Research map / Adversarial robustness and security / Certified and empirical robustness
Regularizing Deep Networks Using Efficient Layerwise Adversarial Training
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 layerwise adversarial perturbations from gradients cached from the previous batch | VGG, Resnet-20, and Resnet-56 · CIFAR-10 and CIFAR-100 | accuracy: 92.1 $\pm$ 0.3 -> 92.65 $\pm$ 0.2; 90.27 $\pm$ 0.4 -> 91.1 $\pm$ 0.3; 91.53 $\pm$ 0.3 -> 94.1 $\pm$ 0.2; 69.8 $\pm$ 0.5 | Helped |
| Add layerwise adversarial training | AlexNet · ImageNet val | accuracy: 56.1\% -> 59.2\% | Helped |
| Train with layerwise perturbations and joint loss | WRN-28-10 · CIFAR-10 and CIFAR-100 | error rate: 4.00 -> 3.62 $\pm$ 0.05; 19.25 -> 17.1 $\pm$ 0.1 | Helped |
| Add layerwise adversarial training without dropout (ablation) | VGG · CIFAR-10 | accuracy: 92.1 -> 92.7 | Helped |
| Add layerwise adversarial training without dropout (ablation) | VGG · CIFAR-100 | accuracy: 69.8 -> 72.3 | Helped |
| Add dropout to layerwise adversarial training (ablation) | VGG · CIFAR-10 and CIFAR-100 | accuracy: 92.7 -> 93.2; 72.3 -> 73.1 | Helped |
| Extend gradient accumulation perturbations from shallower through deeper layers (ablation) | VGG · CIFAR-10 | accuracy: 89.5 -> 89.62 -> 90.24 -> 91.1 -> 91.3 | Helped |
| Use layerwise perturbations rather than input-only FGS training (ablation) | CIFAR-10 models evaluated on FGS-perturbed test inputs · CIFAR-10 | accuracy: At $\epsilon=0.1$: 77.1 -> 69.92 (FGS-orig to FGS-inter); 77.1 -> 68.37 (FGS-orig to Ours-orig); 77.1 -> 64.76 (FGS-orig | Hurt |
| Use input-only FGS adversarial training instead of layerwise training (ablation) | CIFAR-10 models · CIFAR-10 clean test data | accuracy: At $\epsilon=0$: FGS-orig 88.7; FGS-inter 90.9; Ours-orig 91.2; Ours-joint 91.5 | Mixed |
Closest papers by meaning
- Consistency Regularization for Adversarial Robustness (AAAI 2022 · similarity 0.90)
- Improving Adversarial Robustness by Putting More Regularizations on Less Robust Samples (ICML 2023 · similarity 0.90)
- Towards Understanding the Regularization of Adversarial Robustness on Neural Networks (ICML 2020 · similarity 0.89)
- Improving the Adversarial Robustness and Interpretability of Deep Neural Networks by Regularizing Their Input Gradients (AAAI 2018 · similarity 0.89)
- Boosting Adversarial Training via Fisher-Rao Norm-based Regularization (CVPR 2024 · similarity 0.87)
- Efficient local linearity regularization to overcome catastrophic overfitting (ICLR 2024 · 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.