Research map / Adversarial robustness and security / Certified and empirical robustness
Improving DNN Robustness to Adversarial Attacks using Jacobian Regularization
Daniel Jakubovitz, Raja Giryes
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 |
|---|---|---|---|
| add Jacobian regularization as post-processing | CNN · MNIST | DeepFool rho_adv: 20.67 -> 34.24 | Improved |
| add Jacobian regularization as post-processing | CNN · CIFAR-10 | DeepFool rho_adv: 1.21 -> 3.42 | Improved |
| add Jacobian regularization as post-processing | CNN · CIFAR-100 | DeepFool rho_adv: 0.61 -> 1.55 | Improved |
| add adversarial training on top of Jacobian regularization | Jacobian regularization · CIFAR-10 | DeepFool rho_adv: 3.42 -> 6.03 | Improved |
| apply Jacobian regularization as post-processing instead of during regular training (ablation) | Jacobian regularization · MNIST | DeepFool rho_adv: 32.89 -> 34.24 | Improved |
| regularize last-layer Jacobian instead of L-1 layer (fixed W) (ablation) | Jacobian regularization · MNIST | DeepFool rho_adv: 29.61 -> 34.24 | Improved |
| regularize last-layer Jacobian instead of L-2 layer (fixed weights) (ablation) | Jacobian regularization · MNIST | DeepFool rho_adv: 22.52 -> 34.24 | Improved |
| raise lambda beyond 0.1 (ablation) | Jacobian regularization · MNIST | DeepFool rho_adv | Hurt |
| increase fraction of perturbed examples in adversarial training from 10% to 50% (ablation) | Adversarial Training · MNIST | DeepFool rho_adv: 23.73 -> 18.74 | Hurt |
| Jacobian regularization under JSMA attack versus Input Gradient regularization | Jacobian regularization · MNIST | JSMA test accuracy | Hurt |
| Jacobian regularization under JSMA attack versus other defenses | Jacobian regularization · CIFAR-10 | JSMA test accuracy | Improved |
| add Jacobian regularization | CNN · MNIST | Jacobian Frobenius norm on perturbed inputs: 0.1877 -> 0.055 | Improved |
Closest papers by meaning
- Improving Adversarial Robustness by Putting More Regularizations on Less Robust Samples (ICML 2023 · similarity 0.93)
- On Connections Between Regularizations for Improving DNN Robustness (TPAMI 2020 · similarity 0.92)
- Jacobian Adversarially Regularized Networks for Robustness (ICLR 2020 · similarity 0.91)
- Towards Understanding the Regularization of Adversarial Robustness on Neural Networks (ICML 2020 · similarity 0.91)
- Consistency Regularization for Adversarial Robustness (AAAI 2022 · similarity 0.91)
- 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.
