Research map / Adversarial robustness and security / Certified and empirical robustness
Lipschitz-Margin Training: Scalable Certification of Perturbation Invariance for Deep Neural Networks
Yusuke Tsuzuku, Issei Sato, Masashi Sugiyama
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 |
|---|---|---|---|
| Use Lipschitz-margin training instead of weight decay | Wide ResNet-16-4 · SVHN test, C&W attack at perturbation sizes 0.2, 0.5, and 1.0 | accuracy: 72.38 -> 86.90; 20.98 -> 55.11; 2.02 -> 17.69 | Improved |
| Use Lipschitz-margin training instead of weight decay | Wide ResNet-16-4 · SVHN test, clean | accuracy: 98.31 -> 96.38 | Hurt |
| Use Lipschitz-margin training instead of Parseval regularization | Wide ResNet-16-4 · SVHN test, C&W attack at perturbation sizes 0.2, 0.5, and 1.0 | accuracy: 71.35 -> 86.90; 17.92 -> 55.11; 0.94 -> 17.69 | Improved |
| Use Lipschitz-margin training instead of Parseval regularization | Wide ResNet-16-4 · SVHN test, clean | accuracy: 98.30 -> 96.38 | Hurt |
| Use Lipschitz-margin training instead of spectral norm regularization | Wide ResNet-16-4 · SVHN test, C&W attack at perturbation sizes 0.2, 0.5, and 1.0 | accuracy: 73.66 -> 86.90; 20.35 -> 55.11; 1.39 -> 17.69 | Improved |
| Use Lipschitz-margin training instead of spectral norm regularization | Wide ResNet-16-4 · SVHN test, clean | accuracy: 98.27 -> 96.38 | Hurt |
| Apply Lipschitz-margin training | Small network · Test data | median certified perturbation size (L2): 0.012 -> 1.02 | Improved |
| Replace stride-2, kernel-1 convolution with average-pooling and kernel-1 convolution, and use c=0.1 (ablation) | LMT · SVHN test | median certified perturbation size (L2): larger than 0.08 | Improved |
| Replace stride-2, kernel-1 convolution with average-pooling and kernel-1 convolution, and use c=0.1 (ablation) | LMT · SVHN test | accuracy: 86% | Hurt |
| Apply the margin addition only when the prediction is correct (ablation) | Lipschitz-margin training · Training stability | training stability | Improved |
| Increase c linearly to its target value over the first 5 epochs instead of applying strong regularization from the start (ablation) | Lipschitz-margin training · Training stability | training stability: 5 epochs | Improved |
Closest papers by meaning
- A Recipe for Improved Certifiable Robustness (ICLR 2024 · similarity 0.93)
- Efficient and Accurate Estimation of Lipschitz Constants for Deep Neural Networks (NeurIPS 2019 · similarity 0.92)
- ECLipsE: Efficient Compositional Lipschitz Constant Estimation for Deep Neural Networks (NeurIPS 2024 · similarity 0.90)
- DP-SGD Without Clipping: The Lipschitz Neural Network Way (ICLR 2024 · similarity 0.89)
- Efficient Robust Conformal Prediction via Lipschitz-Bounded Networks (ICML 2025 · similarity 0.89)
- Certified Robustness Under Bounded Levenshtein Distance (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.
