Research map / Label-efficient and robust learning / Recognition and long-tail
A Comprehensive Study on Robustness of Image Classification Models: Benchmarking and Rethinking
Chang Liu, Yinpeng Dong, Wenzhao Xiang, Xiao Hui Yang, Hang Su, Jun Zhu, Yuefeng Chen, Yuan He, Hui Xue, Shibao Zheng
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 RandAugment (ablation) | SwinS AT · ImageNet AutoAttack eps=4/255 | robust accuracy: 48.7 -> 49.7 | Improved |
| add Mixup on top of RandAugment (ablation) | SwinS AT · ImageNet AutoAttack eps=4/255 | robust accuracy: 49.7 -> 54.9 | Improved |
| add label smoothing 0.1 (ablation) | SwinS AT · ImageNet AutoAttack eps=4/255 | robust accuracy: 54.9 -> 55.4 | Improved |
| increase weight decay from 0 to 0.1 (ablation) | SwinS AT · ImageNet AutoAttack eps=4/255 | robust accuracy: 55.4 -> 53.7 | Hurt |
| add EMA weight averaging (ablation) | SwinS AT · ImageNet AutoAttack eps=4/255 | robust accuracy: 55.4 -> 55.6 | Improved |
| initialize AT from ImageNet-21K pre-training (ablation) | SwinB AT · ImageNet AutoAttack eps=4/255 | robust accuracy: 57.3 -> 57.4 | No clear effect |
| initialize AT from SimMIM (ablation) | SwinB AT · ImageNet AutoAttack eps=4/255 | robust accuracy: 57.3 -> 55.9 | No clear effect |
| initialize AT from CLIP image encoder (ablation) | ViTB AT · ImageNet AutoAttack eps=4/255 | robust accuracy: 49.7 -> 42.8 | Hurt |
| apply adversarial training | ViTL/ResNet family · natural robustness average over 10 datasets | average accuracy: 34.1 -> 30.2 | Hurt |
| apply adversarial training | ViTB · natural robustness average over 10 datasets | average accuracy: 35.8 -> 39.1 | Improved |
| add ImageNet-21K pre-training | ViTL · natural robustness average over 10 datasets | average accuracy: 35.4 -> 57.4 | Improved |
| add MAE self-supervised pre-training | ViTL · natural robustness average over 10 datasets | average accuracy: 35.4 -> 53.6 | Improved |
Closest papers by meaning
- Benchmarking Adversarial Robustness on Image Classification (CVPR 2020 · similarity 0.95)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.91)
- Foundation Model-oriented Robustness: Robust Image Model Evaluation with Pretrained Models (ICLR 2024 · similarity 0.90)
- Robust NAS under adversarial training: benchmark, theory, and beyond (ICLR 2024 · similarity 0.90)
- Improving Accuracy-robustness Trade-off via Pixel Reweighted Adversarial Training (ICML 2024 · similarity 0.89)
- Towards Million-Scale Adversarial Robustness Evaluation With Stronger Individual Attacks (CVPR 2025 · similarity 0.89)
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.
