Research map / Adversarial robustness and security / Attacks on text and graphs
ImageNet-E: Benchmarking Neural Network Robustness via Attribute Editing
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 |
|---|---|---|---|
| Increase background texture complexity | ImageNet-E evaluation of normally trained models · ImageNet-E, background λ=20 | top-1 accuracy drop: 9.23% | Hurt |
| Make objects very small (pixel rate 0.05) | ImageNet-E evaluation of normally trained models · ImageNet-E, size 0.05 | top-1 accuracy drop: 18.34% | Hurt |
| Switch from RN50 to the larger Swin-S | ImageNet-E · all attributes average | top-1 accuracy drop: 15.72% -> 10.20% | Helped |
| Switch from Swin-S to the larger Swin-B | ImageNet-E · background λ=20 | top-1 accuracy drop: 10.20% -> 8.99% | Helped |
| Apply adversarial training | RN50 · ImageNet-E, size 0.05 | top-1 accuracy drop: 21.26% -> 25.51% | Hurt |
| Apply Ten-Crop preprocessing | image classifier · ImageNet-E, background and size changes | top-1 accuracy: 0.69% and 1.24% | Helped |
| Use ResNest channel-wise attention instead of RN50 | RN50 · ImageNet-E, all attributes average | top-1 accuracy drop: 15.72% -> 12.57% | Helped |
| Add ViT-S self-attention-like layers to the RN50 feature pipeline (ablation) | RN50 · ImageNet-E, all attributes average | top-1 accuracy drop: 11.00% -> 9.96% | Helped |
| Use MAE masked-image-modeling training (ablation) | ViT-B · ImageNet-E, all attributes average | top-1 accuracy drop: 11.05% -> 9.05% | Helped |
| Use SimMIM training (ablation) | ViT-B · ImageNet-E, all attributes average | top-1 accuracy drop: 11.05% -> 9.82% | Helped |
| Use MoCo-v3 contrastive pretraining (ablation) | ViT-B · ImageNet-E, all attributes average | top-1 accuracy drop: 11.05% -> 10.35% | Helped |
| Train with randomly replaced object backgrounds | RN50 · ImageNet validation | top-1 accuracy: 77.48% -> 79.00% | Helped |
Closest papers by meaning
- Benchmarking Neural Network Robustness to Common Corruptions and Perturbations (ICLR 2019 · similarity 0.91)
- ImageNet-D: Benchmarking Neural Network Robustness on Diffusion Synthetic Object (CVPR 2024 · similarity 0.88)
- Foundation Model-oriented Robustness: Robust Image Model Evaluation with Pretrained Models (ICLR 2024 · similarity 0.88)
- Benchmarking Segmentation Models with Mask-Preserved Attribute Editing (CVPR 2024 · similarity 0.87)
- LAION-C: An Out-of-Distribution Benchmark for Web-Scale Vision Models (ICML 2025 · similarity 0.87)
- Doomed to Re-Annotate, Forever: The ImageNet Story (NeurIPS 2026 · 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.