Research map / Adversarial robustness and security / Certified and empirical robustness
Improved Adversarial Training via Learned Optimizer
Yuanhao Xiong, Cho-Jui Hsieh
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 constant-step PGD inner maximizer with backtracking line search (ablation) | AdvTrain · MNIST | robust accuracy (%): 87.33 -> 88.71 | Improved |
| attack with PGD+BLS instead of PGD (ablation) | AdvTrain model · MNIST | robust accuracy (%): 87.33 -> 86.09 | Improved |
| replace PGD inner maximizer with learned RNN optimizer | AdvTrain · MNIST | min robust accuracy (%): 94.28 -> 95.51 | Improved |
| replace PGD inner maximizer with learned RNN optimizer | AdvTrain · CIFAR-10 (VGG-16) | min robust accuracy (%): 40.75 -> 42.89 | Improved |
| replace PGD inner maximizer with learned RNN optimizer | AdvTrain · CIFAR-10 (WideResNet) | min robust accuracy (%): 45.13 -> 45.98 | Improved |
| replace PGD inner maximizer with learned RNN optimizer | TRADES · CIFAR-10 (VGG-16) | min robust accuracy (%): 45.95 -> 47.23 | Improved |
| replace PGD inner maximizer with learned RNN optimizer | TRADES · CIFAR-10 (WideResNet) | min robust accuracy (%): 52.68 -> 54.11 | Improved |
| replace PGD inner maximizer with learned RNN optimizer | TRADES · MNIST | min robust accuracy (%): 95.50 -> 95.80 | Improved |
| replace PGD inner maximizer with learned RNN optimizer | TRADES · Restricted ImageNet (ResNet-18) | min robust accuracy (%): 4.80 -> 5.79 | Improved |
| use RNN optimizer rather than CNN-based L2LDA attacker | L2LDA · CIFAR-10 (VGG-16) | min robust accuracy (%): 34.27 -> 47.23 | Improved |
| extend learned RNN attacker from 10 to 40 steps | RNN-TRADES · CIFAR-10 (VGG-16) | robust accuracy (%): 49.49 -> 49.28 | Improved |
| replace PGD training with learned RNN optimizer on natural accuracy | TRADES · CIFAR-10 (WideResNet) | natural accuracy (%): 85.89 -> 84.21 | Hurt |
Closest papers by meaning
- Attacks Which Do Not Kill Training Make Adversarial Learning Stronger (ICML 2020 · similarity 0.90)
- Adversarial Attack Generation Empowered by Min-Max Optimization (NeurIPS 2021 · similarity 0.90)
- Understanding and Increasing Efficiency of Frank-Wolfe Adversarial Training (CVPR 2022 · similarity 0.89)
- Adversarial Distributional Training for Robust Deep Learning (NeurIPS 2020 · similarity 0.89)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.89)
- Clarifying the Behavior and the Difficulty of Adversarial Training (AAAI 2024 · similarity 0.88)
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.
