Research map / Adversarial robustness and security / Certified and empirical robustness
DVERGE: Diversifying Vulnerabilities for Enhanced Robust Generation of Ensembles
Huanrui Yang, Jingyang Zhang, Hongliang Dong, Nathan Inkawhich, Andrew Gardner, Andrew Touchet, Wesley Wilkes, Heath Berry, Hai Li
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 |
|---|---|---|---|
| train ensemble with DVERGE vulnerability diversification instead of standard training | ResNet-20 ensemble (3) · CIFAR-10 black-box eps=0.03 | accuracy (%): 0 -> 51.0 | Improved |
| use DVERGE instead of GAL | ResNet-20 ensemble (3) · CIFAR-10 black-box eps=0.03 | accuracy (%): 39.8 -> 51.0 | Improved |
| use DVERGE vs baseline on clean data | ResNet-20 ensemble (3) · CIFAR-10 clean | accuracy (%): 93.9 -> 91.3 | Hurt |
| add more sub-models to DVERGE ensemble (3 to 8) (ablation) | DVERGE · CIFAR-10 black-box eps=0.03 | accuracy (%): 51.0 -> 65.0 | Improved |
| add more sub-models to GAL ensemble (3 to 8) | GAL · CIFAR-10 black-box eps=0.03 | accuracy (%): 39.8 -> 21.7 | Hurt |
| increase training eps from 0.03 to 0.07 (ablation) | DVERGE (3 sub-models) · CIFAR-10 black-box eps=0.03 | accuracy (%): 4.2 -> 53.2 | Improved |
| increase training eps from 0.03 to 0.07 (ablation) | DVERGE (3 sub-models) · CIFAR-10 clean | accuracy (%): 92.9 -> 91.4 | Hurt |
| increase training eps from 0.05 to 0.07 (ablation) | DVERGE (5 sub-models) · CIFAR-10 white-box eps=0.01 | accuracy (%): 48.9 -> 42.3 | Hurt |
| add adversarial training objective to DVERGE, compared with AdvT alone (ablation) | DVERGE+AdvT (5) · CIFAR-10 clean | accuracy (%): 78.6 -> 83.4 | Improved |
| add adversarial training objective to DVERGE, compared with AdvT alone (ablation) | DVERGE+AdvT (5) · CIFAR-10 black-box eps=0.03 | accuracy (%): 73.8 -> 76.6 | Improved |
| add adversarial training objective to DVERGE, compared with AdvT alone (ablation) | DVERGE+AdvT (5) · CIFAR-10 black-box eps=0.07 | accuracy (%): 63.3 -> 61.7 | Mixed |
| use DVERGE instead of adversarial training in compute | DVERGE · CIFAR-10 training time | hours: 11.5 -> 10.5 | Improved |
Later papers that compared with this method
1 later accepted paper reports DVERGE/3 as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Ensemble-in-One: Ensemble Learning within Random Gated Networks for Enhanced Adversarial Robustness (AAAI 2023 · better on 5 of 6 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Improving Adversarial Robustness via Promoting Ensemble Diversity (ICML 2019 · similarity 0.91)
- Improving Ensemble Robustness by Collaboratively Promoting and Demoting Adversarial Robustness (AAAI 2021 · similarity 0.90)
- Rethinking Model Ensemble in Transfer-based Adversarial Attacks (ICLR 2024 · similarity 0.89)
- Understanding Model Ensemble in Transferable Adversarial Attack (ICML 2025 · similarity 0.89)
- On the Diversity of Adversarial Ensemble Learning (ICML 2025 · similarity 0.89)
- Enhancing Adversarial Transferability with Checkpoints of a Single Model's Training (CVPR 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.
