Research map / Adversarial robustness and security / Certified and empirical robustness
Are Labels Required for Improving Adversarial Robustness?
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 |
|---|---|---|---|
| Add unlabeled examples with UAT++ instead of relying on the 4K labeled examples alone | Adversarial training on CIFAR-10 · CIFAR-10 | robust accuracy (FGSM20): 32.5% -> 54.1% | Helped |
| Add unlabeled examples with UAT++ instead of relying on the 1K labeled examples alone | Adversarial training on SVHN · SVHN | robust accuracy (FGSM20): 66.0% -> 84.4% | Helped |
| Use unlabeled data rather than ground-truth labels for additional examples | UAT++ · CIFAR-10 | robust accuracy (FGSM20): 54.1%; 1.4% lower | Hurt |
| Combine online-target and fixed-target unsupervised losses (ablation) | UAT · CIFAR-10 and SVHN | robust accuracy | Helped |
| Use fixed pseudo-labels rather than online targets (ablation) | UAT · CIFAR-10, larger unlabeled-data regime | robust accuracy | Helped |
| Use fixed pseudo-labels rather than online targets (ablation) | UAT · SVHN | robust accuracy | No clear effect |
| Use 10-step PGD to generate unsupervised perturbations instead of VAT's one-step FGSM approximation (ablation) | UAT-OT · CIFAR-10 and SVHN | adversarial accuracy | Helped |
| Increase random label noise in fixed pseudo-labels (ablation) | UAT-FT · CIFAR-10 | robust accuracy: less than 10% | Hurt |
| Increase correlated pseudo-label noise (ablation) | UAT-FT · CIFAR-10 | robust accuracy: 6.3% | Hurt |
| Use off-distribution 80m@200K images rather than CIFAR-10 unlabeled examples (ablation) | UAT++ · CIFAR-10, 32K unlabeled examples | robust accuracy: 5.5% | Hurt |
| Scale uncurated unlabeled data from 80m@100K to 80m@200K (ablation) | UAT++ with WRN-34 · CIFAR-10 | adversarial accuracy (FGSM20): 59.41% -> 62.18% | Helped |
| Scale uncurated unlabeled data from 80m@200K to 80m@500K, including more out-of-distribution images (ablation) | UAT++ with WRN-34 · CIFAR-10 | adversarial accuracy (MultiTargeted): 53.35% -> 48.99% | Hurt |
Closest papers by meaning
- Unlabeled Data Improves Adversarial Robustness (NeurIPS 2019 · similarity 0.94)
- Why Do Artificially Generated Data Help Adversarial Robustness (NeurIPS 2022 · similarity 0.89)
- Soften to Defend: Towards Adversarial Robustness via Self-Guided Label Refinement (CVPR 2024 · similarity 0.89)
- Adversarial Self-Supervised Contrastive Learning (NeurIPS 2020 · similarity 0.88)
- Adversarial Unlearning: Reducing Confidence Along Adversarial Directions (NeurIPS 2022 · similarity 0.88)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 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.