Research map / Adversarial robustness and security / Certified and empirical robustness
Deep Defense: Training DNNs with Improved 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 |
|---|---|---|---|
| Add adversarial perturbation-based regularization | DNN classifiers · MNIST, CIFAR-10, ImageNet | robustness to DeepFool (ρ₂): 1.11×10^{-1} -> 2.25×10^{-1}; 2.59×10^{-2} -> 5.32×10^{-2}; 2.98×10^{-3} -> 4.54×10^{-3} | Helped |
| Add adversarial perturbation-based regularization | DNN classifiers · MNIST, CIFAR-10, ImageNet | accuracy on benign examples: 98.31% -> 98.65%; 79.74% -> 81.70%; 56.91% -> 57.11% | Helped |
| Add adversarial perturbation-based regularization | DNN classifiers · MNIST, CIFAR-10, ImageNet | FGS accuracy at 1.0ε_ref: 3.31% -> 50.00%; 23.85% -> 50.00%; 46.05% -> 50.00% | Helped |
| Use a larger c value in the regularizer (ablation) | Deep Defense · MNIST | robustness and benign accuracy | Mixed |
| Use a very small c value, such as c=5 (ablation) | Deep Defense · MNIST | robustness and benign accuracy: c=5 | Mixed |
| Vary λ to increase the regularizer's contribution (ablation) | Deep Defense · MNIST | robustness and benign accuracy: λ in {5, 15, 25} | Mixed |
| Apply the regularizer gradient only to fc1 (ablation) | LeNet · MNIST | ρ₂ and benign accuracy | Helped |
| Apply the regularizer gradient to only one layer rather than all layers (ablation) | LeNet · MNIST | ρ₂ and benign accuracy | Mixed |
| Use an exponential penalty rather than a linear regularizer (ablation) | Deep Defense · training objective | benign accuracy and robustness trade-off | Helped |
| Treat correctly and incorrectly classified samples differently in the regularizer (ablation) | Deep Defense · training objective | benign accuracy and robustness | Helped |
Closest papers by meaning
- A Direct Approach to Robust Deep Learning Using Adversarial Networks (ICLR 2019 · similarity 0.91)
- Defense-GAN: Protecting Classifiers Against Adversarial Attacks Using Generative Models (ICLR 2018 · similarity 0.91)
- Defending Against Universal Attacks Through Selective Feature Regeneration (CVPR 2020 · similarity 0.90)
- You Only Propagate Once: Accelerating Adversarial Training via Maximal Principle (NeurIPS 2019 · similarity 0.89)
- Combating Adversaries with Anti-adversaries (AAAI 2022 · similarity 0.89)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · 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.