Research map / Adversarial robustness and security / Certified and empirical robustness
Confidence-Calibrated Adversarial Training: Generalizing to Unseen Attacks
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 |
|---|---|---|---|
| Replace adversarial one-hot targets with distance-dependent mixtures of the correct label and the uniform distribution, and reject low-confidence inputs | Adversarial training on SVHN · SVHN, unseen L∞ ε=0.06 | confidence-thresholded robust test error (RTE, %): 88.4 -> 53.1 | Helped |
| Replace adversarial one-hot targets with distance-dependent mixtures of the correct label and the uniform distribution, and reject low-confidence inputs | Adversarial training on SVHN · SVHN, unseen L2 ε=2 | confidence-thresholded robust test error (RTE, %): 99.4 -> 29.0 | Helped |
| Replace adversarial one-hot targets with distance-dependent mixtures of the correct label and the uniform distribution, and reject low-confidence inputs | Adversarial training on SVHN · SVHN, unseen L0 ε=10 | confidence-thresholded robust test error (RTE, %): 73.6 -> 3.5 | Helped |
| Use CCAT rather than ordinary adversarial training | Adversarial training on CIFAR10 · CIFAR10, seen L∞ ε=0.03 | confidence-thresholded robust test error (RTE, %): 62.7 -> 68.4 | Hurt |
| Use CCAT rather than TRADES | Adversarial training on CIFAR10 · CIFAR10, seen L∞ ε=0.03 | confidence-thresholded robust test error (RTE, %): 43.5 -> 68.4 | Hurt |
| Use CCAT rather than TRADES | Adversarial training on CIFAR10 · CIFAR10, unseen L2 ε=2 | confidence-thresholded robust test error (RTE, %): 70.9% -> 52.2% | Helped |
| Use CCAT rather than TRADES | Adversarial training on CIFAR10 · CIFAR10, unseen L1 ε=24 | confidence-thresholded robust test error (RTE, %): 96.9% -> 58.8% | Helped |
| Use CCAT rather than TRADES | Adversarial training on CIFAR10 · CIFAR10, unseen L0 ε=10 | confidence-thresholded robust test error (RTE, %): 36.9% -> 23% | Helped |
| Use CCAT rather than normal training | CIFAR10 classifier · CIFAR10-C | test error (%, mean across corruptions): 12.3% -> 8.5% | Helped |
| Increase the confidence threshold from 95% or 98% TPR to 99% TPR | CCAT evaluation · Various benchmarks and threat models | confidence-thresholded robust test error | Mixed |
Closest papers by meaning
- Bag of Tricks for Adversarial Training (ICLR 2021 · similarity 0.89)
- Certified Adversarial Robustness with Additive Noise (NeurIPS 2019 · similarity 0.89)
- Boosting Adversarial Robustness with CLAT: Criticality Leveraged Adversarial Training (ICML 2025 · similarity 0.89)
- On the Scalability of Certified Adversarial Robustness with Generated Data (NeurIPS 2024 · similarity 0.88)
- Towards Certification of Uncertainty Calibration under Adversarial Attacks (ICLR 2025 · similarity 0.88)
- Enhancing Adversarial Robustness with Conformal Prediction: A Framework for Guaranteed Model Reliability (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.