Research map / Adversarial robustness and security / Certified and empirical robustness
Certified Robustness to Label-Flipping Attacks via Randomized Smoothing
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 |
|---|---|---|---|
| Apply label-noise smoothing and certify predictions against training-label flips | Linear least-squares classifier · Dogfish, 25 label flips | certified test accuracy: .0050 -> .8083 | Helped |
| Increase label-noise probability from q=0.3 to q=0.475 (ablation) | Randomized smoothing classifier · MNIST 1/7, 1000 label flips | certified test accuracy: .5726 -> .6674 | Helped |
| Increase label-noise probability from q=0.012 to q=0.1 (ablation) | Randomized smoothing classifier · CIFAR10, 200 label flips | certified test accuracy: .4493 -> .3019 | Hurt |
| Increase label-noise probability from q=0.0001 to q=0.05 (ablation) | Randomized smoothing classifier · Dogfish, 100 label flips | certified test accuracy: .5167 -> .3200 | Hurt |
| Increase label-noise probability from q=0.01 to q=0.1 (ablation) | Randomized smoothing classifier · IMDB Sentiment Analysis, 100 label flips | certified test accuracy: .5392 -> .4788 | No clear effect |
| Add ℓ2 regularization to the computation of α (ablation) | Randomized smoothing classifier · MNIST 1/7 | certified test accuracy | Helped |
| Add ℓ2 regularization to the computation of α (ablation) | Randomized smoothing classifier · MNIST 1/7 | non-robust accuracy | Hurt |
| Use the tight discrete bound instead of the generic KL bound (ablation) | Randomized smoothing certificate · Label-flipping certification | certifiable label flips: approximately twice as many label flips | Helped |
| Reduce label-noise probability (ablation) | Randomized smoothing classifier · Dogfish | certified test accuracy | Helped |
| Use unsupervised RICA features instead of pre-trained deep features (ablation) | Randomized smoothing classifier · Dogfish | certified test accuracy | Hurt |
| Use randomized smoothing with q=10^-4 rather than no label noise | Dogfish classifier · Dogfish, up to 23 label flips | certified test accuracy: less than 1% -> 81.3% | Helped |
Closest papers by meaning
- Treatment of Statistical Estimation Problems in Randomized Smoothing for Adversarial Robustness (NeurIPS 2024 · similarity 0.88)
- Exact Certification of (Graph) Neural Networks Against Label Poisoning (ICLR 2025 · similarity 0.88)
- PatchDEMUX: A Certifiably Robust Framework for Multi-label Classifiers Against Adversarial Patches (CVPR 2025 · similarity 0.88)
- Provably Cost-Sensitive Adversarial Defense via Randomized Smoothing (ICML 2025 · similarity 0.88)
- Incremental Randomized Smoothing Certification (ICLR 2024 · similarity 0.87)
- Certified Adversarial Robustness via Randomized $\alpha$-Smoothing for Regression Models (NeurIPS 2024 · 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.