Research map / Adversarial robustness and security / Certified and empirical robustness
Certified Robust Neural Networks: Generalization and Corruption Resistance
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 KL statistical-error robustness (r>0) to adversarial training | adversarial training (HR) · CIFAR-10 pre-activation ResNet-18 robust overfitting replication | test loss over training epochs | Helped |
| add poisoning robustness (alpha) to the HR loss | HR training · CIFAR-10 with 10% label flips, stat. error + poisoning | mean error rate: HR 25.3% vs ERM 28.8% | Helped |
| set alpha too large in HR | HR training · CIFAR-10 poisoning | test loss | Hurt |
| add evasion robustness (N) via adversarial training | HR training · CIFAR-10 evasion only (eps 0.1 L2) | mean error rate: HR 18.5% vs ERM 35.4%; PGD 20.2% | Helped |
| use TRADES instead of PGD adversarial training | adversarial training · CIFAR-10 evasion only | mean error rate: TRADES 32.8% vs PGD 20.2% | Hurt |
| add statistical-error KL robustness (r) to HR, and poisoning+evasion together, versus KL-only DRO (ablation) | HR training · CIFAR-10 stat. error + poisoning + evasion | mean error rate: HR 36.6% vs KL DRO 49.9% | Helped |
| combine PGD and DPA defenses versus HR | combined defenses · CIFAR-10 stat. error + poisoning + evasion | mean error rate: PGD+DPA 45.2% vs HR 36.6% | Hurt |
| HR training computational overhead relative to PGD adversarial training | HR training · CIFAR-10 ResNet-18 training time | minutes per epoch: 2.75 vs 2.60 min/epoch (<+6%) | No clear effect |
| add statistical-error KL robustness (r) to HR under small data (5-50% subsamples) | HR training · CIFAR-10 subsampled training data | test loss | Helped |
Closest papers by meaning
- Fantastic Robustness Measures: The Secrets of Robust Generalization (NeurIPS 2023 · similarity 0.91)
- Towards Deep Learning Models Resistant to Adversarial Attacks (ICLR 2018 · similarity 0.90)
- Benign Overfitting in Adversarial Training of Neural Networks (ICML 2024 · similarity 0.89)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.89)
- On the Clean Generalization and Robust Overfitting in Adversarial Training from Two Theoretical Views: Representation Complexity and Training Dynamics (ICML 2025 · similarity 0.89)
- Conserve-Update-Revise to Cure Generalization and Robustness Trade-off in Adversarial Training (ICLR 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.