Research map / Adversarial robustness and security / Certified and empirical robustness
Seasoning Model Soups for Robustness to Adversarial and Natural Distribution Shifts
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 |
|---|---|---|---|
| Select soup weights separately for each shifted dataset rather than one soup for average performance | Model soups · ImageNet variants (Mean accuracy) | accuracy: 59.46% -> (64.24%) | Helped |
| Use a dataset-specific soup instead of the closest reported baseline | Model soups · ImageNet-R accuracy | accuracy: 53.04% -> 56.39% | Helped |
| Use a dataset-specific soup instead of adversarial training | Model soups · Conflict Stimuli accuracy | accuracy: 59.45% -> 67.03% | Helped |
| Use robust models trained with εp / 2 radii instead of standard εp radii (ablation) | Single model soup · ImageNet variants (Mean accuracy) | accuracy: 59.46% -> 59.10% | Hurt |
| Use robust models trained with εp / 4 radii instead of standard εp radii (ablation) | Single model soup · ImageNet variants (Mean accuracy) | accuracy: 59.46% -> 59.01% | Hurt |
| Use robust models trained with εp / 2 radii instead of standard εp radii (ablation) | Dataset-specific model soups · ImageNet variants | accuracy: (64.24%) -> (62.83%) | Mixed |
| Use robust models trained with εp / 4 radii instead of standard εp radii (ablation) | Dataset-specific model soups · ImageNet variants | accuracy: (64.24%) -> (62.59%) | Mixed |
| Interpolate the ℓ∞-robust model with its ℓ2-fine-tuned variant | CIFAR-10 and ImageNet ℓ∞-robust classifiers · ℓ∞ adversarial robustness | robust accuracy | Helped |
| Use an intermediate soup instead of the individual ℓ2-specialized model | CIFAR-10 three-threat-model soup · ℓ2 adversarial robustness | robust accuracy | Helped |
| Combine ℓ∞- and ℓ1-robust models without simultaneous training on all attacks | CIFAR-10 three-threat-model soup · Union of ℓ∞, ℓ2 and ℓ1 threats | robust accuracy: 42.0% vs 47.2% | Hurt |
| Increase the number of labeled examples used to select soup weights beyond 100 (ablation) | Few-shot soup selection · ImageNet variants | held-out accuracy | No clear effect |
Closest papers by meaning
- Certifying Some Distributional Robustness with Principled Adversarial Training (ICLR 2018 · similarity 0.90)
- Certified Adversarial Robustness with Additive Noise (NeurIPS 2019 · similarity 0.88)
- Adversarial Training and Robustness for Multiple Perturbations (NeurIPS 2019 · similarity 0.88)
- Adversarial Robustness against Multiple and Single $l_p$-Threat Models via Quick Fine-Tuning of Robust Classifiers (ICML 2022 · similarity 0.88)
- Towards Deep Learning Models Resistant to Adversarial Attacks (ICLR 2018 · similarity 0.88)
- Adversarial Robustness of Supervised Sparse Coding (NeurIPS 2020 · 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.