Research map / Adversarial robustness and security / Attacks on text and graphs
Improving Robustness Against Common Corruptions With Frequency Biased Models
Tonmoy Saikia, Cordelia Schmid, Thomas Brox
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 |
|---|---|---|---|
| add TV regularization on first conv feature maps to AugMix (ablation) | AugMix ResNet18 · ImageNet-C-100 | mCE: 40.9 -> 35.9 | Improved |
| add TV regularization on first conv feature maps to AugMix (ablation) | AugMix ResNet18 · ImageNet-100 clean | clean error: 11.8 -> 14.8 | Hurt |
| finetune with Gaussian noise and blur (HF bias) (ablation) | AugMix ResNet18 · ImageNet-C-100 | mCE: 40.9 -> 32.5 | Improved |
| finetune with Gaussian noise and blur (HF bias) (ablation) | AugMix ResNet18 · ImageNet-100 clean | clean error: 11.8 -> 13.2 | Hurt |
| finetune with contrast augmentation (LF bias) (ablation) | AugMix ResNet18 · ImageNet-C-100 | mCE: 40.9 -> 39.1 | Improved |
| add TV regularization to HF-finetuned model (ablation) | AM-HF · ImageNet-C-100 | mCE: 32.5 -> 31.5 | Improved |
| use HF and LF experts (different biases) instead of same-bias ensemble (ablation) | AM ensemble · ImageNet-C-100 | mCE: 39.1 -> 25.9 | Improved |
| use HF+LF experts instead of ensemble of two same-bias augmented models (ablation) | AM(Gauss,Cont) ensemble · ImageNet-C-100 | mCE: 29.0 -> 25.9 | Improved |
| replace AMDA ensemble with RoHL experts (ablation) | Ensemble(AMDA, AMDA) · ImageNet-C | mCE: 51.9 -> 47.9 | Improved |
| add RoHL HF/LF experts to AMDA | AMDA · ImageNet clean | clean error: 24.2 -> 22.7 | Improved |
| apply TV regularization to deeper layers instead of conv1 (ablation) | TV regularization ResNet18 · ImageNet-C-100 | high-frequency corruption error | Hurt |
| use larger TV factor lambda (>=1e-4) (ablation) | TV regularization · ImageNet-100 | clean and corruption error: 1e-6 to 9e-5 good; larger degrades | Hurt |
Closest papers by meaning
- A Simple Way to Make Neural Networks Robust Against Diverse Image Corruptions (ECCV 2020 · similarity 0.88)
- Does enhanced shape bias improve neural network robustness to common corruptions? (ICLR 2021 · similarity 0.87)
- Improving robustness against common corruptions by covariate shift adaptation (NeurIPS 2020 · similarity 0.87)
- Improving robustness to corruptions with multiplicative weight perturbations (NeurIPS 2024 · similarity 0.86)
- Dynamic Sparse Training versus Dense Training: The Unexpected Winner in Image Corruption Robustness (ICLR 2025 · similarity 0.85)
- Robust Image Denoising through Adversarial Frequency Mixup (CVPR 2024 · similarity 0.85)
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.
