Research map / Adversarial robustness and security / Certified and empirical robustness
Fundamental Tradeoffs between Invariance and Sensitivity to Adversarial Perturbations
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 ℓ∞-PGD training perturbation bound | MNIST classifiers · MNIST ℓ∞ invariance examples | Accuracy against human labels | Hurt |
| Increase ℓ1-PGD training perturbation bound | MNIST classifiers · MNIST ℓ0 invariance examples | Accuracy against human labels | Hurt |
| Train for ℓp-bounded sensitivity robustness | MNIST robust classifiers · MNIST invariance examples | Accuracy against human labels | Hurt |
| Use manual examples rather than automated examples | ℓ∞ invariance attack ε=0.4 · MNIST human study | Success rate: 37% -> 88% | Helped |
| Use manual examples rather than automated examples | ℓ∞ invariance attack ε=0.3 · MNIST human study | Success rate: 21% -> 26% | Helped |
| Increase ℓ∞ perturbation budget from ε=0.3 to ε=0.4 | Automated invariance attack · MNIST human study | Success rate: 21% -> 37% | Helped |
| Add data augmentation over spatial transformations | MNIST models trained against ℓ∞ perturbations · MNIST invariance examples ε=0.4 | Agreement with human labels | Helped |
| Increase ℓ∞ training perturbation bound while using spatial data augmentation | MNIST models · MNIST augmented adversarial training | Robust accuracy | Hurt |
| Use a larger ℓ1-PGD training bound | MNIST classifiers · MNIST, ℓ0-PointwiseAttack ε=25 | Robust accuracy: 41% -> 59% -> 65% | Helped |
| Increase ℓ∞-PGD training bound from ε=0.2 to ε=0.3 | MNIST classifiers · MNIST, PGD ε=0.3 | Robust accuracy: 6% -> 92% | Helped |
| Increase ℓ∞-PGD training bound from ε=0.3 to ε=0.4 | MNIST classifiers · MNIST, PGD ε=0.4 | Robust accuracy: 7% -> 90% | Helped |
Closest papers by meaning
- Excessive Invariance Causes Adversarial Vulnerability (ICLR 2019 · similarity 0.92)
- Theoretical Understanding of Learning from Adversarial Perturbations (ICLR 2024 · similarity 0.89)
- Wide Two-Layer Networks can Learn from Adversarial Perturbations (NeurIPS 2024 · similarity 0.89)
- Doppelgangers and Adversarial Vulnerability (CVPR 2025 · similarity 0.88)
- Precise Accuracy / Robustness Tradeoffs in Regression: Case of General Norms (ICML 2024 · similarity 0.88)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · 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.