Research map / Adversarial robustness and security / Certified and empirical robustness
Adversarial Examples Are Not Real Features
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 |
|---|---|---|---|
| Transfer non-robust features from supervised training to self-supervised paradigms | Non-robust-feature datasets · CIFAR-10 and Tiny-ImageNet-200 | cross-paradigm relative usefulness: CIFAR-10: 0.791 -> 0.307; Tiny-ImageNet-200: 0.672 -> 0.134 | Hurt |
| Use robust rather than non-robust features | Feature datasets · CIFAR-10, supervised relative usefulness | relative usefulness (SL): 0.914 -> 0.880 | Mixed |
| Train naturally on robust-feature data instead of using adversarial training | CIFAR-10 classifiers · CIFAR-10, AutoAttack | robust accuracy: 14.35 ± 1.04 -> 12.37 ± 0.33 | Hurt |
| Remove augmentation from the CL attack configuration (ablation) | CL-to-SL adversarial transfer · CIFAR-10 | robustness (%): 70.31 -> 70.21 | Helped |
| Remove the projector from the CL attack configuration (ablation) | CL-to-SL adversarial transfer · CIFAR-10 | robustness (%): 70.31 -> 71.09 | Hurt |
| Remove both projector and augmentation (ablation) | CL-to-SL adversarial transfer · CIFAR-10 | robustness (%): 70.31 -> 59.76 | Helped |
| Replace InfoNCE with alignment-only loss (ablation) | CL-to-SL adversarial transfer · CIFAR-10 | robustness (%): 70.31 -> 33.20 | Helped |
| Add RandomFlip to training on the robust dataset (ablation) | SL classifier · CIFAR-10, ℓ2 robust accuracy | robust accuracy: 8.24 -> 14.23 | Helped |
| Add RandomCrop to training on the robust dataset (ablation) | SL classifier · CIFAR-10, ℓ2 robust accuracy | robust accuracy: 8.24 -> 18.83 | Helped |
| Add RandomFlip and RandomCrop to training on the robust dataset (ablation) | SL classifier · CIFAR-10, ℓ2 robust accuracy | robust accuracy: 8.24 -> 19.49 | Helped |
| Add RandomColorJitter to Flip + Crop training (ablation) | SL classifier · CIFAR-10, ℓ2 robust accuracy | robust accuracy: 19.49 -> 20.76 | Helped |
| Use the robust dataset generated by a CL model rather than an SL model | Naturally trained encoders · CIFAR-10, cross-paradigm robustness | cross-paradigm robustness: 0.23 -> 1.16 | Helped |
Closest papers by meaning
- Adversarial Examples Are Not Bugs, They Are Features (NeurIPS 2019 · similarity 0.93)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.91)
- On Adversarial Training without Perturbing all Examples (ICLR 2024 · similarity 0.90)
- Are adversarial examples inevitable? (ICLR 2019 · similarity 0.89)
- Robust Models are less Over-Confident (NeurIPS 2022 · similarity 0.89)
- Improving Adversarial Robustness Requires Revisiting Misclassified Examples (ICLR 2020 · 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.