Research map / Adversarial robustness and security / Certified and empirical robustness
Run-off Election: Improved Provable Defense against Data Poisoning Attacks
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 ROE to DPA | DPA · MNIST, B≤500 | certified fraction: 32.42% -> 37.15% (+4.73%) | Helped |
| Apply ROE to DPA | DPA · CIFAR-10, k=50, B≤20 | certified fraction: 19.36% -> 21.30% (+1.94%) | Helped |
| Apply ROE to DPA | DPA · GTSRB, k=100, B≤5 | certified fraction: 46.16% -> 46.09% (--0.07%) | Hurt |
| Apply ROE to FA | FA · MNIST, d=16, B≤500 | certified fraction: 31.97% -> 36.76% (+4.79%) | Helped |
| Apply ROE to FA | FA · GTSRB, k=50, d=16, B≤5 | certified fraction: 82.71% -> 82.59% (--0.12%) | Hurt |
| Apply ROE to DPA* | DPA*, CIFAR-10 k=50 d=32 · CIFAR-10, B≤18 | certified fraction: 51.05% -> 51.83% (+0.78%) | Helped |
| Apply ROE to DPA* | DPA*, GTSRB k=50 d=16 · GTSRB, B≤5 | certified fraction: 83.71% -> 83.67% (--0.04%) | Hurt |
| Use more robust base classifiers by training d submodels per DPA partition and averaging their logits (ablation) | DPA · CIFAR-10, k=50, B≤18 | certified fraction: 27.52% -> 33.42% | Helped |
| Increase the DPA* submodel count from d=16 to d=32 (ablation) | DPA* · CIFAR-10, k=50, B≤18 | certified fraction: 31.41% -> 51.05% | Helped |
| Increase the FA overlap parameter from d=16 to d=32 (ablation) | FA · CIFAR-10, k=50, B≤15 | certified fraction: 34.61% -> 36.03% | Helped |
Closest papers by meaning
- Improved Certified Defenses against Data Poisoning with (Deterministic) Finite Aggregation (ICML 2022 · similarity 0.91)
- Lethal Dose Conjecture on Data Poisoning (NeurIPS 2022 · similarity 0.90)
- Deep Partition Aggregation: Provable Defenses against General Poisoning Attacks (ICLR 2021 · similarity 0.88)
- Not All Poisons are Created Equal: Robust Training against Data Poisoning (ICML 2022 · similarity 0.88)
- Phantom Transfer: Data Poisoning can Survive Data-Level Defences (NeurIPS 2026 · similarity 0.85)
- Activation Gradient based Poisoned Sample Detection Against Backdoor Attacks (ICLR 2025 · similarity 0.85)
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.