Research map / Adversarial robustness and security / Certified and empirical robustness
Overfitting in adversarially robust deep learning
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 |
|---|---|---|---|
| Stop PGD adversarial training at its best checkpoint rather than at convergence | PGD adversarial training · CIFAR-10, ℓ∞ 8/255 | robust test error (%): 51.4{} \pm 0.41{} -> 43.2{} | Helped |
| Stop PGD adversarial training at its best checkpoint rather than at convergence | PGD adversarial training · ImageNet, ℓ2 76/255 | robust test error (%): 94.8{} \pm 1.16{} -> 63.0{} | Helped |
| Stop vanilla PGD training at its best checkpoint | CIFAR-10 Wide ResNet-28-10 · CIFAR-10, ℓ∞ 8/255 | robust test error (%): Vanilla PGD achieves 42.3%; reported TRADES result is 43.4% | Helped |
| Use a 1,000-example hold-out set to choose the stopping checkpoint | PGD adversarial training · CIFAR-10 | robust test error (%): 46.9{} vs 46.7{} | No clear effect |
| Use smoother learning-rate schedules instead of piecewise decay (ablation) | PGD adversarial training · CIFAR-10 | best robust test error | Hurt |
| Add ℓ1 regularization and train to convergence (ablation) | PGD adversarial training · CIFAR-10, ℓ∞ 8/255 | robust test error (%): 46.9 -> 53.0{}\pm 0.39{} | Hurt |
| Add ℓ2 regularization and train to convergence (ablation) | PGD adversarial training · CIFAR-10, ℓ∞ 8/255 | robust test error (%): 46.9 -> 55.2{}\pm 0.4 | Hurt |
| Add cutout augmentation and train to convergence (ablation) | PGD adversarial training · CIFAR-10, ℓ∞ 8/255 | robust test error (%): 46.9 -> 48.8{}\pm 0.79{} | Hurt |
| Add mixup augmentation and train to convergence (ablation) | PGD adversarial training · CIFAR-10, ℓ∞ 8/255 | robust test error (%): 46.9 -> 49.1{}\pm 1.32{} | Hurt |
| Add semi-supervised augmentation and train to convergence (ablation) | PGD adversarial training · CIFAR-10, ℓ∞ 8/255 | robust test error (%): 46.9 vs 47.1\pm 4.32 | No clear effect |
| Add semi-supervised augmentation and select its best checkpoint (ablation) | PGD adversarial training · CIFAR-10, ℓ∞ 8/255 | robust test error (%): 46.9 -> 40.2 | Helped |
| Increase Wide ResNet width factor from 10 to 20 (ablation) | PGD adversarial training · CIFAR-10, ℓ∞ 8/255 | best robust test error (%): 43.2{} -> 41.8{}\% | Helped |
Closest papers by meaning
- Benign Overfitting in Adversarial Training of Neural Networks (ICML 2024 · similarity 0.92)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.92)
- Theoretical Analysis of Robust Overfitting for Wide DNNs: An NTK Approach (ICLR 2024 · similarity 0.90)
- Layer-Aware Analysis of Catastrophic Overfitting: Revealing the Pseudo-Robust Shortcut Dependency (ICML 2024 · similarity 0.90)
- On the Clean Generalization and Robust Overfitting in Adversarial Training from Two Theoretical Views: Representation Complexity and Training Dynamics (ICML 2025 · similarity 0.90)
- The Effectiveness of Random Forgetting for Robust Generalization (ICLR 2024 · similarity 0.90)
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.