 50.31% -> 51.23% | Helped |
| Combine ST with AWP | ResNet-18 adversarial training · CIFAR-100 | AA accuracy: 26.06% -> 26.86% | Helped |
| Combine ST with AWP | ResNet-18 adversarial training · SVHN | AA accuracy: 52.54% -> 53.37% | Helped |
| Combine ST with RST and unlabeled data | WRN-28-10 · CIFAR-10 | AA accuracy: 60.36% -> 60.75% | Helped |
| Use symmetric KL as the ST discrepancy function (ablation) | ST · CIFAR-10 | AA accuracy | Helped |
| Use JS divergence or squared l2 distance instead of symmetric KL (ablation) | ST · CIFAR-10 | clean and robust accuracy trade-off | Hurt |
| Reduce the number of inner optimization steps by half (ablation) | ST · CIFAR-10 | AA accuracy: 50.50% -> 49.52% | Hurt |
| Reduce the number of inner optimization steps by half (ablation) | ST · CIFAR-10 | clean accuracy: 83.10% -> 84.37% | Helped |
| Replace ST's paired worst-case discrepancy with a two-example mean regularization in TRADES (ablation) | TRADES† · SVHN | AA accuracy: 48.81% vs 49.51% | Hurt |
Closest papers by meaning
- Improving Adversarial Robustness Requires Revisiting Misclassified Examples (ICLR 2020 · similarity 0.92)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.92)
- On Adversarial Training without Perturbing all Examples (ICLR 2024 · similarity 0.91)
- Over-parameterized Adversarial Training: An Analysis Overcoming the Curse of Dimensionality (NeurIPS 2020 · similarity 0.89)
- Convergence of Adversarial Training in Overparametrized Neural Networks (NeurIPS 2019 · similarity 0.89)
- A Closer Look at Curriculum Adversarial Training: From an Online Perspective (AAAI 2024 · similarity 0.89)
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.