Research map / Adversarial robustness and security / Certified and empirical robustness
Subspace Adversarial Training
Check 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 |
|---|---|---|---|
| Constrain single-step GAT to the extracted subspace | GAT on CIFAR-10 · CIFAR-10, PGD-50, best robust accuracy | robust accuracy: 50.03 -> 51.15 | Helped |
| Constrain single-step GradAlign to the extracted subspace | GradAlign on CIFAR-100 · CIFAR-100, PGD-50, best robust accuracy | robust accuracy: 23.80 -> 25.69 | Helped |
| Constrain PGD-10 AT to the extracted subspace | PGD-10 AT on CIFAR-10 · CIFAR-10, PGD-20, best robust accuracy, l-infinity epsilon=8/255 | robust accuracy: 51.09 -> 52.79 | Helped |
| Constrain PGD-10 AT to the extracted subspace | PGD-10 AT on CIFAR-10 · CIFAR-10, Auto-Attack, best robust accuracy, l-infinity epsilon=8/255 | robust accuracy: 47.29 -> 48.37 | Helped |
| Increase Fast Sub-AT training radius from 8/255 to 12/255 (ablation) | Fast Sub-AT on CIFAR-10 · CIFAR-10, PGD-50, best robust accuracy | robust accuracy: 48.22 -> 50.38 | Helped |
| Increase Fast Sub-AT training radius from 8/255 to 16/255 (ablation) | Fast Sub-AT on CIFAR-10 · CIFAR-10, PGD-50, best robust accuracy | robust accuracy: 48.22 -> 51.46 | Helped |
| Increase Fast Sub-AT training radius from 8/255 to 16/255 (ablation) | Fast Sub-AT in the PGD-10 AT subspace on CIFAR-10 · CIFAR-10, PGD-50, final robust accuracy | robust accuracy: 52.01 -> 54.11 | Helped |
| Increase Fast Sub-AT training radius from 8/255 to 16/255 (ablation) | Fast Sub-AT on Tiny-ImageNet · Tiny-ImageNet, PGD-20, best robust accuracy | robust accuracy: 16.75 -> 18.79 | Helped |
| Vary the subspace dimension from 50 to 100 (ablation) | Fast Sub-AT · CIFAR-10, PGD-20, best robust accuracy | robust accuracy: <1% | No clear effect |
| Replace the full PGD-10 AT training process with Sub-AT optimization | PGD-10 AT on Tiny-ImageNet · Tiny-ImageNet, PGD-20, best robust accuracy | robust accuracy: 19.84 -> 20.52 | Helped |
Closest papers by meaning
- Taxonomy Driven Fast Adversarial Training (AAAI 2024 · similarity 0.89)
- Kernel Learning with Adversarial Features: Numerical Efficiency and Adaptive Regularization (NeurIPS 2025 · similarity 0.87)
- Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off (ICLR 2025 · similarity 0.87)
- Adversarial Initialization with Universal Adversarial Perturbation: A New Approach to Fast Adversarial Training (AAAI 2024 · similarity 0.87)
- Enhancing the Adversarial Robustness via Manifold Projection (AAAI 2025 · similarity 0.87)
- Efficient local linearity regularization to overcome catastrophic overfitting (ICLR 2024 · similarity 0.87)
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.