Research map / Adversarial robustness and security / Certified and empirical robustness
Self-Ensemble Protection: Training Checkpoints Are Good Data Protectors
Sizhe Chen, Geng Yuan, Xinwen Cheng, Yifan Gong, Minghai Qin, Yanzhi Wang, Xiaolin Huang
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 self-ensemble checkpoint gradients to craft protective perturbations instead of the single-model AdvPoison approach (ablation) | CIFAR-10 protection, ResNet18 appropriator · CIFAR-10, ℓ∞=2/255 | testing accuracy: 41.35 -> 19.92 | Improved |
| Add feature alignment to SEP (ablation) | SEP · CIFAR-10, ℓ∞=2/255, ResNet18 appropriator | testing accuracy: 19.92 -> 16.91 | Improved |
| Add variance reduction to SEP-FA (ablation) | SEP-FA · CIFAR-10, ℓ∞=2/255, ResNet18 appropriator | testing accuracy: 16.91 -> 14.68 | Improved |
| Use self-ensemble checkpoints instead of training a conventional ensemble of 30 models (ablation) | CIFAR-10 protection · CIFAR-10, ℓ∞=2/255, ResNet18 appropriator | testing accuracy: 16.32 -> 19.92 | Hurt |
| Add feature alignment to SEP (ablation) | SEP · CIFAR-10, ℓ∞=2/255, VGG16 appropriator | testing accuracy: 28.12 -> 26.82 | Improved |
| Add variance reduction to SEP-FA (ablation) | SEP-FA · CIFAR-10, ℓ∞=2/255, VGG16 appropriator | testing accuracy: 26.82 -> 23.66 | Improved |
| Increase the number of self-ensemble checkpoints from 30 to 120 (ablation) | SEP · CIFAR-10 | testing accuracy | No clear effect |
| Use ℓ∞+ℓ0 perturbations instead of ℓ∞ perturbations | SEP-FA-VR under ℓ∞ adversarial training · CIFAR-10, ResNet18 | testing accuracy: 82.51 -> 13.19 | Mixed |
Closest papers by meaning
- Self-ensemble Adversarial Training for Improved Robustness (ICLR 2022 · similarity 0.88)
- Boost Neural Networks by Checkpoints (NeurIPS 2021 · similarity 0.88)
- MORA: Improving Ensemble Robustness Evaluation with Model Reweighing Attack (NeurIPS 2022 · similarity 0.86)
- Ensemble Adversarial Training: Attacks and Defenses (ICLR 2018 · similarity 0.86)
- Enhancing Adversarial Transferability with Checkpoints of a Single Model's Training (CVPR 2025 · similarity 0.85)
- 3SAT: A Simple Self-Supervised Adversarial Training Framework (AAAI 2025 · similarity 0.84)
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.
