Research map / Adversarial robustness and security / Certified and empirical robustness
Rethinking Softmax Cross-Entropy Loss for Adversarial Robustness
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 SCE with MMC-10 | ResNet-32 · CIFAR-10, adaptive PGD-50 targeted, epsilon=8/255 | accuracy: 8.9 -> 26.6 | Helped |
| Replace SCE with MMC-10 | ResNet-32 · CIFAR-10, adaptive PGD-50 untargeted, epsilon=8/255 | accuracy: 4.9 -> 24.8 | Helped |
| Replace SCE with MMC-10 | ResNet-32 · CIFAR-10, clean test | accuracy: 92.9 -> 92.7 | No clear effect |
| Replace SCE with MMC-10 | ResNet-32 · CIFAR-10, adaptive C&W targeted | l2 distortion: 0.12 -> 0.34 | Helped |
| Replace SCE with MMC-10 | ResNet-32 · CIFAR-10, SPSA-10 untargeted, epsilon=8/255 | accuracy: 1.2 -> 56.9 | Helped |
| Replace SCE with MMC-10 | ResNet-32 · CIFAR-100, adaptive PGD-10 untargeted, epsilon=8/255 | accuracy: 8.0 -> 23.4 | Helped |
| Replace SCE with MMC-10 | ResNet-32 · CIFAR-100, clean test | accuracy: 72.9 -> 71.9 | No clear effect |
| Use optimal Max-Mahalanobis centers rather than uniformly sampled hypersphere centers (ablation) | MMC-10 · CIFAR-10, adaptive PGD-50 targeted, epsilon=8/255 | accuracy: 20.9 -> 26.6 | Helped |
| Use optimal Max-Mahalanobis centers rather than uniformly sampled hypersphere centers (ablation) | MMC-10 · CIFAR-10, adaptive PGD-50 untargeted, epsilon=8/255 | accuracy: 18.4 -> 24.8 | Helped |
| Add MMC-10 loss to targeted adversarial training instead of using SCE | ResNet-32 adversarial training with targeted PGD-10 · CIFAR-10, adaptive PGD-50 untargeted, epsilon=16/255 | accuracy: 16.0 -> 45.1 | Helped |
| Add MMC-10 loss to untargeted adversarial training instead of using SCE | ResNet-32 adversarial training with untargeted PGD-10 · CIFAR-10, adaptive PGD-50 untargeted, epsilon=16/255 | accuracy: 21.5 -> 27.7 | Helped |
| Train with MMC rather than SCE | ResNet-32 · CIFAR-10, rotation transformation | accuracy: 83.5 -> 87.2 | Helped |
Closest papers by meaning
- Building Robust Ensembles via Margin Boosting (ICML 2022 · similarity 0.90)
- Fantastic Robustness Measures: The Secrets of Robust Generalization (NeurIPS 2023 · similarity 0.89)
- Expressive Losses for Verified Robustness via Convex Combinations (ICLR 2024 · similarity 0.88)
- Revisiting Adversarial Training Under Long-Tailed Distributions (CVPR 2024 · similarity 0.87)
- Soften to Defend: Towards Adversarial Robustness via Self-Guided Label Refinement (CVPR 2024 · similarity 0.87)
- On the Clean Generalization and Robust Overfitting in Adversarial Training from Two Theoretical Views: Representation Complexity and Training Dynamics (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.