Research map / Adversarial robustness and security / Certified and empirical robustness
Revisiting Outer Optimization in Adversarial Training
Ali Dabouei, Fariborz Taherkhani, Sobhan Soleymani, Nasser M. Nasrabadi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 MSGD with ENGM as outer optimizer | vanilla AT · CIFAR-10, ResNet-18 | AutoAttack accuracy: 46.77 -> 49.24 | Improved |
| replace MSGD with ENGM as outer optimizer | vanilla AT · CIFAR-10, ResNet-18 | robust overfitting: 13.2 -> 3.9 | Improved |
| replace MSGD with ENGM as outer optimizer | vanilla AT · CIFAR-10, ResNet-18 | natural accuracy: 84.70 -> 83.44 | Hurt |
| add gradient norm clipping to MSGD (ablation) | MSGD · CIFAR-10, ResNet-18 | AutoAttack accuracy: 46.77 -> 47.59 | Improved |
| replace per-example bounding with normalizing to constant (F-ENGM) (ablation) | ENGM · CIFAR-10, ResNet-18 | AutoAttack accuracy: 49.24 -> 46.54 | Hurt |
| use ENGM instead of naive approximation N-ENGM (ablation) | ENGM · CIFAR-10, ResNet-18 | AutoAttack accuracy: 48.06 -> 49.24 | Improved |
| use ENGM with exact norms instead of A-ENGM (ablation) | ENGM · CIFAR-10, ResNet-18 | execution time (sec./iter): 0.83 -> 5.06 | Hurt |
| replace MSGD with ENGM on TRADES | TRADES · CIFAR-100, ResNet-18 | AutoAttack accuracy: 26.93 -> 29.20 | Improved |
| replace MSGD with ENGM on AWP | AWP · Tiny-ImageNet, ResNet-18 | AutoAttack accuracy: 13.09 -> 16.02 | Improved |
| replace MSGD with ENGM on vanilla AT | vanilla AT · Tiny-ImageNet, ResNet-18 | natural accuracy: 35.71 -> 29.78 | Hurt |
| replace MSGD with ENGM on AWP | AWP · CIFAR-10, WRN-34-10 | AutoAttack accuracy: 56.17 -> 57.45 | Improved |
| replace MSGD with N-ENGM on AWP | AWP · CIFAR-10, WRN-34-10 | AutoAttack accuracy: 56.17 -> 57.11 | Improved |
Closest papers by meaning
- Improved Adversarial Training via Learned Optimizer (ECCV 2020 · similarity 0.89)
- Fast is better than free: Revisiting adversarial training (ICLR 2020 · similarity 0.89)
- Bag of Tricks for Adversarial Training (ICLR 2021 · similarity 0.88)
- Understanding and Increasing Efficiency of Frank-Wolfe Adversarial Training (CVPR 2022 · similarity 0.88)
- Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off (ICLR 2025 · similarity 0.87)
- Adversarial Initialization with Universal Adversarial Perturbation: A New Approach to Fast Adversarial Training (AAAI 2024 · 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.
