Research map / Adversarial robustness and security / Certified and empirical robustness
Adversarial Robustness Via Fisher-Rao Regularization
Marine Picot, Francisco Javier Messina, Malik Boudiaf, Fabrice Labeau, Ismail Ben Ayed, Pablo Piantanida
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 Fisher-Rao regularizer instead of KL (TRADES) (ablation) | FIRE · CIFAR-10 (eps 8/255, no extra data) | AutoAttack accuracy: 50.47 -> 51.45 | Improved |
| use Fisher-Rao regularizer instead of KL (TRADES) (ablation) | FIRE · CIFAR-100 (eps 8/255) | AutoAttack accuracy: 25.89 -> 26.42 | Improved |
| use Fisher-Rao regularizer instead of KL (TRADES) (ablation) | FIRE · MNIST (eps 0.3) | AutoAttack accuracy: 94.27 -> 94.44 | Improved |
| use Fisher-Rao regularizer instead of Hellinger (ablation) | FIRE · CIFAR-10 (eps 8/255, no extra data) | AutoAttack accuracy: 50.52 -> 51.45 | Improved |
| use Fisher-Rao regularizer instead of Hellinger (ablation) | FIRE · CIFAR-100 (eps 8/255) | AutoAttack accuracy: 25.58 -> 26.42 | Improved |
| use Fisher-Rao regularizer instead of Hellinger (ablation) | FIRE · MNIST (eps 0.3) | AutoAttack accuracy: 94.03 -> 94.44 | Improved |
| use FIRE instead of TRADES to reduce training time | FIRE · CIFAR-10 (eps 8/255, no extra data) | training runtime: 17h49 -> 11h00 | Improved |
| use FIRE instead of Overfitting method for training time | FIRE · CIFAR-10 (eps 8/255, no extra data) | training runtime: 42h01 -> 11h00 | Improved |
| combine FIRE with Self Adaptive | Self Adaptive · CIFAR-10 (eps 8/255, no extra data) | AutoAttack accuracy: 53.11 -> 53.26 | No clear effect |
| use FIRE with 80M-TI extra data instead of RST-adv | FIRE · CIFAR-10 (eps 8/255, with 80M-TI extra data) | AutoAttack accuracy: 59.69 -> 59.97 | Improved |
| use FIRE instead of Overfitting method | WRN-34-10 · CIFAR-100 (eps 8/255) | natural accuracy: 59.22 -> 61.03 | Improved |
| use FIRE instead of Overfitting method | WRN-34-10 · CIFAR-10 (eps 8/255, no extra data) | AutoAttack accuracy: 51.72 -> 51.45 | No clear effect |
Closest papers by meaning
- Boosting Adversarial Training via Fisher-Rao Norm-based Regularization (CVPR 2024 · similarity 0.90)
- Rethinking Softmax Cross-Entropy Loss for Adversarial Robustness (ICLR 2020 · similarity 0.90)
- Distributional Robustness with IPMs and links to Regularization and GANs (NeurIPS 2020 · similarity 0.89)
- Adversarial Robustness of Supervised Sparse Coding (NeurIPS 2020 · similarity 0.89)
- Improving Adversarial Robustness by Putting More Regularizations on Less Robust Samples (ICML 2023 · similarity 0.88)
- Revisiting Hilbert-Schmidt Information Bottleneck for Adversarial Robustness (NeurIPS 2021 · 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.
