Research map / Adversarial robustness and security / Certified and empirical robustness
How Benign is Benign Overfitting ?
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 |
|---|---|---|---|
| Increase label noise while training to zero training error | Toy-MNIST neural network · Toy-MNIST | adversarial error | Hurt |
| Fit randomized labels until zero training error | Naturally trained CNN · MNIST with 15% randomized labels | training accuracy: 100% | Helped |
| Use adversarial training instead of natural training | CNN · MNIST with 15% randomized labels | training misclassified examples: 997; 994 | Mixed |
| Use robust training rather than natural training | AT and TRADES models · MNIST and CIFAR10 training sets with identified noisy labels | accuracy on noisy training labels | Hurt |
| Increase ℓ2 adversarial-training strength from ε=0 to ε=0.25 | ResNet-50 · Clean dataset | Test-Acc (%): 95.25 -> 92.77 | Hurt |
| Increase ℓ2 adversarial-training strength from ε=0.25 to ε=1.0 | ResNet-50 · Clean dataset | Test-Acc (%): 92.77 -> 81.62 | Hurt |
| Use adversarial training instead of natural training | Models trained on synthetic data with rare sub-populations · Synthetic circles | robust accuracy | Helped |
| Train with fine-grained subclass labels, then aggregate subclass predictions | Deep neural network · Synthetic circles | robustness | Helped |
| Train on fine labels and aggregate logits into coarse labels | Fine2Coarse ResNet-50 · CIFAR-100 coarse labels | adversarial risk | Helped |
| Stop training before memorizing noisy labels | Toy-MNIST neural network · Toy-MNIST | adversarial error | Helped |
| Fit noisy labels to zero training error rather than avoid memorizing them | Neural networks trained on noisy data · MNIST and toy-MNIST | adversarial vulnerability | Hurt |
Closest papers by meaning
- Label Noise in Adversarial Training: A Novel Perspective to Study Robust Overfitting (NeurIPS 2022 · similarity 0.91)
- A law of adversarial risk, interpolation, and label noise (ICLR 2023 · similarity 0.91)
- Benign Overfitting in Adversarial Training of Neural Networks (ICML 2024 · similarity 0.90)
- Soften to Defend: Towards Adversarial Robustness via Self-Guided Label Refinement (CVPR 2024 · similarity 0.90)
- Overfitting in adversarially robust deep learning (ICML 2020 · similarity 0.89)
- Rethinking Benign Overfitting in Two-Layer Neural Networks (ICML 2025 · 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.