Research map / Adversarial robustness and security / Certified and empirical robustness
Calibration and Consistency of Adversarial Surrogate Losses
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 |
|---|---|---|---|
| Optimize calibrated shifted ramp instead of the adversarial 0/1 objective | Linear models · Unit Circle | adversarial generalization error: 0.5000 -> 0.5263 | Hurt |
| Optimize calibrated shifted sigmoid instead of the adversarial 0/1 objective | Linear models · Unit Circle | adversarial generalization error: 0.5000 -> 0.5261 | Hurt |
| Use a calibrated surrogate without surrogate-risk realizability | Linear models · Segments | adversarial generalization error: 0 -> 0.0781 | Hurt |
| Use calibrated shifted sigmoid without surrogate-risk realizability | Linear models · Segments | adversarial generalization error: 0 -> 0.0777 | Hurt |
| Use calibrated ρ-margin ramp φ2 under the stated realizability conditions | Linear models · Segments | adversarial generalization error: 0.0111 -> 0 | Helped |
| Use convex loss φ1, whose surrogate risk has multiple zero-risk classifiers | Linear models · Segments | adversarial generalization error: 0 -> 0.0111 | Hurt |
| Replace convex margin losses with convex supremum-based losses | Generalized linear models and one-layer neural networks · Adversarial 0/1 loss | H-calibration | Hurt |
| Use a continuous margin-based surrogate without distributional assumptions | Linear models · Adversarial 0/1 loss | H-consistency | Hurt |
| Use a continuous non-increasing supremum-based surrogate without distributional assumptions | Linear models · Adversarial 0/1 loss | H-consistency | Hurt |
| Require zero adversarial Bayes risk and an in-class surrogate-risk minimizer | Calibrated ρ-margin ramp · Generalized linear models and ReLU-based models | H-consistency | Helped |
| Allow unbounded network output capacity (Λ=+∞) for the supremum-based ρ-margin loss | One-layer neural networks · Adversarial 0/1 loss | H-consistency | Helped |
Closest papers by meaning
- $H$-Consistency Guarantees for Regression (ICML 2024 · similarity 0.91)
- Realizable $H$-Consistent and Bayes-Consistent Loss Functions for Learning to Defer (NeurIPS 2024 · similarity 0.89)
- Trading off Consistency and Dimensionality of Convex Surrogates for Multiclass Classification (NeurIPS 2024 · similarity 0.88)
- Consistency Conditions for Differentiable Surrogate Losses (NeurIPS 2025 · similarity 0.88)
- Expressive Losses for Verified Robustness via Convex Combinations (ICLR 2024 · similarity 0.87)
- Towards Certification of Uncertainty Calibration under Adversarial Attacks (ICLR 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.