Research map / Adversarial robustness and security / Certified and empirical robustness
Revisiting adapters with adversarial training
Sylvestre-Alvise Rebuffi, Francesco Croce, Sven Gowal
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 |
|---|---|---|---|
| Use clean and adversarial classification-token adapters for co-training | ViT-B16 on ImageNet · ImageNet | clean accuracy: 82.64% -> 83.56% | Improved |
| Use clean and adversarial classification-token adapters for co-training | ViT-B16 on ImageNet · ImageNet | robust accuracy: 0% -> 49.87% | Improved |
| Add clean/adversarial-specific trainable scale and offset parameters while sharing batch statistics (ablation) | Co-trained normalization layers · ImageNet | robust accuracy | Improved |
| Replace dual BatchNorm statistics with shared statistics but retain domain-specific affine parameters (ablation) | Co-trained BatchNorm · ImageNet | robust accuracy | No clear effect |
| Use only two classification tokens rather than dual LayerNorm scales and offsets (ablation) | ViT-B16 co-training · ImageNet | robust accuracy: 49.87% vs 50.02% | No clear effect |
| Replace PGD2 with one-step Fast-AT (ablation) | ViT-B16 with classification-token adapter · ImageNet | robust accuracy: 49.87% -> 0.00% | Hurt |
| Replace PGD2 with one-step N-FGSM (ablation) | ViT-B16 with classification-token adapter · ImageNet | robust accuracy: 49.87% -> 0.00% | Hurt |
| Increase training attack steps from 2 to 5 (ablation) | ViT-B16 with classification-token adapter · ImageNet | robust accuracy: 49.87% -> 50.27% | Improved |
| Set co-training clean-loss weight α=0.4 instead of α=0.5 (ablation) | ViT-B16 with classification-token adapter · ImageNet | clean accuracy: 83.56% -> 83.76% | Improved |
| Set co-training clean-loss weight α=0.4 instead of α=0.5 (ablation) | ViT-B16 with classification-token adapter · ImageNet | robust accuracy: 49.87% -> 52.19% | Improved |
| Interpolate the co-trained clean and adversarial tokens to form a model soup | ViT-B16 · ImageNet variants, mean accuracy | mean accuracy: 57.23% -> 61.23% | Improved |
| Form a weight soup from independently nominally and adversarially trained models (ablation) | Independent ViT-B16 models · ImageNet | clean accuracy: 82.64% -> 28.72% (β=0.1) | Hurt |
Closest papers by meaning
- Pyramid Adversarial Training Improves ViT Performance (CVPR 2022 · similarity 0.89)
- Towards Robust Vision Transformers: Path Dependency Analysis and a Simple Two-Stage Adversarial Training (CVPR 2026 · similarity 0.88)
- When Adversarial Training Meets Vision Transformers: Recipes from Training to Architecture (NeurIPS 2022 · similarity 0.88)
- Benign Overfitting in Adversarial Training for Vision Transformers (ICML 2026 · similarity 0.87)
- Towards Efficient Adversarial Training on Vision Transformers (ECCV 2022 · similarity 0.87)
- Domain Adaptation with Adversarial Training on Penultimate Activations (AAAI 2023 · 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.
