Research map / Adversarial robustness and security / Certified and empirical robustness
Improving Adversarial Robustness via Promoting Ensemble Diversity
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 ADP regularization (α=2, β=0.5) to the ensemble | Three-member ResNet-20 ensemble · MNIST, CIFAR-10, and CIFAR-100 test sets | classification error rate (%): MNIST: 0.32 -> 0.28; CIFAR-10: 6.78 -> 6.56; CIFAR-100: 30.35 -> 29.80 | Helped |
| Add ADP regularization (α=2, β=0.5) to the ensemble | Three-member ResNet-20 ensemble · MNIST FGSM (ε=0.1), BIM (ε=0.1), and PGD (ε=0.1) | classification accuracy (%): FGSM: 78.3 -> 96.3; BIM: 52.3 -> 88.5; PGD: 50.7 -> 82.8 | Helped |
| Add ADP regularization (α=2, β=0.5) to the ensemble | Three-member ResNet-20 ensemble · CIFAR-10 FGSM (ε=0.02) and PGD (ε=0.01) | classification accuracy (%): FGSM: 36.5 -> 61.7; PGD: 23.4 -> 48.4 | Helped |
| Add ADP regularization to FGSM adversarial training | CIFAR-10 ResNet-20 ensemble · CIFAR-10 FGSM adversarial training (ε=0.04) | classification accuracy (%): 39.3 -> 56.1 | Helped |
| Add ADP regularization to PGD adversarial training | CIFAR-10 ResNet-20 ensemble · CIFAR-10 PGD adversarial training | classification accuracy (%): PGD: 32.8 -> 36.2 | Helped |
| Remove the LED diversity term (set β=0), retaining ensemble entropy (ablation) | ADP-trained ensemble · CIFAR-10 test | classification error rate (%): 6.56 -> 6.74 | Hurt |
| Remove the LED diversity term (set β=0), retaining ensemble entropy (ablation) | ADP-trained ensemble · CIFAR-10 PGD (ε=0.01) | classification accuracy (%): 43.2 -> 48.4 | Hurt |
| Remove both regularizer terms (α=β=0; baseline training) (ablation) | Three-member ResNet-20 ensemble · CIFAR-10 test | classification error rate (%): 6.56 -> 6.78 | Hurt |
| Use ADP instead of baseline training | Individual members of the three-member ensemble · CIFAR-10 test | classification error rate (%): Net 1: 8.30 -> 9.34; Net 2: 8.52 -> 9.34; Net 3: 8.67 -> 9.92 | Hurt |
| Train five members with ADP when K does not divide L−1 | Five-member ResNet-20 ensemble · CIFAR-10 FGSM (ε=0.02) | classification accuracy (%): 42.0 -> 58.4 | Helped |
| Promote diversity among member predictions with ADP training | Ensemble members · CIFAR-10, adversarial examples transferred between members | transferability | Helped |
| Add ADP training to the ensemble | CIFAR-10 ensemble, K=20 · CIFAR-10 training | training time per epoch: 692s/epoch -> 744s/epoch | Hurt |
Closest papers by meaning
- On the Diversity of Adversarial Ensemble Learning (ICML 2025 · similarity 0.93)
- Improving Ensemble Robustness by Collaboratively Promoting and Demoting Adversarial Robustness (AAAI 2021 · similarity 0.92)
- Adversarial Defence by Diversified Simultaneous Training of Deep Ensembles (AAAI 2021 · similarity 0.92)
- Exploiting Joint Robustness to Adversarial Perturbations (CVPR 2020 · similarity 0.91)
- Ensemble Diversity Facilitates Adversarial Transferability (CVPR 2024 · 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.