Research map / Adversarial robustness and security / Certified and empirical robustness
Towards Robust Neural Networks via Random Self-ensemble
Xuanqing Liu, Minhao Cheng, Huan Zhang, Cho-Jui Hsieh
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 noise layers at training time as well as testing time (vs test noise only) (ablation) | RSE · CIFAR-10 VGG16 | accuracy without attack: 20% | Improved |
| add noise at test time with ensemble (vs noise only in training) (ablation) | RSE · CIFAR-10 VGG16 under C&W | accuracy under weak attacks | Improved |
| use 50-ensemble instead of 1-ensemble (ablation) | RSE · CIFAR-10 VGG16 (c<0.4) | accuracy under attack: ~8% | Improved |
| use 50-ensemble instead of 1-ensemble under very strong attack (ablation) | RSE · CIFAR-10 VGG16 (c>1.0) | accuracy under attack | Hurt |
| increase number of ensemble models beyond 10 (ablation) | RSE · CIFAR-10 VGG16 | clean test accuracy | No clear effect |
| add higher noise levels (init, inner) in training and testing (ablation) | RSE · CIFAR-10 VGG16 | robustness vs weak-attack accuracy | Mixed |
| use RSE instead of Robust Opt+BReLU | RSE · CIFAR-10 VGG16, c=0.1 | accuracy under C&W: 20.69% -> 67.19% | Improved |
| use RSE instead of no defense | VGG16 · CIFAR-10, c=0.03 | accuracy under C&W: 8.93% -> 86.06% | Improved |
| use RSE instead of adversarial training (II) | RSE · CIFAR-10 VGG16 | accuracy under C&W | No clear effect |
| use Robust Opt+BReLU | VGG/Model A · STL10 Model A | accuracy under C&W | Hurt |
Closest papers by meaning
- MORA: Improving Ensemble Robustness Evaluation with Model Reweighing Attack (NeurIPS 2022 · similarity 0.90)
- Self-ensemble Adversarial Training for Improved Robustness (ICLR 2022 · similarity 0.89)
- On the Robustness of Randomized Ensembles to Adversarial Perturbations (ICML 2023 · similarity 0.88)
- Adversarial Vulnerability of Randomized Ensembles (ICML 2022 · similarity 0.88)
- On the Role of Randomization in Adversarially Robust Classification (NeurIPS 2023 · similarity 0.88)
- Adversarial Robustness via Deformable Convolution with Stochasticity (ICML 2025 · similarity 0.87)
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.
