Research map / Adversarial robustness and security / Certified and empirical robustness
On the Sensitivity of Adversarial Robustness to Input Data Distributions
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 |
|---|---|---|---|
| Smooth MNIST inputs with progressively larger kernels | PGD-trained LeNet5 · MNIST variants test | robust accuracy (ε=0.3): 95.1 % -> 83.1 % | Hurt |
| Binarize MNIST inputs rather than use original MNIST | PGD-trained LeNet5 · MNIST variants test | robust accuracy (ε=0.3): 95.1 % -> 98.1 % | Helped |
| Saturate CIFAR10 more strongly, from original through level 16 | PGD-trained WideResNet-28-4 · CIFAR10 variants test | robust accuracy (ε=8/255): 43.2 % -> 79.4 % | Helped |
| Saturate CIFAR10 at level 1 instead of using original CIFAR10 | PGD-trained WideResNet-28-4 · CIFAR10 variants test | robust accuracy (ε=8/255): 43.2 % -> 33.0 % | Hurt |
| Use more strongly saturated CIFAR10, level 16, for standard training | Standardly trained WideResNet-28-4 · CIFAR10 variants test | test accuracy: 95.0 % -> 92.9 % | No clear effect |
| Use semantically similar data with different gamma correction | PGD-trained CIFAR10 models · Gamma-mapped CIFAR10 | robust accuracy | Mixed |
| Train on edge-fashion-MNIST rather than fashion-MNIST | PGD-trained models · fashion-MNIST and edge-fashion-MNIST | robust accuracy | Helped |
| Allow PGD attacks beyond the input-domain boundary (ablation) | PGD evaluation on MNIST and CIFAR10 variants · MNIST and CIFAR10 variants | robust accuracy: 95.1 % -> 95.1 %; 43.2 % -> 43.0 % | No clear effect |
| Match inter-class distance using scaled original or binarized MNIST instead of smoothed MNIST (ablation) | PGD-trained models · MNIST variants | robust accuracy: 91.3 % -> 94.6 %; 91.3 % -> 98.6 % | Helped |
| Increase LeNet5 capacity on MNIST variants (ablation) | PGD-trained LeNet5 · MNIST variants test | robust accuracy: original 87.1% -> 95.2%; smooth 8 71.7% -> 83.1% | Mixed |
| Increase WideResNet capacity on CIFAR10 variants (ablation) | PGD-trained WideResNet · CIFAR10 variants test | robust accuracy: original 35.0% -> 43.2%; saturate 16 65.0% -> 79.4% | Mixed |
| Increase training-set size for PGD training (ablation) | PGD-trained models · MNIST and CIFAR10 variants | test robust accuracy | Mixed |
Closest papers by meaning
- Robust Models are less Over-Confident (NeurIPS 2022 · similarity 0.91)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.91)
- On the Vulnerability of Adversarially Trained Models Against Two-faced Attacks (ICLR 2024 · similarity 0.91)
- Explicit Tradeoffs between Adversarial and Natural Distributional Robustness (NeurIPS 2022 · similarity 0.91)
- Certifying Some Distributional Robustness with Principled Adversarial Training (ICLR 2018 · similarity 0.91)
- OODRobustBench: a Benchmark and Large-Scale Analysis of Adversarial Robustness under Distribution Shift (ICML 2024 · similarity 0.89)
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.