Research map / Adversarial robustness and security / Certified and empirical robustness
Defending Against Universal Perturbations With Shared Adversarial Training
Chaithanya Kumar Mummadi, Thomas Brox, Jan Hendrik Metzen
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 |
|---|---|---|---|
| apply standard adversarial training (s=1) to the ERM model | ResNet20 · CIFAR10 universal attack | universal robustness: 14.9 -> 28.14 | Improved |
| raise sharedness from 1 to 64 in adversarial training (ablation) | ResNet20 · CIFAR10 | accuracy at similar robustness: 78.75% -> 89.94% | Improved |
| raise sharedness from 1 to 8 (ablation) | ResNet20 · CIFAR10 | Pareto front | No clear effect |
| raise sharedness beyond 64 (ablation) | ResNet20 · CIFAR10 | robustness | No clear effect |
| use shared adversarial training with s=64 at no accuracy loss | ResNet20 · CIFAR10 | universal robustness: 14.9 -> 22.7 | Improved |
| use shared adversarial training instead of defense of Moosavi-Dezfooli et al. (compute) | ResNet20 · CIFAR10 | time per epoch: 3118s -> 189s | Improved |
| use shared adversarial training instead of defense of Perolat et al. (compute) | ResNet20 · CIFAR10 | time per epoch: 3840s -> 189s | Improved |
| use shared adversarial training s=32 at no accuracy loss | WRN-50-2-bottleneck · ImageNet 200-class subset | universal robustness: 8.4 -> 15.0 | Improved |
| use s=32 instead of s=1 at sigma=1.0, eps=10 (ablation) | WRN-50-2-bottleneck · ImageNet 200-class subset | accuracy: 63.51% -> 72.74% | Improved |
| use s=32 instead of s=1 at sigma=1.0, eps=2 (ablation) | WRN-50-2-bottleneck · ImageNet 200-class subset | universal robustness: 20.17 -> 14.94 | Hurt |
| use shared instead of standard adversarial training (ablation) | FCN-8 · Cityscapes targeted universal attack | universal robustness: 62.5 -> 111.7 | Improved |
| use shared instead of standard adversarial training (ablation) | FCN-8 · Cityscapes untargeted universal attack | universal robustness: 25 -> 47.8 | Improved |
Closest papers by meaning
- Universal Adversarial Perturbations Against Semantic Image Segmentation (ICCV 2017 · similarity 0.95)
- Universal Adversarial Training (AAAI 2020 · similarity 0.94)
- Generate Universal Adversarial Perturbations for Few-Shot Learning (NeurIPS 2024 · similarity 0.89)
- Wide Two-Layer Networks can Learn from Adversarial Perturbations (NeurIPS 2024 · similarity 0.89)
- Democratic Training Against Universal Adversarial Perturbations (ICLR 2025 · similarity 0.88)
- 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.
