Research map / Adversarial robustness and security / Certified and empirical robustness
SAFER: Sharpness Aware layer-selective Finetuning for Enhanced Robustness in vision transformers
Bhavna Gopal, Huanrui Yang, Mark Horton, Yiran Chen
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 |
|---|---|---|---|
| select layers by sharpness instead of randomly (ablation) | SAFER on DeiT-Ti · CIFAR-10 | PGD-20 accuracy: 65.46 -> 68.50 | Improved |
| select layers by sharpness instead of randomly (ablation) | SAFER on ViT-S · CIFAR-10 | Auto Attack accuracy: 46.19 -> 50.12 | Improved |
| select layers by sharpness instead of randomly (ablation) | SAFER on Swin-B · CIFAR-10 | PGD-20 accuracy: 49.72 -> 53.65 | Improved |
| add SAFER layer-selective SAM finetuning | PGD-AT (SGD) DeiT-Ti · CIFAR-10 | PGD-20 accuracy: 48.10 -> 68.50 | Improved |
| add SAFER | PGD-AT (SGD) DeiT-Ti · CIFAR-10 | clean accuracy: 75.46 -> 82.36 | Improved |
| use SAFER instead of naive SAM in PGD-AT | DeiT-S · CIFAR-10 | PGD-20 accuracy: 53.64 -> 70.26 | Improved |
| add SAFER on top of Tian et al. | Tian et al. DeiT-Ti · CIFAR-10 | PGD-20 accuracy: 46.33 -> 48.05 | Improved |
| use CNN critical-layer method CLAT instead of PGD-AT | DeiT-Ti · CIFAR-10 | PGD-20 accuracy: 48.10 -> 46.29 | Hurt |
| use RiFT instead of PGD-AT | DeiT-Ti · CIFAR-10 | PGD-20 accuracy: 48.10 -> 44.32 | Hurt |
| use RiFT instead of PGD-AT | ViT-S · CIFAR-10 | PGD-20 accuracy: 50.86 -> 45.93 | Hurt |
| add SAFER with LoRA | LoRA PGD-AT (SAM) DeiT-Ti · CIFAR-10 | PGD-20 adversarial accuracy: 51.45 -> 63.50 | Improved |
| add SAFER on top of full PGD-AT after pretraining epochs (vs PGD-AT alone) (ablation) | DeiT-Ti · CIFAR-10 | clean accuracy | Improved |
Closest papers by meaning
- Unveil Benign Overfitting for Transformer in Vision: Training Dynamics, Convergence, and Generalization (NeurIPS 2024 · similarity 0.89)
- Vision Transformers Beat WideResNets on Small Scale Datasets Adversarial Robustness (AAAI 2025 · similarity 0.89)
- Improving Interpretation Faithfulness for Vision Transformers (ICML 2024 · similarity 0.89)
- Boosting the Transferability of Adversarial Attack on Vision Transformer with Adaptive Token Tuning (NeurIPS 2024 · similarity 0.89)
- Towards Understanding and Improving Adversarial Robustness of Vision Transformers (CVPR 2024 · similarity 0.89)
- Harnessing the Computation Redundancy in ViTs to Boost Adversarial Transferability (NeurIPS 2025 · 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.
