Research map / Adversarial robustness and security / Certified and empirical robustness
Characterizing the Optimal $0-1$ Loss for Multi-class Classification with a Test-time Attacker
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 pairwise binary-classification bounds instead of truncated conflict hypergraphs | Optimal-loss lower-bound computation · MNIST and CIFAR-10, 10-class | tightness of optimal-loss lower bound | Hurt |
| Limit conflict hyperedges to size 2 rather than including higher-degree hyperedges (ablation) | Optimal-loss lower-bound computation · MNIST and CIFAR-10, 10-class, small epsilon | lower-bound information | No clear effect |
| Add degree-3 hyperedges to the conflict hypergraph (ablation) | Optimal-loss lower-bound computation · CIFAR-10, epsilon 3 | optimal-loss lower bound | No clear effect |
| Add degree-4 hyperedges to the conflict hypergraph (ablation) | Optimal-loss lower-bound computation · MNIST, CIFAR-10, and CIFAR-100 | optimal-loss lower bound | No clear effect |
| Use the Caro-Wei upper bound with weights from L*(2) | Optimal-loss estimation · MNIST and CIFAR-10, small epsilon | gap between optimal-loss lower and upper bounds | Helped |
| Increase the neural-network architecture size (ablation) | TRADES-trained classifiers · MNIST and CIFAR-10, 3-class | 0-1 loss | Mixed |
| Use PGD adversarial training instead of TRADES (ablation) | Adversarially trained classifier · MNIST and CIFAR-10, 3-class | 0-1 loss | Hurt |
| Use current state-of-the-art adversarially trained classifiers rather than the optimal classifier | 3-class robust classification · CIFAR-10, epsilon where optimal loss is near 0 | 0-1 loss | Hurt |
| Use the best certifiably robust models rather than the optimal-loss bound | Robust classifiers · MNIST and CIFAR-10, budgets 1.0–2.0 | 0-1 loss: 0 vs 0.27; 0 vs 0.44; 0 vs 0.6; 0 vs 0.8 | Hurt |
Closest papers by meaning
- Lower Bounds on Adversarial Robustness from Optimal Transport (NeurIPS 2019 · similarity 0.90)
- MultiRobustBench: Benchmarking Robustness Against Multiple Attacks (ICML 2023 · similarity 0.89)
- Analyzing the Robustness of Nearest Neighbors to Adversarial Examples (ICML 2018 · similarity 0.88)
- Lower Bounds on Cross-Entropy Loss in the Presence of Test-time Adversaries (ICML 2021 · similarity 0.88)
- Adversarial examples from computational constraints (ICML 2019 · similarity 0.88)
- Gradient Flow Provably Learns Robust Classifiers for Orthonormal GMMs (ICML 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.