Research map / Adversarial robustness and security / Certified and empirical robustness
On the Loss Landscape of Adversarial Training: Identifying Challenges and How to Overcome Them
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 |
|---|---|---|---|
| add eps warmup (cosine) to adversarial training (ablation) | LeNet adversarial training · MNIST eps=0.4, PGD | robust error (%): 8.58 -> 6.64 | Helped |
| add eps warmup (linear) to adversarial training (ablation) | LeNet adversarial training · MNIST eps=0.4, clean | clean error (%): 1.56 -> 1.06 | Helped |
| add eps warmup (cosine) to adversarial training (ablation) | ResNet18, no periodic LR · CIFAR10 eps=8/255, PGD | robust error (%): 55.00 -> 53.95 | Helped |
| add eps warmup (cosine) to adversarial training (ablation) | VGG, no periodic LR · CIFAR10 eps=8/255, PGD | robust error (%): 56.22 -> 56.06 | No clear effect |
| add eps warmup (cosine) with periodic LR ensemble (ablation) | ResNet18 · CIFAR10 eps=8/255, PGD | robust error (%): 48.98 -> 48.57 | Helped |
| add periodic LR / 3-model ensemble (PAS) to constant eps (ablation) | ResNet18 adversarial training · CIFAR10 eps=8/255, PGD | robust error (%): 55.00 -> 48.98 | Helped |
| add periodic LR / ensemble to constant eps (ablation) | ResNet18 adversarial training · CIFAR10 eps=8/255, clean | clean error (%): 18.62 -> 21.00 | Hurt |
| add periodic LR / ensemble to constant eps (ablation) | VGG adversarial training · CIFAR10 eps=8/255, PGD | robust error (%): 56.22 -> 54.24 | Helped |
| increase adversarial budget epsilon | adversarial training · CIFAR10 ResNet18 early training | gradient norm / escape from initial region | Hurt |
| increase adversarial budget epsilon | adversarial training · CIFAR10 ResNet18 Hessian | top Hessian eigenvalues (sharpness) | Hurt |
| add eps warmup (linear/cosine) (ablation) | adversarial training with Adam · MNIST eps=0.4 | robust test error across learning rates | Helped |
| add learning rate warmup instead of eps warmup (ablation) | adversarial training · MNIST | robust error across learning rates | No clear effect |
Closest papers by meaning
- Towards Understanding the Dynamics of the First-Order Adversaries (ICML 2020 · similarity 0.90)
- Clarifying the Behavior and the Difficulty of Adversarial Training (AAAI 2024 · similarity 0.88)
- Over-parameterized Adversarial Training: An Analysis Overcoming the Curse of Dimensionality (NeurIPS 2020 · similarity 0.88)
- Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off (ICLR 2025 · similarity 0.87)
- A Closer Look at Curriculum Adversarial Training: From an Online Perspective (AAAI 2024 · similarity 0.87)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (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.