Research map / Adversarial robustness and security / Certified and empirical robustness
Detecting Overfitting via Adversarial Examples
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 |
|---|---|---|---|
| Use the strongest-perturbation translation AEG rather than the smallest-perturbation or random AEG | Pairwise independence test on ImageNet training data · ImageNet training set | independence-test confidence level: 99.9994% | Helped |
| Apply the test to ImageNet test data | VGG16 with strongest-perturbation translations · ImageNet test set | independence-test p-value: 0.96 | No clear effect |
| Average pairwise statistics across 120 independently trained models | Resnet50 N-model independence test · ImageNet test set | independence-test p-value: p value of 1 | No clear effect |
| Use adversarial-example-based testing at ε between 10 and 50 | Synthetic overfitted linear classifier · Synthetic overfitting experiment | independence-test confidence level: 1 - δ ≈ 100% | Helped |
| Use the N-model test with N=100 instead of testing each training run separately | Synthetic overfitting experiment at ε=6 · Synthetic overfitting experiment | p-value: 0.1153; mean 0.5984; median 0.6385 | Helped |
| Replace the strongest-perturbation AEG with the smallest-perturbation AEG (ablation) | ImageNet training-set independence test · ImageNet training set | overfitting detection | Hurt |
| Replace the strongest-perturbation AEG with random transformations (ablation) | ImageNet training-set independence test · ImageNet training set | overfitting detection | Hurt |
| Use separate confidence intervals instead of the pairwise test (ablation) | ImageNet training-set independence test · ImageNet training set | overfitting detection: 97.5%; 95% | Hurt |
| Use the nearest misclassified neighbor instead of the strongest perturbation (ablation) | Translational AEG for ImageNet classifiers · ImageNet | independence-test p-value | Mixed |
| Choose translations uniformly from non-identity neighbors without conditioning on attack success (ablation) | Random translational AEG · ImageNet | overfitting detection | Hurt |
| Choose translations uniformly from all neighbors, including the original image (ablation) | Random2 translational AEG · ImageNet | overfitting detection | Hurt |
| Increase the synthetic attack strength from the intermediate setting to ε=10 (ablation) | Single-model pairwise test on an overfitted classifier · Synthetic overfitting experiment | p-value distribution | Helped |
Closest papers by meaning
- Overfitting in adversarially robust deep learning (ICML 2020 · similarity 0.90)
- Natural Adversarial Examples (CVPR 2021 · similarity 0.88)
- Benign Overfitting in Adversarial Training of Neural Networks (ICML 2024 · similarity 0.88)
- Not All Wrong is Bad: Using Adversarial Examples for Unlearning (ICML 2025 · similarity 0.86)
- Wide Two-Layer Networks can Learn from Adversarial Perturbations (NeurIPS 2024 · similarity 0.86)
- On Local Overfitting and Forgetting in Deep Neural Networks (AAAI 2025 · similarity 0.86)
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.