Research map / Adversarial robustness and security / Certified and empirical robustness
Robustness via Cross-Domain Ensembles
Teresa Yeo, Oğuzhan Fatih Kar, Amir Zamir
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 8 diverse middle-domain paths with simple probability averaging instead of a single ResNet-50 | ResNet-50 classifier · ImageNet-C | mCE: 76.21 -> 67.85 | Improved |
| Replace same-architecture deep ensemble members (different inits) with middle-domain paths | 8-path ResNet-50 ensemble · ImageNet-C | mCE: 70.43 -> 67.85 | Improved |
| Replace deep ensemble with middle-domain ensemble | 8-path ResNet-50 ensemble · ImageNet (clean) | clean error: 21.50 -> 21.61 | No clear effect |
| Use an ensemble of 8 paths instead of a single ResNet-50 | ResNet-50 classifier · ImageNet (clean) | clean error: 24.37 -> 21.61 | Improved |
| Add inverse variance merging instead of uniform merging (ablation) | Cross-domain ensemble · Taskonomy with Common Corruptions (normals, depth, reshading) | l1 error | Improved |
| Add inverse variance merging instead of network (stacking) merging (ablation) | Cross-domain ensemble · Taskonomy with Common Corruptions | l1 error | No clear effect |
| Add sigma training (ST) and cross-task consistency (X-TC) (ablation) | Cross-domain ensemble · Taskonomy with Common Corruptions | l1 error vs deep ensembles | No clear effect |
| Increase number of paths (middle domains) in the ensemble (ablation) | Cross-domain ensemble vs deep ensembles · Taskonomy with Common Corruptions | average l1 error | Improved |
| Use middle domains instead of random inits in deep ensembles that also have ST, X-TC and merging (ablation) | Deep ensembles with ST and X-TC · Taskonomy with Common Corruptions | l1 error | Improved |
| Use the proposed middle-domain ensemble instead of a deep ensemble under I-FGSM attacks | Cross-domain ensemble · Taskonomy with adversarial I-FGSM attacks | prediction error | Improved |
| Use a different choice of middle domain sets (ablation) | Cross-domain ensemble · Taskonomy with Common Corruptions | average l1 error | No clear effect |
| Add sigma training so uncertainty tracks error under shift (ablation) | Per-path sigma estimates · Taskonomy epistemic uncertainty vs error | sigma vs error correlation | Improved |
Closest papers by meaning
- Training Robust Ensembles Requires Rethinking Lipschitz Continuity (ICLR 2025 · similarity 0.88)
- Improving Adversarial Robustness via Promoting Ensemble Diversity (ICML 2019 · similarity 0.87)
- Training independent subnetworks for robust prediction (ICLR 2021 · similarity 0.87)
- Building Robust Ensembles via Margin Boosting (ICML 2022 · similarity 0.87)
- On the Certified Robustness for Ensemble Models and Beyond (ICLR 2022 · similarity 0.87)
- Deep Ensembles Work, But Are They Necessary? (NeurIPS 2022 · 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.
